Estimating trunk diameter at breast height for scattered ... · tree trunk diameter, more formally...

I

Estimating trunk diameter at breast height for scattered

Eucalyptus trees: a comparison of remote sensing

systems and analysis techniques

Niva Kiran Verma

BSc (Hons.) Vinoba Bhave University

M.Tech. Birla Institute of Technology

A thesis submitted for the degree of Doctor of Philosophy of the University of

New England

October 2014

II

Certification

I certify that the substance of this thesis has not already been submitted for any degree and is not

currently being submitted for any other degree or qualification.

I certify that any help received in preparing this thesis, and all sources used, have been

acknowledged in this thesis.

Niva Kiran Verma

29th

October, 2014

III

Acknowledgements

First and foremost, I would like gratefully acknowledge my principal supervisor Prof. David. W.

Lamb for guidance, support and encouragement at all times. I could not have dreamt of a better

supervisor than you.

I extend my sincere thanks to my co-supervisors Prof. Nick. C. H. Reid and A/Prof. Brian R.

Wilson for their guidance and supervision throughout the research work and providing valuable

comments. Thanks to Prof. Kerrie Mengersen for the valuable comments during our meetings.

This research would not have been possible without support from co-operative research centre for

spatial information (CRCSI). My special thanks to CRCSI for funding the thesis and the generous

support. I am very thankful to Derek Schneider and Ashley Saint for supporting me immensely in

my field work and without their support my field work would not have been possible. My sincere

thanks to Cate McGregor for software - related help and support.

I express my sincere thanks to Jackie Reid and Gregory Falzon for giving me insight into advanced

statistics. My heartfelt thanks to Arjan Wilkies for providing additional datasets. Thanks also go to

my well wishers and friends for supporting me and keeping me always in their prayers.

I also wish to thank my dear friend Dr. Brad Crook for reading one of my chapters and providing

valuable comments.

I express my heartiest and sincere thanks to my siblings and in-laws for their constant support

throughout my journey.

Last but not the least my heartiest thanks to my husband Dr. Priyakant Sinha for encouraging me

to take this journey and being with me and supporting me at all times, and my lovely daughter

Anviti Avatansh Sinha for sharing my dreams and being with me always. I could not have done

this without you both.

This thesis is dedicated to the memory of my parents:

The late, Shri Ramjee Verma and the late, Bimla Verma

IV

Note to Examiners

This thesis has been written in journal-article format. I have attempted to minimize the duplication

of material between chapters. However, some repetition remains, particularly in the study area,

data description and methodology sections of the articles as these were independent publications.

Although effort has been made to ensure consistency in the format for the purposes of this thesis, I

acknowledge that some inconsistencies remain because of the requirements of each of the journals

to which the separate papers were submitted.

V

Abstract

‘Farmscapes’ are farming landscapes that comprise combinations of forests and scattered remnant

vegetation (trees), natural and improved grasslands and pastures and crops. Scattered eucalypt

trees are a particular feature of Australian farmscapes. There is a growing need to assess carbon

and biomass stocks in these farmscapes in order to fully quantify the carbon storage change in

response to management practices and provide evidence-based support for carbon inventory. Since

tree trunk diameter, more formally known as diameter at breast height (DBH), is correlated with

tree biomass and associated carbon stocks, DBH is accepted as a means inferring the biomass–

carbon stocks of trees. On ground measurement of DBH is straightforward but often time

consuming and difficult in inaccessible terrain and certainly inefficient when seeking to infer

stocks over large tracts of land. The aim of this research was to investigate various avenues of

estimating DBH using synoptic remote sensing techniques. Tree parameters like crown projected

area, tree height and crown diameter are all potentially related to DBH. This thesis first uses on–

ground measurements to establish the fundamental allometric relationships between such

parameters and DBH for scattered and clustered Eucalyptus trees on a large, ~3000-ha farm in

north eastern part of New South Wales, Australia. The thesis then goes on to investigate a range of

remote sensing techniques including very high spatial resolution (decicentimetre) airborne

multispectral imagery and satellite imagery and LiDAR to estimate the related parameters.

Overall, the research demonstrated the usefulness of remote sensing of tree parameters such as

crown projection area and canopy volume as a means of inferring DBH on a large scale.

VI

Contents

Chapter 1

Introduction 1

1.1 Context and defining the research challenge 1

1.2 Literature review 2

1.3 Research objectives 6

1.4 Study area descriptions 7

1.5 Format of this thesis 8

Chapter 2

An allometric model for estimating DBH of isolated and clustered

Eucalyptus trees from measurements of crown projection area

Abstract 11

2.1 Introduction 11

2.2 Materials and Methods 14

2.2.1 Study area 14

2.2.2 Field measurements 15

2.3 Model development and validation 17

2.4 Results and Discussions 18

2.4.1 Single trees 18

2.4.2 Tree clusters 21

2.4.3 Combining both single trees and tree cluster datasets 24

2.5 Conclusions 26

2.6 Acknowledgments 27

Chapter 3

A comparative study of land cover classification techniques for

“farmscapes” using very high resolution remote sensed data

Abstract 30

3.1 Introduction 30

3.2 Materials and Method 34

3.2.1 Study Area 34

3.2 Image acquisition and data preparation 36

3.2.2 Field data collection 36

3.3 Image Classification 37

3.3.1 Object-based and pixel-based classifications 37

3.4 Accuracy assessment 39

VII

3.5 Results and discussion 40

3.5.1 Pixel versus object-based classification results 42

3.6 Conclusion 46

3.7 Acknowledgments 46

Chapter 4

Tree cover extraction from 50 cm worldview2 imagery: a comparison of

image processing techniques

Abstract 50

4.1 Introduction 50

4.2 Materials and methods 51


4.4 Conclusions 54

Chapter 5

The use of shadows in high spatial resolution, remotely sensed,

imagery to estimate the height of individual Eucalyptus trees on

undulating farm land

Abstract 59

5.1 Introduction 59

5.2 The relationship between shadow length and tree height 60


5.3.1 Study Area 63

5.3.2 Image data 64

5.3.3 Field Data Collection 64

5.3.4 Image analysis and calculating input parameters 65

5.4 Results and Discussion 68

5.4.1 Tree height 68

5.4.2 Challenging the basic assumptions 70

5.5. Conclusions 74

5.6. Acknowledgments 75

Chapter 6

Estimating crown projected area from remote sensing at different

spatial resolution and its use in estimating DBH

6.1 Introduction 78


6.2.1 Study Area 80

VIII

6.2.2 Remote Sensing Datasets 80

6.2.3 Field Data Collection 80

6.3 Data Analysis 80

6.3.1 Manual method (On screen vectorization) 80

6.3.2 Automated method (Image Segmentation and Classification) 81

6.3.3 Statistical Analysis 81

6.4 Analysis Results 82

6.5 Conclusions 88

Chapter 7

Using tree canopy measurements to infer canopy volume: A

comparison of high resolution remotely sensed images and LiDAR

7.1 Introduction 90


7.2.1 Study Area 94

7.2.2 Field measurements of canopy volume 95

7.2.3 LiDAR data acquisition and post-processing 96

7.2.4 Delineation of tree attributes from WorldView2 data 98

7.2.5 Evaluating the performance of the two techniques 99


7.3.1 Field measurements of tree parameters 99

7.4 Conclusions 104

Chapter 8

Remote Sensing based Stem Density measurements in Tree Clusters

for DBH estimation: comparison of techniques

8.1 Introduction 106


8.2.1 Study Area 108

8.2.2 Tree measurements 108

8.2.3 LiDAR Data 108

8.2 LiDAR data processing for tree stem extraction 109

8.3 Results and discussions 110

8.4 Conclusions 114

Chapter 9

Integration of LiDAR and ADS40 imagery for mapping forest species in

Australian country “farmscape”

9.1 Introduction 116

IX


9.2.1 Study Area 120

9.2.2 Remote Sensing datasets 121

9.3 Methodology 121

9.3.1 Image Segmentation and classification 121

9.3.2 Accuracy Assessment 123


9.5 Conclusions 129

Chapter 10

Conclusions 130

10.1 Summary 130

10.2 Scope for further work 132

References 134

X

List of Tables

Table 2.1: The allometric models tested for crown projection area (CA, m2), tree height

(Ht, m) and DBH (m) for single trees and tree clusters. The subscript ‘av’ denotes the

average value per stem within a cluster. The subscript ‘sum’ as applied to DBH is the sum

of each value for the individual stems and when applied to CA is the size of the canopy

envelope enclosing all the trees within the cluster.

18

Table 2.2: Summary statistics for single trees; n is the number of trees used in both the

model development and validation.

19

Table 2.3: Derived regression parameters (95% confidence intervals) for single trees (n =

86). The MPE values are derived from a separate validation dataset (n = 86). 20

Table 2.4: Summary statistics for tree clusters (all species); n is the total number of trees

used for model development and validation.

21

Table 2.5: Derived regression parameters (95% confidence intervals) for tree clusters (n =

26). The MPE values are derived from a separate validation dataset (n = 26).

21

Table 2.6: Derived regression parameters (95% confidence intervals) for a combined

individual tree–tree cluster model. The parameters DBH* and CA* incorporate data for

both single trees (DBH, CA) and the average values of each tree within the measured

clusters (DBHav, CAav). The parameters DBH+ and CA+ incorporate data for both single

trees (DBH, CA) and the sum of each tree within the measured clusters (DBHsum,

CAsum); n = 112. The MPE values are derived from a separate validation dataset (n =

112). The percentages in brackets indicate the mean relative prediction error.

24

Table 3.1: LULC description of study area (The Australian Land Use and Management (ALUM)

Classification Version 7, May 2010)

35

Table 3.2: Feature objects used in classification 38

Table 3.3: Comparison of LULC classification accuracies using pixel-based and object-

based techniques at different scale factors.

43

Table 4.1 Performance of classification techniques used in this study 52

Table 4.2 Tree area estimation from different classification techniques 53

Table 5.1: Environmental and tree parameters necessary to infer tree height from

shadows.

65

Table 6.1: The crown projected area estimates by the two methods compared to the field

measurements

86

Table 7.1: LiDAR data acquisition parameters 96

Table 7.2: Summary statistics for single trees from the field measurements; n = 79 is the

number of trees used in the model development.

99

XI

Table 7.3: Derived regression parameters for determining canopy volume (CVfield) from

crown projected area (CAfield) and crown diameter (CDfield), based on field measurements

of CAfield and CDfield, and calculation of CVfield from Equation 7.1 using a ‘multiplier’

value of 0.3927 (n = 79).

100

Table 8.1: Physical characteristics of clustered trees; n is the number of tree clusters 108

Table 8.2: Tree stem number detected by the three algorithms along with the field based

measurements.

112

Table 8.3: The effect on corresponding crown area as per the number of stems detected

by three methods.

113

Table 9.1: Segmentation scheme followed in the study 123

Table 9.2 Metrics defined for rule based classification 125

Table 9.3: Comparison of species classification accuracy 127

List of Figures

Figure 1.1 Location Map of the study area 8

Figure 2.1: Location map of the study site in north eastern NSW, Australia (Source: Open access data). 15

Figure 2.2: Tree and tree cluster locations (white circles) overlaid on a grey-scale aerial

image of the field site. 16

Figure 2.3: Scatter plots of DBH versus (a) tree height (Ht) and (b) crown projection area

(CA) for single trees (all species, n = 172). The regression curves are the log-transformed

models for each individual parameter (Table 2.5). 19

Figure 2.4: Scatter plot of crown projection area (CA) against tree height (Ht) for single

trees (n = 172, 5 species). The solid regression line (and R2) is based on a linear regression

between the parameters. 20

Figure 2.5: Scatter plots of DBHav versus (a) average tree height (Htav) and (b) crown

projection area (CAav) for tree clusters (n = 52). The regression curves are the log-

transformed models for each individual parameter (Table 2.5). 21

Figure 2.6: Scatter plot of DBHsum versus the cluster crown projection area (CAsum) for

tree clusters (n = 52). The regression curve is the log-transformed model (Table 2.5).

22

XII

Figure 2.7: Scatter plot of stem density (/ha) versus total crown projection area (CAsum,

m2) for tree clusters. The dashed curve is the envelope for the single tree data, calculated

from the crown projection area (CA). The solid curve is represents the fitted power curve

to the cluster data.

23

Figure 2.8: Scatter plots of (a) measured versus predicted DBH* (comprising DBH of

single trees and DBHav of tree clusters) using both single tree CA and tree cluster CAav

data in the regression model, and (b) measured versus predicted DBH+ (comprising both

DBH of single trees and DBHsum of tree clusters) using both single tree CA and tree

cluster CAsum data in the regression model (n = 112). Solid lines indicate 1:1 equivalence

between predicted and actual values.

25

Figure 3.1 Location Map of the study area 34

Figure 3.2: Field photos of farmscape LULC categories of the study area. 34

Figure 3.3: Image segmentation results of Airborne 15 cm image for two different sites at

the scale of 50 (a); 60 (b); 70 (c); and 80 (d). The image is a false color composite with

Band 1 = IR, Band 2 = Red and Band 3 = Green.

41

Figure 3.4: Mean LULC training samples spectral separability in different bands for (a)

Pixel-based, and (b) object-based techniques. 42

Figure 3.5: Comparison of LULC classification results at two different sites in the study

area using pixel-based and object-based techniques. 45

Figure 4.1: Tree area extracted from different classification techniques (a) Standard FCC,

(b) Object based, (c) Supervised and (d) Unsupervised. 54

Figure 5.1: North-facing slope (ga = 0) and sun due north ( a = 0)

61

Figure 5.2: Shadow projected on a slope (ga) of aspect angle sa and slope ss from the sun

at an elevation angle of e and azimuth of a. Note, the sun is depicted as large in size as

it is portrayed in the semi-foreground.

62

Figure 5.3: Location of the study area in southeastern Australia. 63

Figure 5.4: ‘False colour’ image of the study site showing image transects flown (red

lines), the sampled trees (green circles) and the mosaic seam lines (yellow lines).

64

Figure 5.5. (a) Schematic of a tree canopy (grey shape) and its projected shadow (black

shape) on the ground beneath. The two vectors (black arrows) are used to determine the

azimuth angles from which the shadow azimuth is calculated from the average). (b)

66

XIII

Schematic of a dead tree ‘skeleton’ with the shadow of the trunk clearly projected on the

ground and the vector (black arrow) indicating the trunk shadow azimuth.

Figure 5.6: Sun elevation ( e) /sun azimuth ( a) conversion curve for the study site. Fitted

5th

-order polynomial curve from which the calibration equation was derived has R2 = 1.0.

.

67

Figure 5.7: ‘False colour’ image of a portion of the study site. The individual trees

selected for evaluation/analysis are mark with green dots, which also indicate the assigned

geometric centre of the tree canopies. The vector describing the average shadow azimuth

(yellow line), emanates from each green dot circle. The large, hollow red circles indicate

dead trees where it is possible to view the shadow of the trunk and hence its azimuth.

68

Figure 5.8: Scatter plot of estimated sun azimuth as derived from the azimuth angles of

all the tree canopies. The solid grey columns corresponds sun azimuth angles between

1030 hrs and 1100 hrs AEST, and the black column is the category containing both the

average sun azimuth angles derived from all trees ( a = 42.9o) and that corresponding to

the proported image acquisition time (1045 hr AEST, a = 40.4o).

69

Figure 5.9: Scatter plot of tree height estimates from shadow using sun elevation angle

derived from (a) sun azimuth values derived from the individual shadows themselves

(Figure 8), (b) the average sun azimuth from all trees (42.9o) converted to sun elevation

angle (n = 180).

70

Figure 5.10: (a) Schematic diagram of the sun’s tangential ray that defines the shadow

length, l, on the ground, and the shadow length, l’, assumed to represent the tree height.

(b) Geometric representation of the canopy envelope and the resulting increase (l) in

shadow length that results from the tangential ray passing the outer extent of the canopy.

71

Figure 5.11: Scatter plot of tree height estimates from shadows using sun elevation angles

derived from the sun azimuth values for each shadow and the corrected shadow length (a)

l’= 0.69r and (b) l’= 0.60r. (n = 180).

74

Figure 6.1: shows the trees generated by automated methods a) Color Infra red image b)

World View2 image c) ADS 40 image

83

Figure 6.2: Scatter plots of the derived CA from each of the images versus the field-

measured CAfield. (a) MS4100, (b) ADS40, and (c) WV2. The image-derived CA was

calculated using manual vectorization; n = 172.

84

XIV

Figure 6.3: Scatter plots of the derived CA from each of the images versus the field-

measured CAfield. (a) MS4100, (b) ADS40, and (c) WV2. The image-derived CA was

calculated using image segmentation; n = 172.

85

Figure 6.4: Scatter plots of the derived CA from each of the images using manual

methods versus the segmentation based CA (a) MS4100, (b) ADS40, and (c) WV2 (n-

172)

86

Figure 6.5: Scatter plots of the predicted DBH and the field-measured values for (a)

MS4100, (b) ADS40, and (c) WV2. The image-derived CA was calculated using manual

vectorization; n = 172.

87

Figure 7.1: Flow diagram for canopy volume estimations

.

91

Figure 7.2: Schematic diagram indicating canopy dimensions required to estimate canopy

volume

92

Figure 7.3: Location map of the study site in north eastern NSW, Australia. Source: open

access data 95

Figure 7.4: Tree parameters (total tree height and trunk height) from FUSION/LDV

software. The colours green, yellow and red represents the canopy at different heights.

Blue represents the ground height.

97

Figure 7.5: An example of the (a) derived LiDAR canopy height model (CHM) and (b)

the segmentation results.

98

Figure 7.6. Scatterplot between canopy volume (CVfield), as calculated using Equation

7.1, and measured crown diameter (CDfield). The solid curve is the best-fit, polynomial

regression equation.

99

Figure 7.7. Scatterplot between canopy volume (CVfield), as calculated using Equation

7.1, and measured crown projected area (CAfield) derived from field measurements. The

solid curve is the best-fit, polynomial regression equation.

100

Figure 7.8: Scatterplot of canopy volume from WV2 using (a) crown projected area (CA),

and (b) crown volume (CV) as the predictor variable and field measurements. The dashed

lines are the 1:1 equivalence between measured and predicted values and the solid lines

the best-fit regression curves (power and polynomial, respectively).

101

Figure 7.9. Scatterplot between canopy volume predicted using the LiDAR-derived values

of crown height and crown diameter (Equation 7.1) and the field measured values. The

dashed line is the 1:1 equivalence between measured and predicted values and the solid

line is the best-fit regression curve (polynomial).

102

XV

Figure 7.10. Scatterplot between (a) LiDAR-derived crown diameter and field

measurements, and (b) LiDAR-derived crown height and field measurements. The dashed

lines are the 1:1 equivalence between measured and derived values and the solid lines the

best-fit regression curves (liner and polynomial, respectively).

103

Figure 8.1: Tree crowns in a cluster as detected from TreeVaW algorithm. (a) The tabular

output and (b) detected trees overlaid on the canopy height model

110

Figure 8.2: Tree crowns in a cluster as detected from SAGA GIS software. (a) The

segmentation Output and (b) the segments with each point representing a tree stem 111

Figure 8.3: Tree crowns in a cluster as rendered in the FUSION/LDV software. (a) the

cluster as seen from above and (b) cross section view of the tree cluster. Cooler colours

the trunk and lower part of the crown, while the warmer colours represent the higher end

of the crown.

111

Figure 8.4: Graphical representation of stem numbers as determined by the three different

algorithms 112

Figure 8.5: Scatter plots of calculated versus actual field measured stem number in the 7

tree clusters. The solid line represents the 1:1 line 113

Figure 9.1: (a) ADS 40 multispectral image of the study area (b) CHM of 1m resolution

derived from the LiDAR returns 124

Figure 9.2: Examples of the results of data segmentation (a) ADS 40 multispectral image

(b) CHM of 1m resolution derived from the LiDAR data (c) LiDAR/multispectral

combined

126

Figure 9.3: Classification output based on height ranges of LiDAR

128

Figure 9.4 The species classifications generated using (a) ADS40 multispectral image, (b)

1 m resolution CHM derived from the LiDAR data, and (c) the combined

LiDAR/multispectral datasets.

128

Chapter 1 Introduction

1

Chapter1

Introduction

This research was based on the premise that a ratio or regression method using empirical equations

and allometric models is the most reliable and appropriate non-destructive, field-based method for

quantification of biomass and carbon stored in trees, and that the diameter at breast height (DBH) is

closely related to tree biomass and carbon stocks. As field based methods are laborious, expensive and

time consuming, and often constrained to human-accessible terrain, the aim of this research was to

consider alternative methods to field based measurements of DBH which would help in minimizing

time and cost involved in undertaking such measurements over large scales. The study focused on

remnant vegetation on farming land, in particular scattered Eucalyptus trees owing to their importance

in the Australian landscape. The research focused on using remote sensing data as a plausible

alternative to field based measurements for DBH and other tree parameter estimations at the whole-

farm level.

This introductory chapter begins by articulating the context of this research, introducing relevant tree

characteristics and providing a review of remote sensing technologies and potential benefits in large

scale mapping. This chapter concludes with the specific aims of the research and an outline of the

structure of the thesis.

1.1 Context and defining the research challenge

Carbon dioxide (CO2) is one of the most important components in greenhouse gases as it traps heat

within the atmosphere and causes a global warming effect. Photosynthesis binds CO2 and stores it as

carbon in plants. Plant communities thus act as carbon ‘storehouses’ and through their impact on CO2

flux processes, ultimately play an important role in influencing our climate. With the clearing of

vegetation on agricultural lands, the carbon is released and the land effectively acts as a greenhouse

gas (GHG). When vegetation is restored, carbon is sequestered and the land can be considered a sink

of carbon.

Biomass is an important component of the global carbon cycle (Scurlock et al., 2002): forest and

woodlands contain carbon in the form of biomass (trunks, branches, foliage, roots, etc.) and organic

carbon in the soils, accumulated through growth of trees over the years and increase in soil organic

carbon. In addition to forests, grasslands can also act as a carbon sink due to their ability to store large

amounts of carbon.


2

Agriculture is a major feature of Australian life and farms dominate the Australian landscape. More

than two-thirds of the Australian landmass is devoted to agricultural production (Wells, 2013) with

approximately 90 per cent of farm land used for grazing on native pastures, occurring mostly in the

arid and semi-arid zones (Wells, 2013).

The Northern Tablelands region of northern New South Wales (NSW), Australia, is typical of

Australian farming land dedicated to the production of livestock and wool (Cottle et al., 2013). Such

farming landscapes, or ‘farmscapes’, consist of a range of landscape features such as forests and open

woodlands, pockets of remnant vegetation of varying density, sparsely scattered trees, natural and

improved grasslands and cultivated pastures and forage crops.

Native eucalyptus trees, both individual and in clusters are an important feature of Australian

farmscapes (including the New England Region) and they contribute significantly to above and

below-ground carbon stocks in these landscapes (Baffetta et al., 2011; Soto-Pinto et al., 2010). The

assessment of biomass in eucalypt systems has, to date, been largely restricted to plantation forestry

systems. However there is also a growing need to assess carbon and biomass stocks across our

farmscapes in order to fully quantify carbon storage change in response to management and provide

evidence-based support for carbon inventory and ultimately carbon trading. Such assessments must

also include scattered native trees.

This thesis is ultimately focused on scattered tree communities that exist on these farmscapes as a

necessary adjunct to the considerable volume of scientific knowledge surrounding the measurement of

biomass in forests and established tree communities. Unlike forestry, where often the spatial extent

and composition of constituent trees is well known, farmscapes are as diverse in size as they are in

tree composition and density. Certainly, like in forestry, measurement of tree-related biomass and

carbon in farmscapes through manual techniques is time consuming and impractical, especially given

the potential scale over which measurements must be taken. Furthermore, the spatial heterogeneity of

farmscapes renders them unsuited to extrapolation of small-scale measurements too.

1.2 Literature Review

According to the IPCC (Intergovernmental panel for Climate Change) Good Practice Guidance for

LULUCF (Landuse, Landuse Change and Forestry (2003), the carbon pools of terrestrial ecosystems

involving plant biomass are conceptually divided into above-ground and below-ground biomass, dead

mass and litter. Above-ground biomass is measurable with some accuracy at the broad scale. While

below-ground biomass stores a large part of total carbon stocks, it is still poorly known because it can

only be assessed through in situ measurements that tend to be labour- and time-intensive (particularly


3

for forest ecosystems): currently in most of cases the below-ground component is derived from above-

ground biomass.

There are a number of ways to measure biomass which can be used either alone or in combination.

Conventional physical methods of biomass estimation

Two methods are traditionally available for the determination of tree biomass (Murli et al., 2005). The

destructive sampling which involves physically harvesting the plant material and subsequent

extrapolation to a mass value per hectare (Klinge and Herrera, 1983). The second method involves

allometry to extrapolate sampled data to a larger area based on easily-measured parameters such as

diameter at breast height (DBH), tree height etc. For example, Malimbwi et al., (1994) estimated

forest biomass and volume through direct harvesting in the Miombo woodlands in Tanzania.

Stromgaard (1985) estimated the above-ground tree biomass in the same woodland using multiple

regression analyses of parameters such as measured trunk diameter (at breast height), tree height, and

frequency of trees in a specified area.. Both methods proved comparative but similarly proved time

consuming, costly and generally limited to small areas and small tree sample sizes (Ketterings et al.,

2001; Hyde et al., 2006; 2007). Ultimately they are both labour intensive. In addition, extending this

method to map biomass across a large area is extremely challenging because of ecological differences

and scattered sources of biomass data. Also, the allometric coefficients developed are often site and

species specific, and hence cannot be standardized for all areas (Chave et al., 2005). Nonetheless,

efforts have been made to develop generalized regional and national tree biomass equations (Lambert

et al., 2005; Case and Hall., 2008) with mixed degrees of success.

Non-destructive biomass measurement

One of the recent advances in biomass estimation approaches is the incorporation of parameters

derived from remote sensing. The synoptic view afforded by remotely sensed data offers the

capability of capturing the spatial variability in above-ground vegetation parameters such as tree

height, crown closure etc. Remote sensing data available at different scales (local to global), from

various sources (optical or microwave) and platforms (airborne to space borne), are expected to

provide information which can be related directly, and in different ways, to biomass information

(Rosenqvist et al., 2003; Foody et al., 2003). Numerous studies have been carried out to estimate

tree/forest biomass from remote sensing data (Nelson et al, 1988; Franklin and Hiernaux, 1991;

Ranson and Sun, 1994; Steininger, 2000; Foody et al., 2003; Zheng et al., 2004; Zachary and

Rundolph, 2005; Sun and Ranson, 2009). ().

Among the different remote sensing data types, optical image data have been widely used for forest

biomass estimation with varying degrees of success. These are: Landsat TM (Steininger, 2000; Foody


4

et al., 2003; Calvao and Palmeirim, 2004; Lu 2005), Landsat ETM+ (Zeng et al., 2004; Rahman et al.,

2005), IKONOS (Thenkabail et al., 2004; Asner et al., 2002; Greenberg et al., 2005; Song et al.,

2010), Quickbird (Hyde et al., 2006; Song et al., 2010); Spot-5 (Li et al., 2006; Soenen et al., 2010),

NOAA AVHRR (Barbosa et al. 1999; Dong et al, 2003), MODIS (Baccini et al., 2004), and ASTER

(Muukkonen and Heiskanen, 2005). The commonly used biomass estimation approaches are multiple

regression analysis, k-nearest neighbor, and neural network (Roy and Ravan, 1996; Nelson et al.,

2000; Steininger, 2000; Foody et al., 2003; Zeng et al., 2004). Optical image data allows spatial

stratification of vegetation for possible direct estimates of biomass, generally through empirical

relationships. Different vegetation indices and band ratios from optical image data have been used to

extract biomass by correlating vegetation index values or band ratio values with field estimations

(Dong et al., 2003). Alternatively image data may be used indirectly, for example by determining tree

canopy parameters such as crown diameter using multiple regression analysis or canopy reflectance

models (Phua and Saito 2003, Popescu et al., 2003).

Microwave remote sensing data such as Synthetic aperture radar (SAR) data have been found useful

for forest ecosystem analyses, particularly in areas of frequent overcast conditions. Radar systems

have capability of collecting data in all weather (and light) conditions. The SAR sensor can detect

the H (horizontal) or the V (vertical) component of the backscattered radiation. Significant

correlations have been found between radar backscatter (P and L bands) and someforest parameters,

such as tree age, tree height, DBH, basal area, and total above ground dry biomass (Imhoff et al.,

2000; Sun et al., 2002; Santos et al., 2003). A detailed review on use of radar data for biomass

estimation can be found in Kasischke et al. (1997; 2004). Previous studies showed SAR L-band data

to be more useful for forest biomass estimation (e.g.,Sun et al. 2002) as compared to SAR C-band (Le

Toan et al. 1992). Most of the previous studies were based on radar system of single polarization and

incident angle and of low resolution, however, with the establishment of Phased Array L-band SAR

(PALSAR) and RADARSAT-2 (C-band), the data are now available in different polarizations and

resolutions, and varying incident angles and hence can provide more opportunity to assess the

potential of SAR data in forest biomass estimation.

The two dimensional nature of optical and SAR data limit their ability of quantifying some vegetation

characteristics like tree height, canopy height, volume directly. Light detection and ranging (LiDAR)

is a relatively new and sophisticated technology which helps overcome this limitation due to its ability

to extend the spatial analysis to a third dimension. A detailed review of LiDAR data application in

forestry can be found in Lim et al., (2003). The three dimensional LiDAR points represent latitude,

longitude and ellipsoidal height based on the WGS84 reference ellipsoid. There are currently two

types of LiDAR in operation today. 1) Discrete return LiDAR (small footprint), and 2) full waveform

LiDAR (large footprint) (Todd et al., 2003). Both are generally calibrated to operate in the 900-

1064nm wavelengths, where vegetation reflectance is highest (Lefsky et al., 2002). The structural

forest canopy measurements permit the accurate estimation of leaf area index (LAI), net primary


5

productivity (NPP), and above ground biomass (Lefsky et al., 2002). For tropical and deciduous forest

biomass estimation, large footprint waveform systems have shown to provide accurate estimates (e.g.,

Drake et al., 2002). The DEM’s generated from airborne LiDAR data can be very accurate and are

often used in forest mapping and tree parameters estimations. It captures elevation information from

the forest canopy as well as the ground beneath and can be used to assess the complex 3D patterns of

canopy and forest stand structure such as tree density, stand height, basal area, leaf area index (LAI)

and forest biomass and volume (e.g., Lefsky et al., 2002; Næsset and Økland, 2002). The estimation

of biomass is generally based on regression equations relating vegetation biomass to LiDAR derived

variables. Though LiDAR data was found very useful in forest biomass estimation, particularly in

areas of frequent cloud conditions; expensive data acquisition process, complexity in data analysis,

software requirement, are few issues that restrict its use for limited applications.

Tree-based biomass studies carried out in Australia

Forests (plantations, commercial forests and conservation forests) cover about 21 per cent of Australia

and store an estimated 10.5 billion tonnes of carbon (Forest and Wood products Research and

Development Corporation). The Australian Greenhouse is one of the leading organizations working in

the field of carbon sequestration and carbon monitoring. In a collaborative effort through National

carbon accounting System, the Australian greenhouse Office has developed a national wood products

carbon accounting model that tracks the flow of carbon and contributes to the Australian National

Greenhouse Gas Inventory. Numerous Researchers have contributed towards the study of biomass and

carbon in Australian Forestry system by developing allometric equations for different regions for the

predominant species found in the region. A detailed report on Review of Allometric equations for

woody regions across Australia ha been prepared by Australian Green house Office (Keith et al.,

2000; Eamus et al., 2000). Harrington (1979) carried out a study in Cobar, New South Wales for

estimation of above-ground Bbiomass of trees and shrubs. Burrows et al. (2000; 2001) studied

allometric relationships to community biomass stocks of Eucalypts and pine. Chen et al. (2003)

reported on carbon balance in tropical savanna in northern Australia.

The use of advanced technologies like LiDAR and SAR in forest biomass studies has also been

numerous. Though the use of SAR images in Australia dates back to 1992-93, fewer studies have

been carried out especially for Eucalyptus forests. Lucas et al. (1990) conducted a Biomass study in

Queensland Australia in which they examined correlations between the biomass of mixed species of

Eucalyptus woodland with SAR backscatter from both airborne and satellite images. Austin et al.

(2002) undertook a study on estimating forest biomass using satellite radar. The results suggested that

biomass for Eucalypts can be estimated using satellite radar, taking into account landscape

characteristics like topography, surface water and forest structure. Turner (2007) presented an


6

overview of Airborne LiDAR applications in New South Wales forests. Lovell et al. (2014) used

Airborne and ground-based LiDARs to probe the structure of forest canopies. Such information is not

readily available from other remote sensing methods but is essential for modern forest inventory in

which growth models and ecological assessment are becoming increasingly important. They

concluded that current laser ranging systems can be used to derive canopy structural parameters such

as height, cover, and foliage profile provided information based on multiple returns or the intensity of

returns is used to minimize the bias induced by the size of the footprint and the detection threshold.

The other LiDAR system which has gained much attention is the near infra red light detection and

ranging system, the Echidna Validation Instrument (EVI) developed by CSIRO Australia. A study

conducted by Strahler et al.(2008), showed that the forest structural parameters like mean diameter at

breast height, stand height, distance to tree, stem count density, Leaf area index and stand foliage

profile could be retrieved quickly with good accuracy. Similar study was also carried out by Jupp et

al. (2008) for Estimating forest LAI profiles and structural parameters using a ground-based laser

called ‘Echidna’. Lovell et al. (2004) carried out a simulation study for cost effective measurement of

forest inventory parameters. Arroyo et al. (2008) have worked on LiDAR and multispectral data

integration for mapping of riparian environments.

1.3 Research objectives

Given the potential of optical remote sensing techniques to provide a synoptic view of landscapes at a

range of scales and spatial resolutions, this thesis seeks to investigate the use of remote sensing to

infer an important parameter for estimating tree biomass- the trunk diameter at breast height (DBH),

for scattered Eucalyptus trees in a New England (Australia) farmscape. It is acknowledged that much

research has already been done in the field of tree biomass and carbon estimation using both using

image-based remote sensing systems and more recently light detection and ranging (LiDAR) systems.

However much of this work is confined to forestry. LiDAR, in particular is widely applauded for its

applicability to directly measuring tree structural parameters, from which biomass would be

estimated. Yet the technology, especially when deployed in aerial platforms, remains expensive,

sophisticated and the generated data requires specialist software and skilled technicians to process it.

Despite the promise of LiDAR, large scale, regular, operational use of airborne LiDAR for tree

biomass estimations on scattered trees outside the forest (STOF) remains ‘scattered’.

Notwithstanding the historical focus on forestry and on LiDAR, image-based remote sensing systems

are evolving rapidly with metre resolution satellite systems such as WorldView2, and sub-metre

airborne systems now widely available.

The main objective of this thesis was therefore to investigate the potential of very high resolution,

image-based remote sensing data for estimating DBH of our scattered trees, while also examining the


7

use of LiDAR as an adjunct to image-based systems rather than an alternative. The work has been

divided into the following specific aims:

a) Establish allometric equations for estimating the DBH of scattered Eucalyptus trees (including

continuous tree clusters) using regression equations involving physical characteristics such as tree

height and crown dimensions,

b) Examine various image processing methods of delineating tree crowns from very high spatial

resolution satellite and airborne imagery of farmscapes, and quantify the accuracies of these

methods,

c) Evaluate the accuracies of relevant tree crown and tree height parameters as extracted from the

remotely-sensed imagery, and

d) Evaluate the performance of entirely image-based methods of inferring DBH of scattered

Eucalypus trees in our candidate farmscape.

1.4 Study area description

The research area selected for this study was located within the University of New England’s

‘Newholme–Kirby SMART farm’, Armidale, New South Wales, Australia (longitude

151°360΄8.0144˝ E to 151°39´34.1217˝ E and latitude 30°26´31.9827˝ S to 30°24´57.0713˝ S) which

is extensively used for field based research and livestock studies. The farmscape encapsulates an area

of approximately 1500 ha. The study site comprises large tracts of natural eucalypt woodland and

pasture cover. Approximately one third of the area is dense eucalypt woodland, one third open

woodland, and the remainder native pasture. Most of the study site is unimproved pasture grazed by

sheep. The soils within the study site are brown and yellow chromosols, and the mean annual rainfall

is 780 mm. The climate is cool temperate with 60 % of the rain falling in the summer months

(National Parks and Wildlife Service, 2003; BoM, 2014). Figure 1.1 shows the location map of the

study area. The area is dominated by Eucalyptus species with other species present. The dominant

species are E. bridgesiana, E.caliginosa, E. blakelyi, E.viminalis, and E. melliodora


8

Figure 1.1 Location map of the study area

1.5 Format of this thesis

The thesis comprises ten chapters, three of which have been published as peer-reviewed journal or full

conference papers with another one submitted and ‘in review’.

The entire thesis has been structured in the following sequence:

Chapter 2 develops allometric equations for estimating DBH of both individual and clusters of

Eucalyptus trees based on field measurement of a number of physical tree characteristics. This work

was published in the journal Forest Ecology and Management;

Chapter 3 investigates image processing techniques to delineate tree crowns in very high spatial

resolution (10-50 cm), airborne, multispectral, remotely sensed images. The chapter focusses on

comparing feature based tree attribute extraction methods with the conventional pixel based

techniques using a number of subjective measures. This work was published in the Journal of

Photogrammetric Engineering and Remote Sensing;


9

Chapter 4 examines image processing methods for delineating tree cover from other landscape

features in WorldView2 satellite imagery and quantifies the accuracy of the results using a manual

vectorization method. This work has been published as a full paper in IEEE Geoscience and Remote

Sensing Society’s ‘International Geoscience and Remote Sensing Symposium’ (IGARSS 2013);

Chapter 5 develops a theoretical framework for extracting tree height from shadows in remotely

sensed imagery of undulating farmland (i.e. farmscapes) and provides quantitative accuracy data. The

chapter has been submitted for publication in the Remote Sensing of Environment;

Chapter 6 illustrates and discusses the effect of sensor spatial resolution on tree crown area

estimation and identifies an optimum spatial resolution that provides a tree crown area estimation that

is comparable to field based measurement methods;

Chapter 7 investigates the use of image-based remote sensing to infer tree canopy volume and

determines whether image-based remote is a viable alternative to LiDAR in large scale canopy

volume estimations;

Chapter 8 determines whether LiDAR can be deployed, as an adjunct to image-based remote sensing,

to estimate the stem density under the closed canopies of tree clusters, and hence provide an average

DBH value of each stem under these clusters;

Chapter 9 examines whether LiDAR, and very high spatial resolution airborne multispectral imagery

can be used to delineate Eucalyptus species on our farmscape; and finally,

Chapter 10 summarizes the key findings of the thesis, presents a set of general conclusions and

highlights future research needs

Chapter 2 Allometric Model

10

Chapter 2

An allometric model for estimating DBH of isolated

and clustered Eucalyptus trees from measurements

of crown projection area

This chapter has been published as:

Verma, N.K., Lamb, D.W., Reid, N., and Wilson, B. (2014). An allometric model for

estimating DBHofisolated and clustered Eucalyptus trees from measurements of crown

projection area. Forest Ecology and Management, 326, 125-132.


11

Abstract

Owing to its relevance to remotely-sensed imagery of landscapes, this paper investigates the ability to

infer diameter at breast height (DBH) (ie at a height of 1.3 m from the base point along the axis of the

stem) for five species of Australian native Eucalyptus from measurements of tree height and crown

projection area. In this study regression models were developed for both single trees and clusters from

2 to 27 stems (maximum density 536 stems per ha) of Eucalyptus bridgesiana, Eucalyptus caliginosa,

Eucalyptus blakelyi, Eucalyptus viminalis, and Eucalyptus melliodora. Crown projection area and tree

height were strongly correlated for single trees, and the log-transformed crown projection area

explained the most variance in DBH (R2 = 0.68, mean prediction error ± 16 cm). Including tree height

as a descriptor did not significantly alter the model performance and is a viable alternative to using

crown projection area. The total crown projection area of tree clusters explained only 34% of the

variance in the total (sum of) the DBH within the clusters. However average crown projection area per

stem of entire tree clusters explained 67% of the variance in the average (per stem) DBH of the

constituent trees with a mean prediction error ± 8 cm. Both the single tree and tree cluster models

were statistically similar and a combined model to predict average stem DBH yielded R2 = 0.71 with a

mean prediction error (average DBH per stem) of ± 13 cm within the range of 0.28 – 0.84 m. A single

model to infer DBH for both single trees and clusters comprising up to 27 stems offers a pathway for

using remote sensing to infer DBH provided a means of determining the number of stems within

cluster boundaries is included.

Keywords: diameter at breast height, crown projection area, scattered trees, tree clusters, remnant

vegetation, Eucalyptus, farm land

2.1 Introduction

Much of eastern Australia is characterised by diverse farming landscapes, or ‘farmscapes’ containing

a range of land-use systems including crops, native and sown pasture, remnant vegetation and trees at

various densities and configurations. Native eucalyptus trees, both individual and in clusters are an

important feature of Australian farmscapes, and provide shade and shelter for livestock (Bird et al.,

1992). They also have considerable value through their influence on soils (Wilson, 2002; Goh et al.,

1996; Barnes et al., 2011a), biodiversity (Oliver et al., 2006) and their indirect role in pasture quantity

and quality (Barnes et al., 2011b). Scattered trees outside the forest (STOF) contribute significantly to

above and below-ground carbon stocks in these landscapes (Baffetta et al., 2011; Soto-Pinto et al.,

2010). A summary of the role of scattered trees in landscapes is given in Manning et al. (2006). The

assessment of biomass in eucalypt systems has, to date, been largely restricted to plantation forestry

systems. However there is also a growing need to assess carbon and biomass stocks across our


12

farmscapes in order to fully quantify carbon storage change in response to management and provide

evidence-based support for carbon inventory and ultimately carbon trading. Such assessments must

also include scattered native trees.

Destructive sampling is considered the most reliable method for determining biomass in grass and

shrub vegetation but it is rarely used for agroforest ecosystems (Snowdon et al., 2002). The biomass

of large vegetative structures like forests and open woodland is usually estimated by applying ratio or

regression methods using empirical equations and allometric models (Snowdon et al., 2002;

Houghton, 2005; Makungwa et al., 2013). Trunk diameter, more formally known as diameter at breast

height (DBH), is observed to be closely related to tree biomass and carbon stocks (Ter-Mikaelian and

Korzukhin, 1997; Snowdon et al., 2002; Sanquetta et al., 2011a; Kuyah et al., 2012; Beets et al.,

2012).

DBH is an important tree characteristic in its own right and is straight forward to measure on the

ground. The relationships between tree canopy characteristics such as diameter, projection area and

coverage, and DBH is of considerable interest as DBH can then be used to infer canopy attributes, for

example to quantify canopy competition in response to planting (stem) density, growth potential or

habitat modelling (Bella et al., 1971; Cade 1997; Grote 2003). Of course, large-scale collection of

DBH data can be time consuming and now satellite or airborne remote sensing, including LIDAR can

be used to infer canopy characteristics such as crown projection area (e.g. Franklin and Strahler, 1988;

Leckie et al., 2003; Popescu and Wynne, 2004; Lee and Lucas, 2007).

Measuring the canopy diameter or crown projection area from the ground involves measuring the

crown projection across different angular segments of the canopy. These angular segments can either

be based on fixed (pre-set) directions (Röhle and Huber, 1985), or in the case of highly asymmetric

canopies the directions can be adapted to adequately characterise the actual tree dimensions (Hemery

et al., 2005; Fleck et al., 2011). Based on an investigation involving 161 trees in an old-growth forest

(approx. density 392 stems per hectare), Fleck et al. (2011) recommend the 8-point, flexible approach

be used to estimate crown projection area in order to minimise errors due to deviations from canopy

symmetry. Other workers have concluded that two, orthogonal diameter measurements (4 radii) are

suitable for computing crown diameter (Hemery et al., 2005), from which crown projection area can

be subsequently derived.

The relationship between DBH and crown diameter or projected area has been the subject of

numerous papers in recent years, of which a few exemplars will be discussed here. Unsurprisingly, the

exact form of the relationship is driven by competition with neighbouring trees and understorey and

much of this driven by factors such as shade tolerance. Many of the relationships are ‘almost’ linear,

with observed departures from linearity, for example logarithmic or square-root dependence, often at


13

smaller DBH (Hall et al, 1989; Hemery et al, 2005). Arzai and Aliyu (2010) found statistically

significant linear relationships between crown diameter and DBH in some Eucalypt and other tree

species from the savana zone in Nigeria (R2 0.23 – 0.82), as did Sanquetta et al. (2011b) for Araucaria

(Araucaria angustifolia), Imbuia (Ocotea porosa) and Canelas (Nectandra grandiflora) in the mid

southern Parana State of Brazil (R2 0.47 – 0.78). On the other hand, O’Brien et al. (1995) observed

the good species-dependent predictions using log–log transformed data (R2>0.86), and Sanquetta et al.

(2011b) also improved their prediction, but only for the combined species dataset, by a log-log

transformation (R2 from 0.23 to 0.52). Smith (2008) observed a strong linear relationship between

canopy area and trunk cross-sectional area (proportional to DBH2) for native pecan trees grown in

managed groves (Smith, 2008) and Gill et al. (2000) observed DBH and DBH2, when used as a sole

independent variable, to be reasonably good at predicting canopy radius in forest planted conifer trees

(R2>0.45). Ultimately, models relating crown projection area and DBH are genus dependant, and

often species dependent owing to the range of tree architecture between species. Moreover, little is

known about the transferability of models derived for single trees (for example in open woodlands) as

compared to clusters of trees (i.e., multiple stems) in light of possible competition in growth (e.g.

Biging and Dobbertin, 1992; 1995; Moeur 1997).

Crown projection area could therefore be used to infer DBH. This offers an important pathway to

deploy remote sensing tools for the large-scale assessment of above ground or biomass stocks across

entire landscapes. From a remote sensing perspective, crown projection area is more easily measured

than canopy diameter for the simple reason that the extracted crown diameter parameter is influenced

by the direction of the measurement vector. Assuming image pixels, or objects are correctly classified

as a particular tree crown, and the tree crown is at the nadir viewpoint (or close to it), it remains to

sum the pixels (of known dimensions) within that crown to determine the projected area. This is not to

say, however that crown diameter is irrelevant to the remote estimation of DBH. Numerous workers

have found strong relationships between crown diameter and DBH (for example Hall et al., 1989;

Gering and May, 1995) and given established sampling regimes for determining crown diameter (or

canopy extent) are aimed at minimising the effect of crown asymmetry (Fleck et al., 2011), crown

diameter could potentially be derived from the remotely sensed canopy projection area measurements

by applying appropriate shape parameter such as discussed in the various work of Nelson (1997),

Grote (2003) and Fleck et al. (2011). Armston et al (2009), in the state wide land cover and trees

survey in Queensland used the allometric relationship between basal area and foliage projected cover.

The present study aims to develop an allometric model for predicting DBH of single trees and trees in

clusters using the tree/cluster characteristics of crown projection area and tree height. The overall

context of this study is remnant, native vegetation that exists in typical Australian farming landscapes.

In this study, we examine trees from the genus Eucalyptus, owing to its overall significance in


14

Australian landscapes (Commonwealth of Australia, 1999) and in particular its prevalence in

Australian farmscapes (Attiwill and Adams, 1996). In this study we hypothesize that DBH can be

predicted using field measurements of tree height and crown projection area for both single trees as

well as clusters of trees. We also wish to test the hypothesis that the same model is applicable for

both single trees and isolated clusters of trees. The definition of a cluster can be somewhat arbitrary,

and proximity of trees to one another will influence more than just competition for growth. For

example Barnes et al. (2011b) demonstrated the extent to which the litterfall from an isolated tree will

influence soil nutrients beyond the immediate canopy envelope. Moreover, given the context of this

particular work in relation to the use of large scale remote sensing tools to possibly infer DBH,

consideration should be given to the spatial resolution of such large scale sensing systems. To this end

we define a cluster as a group of trees with canopy envelopes within 3 m proximity to each other.

2.2 Materials and Methods

2.2.1 Study area

The study site was located within the University of New England’s, ‘Newholme-Kirby SMART

farm’, Armidale, New South Wales, Australia (longitude 151°35’40” E to 151°37’12” E and latitude

30°26´09´´S to 30°25´12´´S) (Fig. 2.1). The 662 ha site comprises large tracts of natural eucalypt

woodland and pasture cover. Approximately one third of the study area is dense eucalypt woodland,

one third open woodland, and the remainder native pasture. Most of the study site is unimproved

pasture grazed by sheep. The soils within the study site vary from brown and yellow chromosols, and

the mean annual rainfall is 780 mm. It has a cool temperate climate with the majority of rain falling in

the summer months (National Parks and Wildlife Service, 2003; BoM 2014).


15

Figure 2.1: Location map of the study site in north eastern NSW, Australia

2.2.2 Field measurements

High spatial resolution (15 cm), color infrared (CIR) airborne imagery of the study area was first used

to identify single trees and tree clusters in the field. A total of 52 tree clusters and 172 individual trees

were identified (Fig. 2.2). Within the tree clusters the number of stems ranged from two to twenty

seven with a density ranging from 38 to 536 stems per ha (SD 4.2 trees per cluster, 105 stems per ha).

The tree/cluster locations were extracted from the orthorectified, georeferenced imagery for

subsequent field visitation, aided by a DGPS (GPS Pathfinder® Pro XRS receiver, Ranger TSC2

model, Trimble, California). The horizontal accuracy of GPS was ~0.5 m allowing unambiguous

identification of target trees/clusters on the ground. On-ground visitation and measurement of selected

trees and tree clusters was conducted during the period September -December 2012.


16

Figure 2.2: Tree and tree cluster locations (white circles) overlaid on a grey-scale aerial image of the

field site.

The DBH of individual trees and those in clusters was derived from the measured trunk circumference

at 1.3 m above local ground level. In the case of individual trees (as delineated by a single root-ball)

with multiple stems at 1.3 m above the ground, the DBH of each stem was measured and tree DBH

calculated using:

(1)

where d1, d2, d3 were the diameters of each stem (Pretzsch, 2009). The DBH parameters investigated

for tree clusters were the average DBH of the individual trees within a cluster (DBHav) and the sum of

the DBH values (DBHsum).

For individual trees, the height (Ht) was measured using a combination of laser rangefinder and

clinometer (MDL LaserAce 300, Measurement Devices Ltd. Scotland, UK). The height of a tree was

taken to be the vertical distance between the base point and the highest point of the tree, following

Gschwantner et al (2009). Cluster height was determined by first measuring the height of each tree

within the cluster, and the average height (Htav) calculated. In those clusters where the individual tree

crowns could not be delineated, Htav was calculated from six measurements of the entire canopy

envelope acquired from different azimuthal viewer positions relative to the cluster.

Canopy cover is defined as the proportion of ground covered by the vertical projection of the tree

crowns (Jennings et al., 1999). The crown projection area (CA) of a tree is the area of the vertical

projection of the outermost perimeter of the crown on the horizontal plane (Gschwantner et al; 2009 )

In order to measure crown projection area (CA) of both single trees and tree clusters, we adopted a

hybrid version of the 8-point crown projection method recommended by Fleck et al. (2011) and that

investigated by Röhle and Huber (1985). Firstly a pair of vertical range poles were placed in the


17

ground to delineate the edges of the tree/cluster canopy along a cardinal direction, for example N-S,

passing through the tree stem or the estimated centre of the cluster. The crown periphery for locating

the pole positions was located using a clinometer set to ‘vertical view, in effect acting as a crown

mirror. The crown diameter (d) is defined as the horizontal width of the crown, taken from ‘dripline’

to ‘dripline’ as one moves around the crown(Gschwantner et al; 2009). In this case the ‘dripline’ is

effectively the canopy edge as delineated using the clinometer. The crown diameter along the given

direction was then measured using a laser range finder positioned at right-angles to, and a known

distance well back from, the line between the poles. This measurement avoided errors that would

otherwise be incurred by using tapes to measure the straight-line distance between the poles with tree

stems in between. This measurement was undertaken for six cardinal directions namely N, ENE,

ESE, S, WSW and WNW, respectively and the average diameter, which is effectively the average

crown spread, d, calculated (Sumida and Komiyama, 1997). The crown projection area was then

calculated using:

CA = π d2 /4. (2)

The crown projection area parameters investigated for tree clusters were the average of the individual

trees within a cluster (CAav) and the total area of the cluster (CAsum).

The information about tree species for each tree /cluster was noted as was the number of tree stems in

a cluster. The dimensions of five different Eucalyptus species, namely Apple Box (AB, Eucalyptus

bridgesiana), Stringy Bark (SB, Eucalyptus caliginosa ), Red Gum (RG, Eucalyptus blakelyi ), White

Gum (WG, Eucalyptus viminalis ) and Yellow Box (YB, Eucalyptus melliodora ) were sampled in

this way.

2.3 Model development and validation

The allometric models tested for crown projection area (CA, m2), tree height (Ht, m) and DBH (m) for

single trees and tree clusters are listed in Table 1 and follow the forms reviewed and listed in Hall et

al. (1989). The subscript ‘av’ denotes measurements for tree clusters where DBHav, CAav and Htav are

effectively the average value per stem within the cluster. Similarly the subscript ‘sum’, as applied to

DBHsum is the sum of each value for the individual stems and the parameter CAsum is the total size of

the canopy envelope enclosing all the trees within the cluster. In the case of non-normal data

distributions (Shapiro–Wilk Test), a logarithmic transformation was first carried out and the

transformed data tested for species-specific variations. Linear regression models were evaluated using

the statistical software R (Studio Version 0.97.318). In all the models, the interaction terms between

the parameters were also tested.


18

Table 2.1 : The allometric models tested for crown projection area (CA, m2), tree height (Ht, m) and

DBH (m) for single trees and tree clusters. The subscript ‘av’ denotes the average value per stem

within a cluster. The subscript ‘sum’ as applied to DBH is the sum of each value for the individual

stems and when applied to CA is the size of the canopy envelope enclosing all the trees within the

cluster.

Regression models

Individual trees Tree Clusters

DBH = 0 + 1.(Ht,CA) DBHav = β0 + β1.(Ht, CA)av

ln(DBH) = 0 + 1.ln(Ht, CA) ln(DBHav) = β0 + β1.ln(Ht, CA)av

ln(DBH) = 0 + 1.ln(Ht) + 2.ln(CA) ln(DBHav) = β0 + β1.ln(Htav) + β2.ln(CAav)

DBHsum = β0 + β1.CAsum

ln(DBHsum) = β0 + β1.ln(CAsum)

The data were tested for normality using Shapiro–Wilk test and influential outliers, if any, were

detected by means of Cooks distance statistics of the residuals. Any data that had a residual Cook’s

distance value of ≥ 2 was cross-checked with the original dataset to validate its precision and impact

on the model. To satisfy the assumptions of linear regression analysis, scatter plots of residuals were

checked for linearity, homoscedasticity and normality. The strength of the underlying relationship of

the predictor and response variables was evaluated by analyzing the regression coefficients of the

fitted models. The coefficient of determination (R2) was used to evaluate the level of variance in DBH

explained by the variables. For each model, half of the samples, namely 86 for single trees and 26 for

tree clusters, respectively, were withheld from the initial calibration for subsequent validation of the

model. The prediction performance of each model was quantified using a mean prediction error

(MPE) given by MPE= DBHpredicted-DBHactual

.

2.4 Results and Discussions

2.4.1 Single trees

Table 2.2 lists the descriptive statistics for the single trees. With all the species combined, the dataset

was found to be non-normal, in particular the subset comprising White Gum and Yellow Box

(Shapiro-Wilk test W = 0.97 p = 0.004). A logarithm transformation was sufficient to normalize the

data (W = 0.992, p = 0.389).


19

Table 2.2: Summary statistics for single trees; n is the number of trees used in both the model

development and validation.

Tree characteristics/species n Min Max Mean SD

DBH

Apple Box 23 0.34 1.65 0.840 0.310

Red gum 5 0.5 0.83 0.690 0.151

Stringy bark 51 0.36 1.33 0.839 0.221

White Gum 28 0.36 1.92 0.911 0.391

Yellow Box 65 0.28 1.47 0.807 0.244

CROWN PROJECTION

AREA (CA)

Apple Box

26.26 750 212 183.5

Red gum

70.85 167 136 41.6

Stringy bark

42.41 413 195 100.4

White Gum

36.83 732 230 164.1

Yellow Box

9.34 683 222 132.3

TREE HEIGHT (Ht)

Apple Box

9.1 40.5 17.6 7.28

Red gum

12.7 23.6 18.2 4.12

Stringy bark

13.5 30.4 21.4 4.59

White Gum

11.8 44.6 20.7 8.05

Yellow Box

9.5 42.1 21.9 6.42

Scatter plots of DBH versus individual tree height (Ht) and crown projection area (CA) parameters are

given in Fig. 2.3, along with regression curves of the log-transformed models. The regression

statistics are summarised in Table 2.3. When the models were validated against the 86 trees retained

from the data for this purpose, the log-transformed models yielded a MPE of 16 cm. A multi-linear

regression model involving both log-transformed Ht and CA parameters combined gave a slightly

improved prediction of DBH with a MPE of 14 cm.

(a)

(b)

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

0 10 20 30 40 50

DB

H (

m)

Ht(m)

R2 = 0.31

MPE= 0.16 m0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

0 200 400 600 800

DB

H (

m)

CA(sq m)

R2 = 0.68

MPE= 0.16 m


20

Figure 2.3: Scatter plots of DBH versus (a) tree height (Ht) and (b) crown projection area (CA) for

single trees (all species, n = 172). The regression curves are the log-transformed models for each

individual parameter (Table 2.5).

Table 2.3: Derived regression parameters (95% confidence intervals) for single trees (n = 86). The

MPE values are derived from a separate validation dataset (n = 86).

Equation R2

F-stat p MPE (m)

ln(DBH) = –2.10229 + 0.61742 ln(Ht) 0.31 37.4 <0.0001 0.16

ln(DBH) = –2.40568 + 0.42616 ln(CA) 0.68 181.6

<0.0001

0.16

ln(DBH) = –2.64742 + 0.15142 ln(Ht) + 0.38002 ln(CA) 0.59 60.1 <0.0001 0.14

Interestingly, while log-transformed crown projection area on its own explains significantly more

variance in log-transformed DBH than tree height (R2 = 0.68 compared to 0.31), both tree height and

crown projection area predict DBH with a similar MPE (16 cm). This is a reflection of the strong

inter-relationship between crown projection area and tree height. The scatter plot of crown projection

area versus tree height for all species (Fig. 2.4) illustrates this, with tree height explaining 30% of the

variance in crown projection area.

Figure 2.4: Scatter plot of crown projection area (CA) against tree height (Ht) for single trees (n =

172, 5 species). The solid regression line (and R2) is based on a linear regression between the

parameters.

It can be concluded at this point that either Ht or CA could be used to infer DBH for single trees,

although based on the level of variance explained in the DBH by CA, this latter parameter is likely to

provide better precision in predicting DBH on a tree by tree basis.

0

100

200

300

400

500

600

700

800

5 15 25 35 45

CA

(sq

m)

Ht (m)

Apple Box

Red Gum

Stringy Bark

White Gum

Yellow Box

R2 = 0.30


21

2.4.2 Tree clusters

Table 2.4 lists the descriptive statistics for the tree clusters. For the tree clusters, scatter plots of

average DBH (DBHav) versus average tree height (Htav) and average crown projection area (CAav) are

given in Fig. 2.5, along with the curves based on the derived regression models. Regression statistics

are summarised in Table 2.5.

Table 2.4: Summary statistics for tree clusters (all species); n is the total number of trees used for

model development and validation.

Tree characteristics n Min Max Mean SD

DBHav (m) 52 0.28 1.10 0.56 0.17

Htav (m)

9.12 26.50 18.23 4.10

CAav (m2)

18.66 443.45 111.75 78.30

Number of stems

2 27 5.71 4.20

Stem density (/ha)

37.91 535.91 150.71 105.11

(a)

(b)

Figure 2.5: Scatter plots of DBHav versus (a) average tree height (Htav) and (b) crown projection area

(CAav) for tree clusters (n = 52). The regression curves are the log-transformed models for each

individual parameter (Table 2. 5).

Table 2.5: Derived regression parameters (95% confidence intervals) for tree clusters (n = 26). The

MPE values are derived from a separate validation dataset (n = 26).

Equation R2

F-stat p MPE (m)

ln(DBHav) = –1.85617 + 0.42221 ln(Htav) 0.14 3.8

0.07 0.10

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.0 5.0 10.0 15.0 20.0 25.0 30.0

DB

Ha

v(m

)

Htav(m)

R2 = 0.14

MPE= 0.10 m

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 50 100 150 200 250 300

DB

Ha

v(m

)

CAav (sq m)

R2 = 0.67

MPE= 0.08 m


22

ln(DBHav) = –2.13471+ 0.335344 ln(CAav) 0.67 46.2 <0.0001 0.08

ln(DBHav) = –2.5813 + 0.18246 ln(Htav) + 0.31712

ln(CAav)

0.69

24.7

<0.0001

0.07

ln(DBHsum) = –2.17648+ 0.523622 ln(CAsum) 0.34 11.9 <0.01 0.66

An assessment of the datasets showed that, again DBHav was not normally distributed (Shapiro-Wilk

W = 0.953, p = 0.03147). Once log-transformed, the parameter CAav explained a significantly larger

proportion of the variance observed in the average DBH per stem in each cluster as compared to the

average tree height within the cluster (R2 = 0.67 compared to 0.14). Unlike the single trees, Htav and

CAav were not strongly correlated (R2 = 0.04, data not shown ), which suggests that a combination of

both parameters in a multi-linear regression model (log-transformed inputs) may be desirable, even

though this yields only modest gains in the level of variance explained in DBHav (R2 = 0.69

compared to 0.67) and precision in predicting values of DBHav (0.07 m compared to 0.08 m). The

CAav parameter alone was found to yield a MPE of 0.08 m when predicting DBHav in the range from

0.28 to 0.84 m (~28.5% and 9.5% error, respectively)

The regression statistics for the sum of the DBH values within the clusters ( DBHsum) and the total

crown projected area (CAsum) of the clusters are also listed in Table 2.5 and a scatter plot of the of

DBHsum versus CAsum is given in Fig. 2.6. Again a log transformation provided the best regression

model, although the total crown projected area only explained 34% of the variance in the total DBH

(R2 = 0.34) with a MPE of 0.66 m where the DBHsum ranged from 1.06 to 8.91 m (~63% and 7.4%

error, respectively).

Figure 2.6: Scatter plot of DBHsum versus the cluster crown projection area (CAsum) for tree clusters (n

= 52). The regression curve is the log-transformed model (Table 2.5).

0

1

2

3

4

5

6

7

8

9

10

0 200 400 600 800 1000 1200

DB

Hsu

m(m

)

CAsum (sq m)

R2 = 0.34

MPE= 0.66 m


23

The fact that the net crown projected area for a cluster is only weakly related to the sum of DBH

values within that cluster is evidence of competition effects (Bella, 1971; White 1981). Fig. 2.6

exhibits increasing scatter at higher CAsum. In the clusters investigated in this work, a higher CAsum is

correlated with an increase in the number of stems (R = 0.55). At the same time we observe the stem

density (stems/ha) - CAsum relationship (Fig. 7) to take the form:

= - (3)

The larger tree clusters in our farmscape are made up of fewer, older trees, consistent with the self-

thinning rule discussed by White (1981). Superimposed on Fig. 2.7 is the curve generated by the

single tree data; the so-called ‘single tree envelope’. Here the stem density (/ha) for the single trees

was calculated using the ratio 10,000 m2 /CA, effectively assuming the inter-stem distance is limited

to the unperturbed canopy envelope of the individual trees themselves; namely zero influence overlap

(Bella 1971). It is evident that in our scattered tree clusters, the self-thinning mechanism is resulting

in trees expressing similar canopy dimensions as isolated trees. With these older trees there is

expected to be an increase in variability in canopy extent due to effects of weathering, pests and

disease (for example Landsberg and Ohmart, 1989; Elliot et al., 1993; Neumann, 1993; Köstner et al.,

2002) and this may also explain the increasing variability observed at higher values of CAsum.

Summing the DBH in Fig. 2.6 effectively accumulates the effects of individual tree competition, and

the resulting departure from the behaviour of single, isolated trees. The act of taking average (per

stem) crown projected area and average DBH in a cluster most likely reduces this accumulating error.

Figure 2.7: Scatter plot of stem density (/ha) versus total crown projection area (CAsum, m2) for tree

clusters. The dashed curve is the envelope for the single tree data, calculated from the crown

projection area (CA). The solid curve is represents the fitted power curve to the cluster data.

stem density = 6007.1 x (CAsum)-0.652

R² = 0.35

0

200

400

600

800

0 200 400 600 800 1000 1200

Ste

m d

ensi

ty (

/ha

)

CAsum (sq m)

single tree envelope


24

2.4.3 Combining both single trees and tree cluster datasets

A comparison of the single tree and tree cluster regression models based on crown projected area

alone (DBH = f(CA); Tables 2.3 and 2.5) suggests the parameters derived for single trees, and those

derived on an per-tree basis from tree clusters (stem numbers ranging from 2 – 27), are similar. A

comparison of the derived regression equations for the individual trees and tree clusters shows the

interaction terms to be non–significant (p = 0.79). This implies that the slopes and the intercepts of the

two regression models do not differ significantly; statistically the two equations generate similar

estimates of DBH.

The log-transformed regression models derived using the combined datasets are given in Table 2.6

and scatter plots of the model-predicted versus actual DBH are given in Fig. 8. Here the models are

based on the log-transformed DBH (DBH*, DBH+) and the log-transformed crown projection area

(CA*, CA+) values. The superscript ‘*’ denotes the fact that each parameter involves data for both

single trees (DBH, CA), and the average values of each tree within the measured clusters (DBHav,

CAav). The superscipt ‘+’ denotes the fact that each parameter involves data for both single trees

(DBH, CA), and the net cluster projected canopy area and sum of DBH of each tree within the

measured clusters (DBHsum, CAsum). Both models were created using a random selection of half the

single and cluster data for calibration (n = 112) and the remaining data withheld for subsequent

validation (n = 112).

Table 2.6: Derived regression parameters (95% confidence intervals) for a combined individual tree–

tree cluster model. The parameters DBH* and CA* incorporate data for both single trees (DBH, CA)

and the average values of each tree within the measured clusters (DBHav, CAav). The parameters

DBH+ and CA

+ incorporate data for both single trees (DBH, CA) and the sum of each tree within the

measured clusters (DBHsum, CAsum); n = 112. The MPE values are derived from a separate validation

dataset (n = 112). The percentages in brackets indicate the mean relative prediction error.

Equation R2

F-stat p MPE (m)

ln(DBH*) = –2.46441 + 0.426425 ln(CA*)

0.71

265.0

<0.0001

0.13

(17%)

ln(DBH+) = –3.73606 + 0.70211 ln(CA

+)

0.59

157.0

<0.0001

0.43

(31%)

The data for the single tree and tree cluster datasets are shown as separate symbols. The effect of the

error in the DBHsum versus CAsum, discussed earlier is again evident in Fig. 2.8 (b), and it is clear that

a single model for both individual trees and tree clusters breaks down when the DBHsum values are

included in DBH+. In the DBH* - CA* model that incorporates both DBHav and CAav values of


25

clusters with the data for single trees, the model performs well; it is noteworthy to observe the tree

cluster data to be distributed amongst the single tree data points. The MPE in DBH for the tree-

averaged data is 0.13 m. While encouraging, this equates to a mean relative prediction error

approximately 17% of the DBH values encountered in the sampling. An investigation of the relative

prediction error on a sample by sample basis did not show any systematic trend towards increasing

prediction error with increasing DBH except for values exceeding 1.2 m.

(a)

(b)

Figure 2.8: Scatter plots of (a) measured versus predicted DBH* (comprising DBH of single trees and

DBHav of tree clusters) using both single tree CA and tree cluster CAav data in the regression model,

and (b) measured versus predicted DBH+ (comprising both DBH of single trees and DBHsum of tree

clusters) using both single tree CA and tree cluster CAsum data in the regression model (n = 112).

Solid lines indicate 1:1 equivalence between predicted and actual values.

2.5 Conclusions

Simple regression models involving crown projection area of Eucalyptus trees (six species), both

isolated and in clusters of up to 27 stems ranging from 38 to 536 stems per ha), explained 67% and

68%, respectively of the variance in stem DBH. A single model involving both single trees and the

tree clusters to predict average stem DBH had similar explanatory power (R2 = 0.71) and yielded a

mean prediction error in average DBH per stem of ± 13 cm. We conclude that it is sufficient to use

crown projection area to infer DBH for these species (and these stem densities and stem numbers).

While the results appear encouraging, we acknowledge that the landscape investigated here was only

662 ha, and while encapsulating considerable variation in soils, elevation and aspect (for example as

reported in Garraway and Lamb, 2011), it would be expected that the robustness and precision of the

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

DB

H*

pre

dic

ted

(m

)

DBH* actual (m)

single trees

tree clusters

R2 = 0.71

MPE= 0.13 m

0

0.5

1

1.5

2

2.5

3

3.5

4

0 0.5 1 1.5 2 2.5 3 3.5 4

DB

H+

pre

dic

ted

(m

)

DBH+ actual (m)

single trees

tree clusters

R2 = 0.59

MPE= 0.43 m


26

model could potentially be enhanced by including other landscape parameters. This is the subject of

further work. Nevertheless, the use of larger scale data sources such as airborne or satellite imagery to

infer DBH from derived values of crown projection area for both single Eucalypt trees and clusters up

to 27 stems (between 22 and 536 stems per ha) appears feasible and worthy of further investigation.

Admittedly, the single model developed here does require knowledge of the number of stems within a

given tree cluster. However the increasing use of other sensing technologies like LIDAR to both infer

total crown projection area and delineate and count the number of stems within a canopy (for example

Popescu et al., 2004) potentially offers the means to achieve this.

2.6 Acknowledgments

This work was partially funded by the CRC for Spatial Information (CRCSI), established and

supported under the Australian Government Cooperative Research Centres Programme. One of the

authors (NKV) wishes to acknowledge the receipt of a Postgraduate ‘Top-up’ Scholarship from the

CRCSI. We would like to thank Ashley Saint and Derek Schneider (UNE-PARG) for their assistance

in conducting the field work and Drs Gregory Falzon (UNE-C4D), Robin Dobos (NSW DPI) and

Jackie Reid (UNE) for their helpful comments on the statistical analysis.


27

Journal-Article Format for PhD Theses at the University of New England

STATEMENT OF AUTHORS’ CONTRIBUTION

(To appear at the end of each thesis chapter submitted as an article/paper)

We, the PhD candidate and the candidate’s Principal Supervisor, certify that all co-authors have

consented to their work being included in the thesis and they have accepted the candidate’s

contribution as indicated in the Statement of Originality.

Author’s Name (please print clearly) % of contribution

Candidate Niva Kiran Verma

75

Other Authors David. W. Lamb 15

Nick Reid

5

Brian Wilson

5

Name of Candidate: Niva Kiran Verma

Name/title of Principal Supervisor: Prof. David. W. Lamb

Candidate Date: 29.10.2014

Principal Supervisor Date: 29.10.2014


28


STATEMENT OF ORIGINALITY


We, the PhD candidate and the candidate’s Principal Supervisor, certify that the following text,

figures and diagrams are the candidate’s original work.

Type of work Page number/s

All text All pages

All figures and diagrams All pages





Chapter 3 Classification Comparison

29

Chapter 3

A comparative study of land cover classification

techniques for “farmscapes” using very high

resolution remotely sensed data


Verma, N.K., Lamb, D.W., Reid, N., and Wilson, B. (2014). A comparative study of land

cover classification techniques for “farmscapes” using very high resolution remotely sensed

data. Photogrammetric Engineering and Remote Sensing, 80 (5), 461–470.


30

Abstract

High spatial resolution images (~10 cm) are routinely available from airborne platforms. Few studies

have examined the applicability of using such data to characterize land cover in ‘farmscapes’

comprising open pasture and remnant vegetation communities of varying density. Very high spatial

resolution remotely sensed imagery has been used to classify land cover classes on a ~5000 ha

extensive grazing farm in Australia. This ‘farmscape’ consisted of open pasture fields, scattered trees

and remnant vegetation (woodlands). The relative performances of object-based and pixel-based

approaches to classification were tested for accuracy and applicability. Maximum likelihood

classification (MLC) was used for pixel-based classification while the k-nearest neighbor (k-NN)

technique was used for object-based classification. A range of image sampling scales was tested for

image segmentation. At an optimal sampling scale, pixel-based classification resulted in an overall

accuracy of 77%, while object-based classification achieved an overall accuracy of 86%. Whilst both

the object and pixel-based classification techniques yielded higher quantitative accuracies, a ‘more

realistic’ land cover classification, with few errors due to intermixing of similar classes, was achieved

using the object-based method.

3.1 Introduction

Remote sensing provides a useful source of data to extract accurate land use and land cover (LULC)

information for planning and implementation of different land use practices (Xiuwan, 2002; Falcucci

et al., 2007). The demand for accurate and up-to-date LULC information, along with historical change

as well as future trajectories, has been acknowledged by various researchers (e.g. Sobrino and

Raissouni, 2000; Hester et al., 2010). Most studies in the past two decades were based on medium

resolution remote sensing data such as Landsat TM (30 m), SPOT (20 m) etc. These were found

useful in regional and medium scale land cover mapping and change detection analysis (e.g.,

Robertson and King, 2011). In recent years, numerous studies have used high-resolution images of

metre to sub-metre spatial resolution from satellite systems such as IKONOS and Quickbird, to

identify small-scale features in a time and cost-effective way (e.g. Puissant et al., 2005; Johansen et

al., 2007). Today, very high spatial-resolution data, in the order of tens of centimetre, is now routinely

available from airborne sources, offering the possibility of creating land cover maps of greater detail

for planning applications (Dehvari and Heck, 2009) as well as for conducting above-ground biomass

or carbon stock assessments (Brown et al., 2005; Hester et al., 2010).

High-resolution imagery increases the information available on land cover at both local and national

scales (Aplin et al., 1997), allowing improved delineation between features (Thomas et al., 2003).

Landscapes, including those used for extensive farming practices, so-called ‘farmscapes’, are

inherently spatially variable. Taken at the individual farm level, farmscapes can include agricultural


31

fields (crops and pastures), remnant native vegetation (trees, open woodland), water features, roads,

buildings, orchards and other developments. In Australia, a single farm can include a diverse range of

LULC classes and can range in size from 10 ha to 10 000 ha. Despite the considerable range in spatial

extent, enterprise-relevant farm management tools to assess biomass in its many forms start at the

individual field scale and land use classification must necessarily occur at the sub-field scale, namely,

at the order of tens of metre (Bramley et al., 2008; Cook and Bramley, 2000; Lamb, 2000). At this

spatial scale, medium-resolution data may give erroneous results when classifying boundary pixels.

Sampling resolution must always exceed that of the proposed delineation (Woodcock and Strahler,

1987). This increases the probability of having purer pixels (end-member pixels) available for

whatever classification procedure is in use (Mundt et al., 2006) as well as reducing co-registration

errors (Weber et al., 2008).

For many years, pixel-based classifications (PBC) have been used successfully in various

applications. The conventional pixel based supervised methods such as maximum likelihood classifier

(MLC), minimum distance from means (MDM), and parallelepiped all examine only spectral

information of the image to produce a classification. Such parametric classifiers work on two

assumptions: (i) that the image data is normally distributed, and (ii) that the training samples’

statistical parameters (e.g. mean vector and covariance matrix) truly represent the corresponding land

cover class. However, the image-derived parameters are not always normally distributed, especially in

complex landscapes, and uncertainty in image classification can be exacerbated by a lack of sufficient

training data and multimodal training samples. Moreover, with high spatial resolution multispectral

images, classification can also be confounded by spectral similarities between land cover classes and

the ‘salt and pepper’ effect often resulting from classification also degrades the accuracy of the end

product. Ironically, the increase in spatial resolution available from remote sensing systems actually

increases the complexity of image classification into homogeneous vegetation structural classes

(Johansen et al., 2007).

With smaller pixel size, more information actually resides in surrounding pixels – so-called

contextual information – and this challenges traditional pixel-based approaches (Jensen, 2009). Also

for very high spatial-resolution data, the increased spatial resolution often comes at the expense of

spectral information. The majority of centimetre-resolution systems are multispectral, comprising

only three or four spectral bands. It could therefore be argued that land cover classification processes

involving very high spatial-resolution imaging systems should rely more heavily on the contextual

component of features rather than a limited spectral component. Therefore, spatial information such

as texture and context must be exploited to produce accurate classification maps (Shackelford and

Davis, 2003), that allow land use categories to become a combination of different and spectrally

distinct land cover types (e.g. Zhang and Wang, 2003).


32

Object-based classification (OBC) considers not only the spectral properties of pixels in account but

also the shape, texture and context information during classification process. The OBC starts by

segmenting the image into meaningful objects based on both spectral and spatial information, and

then classifying objects to produce more useful thematic maps (Fisher, 1997; Blaschke and Strobl,

2001; Goodchild et al., 2007; Robertson and King, 2011). The key differences between pixel and

object-based image classification are that : (1) the basic processing units in OBC are image objects

(segments) rather than individual pixels and classification is performed on image objects, and (2) in

most cases the OBC uses soft classifiers (non-parametric) and not hard classifiers (parametric) such as

MLC, particularly when classification is carried out in commercially available software like

eCognition (Tadesse et al., 2003; eCognition User’s Manual, 2004). There are studies that have used a

supervised classifier such as MLC for the subsequent classification of segmented objects (e.g. Platt

and Rapoza, 2008), but this is not common. One often reported advantage of using object-based

classification is that the results often render landscape entities more ‘realistic’ in both shape and

classification and many researchers consequently report ‘more satisfactory’ results using OBC

compared to PBC (e.g. Flanders et al., 2003; Robertson and King, 2011; Yan et al., 2006; Yu et al.,

2006). A detailed review of OBIA techniques can be found in Blaschke (2010).

Image segmentation, a critical first step in OBC, is a semi-automated, hierarchical process that

reduces an image into discrete regions or objects that are spectrally and spatially homogeneous

(Ryherd and Woodcock, 1996; Wong et al., 2003). The object-oriented paradigm often has

components not typically used in traditional pixel-based classification methods: (1) the segmentation

procedure, (2) nearest neighbor classifier, (3) the integration of expert knowledge, and (4) feature

space optimization. A majority of image segmentation algorithms such as the Fractal Net Evolution

Approach (FNEA) are based on region growing methods, which take some pixels as seeds and grow

the regions around them based on certain homogeneity criteria. Image objects are then pairwise

merged one by one to form bigger objects by finding areas of minimum spectral and spatial

heterogeneity (Baatz and Schape, 2000; Benz et al., 2004). Two parameters are used to determine

these: (i) color versus form (homogeneity) and (ii) scale (heterogeneity). The four criteria that define

the relative homogeneity of image objects are grouped into two pairs: (i) color versus shape, and (ii)

smoothness versus compactness. The color and shape in the first pair quantifies the spectral and

spatial homogeneity, respectively, and affect the objects being created during segmentation. The color

and shape, inversely proportional to each other, have weightings ranging from 0 to 1 and, determine

the contribution of spectral heterogeneity (in this case green, red and near infrared) and shape to

overall heterogeneity that is to be minimized. The smoothness and compactness parameters are

additional weights between 0 and 1 that determine how shape is calculated. Spectral heterogeneity is

defined as the sum of standard deviations of each image band. Minimizing only spectral heterogeneity

results in objects that are spectrally similar, but that might have fractally shaped borders or many


33

branched segments (Baatz et al., 2004). To address this issue, the segmentation process also

incorporates shape in terms of compactness or smoothness. Compactness is defined as the ratio of the

border length and the square root of the number of object pixels. Smoothness is defined as the ratio of

the border length and the shortest possible border length derived from the bounding box of an image

object (Baatz et al., 2004). The ‘best’ compactness and smoothness parameters depend on the size and

types of objects to be extracted. The size of the image objects is determined by a ‘scale parameter’, a

unitless number related to the image resolution that describes the maximum allowable heterogeneity

of image objects. As the scale parameter increases, the size of the image objects also increases (Benz

et al., 2004). The selections of segmentation parameters are largely subjectively determined through a

combination of trial and error, and ultimately user experience. Parameters that work well for one

image may not work well for another, even if the images are similar. The segmentation process can be

time consuming as complex features are involved in analyzing polygons (segments); however,

subsequent image processing time can be reduced by applying preconfigured segmentation descriptors

in batch analysis (Laymon, 2005). Obviously, for the reason stated above care must be exercised in

this approach. Non-parametric classifiers have been found more attractive for classification of

segmented objects as they can be used with arbitrary data without any statistical parameters to

separate image classes (Lu and Weng, 2007) and also easily incorporate non-spectral data into a

classification procedure. Several studies have demonstrated improved classification outcomes using

non-parametric classifiers compared to parametric classifiers, especially in complex landscapes (e.g.

Baatz and Schäpe, 2000; Willhauck et al., 2000; Hay et al., 2005; Yan et al., 2006). The most

commonly used non-parametric classification approaches are artificial neural networks (ANN),

decision trees (DT), support vector machines (SVM), and k-Nearest Neighbor (k-NN) (e.g. Lu and

Weng, 2007). Recently, Random Forest (RF) has also gained popularity as a non-parametric method

for image classification because of its accuracy, robustness against noise and also its simplicity

compared to other non-parametric methods (Brieman, 2001).

Notwithstanding the fact that numerous researchers have reported improvements in classification

accuracies using object-based methods, these studies have utilized data with spatial resolution in the

range of 30 m to 1 m (e.g. Duveiller et al., 2008; Jobin et al., 2008; Robertson and King, 2011, to

name a few) and have primarily focussed on urban scenes and to our knowledge have not been tested

in complex landscapes such as we define to be ‘farmscapes’. It is the aim of this paper to compare

pixel and object-based image classification procedures on very high spatial resolution images (15 cm)

of a complex, yet relatively open landscape, namely a farmscape. In this context, this paper seeks to

compare the land use and land cover (LULC) classification accuracy of the two approaches. The

hypothesis for this study is that spectral separability between LULC categories can be enhanced at a

certain object-level as opposed to the pixel-level.


34

3.2 Materials and Method

3.2.1 Study Area

Figure 3.1: Location map of the study area.

Figure 3.2: Field photos of farmscape LULC categories of the study area.


35

The study area was the ‘SMART’ farm (Lamb et al., 2013) owned and operated by the University of

New England (UNE), Armidale, New South Wales, Australia (longitude 151°35´40´É to

151°37´12´É and latitude 30°26´09´´S to 30°25´12´´S) (Figure 3.1). Comprising a total area of 4837

ha the extensive grazing system farm includes large tracts of natural forest cover (with several forest

types) and is partitioned into grazed and ungrazed components in both the woodland and open

pasture. Approximately one third of the study area is forested; a third is woodland, and the remainder

native pasture. Most of the property is managed in an agriculturally un-manipulated manner, other

than grazing. The pasture is a mixture of native grasses with broadcast clover species and some sown

rye-grass and fescue combinations. Productivity on a farm of this kind is highly dependent on pasture

management because pasture provides the main food source for the livestock.

For this particular study a 445 ha subset of the farm was used (Figure 3.1). Seven land use classes

were identified and allocated in the study area: Natural Pasture-moderate to high density (NP-MHD),

Natural Pasture-low density (NP-LD), Degraded Pasture (DP), Scattered Trees (ST), Bare Soil (BS),

Outcrop or Rocky area (OC), Roads (RD) and Waterbodies (WB). A description of each class is given

in Table 3.1 and photographs are provided in Figure 3.1.

Table 3.1. LULC description of study area (according to the The Australian Land Use and Management

(ALUM) Classification Version 7, May 2010) LULC Class Description

Natural Pasture Medium to High Density

(NP-MHD)

Pastures with native grasses and many other native

herbs and shrubs where the density was medium to

high, with an estimated LAI > 2 (mixed soil and plant)

Natural Pasture Low Density (NP-LD) Native pastures dominated by native grasses. These

pastures contain native grasses and many other native

herbs and shrubs except that the density was less, with

an estimated LAI ≤ 2 (pure plant)

Degraded Pasture

(DP)

Degradation form of pasture which needs management

to restore the desired pasture composition, Its native

grass with approximately 50 percent dead grass.

Scattered Trees

(ST)

Areas dominated by woody vegetation predominantly

evergreen trees

Bare Soil

(BS)

Land containing less than one third vegetated. It

usually consists of sites visually dominated by

considerable areas of exposed bare rock, and sand

with low herbaceous and shrubby plants.

Outcrop

(OC)

The Bare Exposed Rock category includes areas of

bedrock exposure, volcanic material other

accumulations of rock without vegetative cover.

Roads (RD) For transport and communication

Waterbody

(WB)

All types of waterbodies


36

3.2 Image acquisition and data preparation

Multispectral airborne imagery of the study area was acquired on 11 April 2008 using a Duncan Tech

MS4100 camera system mounted in a Cessna 172 aircraft. Flown at an altitude of 245 m above

ground level (AGL), 8-bit images were acquired with a spatial resolution of approximately 15 cm in

three spectral bands: Band 1 (NIR 0.7–1 μm), Band 2 (Red 0.6–0.7 μm), and Band 3 (Green 0.5–0.6

μm). The images were obtained in adjacent transects of approximately 300 m width, which were

referenced to the WGS 84 UTM Zone 56 S projection system and combined (mosaiced) to produce a

single image covering the entire study area.

3.2.2 Field data collection

Fieldwork sampling for training and validation was conducted in May 2012. Following the minimum

sample size recommendation of Congalton and Green (2009) between 75 and 160 sample points were

randomly selected and stratified according to each of the seven land cover classes inferred from the

raw 15-cm airborne imagery. The variations in sample size for different LULC categories were based

on class abundance and distribution in the study area. However, care was taken that the samples well

distributed across the area and also fulfilled the minimum number required for valid accuracy

evaluation process. Given that the positional accuracy of locations extracted from high resolution

imageries can be degraded by off nadir acquisition and image distortion (Congalton and Green, 2009),

the selected sample points were converted to 5x5 pixels to account for positional error. The geometric

centre of the 5x5 clusters were computed and using in conjunction with differential a GPS (The GPS

Pathfinder® Pro XRS receiver, Ranger TSC2 model, Trimble, California) of 50-cm accuracy to guide

subsequent field verification of the land class. Each site was checked with respect to its assigned class

and, its class re-specified if necessary. A detailed description of each class was recorded and site

photos were taken to assist the training and accuracy assessment process. A total of 802 reference

locations were confirmed for the identified land cover classes (Table 3.3). The number and spatial

dispersion of reference points for each land cover class was kept high in order to optimize

classification using nearest-neighbor classifiers (used for OBC), for training the parametric classifiers

such as the MLC (used for PBC) and to allow a comparison of the parametric and non parametric

methods (e.g. Budreski et al., 2007; Hardin, 1994). The other important consideration in sample

design for the current study was minimum size of features to be extracted. At farmscape level,

individual tree crowns were considered the smallest objects of interest, which could be represented by

50 to 130 pixels, depending upon size of tree and crown form, on a 15-cm image. Thus sample of 5x5

homogenous pixels (~75 cm2) seemed appropriate to train the object size of ≥ 5 m, provided the

clusters were completely within the object.


37

3.3 Image Classification

3.3.1 Object-based and pixel-based classifications

Image segmentation was carried out using the fractal net evolution approach (FNEA), a multifractal

approach implemented in eCognition image processing software (Baaz and Schaepe, 2000; Baatz et

al., 2004; Definiens, 2004). Given the importance of the ‘scale parameter’ in determining the object

size and granularity in an image, scale parameters of 50, 60, 70, 80 and 90 were tested for the

segmentation. After iteratively testing many parameters and observing the effect on rendering known

objects including field boundaries, trees and contiguous exposed rocks and soil features in the image,

the final analysis involved the parameters of: scale, 60; color, 0.7; shape, 0.3; smoothness, 0.5; and

compactness, 0.5. The parameters that offered the best results at 60 scale were used for segmentation

of other scale factors to test the effect of object granularity and size on scale by keeping other

parameters constant.

Three spectral bands along with other contextual information were used to statistically derive twenty-

nine features for each object that best separated the LULC classes: (a) for each of the three spectral

bands, the minimum, maximum, mean and standard deviation (SD), respectively, were calculated

from all pixels forming an object, (b) five geometric features (area, length, compactness, shape and

number of edges), and (c) eight texture features, out of which six GLCM (Grey Level Co-occurrence

Matrix) (homogeneity, contrast, entropy, dissimilarity, standard deviation, correlation) and two

GLDV (Grey-Level Difference Vector) (entropy and contrast) of the visible and near infrared bands.

GLCM describes how different gray level combinations of two pixels occur with respect to their

relative position (Baatz et al., 2004), while GLDV is the sum of the diagonals of the GLCM, another

way to measure texture (Yu et al., 2006). The details of these features can be found in Haralick et al.

(1973) and Definiens (2004). A tree structured classifier, CART, was used to select a subset of

features for classification in a stepwise manner. CART is a recursive and iterative procedure that

partitions the feature space into smaller and smaller parts within which the class distribution becomes

progressively more homogeneous (Breiman et al., 1984; Heikkonen and Varfis, 1998). At the object

level, twelve features best separated the LULC classes and hence were selected for object

classification. The bands beyond these twelve features did little to help separate classes and so were

not used in the classification. All the features were linearly rescaled to the same range. The details of

these features are given in Table 3.2.


38

Table 3.2. Feature objects used in classification Feature objects used in classification Description

GLCM of mean of band 1 (green) Measures the mean frequency of pixel values in combination

of neighbouring pixels values in the green band

GLCM of band 2 (red) Correlation between the values of neighbouring pixels in the

red band

Shape index Ratio of perimeter to four times the square root of the area of

an object

Maximum pixel value of band 3

(infrared)

Measures of vegetation greenness

GLDV entropy of band 2 (red) Measures whether pixels have similar brightness levels in red

band

GLDC contrast of band 3 (infrared) Measures amount of variations in the infrared band within an

object

GLCM SD of band 4 (infrared) Measures the dispersion of variation in the infrared band

within an object

GLCM homogeneity of band 2 (red) Measures he degree that the object displays a lack of variation

in the red band

GLCM homogeneity of band 1

(green)

Measures he degree that the object displays a lack of variation

in the green band

GLDV entropy of band 3 (red) Measures whether pixels have similar brightness levels in

infrared band

Area of object The number of pixels forming an image object.

No. of edges The ratio of the lengths of minor and major axes of an ellipse

approximation of the object.

GLCM SD of band 1 (green) Measures the dispersion of variation in the green band within

an object

For details on each of the parameters, please refer to Haralick et al. (1973) and Definiens (2004).

The object-based classification involved supervised classification using objects selected as training

data based on their class as determined by a combination of field observation and aerial photo

interpretation. Samples for each class were selected from the image objects to act as training areas for

the classification. For the purposes of training the classifier, between 20–50 sample points per class,

depending upon class homogeneity, were randomly sub-sampled from the total sample of 802 points

and the remaining points were retained for accuracy evaluation of each classification. The selected


39

training points were first converted into shape files and then imported into eCognition (e Cognition

Developer 8, Munich, Germany, GmbH) for classification of image objects generated from the

segmentation process. In this study, the non-parametric k-NN technique was used for object-based

classification while the conventional parametric MLC was used in a pixel-based classification. Many

studies classified the segmented objects based on the traditional non-parametric techniques (k-NN)

because it is easily implemented and is sually integrated within the object based classification

software used (e.g., Platt and Rapoza, 2008). To classify an object, k-NN finds the k-neighbors nearest

to the new sample from the training space based on a suitable similarity or distance metric. Unlike

MLC, where training data are statistically condensed into covariance matrices and mean vectors, the

k-NN classifier requires that the actual training vectors participate in each classification. For k-NN, a

three-dimensional feature space was defined for the three spectral bands, rendering each image object

as a point. Since the training samples of each class occupy a spatially clustered location, the final

assignment of an object goes into the class that has the sample nearest to the object in the given

feature space. In this manner, a thematic map was generated and classification accuracy using pixels

as the spatial unit was compared to that using MLC on the same test set.

For comparison, we used the same training set to perform the pixel-based MLC except that we

removed the features specific to objects, such as geometric features and standard deviations. The

classification was carried out in ENVI 4.8 (ITT Visual Information Solution, US). As individual tree

crowns were considered the smallest objects of interest at the farmscape level, the mean of 5 × 5

homogenous pixels provided more accurate class signature representation compared to single pixel

values. Finally, an MLC method was applied via computation of the statistical probability of each

pixel value belonging to a particular land cover class using the mean vector and the covariance matrix.

The candidate pixel was then assigned to the most likely class.

3.4 Accuracy assessment

The remaining sample points, representing a substantial sample sizes for each class, were used for

evaluating the accuracy of the OBC and PBC. Classification accuracy was expressed in the form of an

error matrix of producer’s error (PA) (error of omission), user’s error (UA) (error of inclusion or

commission) and overall accuracy (OA) (Congalton and Green, 2009) along with the Kappa

Coefficient (Congalton, 1991). Following Foody (2004), the statistical significance of the difference

between two classes was evaluated using the McNemar Test, a non-parametric test using the

standardized normal test statistic:

where fij is the frequency of the validation data at row i, column j of the 2 × 2 matrix generated by


40

dissolving the error matrix into two categories (correct and not correct). Where f12 and f21 are the

number of pixels correctly classified by one method as compared to the number of pixels the other

method incorrectly classified (Foody, 2004). The test bases its evaluation on the chi-squared

(χ2) distribution, where the square of Z follows a chi-squared distribution with one degree of freedom

(Agresti, 1996; Foody, 2004) as:

The derived value was compared against tabulated χ2 values to indicate statistical significance at the

95% confidence level.

3.5 Results and discussion

The variation in scale parameter was related to the variable number of objects in each class. The

number of objects in heterogeneous classes like natural pasture low density (NP-LD) and scattered

trees (ST) was greater than in homogenous classes like outcrop (OC) and waterbody (WB) (Table

3.1). The comparison facilitated selection of the optimal scale parameter in this study. Smaller scales

result in higher numbers of objects and increased computation time. Larger scales aggregated

neighboring classes resulting in loss of information.


41

Trees occurring in groups were difficult to classify compared to individual trees because of the

heterogeneous effect (not continuous). On the selected scale parameters (50, 60, 70, and as 80) the

segmentation classifications were compared to determine the optimal scale in terms of accuracies

achieved and also by visual inspection of the image to confirm the agreement with the study area

Site1

Site2

Figure 3.3: Image segmentation results of Airborne 15 cm image for two different sites at the

scale of 50 (a); 60 (b); 70 (c); and 80 (d). The image is a false color composite with Band 1 =

IR, Band 2 = Red and Band 3 = Green


42

features. The segmentation results from different scale factors for two sites are shown in Figure 3.3.

3.5.1 Pixel versus object-based classification results

Figures 3.4 (a and b) shows the mean spectral values for each of the land cover pixel-based training

and the object-based training signatures, respectively, in the three spectral bands. For both pixel-based

and object-based training data, the mean value for natural pasture (MHD) was highest in the infrared

(Band 1) while it was lowest for the feature ‘water’. In pixel-based LULC signatures, the comparable

mean values among the vegetation classes in infrared and red bands and non-vegetated classes such as

bare soil and outcrop in red band showed low class separability and hence intermixing of these classes

can be expected during classification. In object-based LULC signatures, the vegetation classes were

once again not separable in infrared band, however, the classes were found separable in red and green

bands, an improvement over corresponding pixel-based signatures. Since the rural farmscape of study

area is characterized by trees and natural pastures of varying forms and information on vegetation are

mainly obtained through infrared bands, the two spectral signatures dataset seemed inappropriate in

discriminating different vegetation types in the study area. Thus additional information in terms of

shape, texture and context information, along with spectral information, is required to minimize these

intermixing between classes and improve classification accuracy.

Table 3.3 shows the land-cover classification accuracies from PBC and OBC using different scales

factors. The LULC classification accuracies from both OBC at different scales and PBC yielded

overall classification accuracies greater than 77.4 %, with varying degrees of intermixing between the

Figure 3.4 : Mean LULC training samples spectral separability in different bands for (a) Pixel-based, and (b) object-based

techniques.

0

20

40

60

80

100

120

140

160

Band1(IR) Band2(RED) Band3(GREEN)

DN

va

lues

Spectral Bands(a)

Natural Pasture(MHD)

Natural Pasture(LD)

Degraded Pasture

Tree

BareSoil

Outcrop

Water

Road0

20

40

60

80

100

120

140

160


DN

va

lues

Spectral Bands(b)

Natural Pasture(MHD)

Natural Pasture(LD)

Degraded Pasture

Tree

BareSoil

Outcrop

Water

Road

020406080

100120140160


DN

va

lues

Spectral Bands(b)

Natural Pasture(MHD) Natural Pasture(LD) Degraded Pasture

Tree BareSoil Outcrop

Water Road


43

classes. In PBC, more mixing was observed between different types of pasture classes and also

between bare soil, road and outcrop, due to similar spectral responses from these classes (Figure 3.4).

The ‘salt and pepper’ effect further reduced the classification accuracy. In OBC, this effects was not

evident. OBC achieved higher overall classification accuracies for scale parameters of 60 and 70

(86.0%, Kappa = 0.83, and 84.6%, Kappa = 0.81) respectively. The scale factors 50 and 80 reduced

the accuracy to 78.8% and 81.4%, respectively. The difference in classification accuracies can be

attributed to the different nature of the two classifiers, the non-parametric nature of the k-NN, which

is based on the euclidean distance for object classification, and the parametric MLC which assumes

signature to be normally distributed, which is not necessarily the case as this depends upon spectral

characteristics of the feature class. In this study, the signatures for non-homogenous classes such as

degraded pasture and trees along with shadows were found not normal which resulted in intermixing

with other classes in case of MLC and hence degraded the accuracy. The improvements in accuracies

in this case may also be due to different mapping unit involved in classification process, segments or

objects generated through a combination of spectral and contextual information for OBC, instead of

individual pixels spectral value in case of PBC.

Table 3.3. Comparison of LULC classification accuracies using pixel-based and object-based techniques at different scale factors.

The McNemar test confirmed that the classification accuracies derived from PBC and OBC were

statistically different. Predictably, the scale factor effected a change in the granularity of the classified

imagery, this ‘salt and pepper’ effect for the lower scale factor reducing the overall accuracy. It is

apparent that while setting the scale parameter to reduce granularity in the final product improves the

classification accuracy of OBC, a visual assessment of the derived product indicates that this also

improves the structural detail in the shape boundaries. On the whole there was an improvement in

LULC

Class Sample size

PBC

(MLC)

(%)

OBC Scale

50 (%)

OBC Scale

60 (%)

OBC Scale

70 (%)

OBC Scale

80 (%)

AA Tr UA PA UA PA UA PA UA PA UA PA

NP-MHD 76 36 78.2 71.4 87.6 78.8 94.4 79.0 91.3 78.3 93.1 80.1

NP-LD 65 38 76.4 70.3 74.3 82.6 83.1 88.3 93.0 83.1 84.2 82.3

DP 62 33 68.7 91.7 66.3 100.0 83.4 100.0 77.6 100.0 61.3 100.0

ST 71 46 77.1 83.0 81.2 88.4 81.8 94.2 75.1 86.3 73.2 81.3

BS 84 31 81.3 57.1 74.3 38.5 93.4 65.7 94.8 58.8 91.0 47.6

OC 75 28 74.9 84.9 76.8 84.5 85.93 94.6 73.2 94.7 76.8 87.3

RD 34 21 63.2 80.6 71.2 81.6 72.3 82.7 62.1 79.1 58.5 74.6

WB 65 37 92.5 78.9 97.1 85.8 100.0 91.4 100.0 94.7 100.0 95.7

Overall accuracy (OA) 77.4% 78.8% 86.0% 84.6% 81.4%

Kappa 0.75 0.76 0.83 0.81 0.78

AA = accuracy assessment; Tr = training; UA = user's accuracy; PA = producer's accuracy. See Table 1

for class description


44

LULC classification accuracy of ~8.5% with OBC as compared to the PBC as well as OBC

classification yielded visual products that appeared to be more representative of actual conditions on

the ground.

The classification results for two sites in the study area are shown in Figures 3.5 (a) and 3.5(b).

Maximum contrast with less intermixing was observed in ‘waterbody’ and ‘outcrop’ categories.

Waterbody appeared to be mixed with natural pasture in the PBC while the same area was found

accurately (and realistically) delineated in the OBC with very little intermixing. However, the

‘outcrop’ class appeared more real in the PBC and this is likely the result of the spatial

homogeneity within the feature boundaries. Though the segmentation process showed a clear

demarcation of these areas, it is k-NN that was found unable to separate them from neighborhood

classes. In other words, if an object is surrounded by majority of other classes, k-NN was unable to

locate these objects and separate them from surrounding classes. Pixel-based MLC, however, due

to its parametric nature, was found efficient in such situation and clearly classifies these object

from their surroundings objects. However, in the case of features with similar spectral response, for

example a large zone of degraded pasture interspersed within bare soil areas, the degraded pasture

class was not clearly delineated in the PBC. The OBC had difficulty in assigning classes consisting

of very few pixels. For example, with 15 cm resolution data and in the context of utilizing the

derived product as the basis for agricultural land management, classified zones consisting of less

than 25 pixels (0.5 m2) are likely to be meaningless (Yemefac, 2005; Smith and Halvorsen, 2011),

and therefore necessary steps should be taken to merge these pixels into the dominant feature class

in the surroundings. The OBC, however, takes into account the surrounding features, thus providing

more meaningful results in this context. The scattered trees, for example, and owing to their

potential value as a land class in their own right (e.g., Manning et al., 2006), appeared more

realistic in OBC compared to PBC.

Owing to the increasing complexity of spatial structure and spectral composition at smaller spatial

scales, the classification of very high resolution imagery is challenging. Existing pixel-based

classification methods involving only spectral information is prone to yield spatially granular

products. A supervised PBC, typically used as a standard for classifying low or medium resolution

imagery, proved to be of mixed value when applied to this test imagery with a spatial resolution 15-

cm. Alternatively, the OBC proved appropriate for this type of imagery because of the ability to

integrate shape, texture and context information into the process. The critical step in applying OBC is

to decide the specific scale at which to segment an image into objects. The improved segmentation

resulting from OBC in this LULC classification as compared to PBC confirms our hypothesis that

spectral separability between LULC categories can be enhanced at the object level as opposed to the

pixel level.


45

The study has served to highlight the relative advantages and disadvantages to both the pixel and

object-based classification methods as applied to a farmscape: landscape comprising open fields

(pasture) and remnant vegetation. Although the pixel-based method retains the spectral information of

the original image, did not provide a good differentiation between most of the LULC categories. On

the other hand, due to the generalization of spectral information into segments based on the local

homogeneity criterion, the object-based method proved more capable of handling spectrally mixed

classes; a finding echoed on other work (e,g., Wang et al., 2004). However, the OBC process does

risk including neighboring pixels of a different class into a defined object and this can reduce the

classification accuracy. Some feature mixing was observed within the same category (e.g., natural

pasture), however, the different pastures types were expected to be separated clearly on a very high

resolution image of the order of 15 cm. With the optimal segmentation parameters, an improvement in

LULC classification accuracy of ~8.5% with object-based method was obtained over the traditional

Airborne Imagery (Single Band) Pixel-based Classification (PBC) Object-based Classification (OBC)

Site1 (a)

Site2 (b)

Figure 3.5 : Comparison of LULC classification results at two different sites in the study area using pixel-based and object-based techniques.


46

pixel-based method, which was a more realistic visual representation of the actual conditions on the

ground. Thus object-based method was found more suitable for LULC mapping in this candidate

‘farmscape’.

Although the study has shown the object-based method to be more appropriate for using very high

resolution satellite data for LULC classification at farmscape level, the analysis process does have its

limitations. Since the results obtained from this depend partially on the spatial and spectral resolution

of the image used, scene composition, classification system, nature of classifier, training and reference

sample size, and the segmentation parameters, care should be taken to extrapolate the performance

conclusion to any other environment. Nevertheless, the study demonstrated a means to obtain an

accurate LULC map at farmscape level using very high resolution remote sensing data; indeed if

possible a user is encouraged to integrate both PBC and OBC methods and validate a net outcome on

the basis of both techniques. This is an area of future work.

3.6 Conclusion

This study directly compared pixel (PBC) and object-based (OBC) image classification procedures on

a very high spatial resolution (15 cm) airborne image of a demonstrator ‘farmscape’ in Australia. On

the whole, there was an improvement in land use and land cover (LULC) classification accuracy of

~8.5% with the OBC as compared to the PBC. Moreover the OBC process yielded visual products

that appeared to be more representative of actual features on the ground. The improved classification

accuracy of the OBC process, coupled with a more realistic visual representation of ground features

means OBC is a promising technique for the classification of very high spatial resolution imagery of

farmscapes.

3.7 Acknowledgments




CRCSI.


47








Candidate Niva Kiran Verma 80


Nick Reid

5

Brian Wilson

5


Name/title of Principal Supervisor: Prof David. W. Lamb




48







All text All pages





Principal Supervisor Date:29.10.2014

Chapter 4 Tree Cover

49

Chapter 4

Tree cover extraction from 50 cm worldview2

imagery: a comparison of image processing

techniques


Verma, N.K., Lamb, D.W., Reid, N., and Wilson, B. (2014). Tree cover extraction from 50

cmworldview2 imagery: a comparison of image processing techniques. 978-1-4799-1114-

1/13/$31.00 ©2013 IEEE IGARSS 2013, pages 192-195


50

Abstract

High resolution remote sensing is a valuable tool for quantifying the distribution and density of trees

with applications ranging from forest inventory, mapping urban parklands to understanding impacts

on soil nutrient and carbon dynamics in farming land. The present study aims to compare the accuracy

of different remote sensing techniques for delineating the tree cover in 50cm resolution WorldView2

imagery of farmland. An image of farmland comprising pastures, remnant vegetation and woodland

was initially classified into six classes, namely tree cover, bare soil, rock outcrop, natural pasture,

degraded pasture and water body using different techniques. Pixel based classification based on all

four available wavebands, were tested and an overall classification accuracy of 96.8% and 72.9 %

were achieved for supervised and unsupervised techniques. Object based segmentation and

subsequent classification yielded an improved overall classification accuracy of 98.3%. Addition of a

fifth NDVI layer to the available wavebands did improve the accuracy but not significantly (98.1%,

approx 1.3%). In addition to the improvements in overall classification accuracy, a visual inspections

of results from the different methods indicated the object based method to yield a more ‘realistic’

result, avoiding the ‘salt and pepper’ effects apparent in the pixel-based methods. Overall, object

based classification hence is considered more suitable for tree cover extraction from high resolution

images.

4.1 Introduction

As discussed in Chapter 1.2, information on tree cover is important for forest inventory. It plays a key

role in all levels of forest management and planning which is crucial for forest conservation and

environmental management (Leckie et al., 1995). Remote sensing based information on land

cover/forest has been found very useful in forest inventory at various scales. It is also becoming more

important in open farming landscapes, where scattered trees play an important role in providing

shelter for livestock and native animals as well as pasture condition (Barnes et al., 2011 a,b) and soil

chemistry (Graham et al., 2004; Wilson et al., 2007). Contemporary remote sensing systems now offer

spatial resolution in the tens of centimetres, and this offers opportunities for deploying object-based

classification procedures. It has long been anticipated that manual forest inventory procedures will be

superseded by semi-automated and digital remote sensing approaches that promise greater efficiency

and consistency (Bergen et al., 2000; Wulder et al., 2004). Though extraction of tree cover using

conventional visual techniques is accurate, it is time consuming and labor intensive. It has long been

anticipated that manual forest inventory procedures will be superseded by semi-automated and digital

remote sensing approaches that promise greater efficiency and consistency (Bergen et al., 2000;

Caylor, 2000; Pitt et al., 1997). More recently object based approaches have been developed where

results closely match visual interpretation techniques and appear to be more realistic (Blaschke,


51

2010). There is a growing interest in developing tools to inventory scattered trees and tree

communities in agricultural lands. This study aims to investigate the suitability of feature extraction

techniques to delineate trees in high resolution, remotely sensed imagery of an agricultural landscape.

4.2 Materials and methods

The study area is a part of ‘SMART farm’ owned by the University of New England (UNE),

Armidale, New South Wales, Australia (151°36´34´É to 151°38´25´É and 30°26´32´´S to

30°25´22´´S). The farm comprises a total area of 767 hectares which includes large tracts of tree cover

ranging from scattered through to dense woodland. A multispectral, PAN sharpened, WorldView2

image of 50 cm spatial resolution was acquired on January 1, 2012. The image was initially classified

into six landuse/landcover classes namely Tree cover (TC), Bare Soil (BS), Outcrop (OC), Natural

Pasture (NP), Degraded Pasture (DP) and Waterbody (WB) using different techniques, from which

the tree class was subsequently mapped. A pixel-based, supervised classification was performed using

a Maximum likelihood classification (MLC) algorithm by assigning signatures for each class and then

training the classification based on signature statistics. Isodata clustering was used for the

unsupervised classification, by assigning 30 classes for initial classification which were refined and

merged into 6 classes. Georeferenced ground truth samples were used for the accuracy evaluations of

these classifications.

The object based segmentation and subsequent classification was completed in two steps. The initial

segmentation was performed at different ‘scales’ based on the physical size of the features of interest

in the imagery. Numerous objective procedures have been developed as the basis of automating the

feature delineation process (e.g., Linderberg et al., 1998), however, a manual trial and error process

was selected following Mathieu et al.(2007). Scale was considered to be an important parameter in

image segmentation since it determined the size of image objects and denoted the maximum

allowable standard deviation to be utilized in a segmentation procedure (Definiens, 2004). A scale of

40 was considered most appropriate for this study. The effects of ‘color’ and ‘shape’ parameters on

objects created during segmentation were also considered as compactness and smoothness of objects

were associated with its shape. The color and shape factors, inversely proportional to each other, had

potential weightings ranging from 0 to 1. For this study, the values for shape and color were selected

as 0.1 and 0.9, respectively, while value of compactness was retained at the default value of 0.5.

Supervised classification using k-nearest neighbourhood (k-NN) technique was performed on these

objects using the same training samples used in the earlier pixel-based method. Since NDVI is a

global vegetation index and used in many studies in vegetation extraction, an attempt was also made

to assess the usefulness of the NDVI ‘layer’ in the tree extraction process. The pixel and object-based

classification procedures were repeated with the NDVI layer included. The accuracy of all the


52

classified regions was evaluated against a detailed, digitised image sub-scene. The Tree cover class

(TC) of this image sub-scene was carefully, and manually digitised using a combination of visual

assessment and detailed, (GPS)-based, field visitation of selected locations.


Overall, high classification accuracies were obtained from all of the classification techniques used in

this study except from the unsupervised classification. Table 4.1 shows Landuse/Landcover (LULC)

class accuracies in terms of producer, user, overall and Kappa coefficients for all the classifications by

comparing the location and class of each ground-truthed pixels with the corresponding location and

class in the classified images. The highest overall accuracy was achieved in object based technique

(98.3%, Kappa = 0.97), followed by the pixel-based, supervised classification technique, while it was

much lower in the case of the unsupervised classification (72.4%, Kappa = 0.66). In the case of the

object based classification, all of the classes were separable, producing high accuracies, except for

some intermixing of the bare soil with the degraded pasture and rock outcrops (Figure 4.1b). In the

supervised classification procedure, the tree cover and natural pasture classes proved difficult to

separate, although the two classes proved spectrally separable when the NDVI layer was included,

increasing the overall accuracy from 96.8% to 98.1%. Both supervised and object based classification

methods were comparable, and either of these could conceivably be used for tree feature extraction.

Table 4.2 shows the tree area computed from different techniques as compared to digitized tree cover

area which was kept as reference for comparison. Though the classification accuracy was very low for

Table 4.1: Performance of classification techniques used in this study

Supervised

(B1-3 + NDVI)

Supervised

(B1-3)

Object based

(B1-3)

Unsupervised

(B1-3)

LULC Class PA UA PA UA PA UA PA UA

Trees cover (TC) 99.4 99.4 96.4 100.0 99.4 100.0 78.4 76.6

Bare Soil (BS) 91.8 100.0 90.7 100.0 93.0 100.0

84.8

76.8

Natural Pasture

(NP) 100.0 96.8 100.0 93.9 100.0 96.1 100.0 65.7

Outcrop (OC) 100.0 86.4 100.0 78.5 100.0 89.5 25.5 86.6

Degraded Pasture

(DP) 97.5 100.0 96.0 100.0 97.5 99.5 48.0 60.3

Waterbody (WB) 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Overall Accuracy 98.1% 96.8% 98.3% 72.9%

Kappa 0.97 0.96 0.97 0.66

PA, producer accuracy; UA, user accuracy


53

the unsupervised technique, the tree area estimated from it (235 ha) is comparable with the digitized

area (256 ha, 8.2% less), probably due to the formation of a large number of smaller polygons (Figure

4.1d). The area of the Tree cover class derived from the object based classification method was 269

ha, 13 ha more than the digitized area, while area derived from supervised method was 244 ha, 12 ha

less than the digitized area. The reason for the discrepancy illustrates a number of key points: In the

unsupervised classification, spectral data are organised into spectral classes using a clustering

algorithm and each cluster is then associated with a particular landcover class (Jensen, 2005). The

outcome of the unsupervised classification largely depends upon the number of clusters defined at the

beginning of the classification process. A smaller number of clusters generally mean more

intermixing of clusters owing to the intermixing of spectral end-members. Higher number of clusters

improves the purity of pixels, but a larger of smaller polygons result. A supervised classification

procedure allows for the inclusion of mixed pixels as end-members but it then becomes a

classification problem of which of the mixed-pixel end-members a given pixel represents. The results

achieved largely depend on the training pixels used. As far as the object-based method in concerned,

the primary source of degradation is in the boundary delineation. Again this is driven by the nature of

the training data used in the second, supervised classification step. A visual assessment of the

classified products indicates the object based method provides a more realistic and ‘spatially-

smoother’ classification (Figure 4.1a). Both the supervised and unsupervised pixel-based methods

produced the well-known ‘salt and pepper’ effects. Certainly the tree areas extracted from object and

supervised methods were very close to the digitized area used as a reference, but regions delineated by

the object based methods were more aggregated and less susceptible to classification ‘noise’ resulting

from segments of shadow and highlighting within the canopy regions that results from sub-region

differences in surface structure (leaf orientation and density) within the canopy outline.

Table 4.2: Tree area estimated from different classification techniques

Classification techniques

Total Area in

hectares

Difference

in ha

%Difference

Supervised 244 -12 Underestimation 4.6%

Unsupervised 235 -21 Underestimation 8.2%

Object Based 269 +13 Overestimation 5.0%

Visual Interpretation 256


54

Figure 4.1: Tree area extracted from different classification techniques (a) Standard FCC, (b) Object

based, (c) Supervised and (d) Unsupervised.

4.4 Conclusions

The study investigated the relative performance of various landcover classification techniques in

extracting a tree cover class from using high resolution satellite imagery. The results suggest both the

object based and supervised (with and without an NDVI layer included) classification procedures to

be comparable, while the unsupervised method demonstrated the poorest overall classification

accuracy. In addition, the object based method resulted more realistic and smooth classification as

against salt and pepper effects for pixel based supervised and unsupervised methods. Though the tree

area extracted from object and supervised methods were very close to the digitized area used as

reference, tree extracted from object based method was found more aggregated and realistic in terms

of lesser number of spurious polygons formed as in case of supervised method. Therefore, the object

based image segmentation and classification method was found to be the most suitable method for

(a) False color composite (b) Tree from OBC

(c) Tree from supervised classification (d) Tree from unsupervised classification


55

tree feature extraction in the study area. The study demonstrated the means for tree feature extraction

using high resolution remote sensing data which can be used as an important input for forest inventory

as a part of forest conservation and management.


56










Nick Reid

5

Brian Wilson

5






57







All text All pages






Chapter 5 Tree Height from Shadow

58

Chapter 5

The use of shadows in high spatial resolution,

remotely sensed, imagery to estimate the height of

individual Eucalyptus trees on undulating farm land

This chapter has been communicated to The Rangeland Journal as:

Verma, N.K., Lamb, D.W. (2014). The use of shadows in high spatial resolution, remotely sensed,

imagery to estimate the height of individual Eucalyptus trees on undulating farm land.


59

Abstract

The shadows cast by 180 individual Eucalyptus trees, of varying canopy condition, on farm land in

south eastern Australia were used to infer their heights from 50 cm spatial resolution, multispectral

aerial imagery. A geometrical shadow model was developed incorporating the local slope and aspect

of the ground from a DEM at each tree location. A method of deriving ‘local tree time’ to infer the

solar elevation angle, in situations where the image acquisition time is not available, was also

developed. Based on a measurement of the shadow length from the geometric centre of the tree

crowns, and ignoring the role of the crown periphery in distorting the shadow shape, the tree heights

were overestimated by as much as 50%. A geometric correction for shadow distortion assuming

spherical crown geometry provided an improved estimate with a mean prediction error (MPE) error of

±4.8 m (~± 20%).

Keywords: tree height, shadow, allometry, scattered trees, Eucalyptus, farm land.

5.1 Introduction

Recent work on scattered Eucalyptus trees in Australian ‘farmscapes’ indicated that either height or

crown projected area could be used to infer the diameter at breast height (DBH) for single trees, even

though crown projected area was likely to yield to an improved precision in predicting DBH on a tree

by tree basis (Verma et al., 2014a). However, there are some situations where the use of crown

projection area is not feasible, for example where the canopy has been recently degraded by fire, pests

or disease, or it is difficult to discriminate the canopy boundary from the surrounding understorey

(Verma et al., 2013). In these situations tree height may be the only alternative to remotely deriving

values of DBH.

The use of tree shadows to infer tree canopy structural attributes has been described by numerous

workers. For example, Asner and Warner (2003) examined the magnitude and variability of shadow

fraction across tropical forests, savanna and pasture landscapes using IKONOS imagery, concluding

the importance and potential of this approach for divining canopy structural attributes. Greenberg et

al. (2005) demonstrated the concept of ‘shadow allometry’ whereby they estimated DBH and crown

projected area from the dimensions of the canopy shadows in pan-sharpened IKONOS imagery, and

Ozdemir (2008) estimated tree stem volume from a combination of crown area and tree shadow area

using QuickBird imagery.

The shadows cast by trees potentially offer a means of determining their height. With the advent of

very high spatial resolution remotely sensed imagery, tree shadows are often clearly delineated and


60

invariably must be accounted for by their own class in land cover delineation (Benediktsson et al,

2003; Arpan et al., 1997; Dare 2005; Shahtahmassebi et al., 2013; Leblon et al., 1996; Lu 2006;

Verma et al., 2014b). While shadows have been used in estimating the height of buildings (for

example Huertas and Nevatia, 1988; Irvin and McKeown, 1989; Liow and Pavlidis, 1990; Dare

2005), there appears to be little work reported in the literature concerning the use of shadows to

determine the height of trees. One exception is the work of Shettigara and Sumerling (1998), which

used tree shadows, and their corresponding known heights to calibrate an algorithm for estimating the

heights of buildings. The technique was designed for situations where the spatial resolution of the

imagery used is comparable to that of the dimensions of the shadow zone. The technique is applicable

only to extended shadows, such as lines of trees, and assumed the objects in question, were situated

on flat ground. Assuming it is possible to delineate shadows in remotely sensed imagery, the use of

shadows to estimate the height of single trees located on ground of any known slope and aspect is

based purely on geometric considerations; namely the sun and sensor geometry as well as the aspect

and slope of the surface on which they are cast. The aim of this paper is to assess how effectively the

shadows of scattered Eucalyptus trees, as rendered in high-resolution remotely sensed imagery, can be

used to estimate their height on undulating land.

5.2 The relationship between shadow length and tree

height

The total height of a tree can be defined as the distance between the tree base and the highest vertical

extent (tip) of the tree (Leverett 2010; Hunter et al., 2013). In order to understand the relationship

between the length of the projected shadow on the ground surface and the tree height, we must apply a

similar definition to the shadow length. However, it must be recognized that the physical extent of the

shadow on the ground may not reflect the vertical extent of the tree. The asymmetric nature of the

canopy may create a distorted shadow envelope on the ground and so a-priori knowledge of the solar

azimuth angle is important to first identify the direction of the shadow from the tree stem along which

to measure. This can be determined either by accessing the metadata often accompanying remotely

sensed imagery (image time, latitude/longitude), or it may be possible to infer it from within the

images themselves using nearby vertical structures such as power poles and buildings. The

dimensions of shadows on the ground is determined by surface topography of the surface on which

they are projected (Dare, 2005).

Consider a tree on a north-facing hill slope of angle gs with the sun also due north at a solar elevation

angle e (Figure 5.1)


61

Figure 5.1: North-facing slope (ga = 0) and sun due north ( a = 0)

According to Figure 5.1, the elevation angle of the sun relative to the hill slope is . The

height of the tree, h, is related to the shadow length on the sloping ground, l, according to the sine rule

by

.

Hence

, (1)

and

. (2)

Now, if the sun is not due north, but with an azimuth angle a, and if the slope is not directly facing

the sun, but also with an aspect (azimuth), ga, then the tree shadow will lie on a different slope, ss, to

that of the actual hill slope (Figure 5.2). The sine rule in Equation 1 is rewritten to be

and


62

(3)

Figure 5. 2: Shadow projected on a slope (ga) of aspect angle sa and slope ss from the sun at an

elevation angle of e and azimuth of a. Note, the sun is depicted as large in size as it is portrayed in

the semi-foreground.

In order to calculate the shadow slope, ss, the azimuth angle of the sun relative to the slope aspect is

and the shadow slope, ss is therefore given by

,

and thus

, (4)

where a = 180- sa, in other words the shadow aspect and sun azimuth are directly opposite. The solar

azimuth ( a) can be calculated using appropriate solar almanacs (Walraven, 1978) and from this the

shadow azimuth calculated using sa = 180- a.

The estimation of tree height from shadow measurements as described above is based on four

assumptions, namely:

(1) that the height of the tree, h is measured vertically from the base of the stem;


63

(2) the shadow length, l, is likewise measured from the geometric centre of the base of the stem

on the ground;

(3) the geometric centre of the tree canopy will be taken to be the same as the geometric centre of

the base of the stem on the ground; and

(4) the tangent of the sun’s rays at the tree canopy, that determines the shadow extent on the

ground, passes through the top of the tree canopy.


5.3.1 Study Area

The model described in Section 2 was evaluated using remotely sensed imagery from the ‘Newholme-

Kirby’ SMART Farm, Armidale, New South Wales, Australia (longitude 151°35´40” E to 151°37´12”

E and latitude 30°26’09”S to 30°25’12”S) (Figure 5.3). The area was dominated by three major

landcover classes namely forested area, woodland and native pastures. Approximately one third of the

study area was forested; a third was woodland, and the remainder native pasture. Most of the property

is managed in an agriculturally ‘un-manipulated’ manner, other than grazing. The property

encapsulates the Mt Duval Nature Reserve, supporting old-growth native forest, and a 300 ha

conservation zone ‘Mountain Paddock’. The area is characterized by low undulating terrain with

dominant outcrops. The slope had an elevation ranging from 1000 – 1500 metres.

Figure 5.3. Location of the study area in southeastern Australia.


64

5.3.2 Image data

Multispectral imagery of the study area was acquired at approximately 1045 hrs (AEST) on 3

November 2011 using an ADS40 airborne digital scanner (West and Glasbury 2010; Sandau et al.,

2000). Flown at an altitude of 1920 m above ground level (AGL), the 8-bit images were acquired with

a spatial resolution of approximately 50 cm in five spectral bands: Band 1 (NIR 0.7-1 , Band 2

(Red 0.6-0.7 ), and Band 3 (Green 0.5-0.6 ), Band 4 (Blue 0.4-0.5 ). The image transects

were mosaiced and the complete image geo-referenced using ground control points (Figure 5.4).

Figure 5.4: ‘False colour’ image of the study site showing image transects flown (red lines), the

sampled trees (green circles) and the mosaic seam lines (yellow lines).

5.3.3 Field Data Collection

A total of 180 trees were sampled for the study, comprising of 121 ‘live’ trees with foliated canopies

and 59 ‘dead tree’s with skeletal (defoliated) canopy structures. A random sampling design was used

in selecting the trees from the image of the study area. The locations of all candidate trees first were

extracted from the image. Given the positional accuracy of these locations could be degraded on the

off-nadir ‘limbs’ of the scan transects (Congalton and Green, 2009), the pixels ascribed to individual


65

trees were first converted to 5x5 pixels to account for positional error. The geometric centre of the

5x5 clusters was then computed and this location was used using in conjunction with a differential

GPS (GPS Pathfinder® Pro XRS receiver, Ranger TSC2 23 model, Trimble, California) to guide

subsequent field visitation of each respective tree.

The heights of each selected tree was measured using a handheld laser range finder (MDL LaserAce

300, Measurement Devices Ltd. Scotland, UK) following Asner et al. (2002). A handheld clinometer

was also used to measure each tree height and the results compared and found comparable to within ±

2-3 cm (~ ±0.1%).

5.3.4 Image analysis and calculating input parameters

The required inputs necessary to determine the height of a tree from the model of Section 5.2 are

summarized in Table 5.1. These are listed in order of derivation.

Table 5.1. Environmental and tree parameters necessary to infer tree height from shadows.

Input Description Symbol How determined…

Shadow azimuth sa Directly from sun azimuth (a) if known from metadata, or directly

inferred from the shadow in image.

Sun azimuth a Directly from image metadata (time, date and location) or inferred

from shadow direction in image.

Ground slope gs Digital elevation model

Ground aspect ga Digital elevation model

Shadow slope ss Equation 4

Sun elevation e Directly from image metadata (time, date and location) or can be

inferred by applying image-derived sun azimuth to an almanac of

known day and location to determine the corresponding elevation

angle

Shadow length l Extracted from fitting shadow azimuth line from geometric centre


66

of the tree base.

Tree height h Equation 3

Shadow and sun azimuth (sa, a)

As the image scene is a mosaic of successive image transects, the acquisition time, hence the solar

elevation angle and solar azimuth is different for different segments of the mosaiced image. In this

work, the image acquisition time, and hence the sun azimuth was known (a = 55.7o). However, in

addition to this value, and for comparison purposes the sun azimuth was calculated for each tree from

the shadow azimuth of the respective shadows. The observed shadow azimuth was defined as the

angle measured clockwise from the true North. Two approaches were tested for extracting the shadow

azimuth. For full tree canopies, the tree canopy partially obscures the projected shadow on the ground

and the canopy itself (and shadow) may not be symmetrical about the centre of the tree canopy nor

tree stem. In order to minimize the errors arising from this, two tangential vectors were drawn joining

the visible crown periphery and the corresponding shadow periphery (Figure 5(a)) and the two

azimuth angles (sa1, sa2) measured. From these the average angle was determined (sa).

Figure 5.5: (a) Schematic of a tree canopy (grey shape) and its projected shadow (black shape) on the

ground beneath. The two vectors (black arrows) are used to determine the azimuth angles from which

the shadow azimuth is calculated from the average). (b) Schematic of a dead tree ‘skeleton’ with the

shadow of the trunk clearly projected on the ground and the vector (black arrow) indicating the trunk

shadow azimuth.

sa1

sa2

tree canopy envelope

projected shadow

on ground

North

sa

tree canopy skeleton

projected shadow

of trunk on

ground


67

The second approach worked only on dead trees where the canopy allowed a view of the shadow of

the tree trunk on the ground. In this method a single vector was used to determine the shadow azimuth

(Figure 5(b)). From both types of the shadow azimuth measurements, the solar azimuth for particular

tree was directly calculated.

Sun elevation angle, e

Given the possibility that different trees may have different image times, the sun elevation was

derived from the sun azimuth (per tree) using a calibration curve (Figure 5.6) derived from an

almanac (http://aa.usno.navy.mil/data/docs/AltAz.php, accessed January 2014) generated using the

known latitude/longitude coordinates of the study site as no solar azimuth was available.

Figure 5.6: Sun elevation (e) /sun azimuth (a) conversion curve for the study site. Fitted 5th-order

polynomial curve from which the calibration equation was derived has R2 = 1.0.

The conversion equation generated from Figure 5.6 was:

. (5)

Ground slope (gs) and aspect (ga)

The ground aspect and slope at each sample tree was calculated from a vector indicating the steepest

downslope direction at each location on the ground surface. A slope and aspect map for the study site

was derived from an 8 m digital elevation model (DEM) of the site. A buffer of width equal to the

30

35

40

45

50

55

60

65

70

75

80

-60 -40 -20 0 20 40 60 80 100

e

(deg

rees

)

a (degrees, relative True N)

http://aa.usno.navy.mil/data/docs/AltAz.php


68

projected canopy diameter of each tree was created around its projected shadow and this region was

then intersected with the DEM. From the intersected regions both the ground slope (gs, percent) and

the ground aspect (ga, degrees) of each tree was calculated in ArcGIS 10 (ESRI, Redlands CA USA).

Quantifying model performance

The difference between the calculated (predicted) tree height and the measured (actual) values was

quantified using a mean prediction error (MPE) given by

.

5.4 Results and Discussion

5.4.1 Tree height

An example of trees selected for analysis is given in Figure 5.7. The tree centre was manually selected

in the airborne imagery and a line vector was generated in direction of the shadow. The length of the

line vectors were calculated using the ‘Calculate Geometry’ Command in the tables in ArcGIS 10.

Figure 5.7: ‘False colour’ image of a portion of the study site. The individual trees selected for

evaluation/analysis are mark with green dots, which also indicate the assigned geometric centre of the

tree canopies. The vector describing the average shadow azimuth (yellow line), emanates from each

green dot circle. The large, hollow red circles indicate dead trees where it is possible to view the

shadow of the trunk and hence its azimuth.


69

The average height of the 180 trees was 20.68 m with a standard deviation of 6.39 m. Similarly, the

average shadow length was 11.93 m with a standard deviation of 3.44 m. A histogram of the derived

sun azimuth angle, e from the shadows of all the trees is given in Figure 5.8.

Figure 5.8: Scatter plot of estimated sun azimuth as derived from the azimuth angles of all the tree

canopies. The solid grey columns corresponds sun azimuth angles between 1030 hrs and 1100 hrs

AEST, and the black column is the category containing both the average sun azimuth angles derived

from all trees (a = 42.9o) and that corresponding to the proported image acquisition time (1045 hr

AEST, a = 40.4o).

If the tree canopies were all located within a single image (a subset of the generated mosaic) acquired

at the same time, namely 1045 hrs AEST then it would be expected that all the derived sun azimuth

angles would be the same, approximately 40.4o. This would have produced, in turn a single value of

solar elevation angle, namely 70.4o. If, for example a mosaic seam intersected the image scene

(vertically, given the N-S-N transects of the aircraft in Figure 5.4), then a bi-modal distribution would

likely have resulted, and likewise a spread of angles would indicate variable acquisition times for

mosaiced image parts. The fact that we have the spread indicated in Figure 5.8 for a supposedly single

image reflects the inaccuracies in deriving actual sun azimuth angle using shadows. The average

azimuth angle in Figure 5.8 is, in fact 42.9o. Given the small uncertainty in the actual image

acquisition time and the resolution of the latitude and longitude values accepted in the almanac used

to calculate the time-sun azimuth-sun elevation profile of Figure 5.6 (rounded down to whole minutes

of arc), these angles can be considered equivalent within uncertainty. Both values lie within the same

solid black column in Figure 5.8 and the average value, a = 42.9o will be hitherto used. The sun

elevation angle resulting from Equation 5 is 69.7o.

Figure 5.9(a) is a scatter plot of the estimated tree height using the tree canopy shadows and their

‘respective azimuth angles’, as derived from the individual shadows, to infer the sun elevation angle,

against the field-measured values. The mean prediction error (MPE) between the predicted and actual

0

2

4

6

8

10

12

14

16

18

20

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75

Pro

po

rtio

n o

f tr

ees

mea

sure

d (

%)

Estimated sun azimuth angle, a


70

values is ± 10.61 m, or 21% of the maximum value measured. Figure 5.9(b) is the scatter plot

resulting from applying the sun elevation angle (e = 69.7o derived from the average azimuth angle).

The MPE value is slightly increased. Both plots indicate the tree heights to be overestimated by as

much as 50%.

(a)

(b)

Figure 5.9: Scatter plot of tree height estimates from shadow using sun elevation angle derived from

(a) sun azimuth values derived from the individual shadows themselves (Figure 5.8), (b) the average

sun azimuth from all trees (42.9o) converted to sun elevation angle (n = 180).

5.4.2 Challenging the basic assumptions

There are two key assumptions discussed earlier in Section 5.3 that could contribute to a systematic

overestimation of the tree height; namely:

Assumption 3: the geometric centre of the tree canopy will be taken to be the same as the geometric

centre of the base of the stem on the ground; and

Assumption 4: the tangent of the sun’s rays at the tree canopy, that determines the shadow extent on

the ground, passes through the top of the tree canopy.

For the first assumption, it is noted that all the shadows are projected in a direction south-west of the

tree base (in Figure 5.7). The shadow length is defined as the distance from the geometric centre of

the canopy envelop, which is assumed to be the same location as the base of the stem, and the tip of

the shadow on the ground in the direction determined by the derived shadow azimuth. If the trees are

0

10

20

30

40

50

0 10 20 30 40 50

Est

ima

ted

tre

e h

eig

ht

(m)

Actual tree height (m)

RMSE = 10.61 m

0

10

20

30

40

50

0 10 20 30 40 50

Est

ima

ted

tre

e h

eig

ht

(m)


RMSE = 11.3 m

MPE

MPE


71

imaged from off-nadir, the top of the tree canopy can be shifted relative to the centre of the stem on

the ground. For the shadow length to be overestimated, this would require that the tree canopies be

shifted east relative to the westerly shadow envelope on the ground and this, in turn, would require

that the trees would have to lie on the eastern side of the aircraft track. Moreover, the scattered data in

Figures 5.9(a,b) and the 1:1 equivalence lines indicates the heights to be overestimated by

approximately 50%, which following Equation 3 means a 50% overestimation of shadow length. It

can be seen in Figure 5.7 that as much as 75% of the shadows are obscured by their respective

canopies. Assuming 75% ‘coverage’ by the canopies, an average tree height of 20.68 m, and an

average shadow length of 11.93 m, shifting the top of the canopy eastward from the nadir view by

50% of the net shadow length requires a view angle of 17o off nadir. This is possible given the across-

track field of view of the sensor is 23o, however the image of Figure 5.4 confirms the trees to be

spread uniformly across the field of view. In fact while this rules out this effect as a contributor to the

systematic overestimation of tree height, it explains the spread in values Figure 5.9 (a,b).

The second assumption is depicted schematically in Figure 5.10.

(a)

(b)

e

h

l'

l

l

e

l

rr'

r

ry'x'

x

y


72

Figure 5.10: (a) Schematic diagram of the sun’s tangential ray that defines the shadow length, l, on the

ground, and the shadow length, l’, assumed to represent the tree height. (b) Geometric representation

of the canopy envelope and the resulting increase (l) in shadow length that results from the

tangential ray passing the outer extent of the canopy.

Figure 5.10 (a) depicts a tree with a circular canopy (dotted circle). The sun’s rays that passes through

the top of the tree canopy, assuming that this is located at the geometric centre of canopy envelope as

viewed from above, actually produce a shadow length, l’. This is different from the measured shadow

length on the ground, l, that results from tangential rays that pass the periphery of the canopy. A

geometric representation of the canopy envelope (dotted circle) is given in Figure 5.10 (b). Here we

assume the geometry of the canopy can be represented by a circle of radius r. The difference between

the two projected shadow lengths is given by l and this is shown in Figure 5.10 (b) to be r’- x’.

The equation (Cartesian coordinates) of the upper right quadrant of the circle depicted in Figure 5.10

(b) is:

(6)

The x coordinate (r’) of the location on this circular envelope that the tangential ray contacts can be

determined as this is the location where the angle of the circular envelope is the same as the sun

elevation angle, namely where the two gradients are the same.; in other words when

This occurs when

(7)

Now, according to Figure 5.10 (b)

(8)

where ry’ is the y coordinate of the location on this circular envelope that the tangential ray contacts.

Moreover, from Equation 6,

(9)

Therefore, combining Equations 7 and 9 into Equation 8 yields


73

(10)

and

(11)

where

is hitherto referred to as the ‘canopy shape correction coefficient’.

Assuming a circular canopy shape, the shadow length correction, l, is dependent of the crown

projected radius, r, and the sun elevation angle through the coefficient, k. For a sun elevation angle of

e = 69.7o, l ≈ 0.69r. Given , and using Equation 11 above, Equation 3 now becomes

(12)

Of course, re-calculation of the reduced shadow length using Equations 11 and 12 requires knowledge

of the canopy radius, r of each tree. Ideally, in keeping with an image-based approach, this would be

derived from the crown projected area for each tree as extracted from imagery, following Verma et al.

(2013). In this work, however, the crown projected area for each of the trees were also physically

measured on the ground (Verma et al., 2014a) and in order to test the validity of the correction factor

approach, the canopy radii values were calculated from these data assuming a circular crown

projected area. Using the derived values of r for each tree, the recalculated tree height is given in

Figure 5.11 (a). The correction factor has resulted in a closer relationship between the estimated tree

height and that measured on the ground, with a MPE of ±5.6 m.

Increasing the value of the canopy shape correction coefficient, k, moves the scattered points

downward. In the earlier Figure 5.9 (a), where the correction factor was not applied, the value of k is

effectively zero. Reducing the value of k from the initial value (circular canopy envelope) of 0.69, by

13% to 0.60 results in the scattered data lying symmetrically about the 1:1 line, with a minimum MPE

of 4.8 m. This process of ‘optimisation’ may sound somewhat arbitrary; however the coefficient

value of 0.69 is derived from the assumption of a circular canopy shape. While a sound assumption

given the trees in question, this may not be the case over unknown landscapes. Alternatively the use

of a ‘peaked’ ellipsoid canopy envelope for Equation 6, that is one with a vertical radius slightly


74

larger than the horizontal radius (for example Medhurst and Beadle, 2001; Rouvinen and

Kuuluvainen, 1997) would also result in reduction in the value of k.

(a)

(b)

Figure 5.11: Scatter plot of tree height estimates from shadows using sun elevation angles derived

from the sun azimuth values for each shadow and the corrected shadow length (a) l’= 0.69r and (b)

l’= 0.60r. (n = 180).

5.5. Conclusions

A method for estimating the height of single eucalyptus trees from a single, high spatial resolution

image scene, has been developed. The objective of this approach was to determine the height of a tree

from its projected shadow on the ground, taking into account ground slope and aspect, at any time and

any location. This of course pre-supposes the ground on which they exist is the same ground on which

the shadows are cast. In effect the process uses the shadow itself to first define a ‘local tree time’

from which to extract the local sun elevation angle. This process is applicable not only to single scene

images where the acquisition time may not be available (e.g., a single satellite or airborne image

scene), but also for mosaiced imagery where information on the acquisition time corresponding to

location individual objects within the mosaic scene is not available.

The accuracy of tree height estimations is reliant on accurate shadow azimuth values, but is also

affected by where the trees lie within the image scene. Off-nadir trees have canopy envelopes

displaced relative to the projected shadows on the ground and this introduces an error in the

0

10

20

30

40

50

0 10 20 30 40 50

Est

ima

ted

tre

e h

eigh

t (m

)


RMSE = 5.6 m

0

10

20

30

40

50

0 10 20 30 40 50E

stim

ate

d t

ree h

eigh

t (m

)Actual tree height (m)

RMSE = 4.8 mMPE MPE


75

estimation of the shadow length; it was the primary reason for the scatter in the plots comparing the

estimated and measured tree heights.

It was necessary to introduce a correction factor to the shadow length to account for the fact that the

shadow length is defined not by the rays that travel through the peak of the canopy, but rather the

canopy periphery. An analytical function for this correction factor was derived from simple geometric

considerations, assuming the trees to have a circular, horizontal canopy shape profile; this correction

factor is a simple function of the sun elevation angle and the radius of the crown projected area. While

in this current work the crown projected areas used in the correction process were derived from

physical on-ground measurements, it is envisage such measures would be extracted from the same

imagery as the shadow measurements.

Ultimately the accuracy of the height estimates in this work (MPE ±5.6 m, with possible optimization

down to ±4.8 m) is the result of accumulating uncertainties in delineating the shadow and extracting

the related shadow azimuth angle and length, the local topographical data (slope and aspect),

extracting the canopy radius (from image) and the applicability of any assumptions around the

geometric form of the horizontal tree canopy profile. Nevertheless the work demonstrates a possible

pathway to inferring the height of individual trees from imagery alone.

5.6. Acknowledgments




CRCSI. We would like to thank Ashley Saint and Derek Schneider (UNE-PARG) for their assistance

in conducting the field work.


76















77







All text All pages


Name of Candidate: Niva Kiran verma




Chapter 6 Crown Projected Area

78

Chapter 6

Estimating crown projected area from remote

sensing at different spatial resolution and its use in

estimating DBH

6.1 Introduction

Previously, Chapter 2 highlighted the potential of crown projected area to infer DBH of the scattered

Eucalyptus trees. Chapter 5 subsequently explores the use of tree shadows to infer tree height. With

the array of very high resolution datasets now available, the question is ‘how well can multispectral

images of different spatial resolution estimate crown projection area?’ This chapter investigates two

different datasets from different remote sensing systems; one of 15 cm (airborne), and the other of 50

cm (airborne and spaceborne) spatial resolution for estimating the crown projection area of scattered

eucalypt species.

Large Scale Photographs (LSP) and photomensuration methods have largely replaced on-ground

methods of estimating planimetric crown area, or more correctly ‘crown projected area’ (Bertolette

and Spotskey, 1999; Clark et al., 2004). However, most of these approaches involved manual

measurement on the derived data.

The choice of an appropriate scale, or spatial resolution, for a particular application depends on many

factors which includes information desired about the ground scene, the analysis methods to be used to

extract the information, and the spatial structure of the scene. Woodcock and Strahler (1987) carried

out a study where they showed that the local variance of a digital image for a scene changed as the

resolution-cell size changes. Their graphical process can help in selecting an appropriate image scale.

These graphs could be obtained by imaging the scene at fine resolution and then sub-sampling (or

block pixel-averaging) the image to successively coarser spatial resolution while calculating the local

variance. Their findings confirmed that the local variance/resolution graphs for the forested,

agricultural, and urban/suburban environments reveal the spatial structure of each type of scene, and

that this is a function of the size and spatial relationships of the objects contained within the scene. At

the spatial resolutions of SPOT and Thematic Mapper imagery, local image variance is relatively high

for forested and urban/suburban environments, suggesting that information-extracting techniques

utilizing texture, context, and mixture modeling are appropriate for these sensor systems. In

agricultural environments, local variance is low, and the more traditional classifiers are appropriate.


79

Of course the spatial resolution of remotely sensed digital datasets can be as high as decimetres for

both airborne (Chapter 2) and spaceborne (Chapter 3) sensors. Unlike Landsat or SPOT satellite data

where a single pixels can encompass many tree crowns, or significant non-crown features, decimetre

spatial resolution data makes crown assessment studies with single trees possible (e.g., Chapter 2 –

Verma et al., 2014b, Chapter 3 – Verma et al., 2013; Palace et al., 2007; Song et al., 2010; Chopping,

2011). Chubey et al. (2006) modelled canopy cover directly based on spectral and spatial features of

the image; on the other hand, Sanquetta et al. (2011) measured canopy area by projecting the canopy

and then transferring to AutoCAD where the spatial location of each tree from the initial coordinate

was adjusted to an adequate scale. Asner et al. (2002) used IKONOS data to map tree crown size and

by comparing the results with ground measurements, they concluded that satellite based observations

have overestimated larger crowns leaving the smaller crowns undetected. Song et al. (2010) studied

the potential of using the behavior of image semivariograms at different spatial resolutions to estimate

tree crown size from IKONOS and Quickbird images and concluded that this approach can provide

estimates of average tree crown size for hardwood stands. They also concluded that the model can be

generalized across sensors and sites. Greenberg et al. (2005) have effectively used IKONOS data

(spatial resolution 4m) in estimating crown projected area, DBH and stem density. Even though high

resolution remote sensing data have successfully been used in many applications, as discussed earlier,

there are some challenges with a very high resolution data, including the increase of intra crown

spectral variance and the low spectral separability between tree crowns and other vegetated surfaces

in the understorey (Chapter 3 – Verma et al., 2014b; Gougeon and Leckie., 2006; Hirschmugl et al.,

2007; Pouliot et al., 2002) limit the identification of tree crowns, particularly with pixel-based spectral

classifiers. However, with introduction of contextual information into the classification process in the

form of object based image analysis (OBIA) have bridged the gap between the increasing amount of

detailed geospatial data and the inefficient results of conventional pixel based classifiers (Chapter 3 –

Verma et al., 2014b; Blaschke, 2010). The details of OBIA were discussed earlier in Chapter 3

(Verma et al., 2014b).

Following Chapter 3 (Verma et al., 2014b), crown projected area can potentially be extracted from

very high spatial resolution imagery by either of the two methods; manual or on screen digitization or

by automated methods, for example segmentation and classification. Both of these methods will be

tested in this chapter, and the compared with the field measurements of crown projected area.

Furthermore, the allometric equation linking crown projected area with DBH developed earlier in

Chapter 2 (Verma et al., 2014a) will also be applied and the performance of the two sensor datasets

evaluated in terms of predicted DBH as compared to the on ground measured values.


80


6.2.1 Study Area

The study area was same as that used in Chapter 5, and is described in section 5.3.1.

6.2.2 Remote Sensing Datasets

The digital imagery used in this study was acquired from some sources already described in earlier

chapters. Multispectral imagery of 15 cm spatial resolution was acquired using the Duncan Tech

MS4100 camera system mounted in a Cessna 172 aircraft (described in Chapter 2, Verma et al,

2014b) and 50 cm multispectral imagery was acquired from the airborne ADS40 sensor (Section

5.4.2). A 50 cm spatial resolution, multispectral, PAN sharpened WorldView2 (WV2) orthoimage

was also acquired on 1 January 2012 with four spectral bands (Blue 0.4-0.5 μm), Band 2 (Green 0.5-

0.6 μm), Band 3 (Red 0.6-0.7 μm) and band 4 (NIR 0.7-1 μm). In order to compare the performance

of the three datasets, only the three spectral bands common to all of the datasets, namely Green (0.5-

0.6 μm), Red (0.6-0.7 μm) and NIR (0.7-1 μm) were used.

The two 50 cm spatial resolution images, namely the airborne ADS40 and spaceborne WV2 were

resampled to 15 cm using nearest neighbor resampling technique to match with the resolution of

MS4100 dataset for pixel wise comparison.

6.2.3 Field Data Collection

The same 172 individual mature Eucalypt trees used in the earlier allometric equation development

(Chapter 2 – Verma et al., 2014a) was used in this study, and their presence in each of the three image

datasets was confirmed. The characteristics of the trees are summarised in Table 2.2, Chapter 2.

The methods by which crown projected area (CA) and DBH were measured are also described in

Section 2. 2.2 of Chapter 2.

6.3 Data Analysis

All three images were registered and georeferenced to WGS 84 UTM Zone 56 S projection systems.

6.3.1 Manual method (On screen vectorization)

On screen digitization of the tree crowns in the imagery involved manually vectorizing the tree crown

envelope based on the visual interpretation. The process of vectorization was performed using the

approach of Gougeon. (1995). This approach treats the brighter and darker pixels as tree crown and


81

shadow, respectively. The crown projected area was delineated from all the three image sets keeping

the scale of the view constant during the vectorization process. This was to avoid any over and under

estimations of the canopy area. Different sets of band combinations were tested to extract the tree

crown projected area. Since tree crown are usually associated with shadows, hence to get a clear

demarcation of shadows with trees both true colour composite (Red, Green and Blue Bands) and false

colour composite (NIR, Red and Green) was tested. The true colour image gave the best interpretation

results as it helped decipher the crown with shadowed areas. These results were based solely on the

visual inspection. Onscreen digitization step similar to Ke and Quackenbush. (2010) was followed by

the area calculation for each polygon representing tree crowns.

6.3.2 Automated method (Image Segmentation and Classification)

The second method of tree crown extraction involved segmentation of image features into objects and

then classification of objects into a given class using a method called object based image analysis

(OBIA). It is an automated approach which takes into account the form, textures and spectral

information of the image, as discussed earlier in Chapter 3 (Verma et al., 2014b). The analysis was

done in eCognition Developer 8 (Munich, Germany, GmbH; Blaschke and Strobl., 2001).

Segmentation and classification was performed on all the three sensor datasets. The key ‘adjustable’

parameters in the segmentation process of scale, colour and shape have been discussed in Chapter 2.

Again, the quality of the segmentation output was visually assessed for all the three images.

Following the segmentation process, a supervised nearest neighbour (NN) classification was used for

the classification of the image objects. Since the prime objective of the study was to only extract the

tree crowns, the images were classified into only two classes, namely ‘tree’ and ‘no trees’. Features

such as mean pixel value, brightness, standard deviation and area (number of pixels) were considered

for the NN classification. Additional arithmetic features like the NDVI and the Ratio image (NIR and

red) were also taken into account. However, no textural features, such as GLCM and GLDV contrasts

were used as after a number of exploratory trial and error tests textural attribute did not appear to help

in delineating the tree crowns. Following classification, the ‘tree’ class was exported as a separate

shape file and then intersected with the sampling locations. The area of the tree crown polygons with

information on sample number was recorded.

6.3.3 Statistical Analysis

The three sets of measured crown projected area were tested for normality (Q-Q plot and Shapiro

Wilk test) and in case of non-normality, a transformation was carried out; this was assessed by

performing the Wilcox test. Scatter plots were created based on the ground and remote sensing based


82

measurements using the statistical software R (Studio Version 0.97.318). The coefficient of

determination (R2) was used to evaluate the level of variance in the estimates.

The performance of each sensor was quantified using a mean prediction error (MPE) given by

MPE= CApredicted-CAactual

calculated.

6.4 Analysis Results

Examples of the classified tree polygons are given in Figure 6.1.

Although the spectral bands for each of the images were the same, the scale parameter which gave the

best visual segmentation results varied among the three sets of images. The parameters shape and

compactness remained the same for three images (shape = 0.7 and compactness = 0.5). The MS4100

image was segmented at a scale of 40 while ADS 40 and WV2 images were segmented at scale of 70

to get the optimum result.


83

Color Infrared (CIR) image (a) Tree extracted from CIR image

World View 2(WV2) image (b) Tree extracted from WV2 image

ADS40 image (c) Tree extracted from ADS40 image

Figure 6.1shows the trees generated by automated methods a) Color Infra red image b) World View2

image c) ADS 40 image


84

Figure 6.2 shows the scatter plots of field measured CA and manually delineated CA from different

sensors. Regression analysis shows a good correlation between crown projected area estimates from

manual method. Although the manual method of CA extraction was accurate WV2 and ADS 40

explained only ~70% of the variance in the CA (R2 in the range of 0.67 to 0.68) with MPE of 55 and

53 m2 respectively (26% and 25% respectively) whereas MS4100 showed a better performance by

explaining 76% of variance with MPE of 48 m2 (error of 22%). The CA measurement results from

segmentation and classification shows similar trend. Figure 6.3 shows the scatter plots of field

measured CA and automatically delineated CA from different sensors. As expected, error was higher

in this case though not very significant. Interestingly WV2 results showed a decrease in MPE slightly

(error down by 1%), whereas ADS40 and MS4100 showed an increased estimation error of 29% and

26% respectively compared to 25% and 22% by manual methods. Figure 6.4 shows the scatter plots

of manual and segmentation based CA estimations in the three sensors. A very good agreement

between crown projected area estimates from manual and automatic methods indicates that either of

the two methods can be used with confidence.

(a) (b)

(c)

Figure 6.2. Scatter plots of the derived CA from each of the images versus the field-measured CAfield.

(a) MS4100, (b) ADS40, and (c) WV2. The image-derived CA was calculated using manual

vectorization; n = 172.


85

(a) (b)

(c)

Figure 6.3. Scatter plots of the derived CA from each of the images versus the field-measured CAfield.

(a) MS4100, (b) ADS40, and (c) WV2. The image-derived CA was calculated using image

segmentation; n = 172.

A comparison of mean CA estimates by the three sensors using the two methods shows variations

which appeared to be non-significant. As against the observed mean field based measurement value

(211), MS4100 image showed a mean value of 183 (underestimation), ADS40 -229 and WV2 a mean

of 221. The mean CA values for both segmentation and manual methods were almost comparable

except for WV2 where the manual method of extraction resulted in overestimation of the CA. For

MS4100, as compared to the manual method, segmentation method showed underestimation of

canopies, whereas ADS40 showed overestimation of canopies but not very significant.

Underestimation by manual method was possibly due to the shadow effect where the border was not

defined, on the contrary segmentation and rule based classification helped in excluding the shadows

more effectively.

The mean crown projected area estimates from each of the sensors, compared to the field-based

measurements are summarised in Table 6.1.


86

Table 6.1.The crown projected area estimates by the two methods compared to the field

measurements

Image source

Mean CAfield

Mean image-derived CA

SD CAfield Manual

vectorization

Segmentation

MS4100

210.96 m2

194.52 m2 182.87 m

2

ADS40 136.39 197.86 m2 228.72 m

2

WV2 243.41 m2 221.34 m

2

(a) (b)

(c)

Figure 6.4. Scatter plots of the derived CA from each of the images using manual methods versus the

segmentation based CA (a) MS4100, (b) ADS40, and (c) WV2 (n-172)

The conversion of image-derived CA (each sensor) into estimates of DBH for the candidate trees,

using equation 1 (Niva et al 2014a) is given in Figure 6.5.


87

) (1)

(a)

(a) (b)

(c)

Figure 6.5. Scatter plots of the predicted DBH and the field-measured values for (a) MS4100, (b)

ADS40, and (c) WV2. The image-derived CA was calculated using manual vectorization; n = 172.

Previous results (Chapter 2) have shown that either height or CA could be used to infer DBH.

However the variance in DBH could be explained better by CA than height, hence CA is likely to

provide better prediction than height. Assessment of the results in the present study showed that CA is

an easily measured variable from high resolution remote sensing data and could be used for DBH

estimation. The CA measured by these three sensors explained 40 - 45 % variance in DBH with MPE

in the range of (error of 19-20% in all the sensors) which was slightly higher than the previous study

where the MPE was ± 13 cm (error 17 %). However, this result was based on field based

measurements only.

(a)


88

6.5 Conclusions

The allometric relationship between DBH and crown projected area (Chapter 2) made ways to CA

estimations using high resolution remote sensing. This study discussed methods where CA can be

extracted from high resolution remote sensing datasets of the order of sub metre resolution. An

experimental analysis on sensor comparisons was performed and presented. The results of CA

estimates from the two methods (manual and automatic) were close. The mean CA estimates from

WV2 orthoimage was higher than the measurements from other two sensors and also than the field

based measurements (211 m2 versus 247 m

2 by manual vectorization method and 211 m

2 versus 221

m2 by the segmentation method). Manually estimated crown projected area from both the ADS40 and

MS1400 images were closer to the field based measurements (211 vs 195, in MS4100 and 211 vs 198

in ADS40) but underestimated, whereas automated method resulted in over and underestimations in

ADS40 and MS4100 image, respectively. The higher mean value of crown projected area in

WorldView2 image can be attributed to the orthorectified dataset (topographically corrected) where

the shadow effect was minimized. However, this minimization of the shadow effect lead to slight

overestimation of the crown projected area because the crown appeared flat and slightly bigger.

ADS40 image helped in better demarcation of the crown projected area but led to the underestimation

due to no distinct line of separation between the crown shadow and actual tree crown. MS4100 image

resulted in lowest performance, though the spatial resolution was highest due to similar reasons. The

most important point that comes out is that any sub metre resolution satellite imagery can be

effectively used in CA estimations. Our outlined hypothesis that crown projected area estimations

would be higher for high resolution images at least in this case, failed and the CA estimates, in fact,

was found independent of the sensor resolutions. Therefore, the outcome supports that the expensive

high resolution airborne datasets like MS4100 and ADS40 imagery which are usually expensive and

lack temporal resolutions can be replaced by spaceborne sensor like WorldView2 for similar studies

which would achieve similar accuracies. For study like this, the use of WV2 would be cost effective

with less data acquisition time and available over a range of dates and time. The CA estimates from

remote sensing also depends largely on the sun angle and in turn the shadow associated with the trees.

This study can be further supported by using sophisticated and advanced remote sensing technology

like LiDAR.

The intention of delineating CA from remotely sensed image was estimation of DBH from these

measurements. The model for DBH prediction from the three sensors predicted DBH with MPE in the

range of ±16 cm (error of 19-20% in all the sensors), as against the MPE of ±13cm (17 %) in case of

field based measurements. This led us to conclude that identification of trees and extracting crown

projected area for estimating DBH is a most promising technique. Field and image- derived crown

projected area and DBH showed good correlations and in turn can be effectively used for estimating


89

other associated parameters like biomass and tree volume. The crown projected area estimates from

remote sensing image was accurate (both by the method of vectorization and image segmentation)

than the DBH estimates as evident from the R2 values. This can be attributed to the fact that crown

projected area, unlike DBH, is directly viewable by the sensor (Greenberg et al., 2005). Nevertheless

it can be said that with the advances in availability of very high resolution remote sensing images and

image analysis techniques, crown projected area can now be estimated with higher accuracy and the

use of the model will help predict DBH using high resolution remote sensing datasets. Hence high

resolution remote sensing can lead us to better understanding and prediction of forest characteristics

and improvements in the forest ecosystem.

Chapter 7 Canopy Volume

90

Chapter 7

Tree canopy measurements to infer canopy volume:

A comparison of high resolution remotely sensed

images and LiDAR

7.1 Introduction

The measurement of 3D volume is an important parameter in assessing the economic value of a tree

(Gertner, 1991). A number of researchers have tried estimating biomass using biomass-volume

relationships (example Fang et al., 1998). Tree volume measurements could include stem volume

(volume of trunk from ground to tip), canopy volume or total tree volume (the sum of the former, i.e.,

volume of the trunk and the branches). Several allometric equations were developed that relate stem

volume as well as the biomass of several tree components to diameter at breast height and/or to tree

height (e.g., TerMikaelian and Korzukhin 1997; Eamus et al. 2000; Keith et al. 2000; Jenkins et al.

2004). This study determines canopy volume of trees in the study area.

Canopy volume includes the entire living canopy of a tree from the base of the crown to the upper

edge of the crown and from the outer edge of the branch tips inward. It does not include dead

branches, above or below the living portion of the canopy, and is an important parameter in the study

of associated ‘yield’ estimations in horticulture (Tumbo et al., 2002). The conventional way of

estimating the canopy volume of a tree canopy is by manual measurements of crown diameter and

canopy height, and applies a number of assumptions appropriate to the 3D shapes of the crown.

Because of varying crown shape, reach, extent and integral positioning of branches, it is difficult to

calculate the tree derivatives and hence most published models have consolidated all the variations in

tree crowns by using calculations for solid geometric objects (Coder, 2000). Canopy volume is

generally calculated using a predefined solid geometric volume formula given in Eq. 7.1 (Coder,

2000).

Canopy Volume = Crown Height Crown Diameter2 Multiplier (7.1)

The choice of multiplier varies with respect to crown profile as different trees have different general

crown profiles and crown shapes and hence will have different volumes (Frank, 2010). A number of

researchers have calculated either stem volume or total tree volume assuming a certain geometric


91

shape. For example Cutini et al. (2013) calculated stem volume by applying Huber’s formula

assuming a cylindrical geometric shape and measuring the diameter at 0.5 m log length. The area of

the top log was estimated assuming a conical geometric shape. Albrigo et al. (1975) computed the

canopy volume of Valencia plots based on spheroid volume formula. Wheaton et al. (1995) studied

Hamlin and Valencia orange cultivars by measuring trunk diameter and tree canopy, and canopy

volume was calculated based on one half of an ellipsoid.

The relationships between canopy volumes with several tree components such as diameter at breast

height and/or to tree height and crown profiles and crown shapes are used for canopy volume

estimation. Figure 7.1 represents the work flow for the calculation of canopy volume based on tree

characteristics measurements and a schematic of the measurable parameters is give in Figure 7.2.

Figure 7.1. Flow diagram for canopy volume estimations

A schematic of the measurable parameters is given in Figure 7.2.

Visual

Assessment

Tree Characteristics Measurement

required for volume extraction

Crown Diameter/Area

Tree Height

Canopy Height

Tree Trunk Height

Crown Shape

Solid Geometric

Volume Formula

Canopy Volume

(a) (b)

(c)

Combined

Using Range finder

(Using Range finder and

Clinometer)


92

Figure 7.2. Schematic diagram indicating canopy dimensions required to estimate canopy volume

However, the development of canopy volume equation based on field based measurements is

laborious and time consuming process. Lack of standardized approach further complicates the

estimation. Therefore there is a need of much easier and convenient means for tree parameter

measurements with similar accuracy. One technique that has attracted lot of attentions in recent years

is through use of remote sensing data for tree parameters estimation and then for volume estimation.

Optical data, LiDAR (light detection and ranging) and SAR (Synthetic Aperture Radar) are the three

possible avenues for using remote sensing to infer canopy volumes. Optical data from both airborne

and spaceborne platforms have been used to determine the relationships between tree height, crown

diameter and crown cover derived from data and forest stand attributes (e.g., Gering and May 1995;

Ozdemir, 2008).

The most direct remote sensing data for tree parameters and its attribute estimation is LiDAR, a

distance (ranging) measuring technology that relies on the principle of ‘time of flight’. Laser pulses

are directed from a source (e.g., mounted on an aircraft) and a portion of the incident beam on the

target is scattered back towards the source. High-speed detectors and electronics calculate the time of

flight between the emission of the pulse and the return of the back-scattered component, and from this

the distance (range) from the source to the target is calculated. LiDAR captures elevation information

from a forest canopy as well as the ground beneath and can be used to assess complex 3D patterns of

canopy and forest stand structure (e.g., Kini and Popescu 2004; Lefsky et al., 2002; Næsset and

Økland, 2002). As LiDAR derived measurements such as tree height, trunk height, and canopy

diameter etc., then can be used to estimate canopy volume based on formulae described in the Eq. 7.1.

Few examples of LiDAR based canopy volume and biomass estimation along with other forest

vegetation characteristics are: percent canopy cover (Nelson et al., 1984; Hyyppa et al., 2008; Lim et

al., 2008; Lefsky et al., 2002), timber volume (Maclean and Krabill, 1986). Small-footprint LiDAR


93

systems are available commercially and research results on their potential for forestry applications are

very promising (Næsset and Bjerknes, 2001; Holmgren et al., 2002; Næsset and Økland, 2002;

Popescu, 2002; McCombs et al., 2003; Popescu and Wynne, 2003). Popescu et al., (2004) used

LiDAR and multispectral data in forest to estimate tree volume and biomass in pine in Virginia USA.

They found good estimations of biomass and tree volume in the case of pine with an RMSE of 29

mg/ha and 47.9 m3/ha respectively. Naesset and Bjerkes (2001) reported that estimation of forest

stand characteristics from airborne laser scanner data focused mostly on old forest stands or forests

where the mean tree height exceeds about 15 m, and then estimated the mean heights of young forest

stands with tree heights < 6 m and the stem numbers from small-footprint airborne laser scanner

measurements such as canopy height and canopy density.

The other laser technology rapidly gaining attention is Terrestrial laser scanning. For forest

applications where information at larger scales is required airborne LiDAR scanning seems

inadequate. Terrestriall LiDAR on the other hand is implemented to obtain detailed information at the

tree or plot scales. However, because of their short measurement range (up to 3 m), this technique is

limited to measurements at organ or potted sapling scales under controlled conditions (Chambelland

et al. 2008).

Although LiDAR derived tree measurements are more accurate and close to field based

measurements, there are some drawbacks associated with LiDAR. With the existing technology

LiDAR do have problems sometimes in seeing the ground and there are places with few or no ground

returns, which make it hard to interpret. Hence while creating a DEM these places gets extrapolated

this may cause the DEMs to be less accurate. In addition the principal challenge facing potential

LiDAR users wishing to derive canopy volume measures for trees is cost and availability (Krogstad

and Schiess, 2004).

Few studies explored the feasibility of 2D optical remote sensing data for tree parameter

measurements and canopy volume estimation. For example, Ozdemir (2008) estimated tree volume

from pan sharpened QuickBird imagery in open Crimean Juniper forests. Greenberg et al. (2005)

presented a novel approach for generating regional scale above ground biomass estimates using

hyperspectral remote sensing imagery. They related the area of shadowed vegetation to tree structural

parameters, DBH and crown area. They measured the crown area assuming the crowns to be

symmetric and found shadow method to be promising technique for estimating DBH and crown area.

Many studies have been carried out to estimate forest biophysical parameters using SAR radiometry

and polarimetry. Gama et al.,(2010) established a relationship between volume and biomass with

interferometric and radiometric SAR (Synthetic Aperture Radar) response from planted Eucalyptus


94

saligna forest stands, using multi-variable regression techniques. X and P band SAR images from the

airborne OrbiSAR-1 sensor. The volume model developed showed that the stand volume was highly

correlated with the interferometric height logarithm (Log10Hint), since Eucalyptus tree volume has a

linear relationship with the vegetation height. This study represents the potential of SAR technology

to help establish Eucalyptus forest inventory for large areas.

Numerous methods have been tested for tree parameters measurements and canopy volume

estimations from remote sensing data with varying success. Chapter 3 has already demonstrated the

ability to delineate tree canopies from remotely sensed imagery, while Chapter 5 described the ability

to infer tree height from imagery and Chapter 6 explained the ability to infer tree crown projected area

from remotely sensed imagery. These chapters explain the potential of image-based measurements of

tree parameters from remote sensing data to infer canopy volume using Eq.7.1, provided the trunk

height is given. Therefore, the question is whether in absence of a remotely derived measure of trunk

height, can we infer canopy volume based on crown projected area and canopy diameter alone? Few

studies have explored this possibility. For example, Ozdemir (2008) estimated tree volume by the

method of regression from pan sharpened QuickBird imagery in open Crimean Juniper forests and

found that volume can be predicted using just the crown projected area with an RMSE of 15.2 %.

However, more studies are required in this area to support this.

This chapter therefore aims to compare the performance of two sensor systems (airborne LiDAR and

spaceborne multispectral systems) and slightly different approaches for inferring tree canopy volume.

The primary objective of this study is to investigate how well canopy volume in our remnant Eucalypt

species can be estimated using LiDAR and satellite imagery, as compared to canopy volume

estimated based on the field-based measurements as a benchmark. Also owing to the complexity

associated with LiDAR data, the study explores the possibility of using multispectral image alone to

estimate canopy volume given that canopy volume is a 3D tree derivative.


7.2.1 Study Area

The area chosen for the collection of laser and ground datasets was the region of the Newholme-Kirby

property described earlier in Chapters 2-5. A subset of the study area, of approximately 200 ha was

used (Figure 7.3 (insert)), limited in size by the LiDAR data acquisition footprint.


95

Figure 7.3. Location map of the study site in north eastern NSW, Australia.

7.2.2 Field measurements of canopy volume

Candidate single Eucalyptus trees were selected at random using the orthorectified imagery described

in Section 2.4, with care was taken that the trees were well distributed across the study area. Field

measurement was conducted in the month of September – December 2012. In order to establish the

relationship between the crown variables with other tree parameters and canopy volume, 64 trees

belonging to the five different Eucalyptus species were sampled. The number of samples for different

species varied depending on the occurrence in the study area. The structural variables of Tree height

(TH), Crown Diameter (CD) and Canopy Height (CH) were manually measured using a laser

rangefinder (MDL LaserAce 300, Measurement Devices Ltd. Scotland, UK) and a measuring tape

following the procedures outlined earlier in Section 2.4.3. The crown height was measured by first

measuring the trunk height, and subtracting it from the total tree height (Figure 7.2). The crown

diameter (CDfield, m) was measured using the protocol described in Chapter 2 (Verma et al., 2014a).

The canopy volume (CVfield, m2) was then calculated using Equation 7.1. Based on a visual

assessment of the tree crowns in the study area, a parabolic profile crown ‘multiplier’ value of 0.3927

was deemed appropriate. The crown projected area (CAfield) was calculated from the crown diameter

values using the equation in Verma et al. (2014a) (Chapter 2, Equation 2). A regression equation was

then developed between canopy volume (CVfield), and crown diameter (CDfield) and crown projected


96

area (CAfield) parameters to allow subsequent determination of canopy volume from the satellite data

which was unable to provide measurements of canopy height (discussed later in Section 7.2.3).

7.2.3 LiDAR data acquisition and post-processing

The Airborne Laser Scanning system used for the project was the Trimble Harrier 68i/G1 system

flown on June 1, 2013. It consists of a Riegl LiDAR scanning instrument, Applanix POS/AV 410

Inertial Motion System and 12 channels and a dual frequency GPS. The full waveform LiDAR data

collected has the following parameters (Table 7.1):

Table 7.1. LiDAR data acquisition parameters

Parameter Value Unit

Scanning Angle 60 degrees

Flight Speed 216 kmhr-1

Flight Height 375 metres

Scan Rate 192 Hz

Pulse Rate 400 kHz

Swath Width 433 metres

Swath Overlap 37 %

Along Track Point Spacing 0.31 m (along track)

Across Track Point Spacing 0.31 m (across track)

Outgoing Pulse Density 10.26 m-2

Cumulative Pulse Density 17.86 m-2

Calculated Spot Footprint 0.19 m

The acquired LiDAR data were provided in LiDAR Exchange Format (LAS), having first been

classified as ground and non-ground points by the data provider using proprietary software

(Terrascan). An intensity image was created from the point clouds. The selected individual trees

measured in the field were then identified in the point cloud data. The tree height and trunk height was

manually measured for each tree using the software FUSION/LDV (Robert J. McGaughey, Pacific

Northwest Research Station, Version 3.10, Build date May 16, 2012) (Figure 7.4). The canopy height

was then calculated from the difference between the tree and trunk heights.


97

Figure 7.4. Tree parameters (total tree height and trunk height) from FUSION/LDV software. The

colours green, yellow and red represents the canopy at different heights. Blue represents the ground

height.

The automatic extraction of crown parameters from LiDAR was a two step process involving 1)

Canopy Height Model generation (CHM), and 2) Segmentation of the CHM ‘image’ into homogenous

objects representing individual tree canopies. CHM generation requires two grid inputs namely a

Digital Terrain Model (DTM) and a Digital Surface Model (DSM). DTM refers to a digital

representation of topographic surface where the height values in the terrain were known (Dash et al.

2004). DSM was a representation of features above the terrain. The canopy height model, also

referred to as a normalized digital surface model (nDSM), with a vertical resolution of 1m was created

by subtracting the DTM from the DSM. The resolution of CHM was important as the tree height

information extraction from canopy height model largely depends on the accuracy of the height

model. The processing was done using ArcGIS version 10. The canopy height model was used to

extract trees using the method of image segmentation. The software eCognition (eCognition

Developer 8, Munich, Germany, GmbH) used for image segmentation and classification offers a wide

range of segmentation algorithms suited for an array of datasets. The derived canopy height model

was a grid which was represented as a single band image, where the objects appeared partitioned into

lighter and darker areas. A ‘contrast split’ segmentation algorithm was employed for tree extraction

which was later was refined based on the tree heights (Figure 7.5).

The extracted tree polygons were exported to ArcGIS 10 and in accordance with the field based

measurements (Section 7.2.2), the six diameters were measured and the average used to specify crown

diameter (CDLiDAR).

Tree

Height

Trunk

Height


98

Canopy heights were estimated by subtracting trunk height from tree height. The canopy volume

(CVLiDAR) was then calculated using Equation 7.1 with the same ‘multiplier’ as used for the field-

based measurements.

(a)

(b)

Figure 7.5. An example of the (a) derived LiDAR canopy height model (CHM) and (b) the

segmentation results.

7.2.4 Delineation of tree attributes from WorldView2 data

A multispectral, PAN sharpened, WorldView2 image (8-bit) of January 1, 2012 was acquired with a

spatial resolution of approximately 50 cm in four spectral bands, Band 1 (NIR 0.7-1 , Band 2 (Red

0.6-0.7 ), Band 3 (Green 0.5-0.6 ) and Band 4(0.4-0.5 . The image reference system was

WGS 84 UTM Zone 56 S projection system.

The crown projected area for each of the trees (CAWV2) was determined following the process

described in Section 6.3; calculated by the manual method of onscreen digitization/segmentation of

the canopies (ArcGIS version 10) and counting of the canopy pixels, assuming a pixel dimension of

50 cm x 50 cm. The crown diameter (CDWV2) of the individual trees was then determined by first

identifying and then measuring the length of the major and minor axes of the individual tree canopy

polygons, and then calculating the average of the two.

In order to estimate canopy volume for the WorldView2 imagery (CVWV2), the regression equation

developed between on-ground measurements of canopy volume (CVfield), and crown diameter (CDfield)

and crown projected area (CAfield) (Section 7.2.2) was applied to the derived values of CDWV2 and

CAWV2.


99

7.2.5 Evaluating the performance of the two techniques

The derived canopy volumes for each remote sensing method (CVLiDAR, CVWV2) were compared to the

field-measured values (CVfield) and a mean prediction error (MPE) given by

MPE= CVpredicted-CVactual

calculated.


7.3.1 Field measurements of tree parameters

Summary statistics of the measured trees are given in Table 7.2.

Table 7.2 Summary statistics for single trees from the field measurements; n = 79 is the number of

trees used in the model development.

Tree characteristics Min Max Mean Std.Dev

crown diameter (CDfield, m) 6.8 30.5 15.2 4.9

crown projected area (CAfield, m) 36.3 731.8 210.4 129.1

tree height (m) 12.7 42.8 21.3 5.3

Canopy height (m) 8.8 30.6 16.1 4.1

canopy volume (CVfield, m3) (Equation 7.1) 217.9 9040.8 1840.2 1533.1

Scatter plots of canopy volume (CVfield) versus crown diameter (CDfield) and crown projected area

(CAfield) are given in Figure 7.6 and Figure 7.7, respectively, along with the best-fit, polynomial

regression curves. The derived regression equations corresponding to these curves are given in Table

7.3.

Figure 7.6. Scatterplot between canopy volume (CVfield), as calculated using Equation 7.1, and

measured crown diameter (CDfield). The solid curve is the best-fit, polynomial regression equation.


100

Figure 7.7. Scatterplot between canopy volume (CVfield), as calculated using Equation 7.1, and

measured crown projected area (CAfield) derived from field measurements. The solid curve is the best-

fit, polynomial regression equation.

Table 7.3. Derived best-fit regression parameters for calculating canopy volume (CVfield) from

Equation 7.1 using field measurements of crown projected area (CAfield) and crown diameter (CDfield).

Multiplier value = 0.3927 (n = 79).

Equation R2

F-stat P

CVfield= 0.008 field+ 6.5673 – 25.199 0.93 993.0 <0.0001

CVfield = 15.11 CDfield2 – 218.58 CD field+1222.6 0.94 415.6 <0.0001

Both Figures 7.6 and 7.7 and the regression statistics on Table 7.3 indicate the canopy volume of the

candidate Eucalyptus trees can be inferred using crown diameter or crown area, without the need for

measuring canopy height and this bodes well for using remotely sensed imaging systems.

The application of the two regression equations in Table 7.3 to estimate the canopy volume is

depicted in the scatter plots of Figure 7.8 (a,b). Here the predicted canopy volumes (CVWV2) are

compared against the field measured values, the latter including the additional crown height parameter

in Equation 7.1.


101

(a)

(b)

Figure 7.8. Scatterplots of canopy volume from WV2 using (a) crown projected area (CA), and (b)

crown diameter (CD) as the predictor variable, and field measurements. The dashed lines are the 1:1

equivalence between measured and predicted values and the solid lines the best-fit regression curves

(power and polynomial, respectively).

Both scatterplots include a 1:1 equivalence line (dashed line) and a ‘best fit’ regression curve. The

accuracy of using the WorldView2 imagery to infer canopy volume via CA and CD is different, as

depicted in Figure 7.8 (a) and (b), respectively. A MPE of 781.03 m3 (error 42 %) was observed when

using CA as the predictor variable, whereas a comparatively lower MPE (MPE 575.5 m3, 31% error)

was observed with crown diameter (CD) as the predictor variable. Both scatter plots include a

regression curve; a power law explains 65% of the variance between the field and WV2-derived

canopy volume when using crown projected area as the sole variable, whereas a second order

polynomial relationship between the field and WV2-derived canopy volume when using the crown

diameter as the variable explains 76% of the variance. Both parameters yield, on average, and

overestimation of the canopy volume compared to the field measurements and this is most likely due

to the visual classification of mixed, boundary, pixels as canopy. As the imagery was orthorectified it

was often difficult to determine whether the shadow fringes observed in the imagery were parts of the

crown, or that cast upon the underlying ground surface. Consequently the shadow fringes were

allocated as part of the tree canopy. In both cases CA and CD would then be overestimated, hence

yielding a higher CV value. It is likely that the CA-derived CVWV2 estimates are higher than the CD-

derived CVWV2 the because of the additive effect of the erroneously included pixels in calculating CA.

The canopy diameter measure is at least an average of the numerous transect measures. Ideally an

objective classification procedure would apportion sub-pixel dimensions according to the level of

mixing, and this is a recommended subject of further work. At higher canopy volumes (> ~4000 m3)

both techniques tend to underestimate the canopy volume, although there is considerable spread in the

predicted versus actual values for these larger canopies.


102

A scatterplot of the LiDAR-dervied canopy volume versus the field measured CV is given in Figure

7.9. The parameter CVLiDAR is calculated using Equation 7.1 and both the derived measurements of

crown height and crown diameter from the LiDAR data. Even though a lower MPE was observed

when crown volume was estimated using the LiDAR measurements (490.8 m3, 26 % error), there does

appear to be a systematic underestimation of the canopy volume, especially for the larger canopy

volumes (> ~ 2000 m3).

Figure 7.9. Scatterplot between canopy volume predicted using the LiDAR-derived values of crown

height and crown diameter (Equation 7.1) and the field measured values. The dashed line is the 1:1

equivalence between measured and predicted values and the solid line is the best-fit regression curve

(polynomial).

The individual parameters, extracted from the LiDAR data, used to calculate the CVLiDAR values are

plotted in Figures 7.10 (a) and (b).


103

(a)

(b)

Figure 7.10. Scatterplot between (a) LiDAR-derived crown diameter and field measurements, and (b)

LiDAR-derived crown height and field measurements. The dashed lines are the 1:1 equivalence

between measured and derived values and the solid lines the best-fit regression curves (liner and

polynomial, respectively).

Many studies have shown to achieve promising results of lidar systems for assessing single tree

heights (Hyyppä and Inkinen, 1999; Andersen et al., 2001; Hyyppä et al., 2001a and 2001b; Persson

et al., 2002; Brandtberg et al., 2003; Holmgren et al., 2003; McCombs et al., 2003; Popescu et al.,

2003; Holmgren and Persson, 2004; Popescu and Wynne, 2004; Yu et al., 2004; Roberts et al., 2005)

and forest plot or stand heights (Næsset, 1997; Magnussen and Boudewyn, 1998; Magnussen et al.,

1999; Means et al., 2000; Næsset and Bjerknes, 2001; Næsset and Økland, 2002). However, studies

from Rönnholm et al., 2004 and Huang et al., 2009 have shown that tree heights are typically under

estimated by small footprint laser scanning system due to varied number of reasons like (a) variability

in the density and coverage of laser pulses, (b) differences in the algorithms used to obtain the canopy

height model, (c) the amount and height of understory vegetation obscuring the ground surface, (d)

differences in algorithms used to calculate the bare ground elevation, (e) the sensitivity of the laser

system and the algorithms used for signal processing, and (f) lastly the tree shape and tree species.

Analysis of results in the present study (Figure 7.10) shows that both parameters (Crown diameter and

canopy height) used in Equation 7.1 tends towards underestimating at higher values. Crown diameter

shows a slight underestimation whereas canopy height shows significant underestimations especially

at height above 20 m resulting in underestimation of canopy volume CVLiDAR versus CVfield. Given the

diameter value is squared in Equation 7.1 , even that little bit of underestimation is exaggerated too.


104

The present study uses full waveform LiDAR with a point density of 10 per m2. Literature review

suggests this point density to be good enough for height and other parameter estimations. Therefore,

the reasons in this case for underestimations can be attributed to the flying height as well as point

spacing which in this case are 0.31 m along and across track which may have had an effect. Other

factors that influence the crown diameter and canopy height measurements may be the number of

pulses actually hitting the tree edge and understory which would fail to provide a clear demarcation

between the trunk height and the start point of the tree canopy.

7.4 Conclusions

Field based measurements provides the best estimate for any forest characteristics measurements but

it is often expensive and labor intensive, hence it was felt that there should be a practical alternative to

the field based estimations. In this line canopy volume for single eucalyptus trees were estimated

using basic crown parameters like crown diameter and canopy area extracted from image based

remote sensing systems like WorldView2. The results from WorldView2 data were found impressive

and beyond doubt it can be positively said that optical remote sensing can also act as a possible

avenue for measuring three dimensional attributes like canopy volume. This remote sensing system

can thus be looked as viable alternative to field and LiDAR based assessments for determining canopy

volume.

It was not surprising that LiDAR performed well and LiDAR measurements agreed very well with the

field based measurements. The LiDAR based measurement method estimated tree heights with an

MPE of 1.44 m (error of 6.5%), and estimated canopy volume with an error of 26% which was

promising and consistent with other LiDAR based studies, for example Maltamo et al., (2004)

estimated timber volume using LiDAR with RMSE being under 30%. McInerney et al., (2010)

estimated canopy height from LiDAR and medium resolution image combined and found the RMSE

to range from 2 – 31% in two separate studies. Though better performance of LiDAR cannot be ruled

out but, but due consideration should be given to the high price and complexity associated with the

acquisition and processing of these datasets which makes them not so feasible. In addition it can be

argued that even though, LiDAR measurements were significantly accurate few samples did differ

from the field based measurements. Reasons, since canopy height was estimated by measuring the

total tree height and trunk height there is a chance of error propagation which explains the percentage

error in canopy volume estimations using LiDAR.

The model based results using WorldView2 appeared quiet encouraging and can be considered a

direct way of estimating canopy volume; however, the predictability of the model could potentially be

enhanced by incorporating terrain characteristics like slope, aspect, rainfall etc that could also have an

impact on the overall volume estimates which can be the topic for future study. The automated tree

crown measurements like crown projected area and crown diameter by segmentation methods offered


105

greater capabilities to LiDAR and multispectral image data analysis. The study can be extended to

future research wherein different resolution remote sensing datasets can be tested against the volume

estimations. The overall outcome of this study strongly supported the hypothesis that high resolution

remote sensing data can be effective in canopy volume estimations without relying on expensive

datasets like LiDAR.

Chapter 8 Stem Density

106

Chapter 8

Remote Sensing based Stem Density measurements

in Tree Clusters for DBH estimation: comparison of

techniques

8.1 Introduction

This study follows on from the results of Chapter 2 which indicated that both tree height and crown

projection area can be used to infer diameter at breast height (DBH) for single trees and tree clusters.

In this earlier work, tree clusters were defined as groups of trees (n = 2-30) growing in proximity.

The average DBH of the clusters, ranging in density from 38 to 536 stems per ha were predicted using

average crown area per stem as the predictor variable (Chapter 2, Figure 2.8, Niva et al., 2014a).

However, in order to be able to apply the relevant allometric equation (for example Table 2.5 or 2.6 in

Chapter 2) to tree clusters in a situation where it is not possible to clearly delineate the number of tree

stems within the cluster (for example as inferred from the shape of the canopy envelope) then

additional information on the number of stems within the canopy envelope is required. Consequently,

this chapter seeks to determine whether LiDAR data can be used to provide this extra information, for

example following Popescu et al. (2003), for the clusters of Eucalypt species occurring in the

farmscape chosen for this thesis.

A number of algorithms have been developed to delineate individual tree crown and tree stems using

LiDAR. Local maximum filtering (Wulder et al., 2000; Popescu et al., 2002) and watershed

segmentation (Wang et al., 2004) are two of the widely used techniques for such studies. The local

maximum filtering method, in case of optical image, assumes tree apex as the highest point of

reflectance of a tree crown, while with LiDAR data, it assumes that among the laser hits for a tree

crown the highest laser elevation value represents the tree apex. Successful identification of the tree

location using the local maxima technique, however, depends on the careful selection of the filter

window size. Inverse watershed segmentation, commonly referred as watershed segmentation, is the

most common method applied to determining locations of individual tree crowns using a Canopy

Height Model (CHM). Tree identification process involves the segmentation of inverted raster canopy

surface into the equivalent of individual hydrologic drainage basins (Andersen, 2009). The inversion

step helps in separating the CHM into distinct tree polygons with raster crown diameter and height

values.


107

Using the above mentioned algorithms tree stems can be automatically detected based on certain pre-

defined criteria. However, the primary requirement is generation of a CHM, also called normalized

digital surface model (nDSM), which aids in estimation of tree stems in a cluster. The CHM is

generated from two image grids, namely a Digital Terrain Model (DTM) and a Digital Surface Model

(DSM). The DTM refers to a digital representation of topographic surface where the height values in

the terrain are known (Dash et al., 2004) and is extracted from the last return signal from the dynamic

time of flight data generated by the LiDAR profile. The DSM is a representation of features above the

terrain, and represents the mean sea level (MSL) elevations of the reflective surfaces of trees,

buildings, and other features elevated above the "Bare Earth". The LiDAR derived CHM plays a very

important role in forest studies especially when attributes of interest is three dimensional in nature like

biomass, volume, tree height etc. Numerous researchers have successfully used LiDAR derived

CHM's for estimating tree attributes with significantly high accuracy. For example, Jung et al. (2011)

estimated crown variables like crown base height, tree height, crown area and crown geometric

volume using airborne and terrestrial laser scanners (ALS and TLS) and concluded TLS to be

performing better than ALS. In another study, Hunter et al. (2013) estimated biomass using tree

heights measured from LiDAR derived CHM across Brazilian Amazon and concluded LiDAR to be

performing well.

Although the above mentioned techniques (local maxima filtering and watershed segmentation) have

been used successfully for estimation of tree attributes with varying accuracies, the selection of one

over another is always important for a desired outcome in given study. In such conditions, it is often

the case when the relative performances of two techniques are compared and the choice is made with

one with superior outcomes. The current study compares the performances of these methods in

determining stem numbers using LiDAR data. A third method using LiDAR point clouds has also

been tested to explore the possibility of alternative to above two techniques. Three different, freely

available software packages, corresponding to one of the three algorithms are tested. These are:

TreeVaW (based on local maximum filtering), SAGA GIS (based on watershed segmentation) and

Fusion/LDV (based on the LiDAR point clouds).

A number of studies have been carried out using these algorithms in tree attributes determination. For

example, Cao et al. (2012) extracted forest structural parameters based on LiDAR data using local

maximum filtering technique and found only little deviation between the extracted and measured tree

locations. Ke and Quackenbush (2008) used three different algorithms for tree crown detection

namely marker controlled watershed segmentation, region growing and valley following approach.

They concluded that different algorithms could be employed in different applications. For example

marker-controlled watershed segmentation which was based on the assumption that tree tops have

highest reflectance and are located at or near the centre of the crown, could be used in delineating


108

trees with circular shape. Edson and Wing., (2011) studied individual stem location and biomass

measurements using three methods namely FUSION, watershed segmentation (ARCGIS) and

TreeVaW. They concluded all the methods to be performing well, however watershed segmentation

was able to detect smaller trees better than TreeVaW and FUSION.

This study investigates the three algorithms described earlier, namely TreeVaw, SAGA GIS and

FUSION for LiDAR data analysis and compares the stem numbers detected by each method with the

field based measurements. IN doing so, this study completes the hypothesis originally posed in

Chapter 2 (Verma et al., 2014a) that remote sensing can exclusively be used for DBH estimation in

tree clusters.


8.2.1 Study Area

The study area was the same used in Chapter 7 (Figure 7.3).

8.2.2 Tree measurements

The trees used in this analysis were the same clusters of eucalypt trees described earlier in Chapter 2

(Verma et al., 2014a). However not all of the 52 tree clusters were used owing to the limited data

capture area of the LiDAR system. A total of 7 tree clusters were utilized for this particular analysis,

ranging from 3 to 15 stems and densities ranging from 15 stems per ha to 52 stems per ha.

8.2.3 LiDAR Data

The same LiDAR point data described earlier in Chapter 7 (Section7.2.2) was also utilised for this

study. The LiDAR data encompassed an area of approximately 200 ha, consisted of mixed Eucalyptus

species with dimensions summarized in Table 8.1.

Table 8.1. Physical characteristics of clustered trees; n is the number of tree clusters

Tree cluster species n Min Max Mean SD

Number of stems 7 3 15 5.71 4.34

Stems per ha 7 15 52 37 13


109

Using the raw point dataset, the following interpolations were created: DTM from the first return

signal, digital surface model (DSM) from the latest return signal and, by the subtraction of the DTM

from the DSM, the CHM, as discussed earlier in Section 7.2.2.

8.2 LiDAR data processing for tree stem extraction

A canopy height model, CHM, or nDSM as it is commonly referred, with a vertical resolution of 1m

was created by subtracting the DTM from the DSM. The CHM was generated using ArcGIS version

10 (Environmental Systems Research Institute, Inc., Redlands, CA).

TreeVaw (1.1) implements the CHM processing software in Interface definition Language (IDL) to

identify trees based on the local maximum filtering technique that uses a search window of variable

size (Kini and Popescu, 2004). The program was designed for conifer forest applications and is based

on the relationship between crown diameter and height. The software program delineates trees by

deriving an appropriate circular size search window to find tree tops from the height model based on

the above mentioned relationship. For TreeVaW the CHM in TIFF file format was used (since

TreeVaW uses ENVI image format). Defining parameters such as minimum crown diameter,

maximum crown diameter, and minimum tree height were constrained to values based on the field

measurements. For example any CHM values less than 7 m were discarded as ‘trees clusters’ as they

were understorey vegetation. The equation pertaining to the crown diameter and tree height

relationship, Crown Diameter = 5.9133 + 0.4489 * Tree Height (from previous work, Niva et al.,

2014a) was used since TreeVaW also allow users to specify the crown height relations as per the area

and species under consideration. The minimum crown diameter was set to 3.5 m and the maximum to

30 m. The output image consists of the X and Y locations of each of the detected trees within a cluster

along with the heights and radii of individual crowns (Popescu and Wynne, 2004). The X and Y

locations of the detected trees were converted into point locations and overlaid on the image using

ArcGIS 10.

SAGA GIS (version 1.1.1, Department of Physical Geography, Göttingen) focuses on Digital

Elevation Models (DEM) and Terrain Analysis to extract information on the tree stem numbers within

clusters. The grid analysis algorithm in SAGA GIS, such as Gaussian Filter, and Watershed

Segmentation helps extract information on tree numbers. The algorithm is one of the most common

methods applied recently for tree crown identification which separates the CHM into distinct tree

polygons with crown diameter and height values. The result in form of an output shape file gives

optimum stem numbers in a cluster. Since the output is in shape file format it is easier to perform

overlay operations directly in ArcGIS 10. The first step in counting trees using SAGA also involves

creation of CHM by subtracting DTM and DSM followed by smoothening of the height model. The

segmentation is performed on the smoothed height model, which simultaneously creates a point layer


110

based in the number of trees detected. This later can be refined by selecting the trees above a certain

height (above 7m in the present study).

FUSION/LDV (version 3.10, Pacific Northwest Research Station) has two interfaces: FUSION and

'LDV'; the latter being a LiDAR data viewer and uses LiDAR point clouds to determine stem

numbers. Among the three algorithms, TreeVaW and SAGA GIS methods automatically delineate the

crowns while FUSION/LDV requires user to manually locate the stems from the LIDAR point cloud

data. The basis of FUSION is the LiDAR point clouds. The point clouds classified as ground and non

ground points by the data provider were taken into FUSION directly for further analysis. Trees were

displayed based on the unique identification numbers which were assigned to each tree during field

sampling. Each tree was selected manually and tree parameters were measured in LDV using a

measurement marker. Tree stems were measured by counting the number of elevated crown in each

cluster. There are a number of parameters available which can be set before the tree measurements,

like the return numbers, the tree locations etc. The field measured stem numbers and those determined

by the three image-based methods were compared.


The final outputs generated by the three algorithms are given in Figures 8.1 through 8.3. A point

coverage generated from TreeVaW representing the tree location is shown in Figure 8.1. The

segmentation results from SAGA GIS and stems numbers extracted using these segments are shown

in Figure 8.2 whereas Figure 8.3 shows the trees in the cluster as rendered in FUSION/LDV.

(a) (b)

Figure 8.1. Tree crowns in a cluster as detected from TreeVaW algorithm. (a) The tabular output and

(b) detected trees overlaid on the canopy height model

(a)


111

(a) (b)

Figure 8.2 Tree crowns in a cluster as detected from SAGA GIS software. (a) The segmentation

Output and (b) the segments with each point representing a tree stem

(a) (b)

Figure 8.3: Tree crowns in a cluster as rendered in the FUSION/LDV software. (a) the cluster as seen

from above and (b) cross section view of the tree cluster. Cooler colours the trunk and lower part of

the crown, while the warmer colours represent the higher end of the crown.

The number of stems in each cluster derived from the three LiDAR based measurements and the

actual field measured values are summarised in Table 8.2 and graphed in Figure 8.4. The algorithm

with the highest MPE was SAGA GIS (MPE = 4) which consistently underestimated the number of

stems in the cluster. In contrast, FUSION exhibited a lower MPE (MPE = 3). It is likely the improved

performance resulted from the manual estimation method as it is a one to one method of measurement

and the chances of error are very less. The algorithm TreeVaW showed the lowest MPE (MPE = 2).

Here the delineation process matched very well with all the candidate clusters except for one location

where the number of stems in the cluster was quiet high (15). Nearly 80% (Fig 8.5) of the TreeVaW

(a) (b)

(a)


112

results matched the field based results. Although TreeVaW results matched very well with field based

results, it was unable to detect smaller trees. This might be the reason for TreeVaW failing in cluster

number C7 where most of the trees within the cluster were below 12 m high. SAGA GIS also showed

inability in detecting smaller trees as evident from Figure 8.2 (b) where no points were detected over

smaller polygons. FUSION method showed a match of nearly 60 % with the field method, whereas

SAGA GIS showed a poor performance with only approx 15 % match (Figure 8.5). However,

overestimation was observed in C4, where a tree stem was detected in place where actually there is no

tree.

Table 8.2. Tree stem number detected by the three algorithms along with the field based

measurements.

Cluster ID

Field measured

number of stems FUSION SAGA GIS TreeVaW

C1 7 5 3 7

C2 4 3 2 3

C3 3 2 2 3

C4 3 2 1 4

C5 5 2 1 4

C6 3 3 2 3

C7 15 3 2 5

Figure 8.4. Graphical representation of stem numbers as determined by the three different algorithms

0

1

2

3

4

5

6

7

8

C1 C2 C3 C4 C5 C6

Nu

mb

er o

f st

ems

Cluster ID

Field TreeVaw

Saga GIS FUSION


113

Figure 8.5. Scatter plots of calculated versus actual field measured stem number in the 7 tree clusters.

The solid line represents the 1:1 line.

The stem numbers detected in each cluster was later used to estimate the crown area per stem for

estimation of average DBH in each cluster. The crown area per stem estimated by the three algorithms

is summarized in Table 8.3. TreeVaW estimated average crown area per stem with a MPE of only

12.5 m2, whereas with FUSION and SAGA GIS resulted in MPE of 42 and 105.5 m

2, respectively.

Table 8.3. The effect on corresponding crown area as per the number of stems detected by three

methods.

Field Based LiDAR Based

Cluster ID

CA/Stems

(Field)

CA/Stems

(TreeVaW)

CA/Stems

(SAGA)

CA/Stems

(Fusion)

C1 95.84 95.84 223.63 134.18

C2 47.75 63.67 95.50 63.67

C3 117.60 117.60 176.40 176.40

C4 92.50 69.38 277.50 138.75

C5 40.02 50.03 200.10 100.05

C6 80.13 80.13 120.20 80.13

C7 18.66 55.98 139.95 93.30


114

8.4 Conclusions

The study demonstrated the effectiveness of LiDAR datasets in determining the number of stems in a

tree cluster for DBH estimation. The results from three methods were reported namely TreeVaw,

SAGA GIS and FUSION in LiDAR data analysis which was satisfactory and supported our

hypothesis that remote sensing can exclusively be used for determining tree stems in a cluster.

However, there was a difference in estimates between the three algorithms. The discrepancies in

number of stems determined from three methods can be due to the complexity in measuring individual

trees with LiDAR data and also nature of different algorithms used. The most important factor which

may have influenced the tree extraction is the LiDAR pulse striking and reflecting off the tree, and its

ability to decipher between the stems. The accuracy is also influenced by other factors like natural

terrain conditions including tree canopy, understorey vegetation, small scale topography, and other

environmental conditions. However the results were consistent with the findings of other researchers

like Edson and Wing (2011).

FUSION is the only process where the trees were detected based on visualization only, and manually

identifying a tree which is represented by a three dimensional array of dots (the LiDAR point cloud)

(Edson and Wing, 2011), is a difficult task. Differentiating small trees in a cluster was also very

difficult and this can be attributed mainly to the LiDAR resolution or the pulse sticking the tree trunk.

However, it is a matter of research to determine the optimum pulse rate which would enable the

identification of every single resolvable tree in a cluster. TreeVaW and SAGA GIS both rely on a

CHM for tree identification and measurement. Hence the accuracy largely depended on the CHM

generation. CHM on the other hand is a derivative of DTM and DSM. Hence there is a likelihood of

error propagation in the process of CHM generation. SAGA GIS which was based on the concept of

watershed segmentation resulted in a poor performance, which can be attributed again to the

resolution of the CHM used. TreeVaW algorithm resulted in the best overall performance, and was

able to correctly detect trees in the cluster except one cluster where the tree numbers were high (15).

However, it appeared to have difficulty finding small trees (<12m). SAGA GIS also was unable to

detect smaller trees. Nearly 80% of the trees in the cluster were correctly detected by TreeVaW

approach and can be thought to be above other tested algorithms. The high performance of TreeVaW

is also because it enables the user to manually incorporate the crown diameter, tree height relationship

based on the area and species under consideration.

The prime objective of stem detection was ultimately to apply the results to DBH estimation in tree

clusters, which uses canopy area per stem as the predictor variable. An investigation of the effect of

these results on relative crown area per stem estimates was carried out, and the results were in

accordance with the assumptions that stem numbers would affect the average crown area and in turn

DBH estimates. SAGA GIS and FUSION resulted in higher values of average crown area in all the


115

clusters, which would lead to overestimation of average DBH if these algorithms were to be used. The

MPE was highest in case of SAGA (greater than 106 m2) followed by FUSION algorithm (40 m

2).

However, to use the model optimally the number of stem numbers should be possibly very correct,

which TreeVaW is able to do with MPE of only 12.5 m2. Therefore, the study illustrates the

effectiveness of LiDAR and TreeVaw algorithm for successfully identifying tree stems in a cluster.

We therefore conclude that for DBH estimates in single standing trees optical 2D data is well suited

without spending on expensive datasets like LiDAR, however if the aim is to estimate average DBH

in a tree clusters the role of LiDAR cannot be ruled out, since LiDAR is considered to be the best

source for tree characteristics measurements which are three dimensional in nature.

Chapter 9 Tree Species

116

Chapter 9

Integration of LiDAR and ADS40 imagery for

mapping tree species in Australian country

“farmscape”

9.1 Introduction

The previous chapters have demonstrated the ability of a number of remote sensing techniques to

estimate DBH from tree crown characteristics and even tree height. The ultimate motivation for this

work lies in the fact that DBH is the starting point from which estimates of biomass/carbon stocks can

be completed (Section 2.1). Identification of tree species is very important in natural resources

management and biodiversity studies and, of course, biomass/carbon studies based on allometry rely

heavily on species composition as these developed equations are often for specific vegetation types

and tree species. The work thus far has indicated that, at least for the 5 eucalyptus species

investigated, there is no species dependence on inferring DBH from crown/height parameters of

single trees or tree clusters. However this may not be the case for converting DBH to other

biomass/carbon related measures, nor can it be assumed that species is not an important covariate

when examining the DBH-tree canopy/height relationships for other species.

Similar to other tree characteristics, the conventional means of identifying tree species for forest

inventory is through field based methods, which is labor intensive and costly. Scientists have

researched in the past to find an alternative means for tree identification, and remote sensing data has

been found to provide a valuable source of information on the spatial extent, composition, and

structure of species (Ke et al., 2010). Numerous studies have been conducted on species mapping

using a variety of remote sensing datasets ranging from multispectral to hyperspectral images (e.g.,

Vieira et al., 2003; Clark et al. 2005; Goodenough et al., 2003), using multi-temporal data and various

spatial resolutions (e.g., Brown de Colstoun et al., 2003; Gerylo et al., 1998), and via multi-sensor

image data fusion (e.g., Goodenough et al., 2005).

Ongoing developments in hyperspectral imaging and very high spatial resolution (VHR) image

acquisition capability of systems have shifted forest based remote sensing studies from regional-scale

(e.g., Wulder et al., 2004) to more detailed forest species mapping at much finer spatial scale. The

latter work is relevant in the context of delineating individual tree objects within ‘farmscapes’


117

(‘farmscapes’ having previously being defined in Chapters 3, Verma et al., 2014a; 2014b; Leckie et

al., 2003). However, extracting information from high spatial resolution imagery is challenging

because of variation in the spectral response of pixels within a target of interest (e.g. trees) (Chapter 3,

Verma et al., 2014b; Gougeon and Leckie, 2006; Hirschmugl et al., 2007; Pouliot et al., 2002). Light

Detection and Ranging (LiDAR) data (described earlier in Section 7.1) has opened a new avenue for

forest species classification in terms of providing 3D information on species-specific vertical crown

structure (Gerylo et al., 1998), limb/branch distribution (Dassot et al., 2010b) and even trunk

dimensions (Bacher and Mayer., 2000). LiDAR has been used in previous work for individual tree

species discrimination (e.g., Brandtberg et al., 2003; Holmgren and Persson, 2004; Liang et al., 2007;

Ke et al., 2010). In the process of tree classification, the individual tree crowns are first delineated and

then LiDAR-derived metrics for crown structure and shape is used to classify individual trees.

For example, Liang et al. (2007) and Reitberger et al. (2006) distinguished coniferous and deciduous

trees with LiDAR data acquired under leaf-off conditions. Here they assumed that, in the leaf-laden

coniferous trees, the first and last return pulse signals were reflected essentially only by tree tops

while, in deciduous trees, the first return pulse would originate from the tree tops whereas the last

return signal would originate from the ground. In this simplistic approach they obtained an overall

89% accuracy in classifying coniferous and deciduous types. In a second study, they used leaf-on data

for both species and obtained 80% accuracy for the same classification.

Hollaus et al. (2009) achieved 83% accuracy in discriminating between spruce, larch and beech trees.

Their approach used ‘geometric information’ such as echo width and backscatter cross section as

extracted from the full-wave form ALS data to identify the candidate tree species.

Ørka et al. (2007) used two intensity metrics from the return pulses, namely intensity and standard

deviation of intensity, to discriminate between birch, European aspen and Norway spruce. They

achieved 68% to 74% accuracy in classifying tree species depending on the number of considered

variables.

Trees constitute distinct targets; however single tree detection requires high sampling densities.

Moreover, individual tree crowns are not always detectable in LiDAR datasets (Persson et al., 2002;

Korpela, 2004). Although optical remote sensing (namely satellite and airborne imagery as discussed

in previous chapters) and LiDAR data have been used on their own for tree species delineation within

forests classification, several studies have sought to combine high resolution multispectral imagery

and LiDAR data to produce a ‘more effective’ species classification within tree communities (e.g.,

Leckie et al., 2003; Hill and Thomson, 2005; Heinzel et al. (2008); van Ewijk et al., 2014).

Combining the two data types effectively merges the spectral information from the optical imagery,

including the objects as derived from OBC discussed earlier in Chapter 3, and the vertical ‘structure’

attributes of each individual tree from the LiDAR data. Holmgren et al. (2008) integrated LiDAR and


118

high spatial resolution aerial imagery and obtained 8% improvements in individual-tree-based

classification from the combined data. Lidar when used alone resulted in overall accuracy of 88%

while Multispectral image for autumn and summer seasons achieved an accuracy of 91% and 84 %

respectively. The combined use of the datasets increased the accuracy further by ~5-8%. By

combining LiDAR and multispectral autumn images the classification accuracy could be improved to

96%, while combining LiDAR with multispectral summer images could increase the accuracy to only

93%. Their study tested the classification accuracies of two coniferous species; Norway spruce and

Scots pine and also deciduous trees. Object based segmentation was first performed which was

followed by grouping of points clouds within each segment belonging to each tree. To separate the

two species variables like height, canopy shape, proportion of pulse types and intensity of LiDAR

returns were derived from the point clouds to capture variations in the crown structure. The relative

crown based height helped in separating Scots pine trees with other tree species group.

Heinzel et al. (2008) investigated the use of laser scanning data and CIR (Colour Infra-Red) aerial

photographs which were captured simultaneously, for classification of oak, beech and coniferous tree

types in Poland. The CIR images were a combination of near infrared, red and green wavebands

which were first separated and further transformed into hue, saturation and intensity channels. First

2D single tree delineation was conducted using the algorithm developed by Koch et al (2006). The

input data were DSM and DTM which allowed grouping trees in different regions based on the height

values. The LiDAR derived polygons were then fitted on the spectral information and the species

were classified and crowns were refined. The overall accuracy of classification was 83%.

Machala and Zejdova (2014) mapped the forest species in the region of South Moravia in the Czech

Republic using a combination of multispectral image and LiDAR data. Three data sets were used for

classification (multispectral image, DEM and DSM). First the vegetation and non vegetation pixels

were distinguished using the NDVI layer derived from the multispectral image. Both these broad

classes were further classified; the vegetation into forest and non-forest and the non vegetation class

into water, clear cut ground and built-up area. A detailed vegetation classification was further

performed using a NN classification algorithm followed by a classification based solely on the DEM.

The DTM and DSM layers extracted from the LiDAR data were then used in determining the heights

of the forest stands, with an overall classification accuracy of more than 80%.

Arroyo et al. (2010) integrated LiDAR and QuickBird imagery for mapping riparian biophysical

parameters and land cover types in Australian tropical savannas and obtained an overall accuracy of

85.6%. They first created four different data layers from LiDAR dataset namely DTM (Digital terrain

Model, TCM (Tree Canopy Model), PPC (Plant projective cover) and a Streambed map. These layers

along with the four-band multispectral image were segmented and classified. Different land cover

types were then classified based on four features, namely the mean and standard deviation of both


119

muti spectral and LiDAR derived information, contextual information such as the relative border to

objects classified in a particular class and the NDVI.

Ke et al. (2010) evaluated the ‘synergistic use’ of high spatial resolution multispectral imagery and

low-posting-density LiDAR data for forest species classification in central New York State. They

examined three different segmentation and classification schemes namely segmentation based solely

on spectral image layers, segmentation based solely on LiDAR derived layers and segmentation based

on both spectral and LiDAR derived layers. Object based Image segmentation was performed.

Decision tree classification was then used; advantages of such as non parametric in nature and rapid

processing, (as discussed earlier in Section 3.3.1) having being identified by the authors. They showed

the integration of spectral and LiDAR data improved the species classification compared to using

either data source independently. The study revealed that each data source had contributions in

species classification. High spatial resolution multispectral imagery helped in defining forest

boundaries and provided spectral separation between forest species. LiDAR derived topographic and

height information helped in reducing within class spectral variation, enhanced the between class

variation due to different height properties among the species and also enhanced the contrast between

coniferous and deciduous stands.

As described earlier in Sections 3.1 and 4.1, image classification is the simplest way of extracting

information from a remotely sensed imagery, but there are ranges of classification algorithms

available which vary from data type, environment and applications. A number of classification

algorithms have been developed for both LiDAR and multispectral image datasets, to be used for

either type of dataset, or as the previously cited work has shown, to combination datasets. The

conventional pixel-based classification method of image classification and information extraction

works well with medium to coarse spatial resolution images, but often found to be not sufficient,

especially when applied on a very high resolution imagery (e.g., Towonshend et al., 2000; Kim and

Madden, 2006; Myint et al., 2011) and LiDAR (Ke et al., 2010). When the pixel size of any data layer

(LiDAR and image data) is significantly smaller than the average size of the object of interest, object-

oriented approach offers an optimal solution for classifying such data (e.g., Kamagata et al., 2006;

Verma et al., 2014; Chapter 9) and been successfully applied to forest species classification (Thomas

et al., 2003; Wulder and Seemann, 2003).

From the previously cited work above, and given that most of the work to date has focused on

delineating vegetation classes, or between broader groups like genus (e.g deciduous from non-

deciduous, or pine from Eucalypt) where there are huge differences rather than species. Tree species

mapping is obviously a complex task using remote sensing data and challenged by the fact that often

not feasible with certain species because of the natural heterogeneity in physical features that often

occurs with tree species (Ruiz et al., 2004). Of the reported species classification work using LiDAR


120

(examples include Holmgren and Persson, 2004; Orka et al., 2009; Brandtberg et al., 2007), the

intensity of the backscattered data is commonly used. Several authors have tried to classify tree

species using positions of laser points within individual tree crowns as well as the intensity data.

(Orka et al., 2009; Holmgren and Persson, 2004). Kim et al. (2009b) reported that by using intensity

and the derived height data (from the LiDAR time of flight) improved the classification of deciduous

and coniferous species. Kim (2010) classified tree species using cluster analysis and two seasonal

LiDAR datasets. The results showed that species with similar tree characteristics seemed to cluster in

a single group while the species with different characteristics were clustered in other groups. They

concluded that the use of two season datasets led to more reasonable clusters than using either one of

the datasets. In the work of Ke et al. (2010) mentioned earlier, the DSM and height data from LiDAR

helped in classifying elevation specific stands like Hemlock. They concluded that the LiDAR derived

topographic features increased the classification accuracy by reducing the within class variation

among the neighbouring objects caused by shadow effects.

So, while optical data has proven useful in forestry applications for differentiating between forest and

non-forest areas (Lehmann et al., 2011), where there is generally a considerable difference in tree

morphology, and for discriminating between major tree species within a forest, such as coniferous

and deciduous (Chastain and Townsend, 2007), they cannot detect features underneath areas of dense

canopy top-cover nor do they provide information on the vertical composition of vegetation-related

attributes. LiDAR data, on the other hand, allow analysts to directly portray forests in a three-

dimensional format over large areas, however, the data have their own shortcomings. LiDAR data

provide multiple return position and intensity measurements, but contain only limited information for

deriving the correspondence to target objects. A review of the rapidly growing literature on LiDAR

applications emphasizes the need for optical data fusion with LiDAR data to improve various feature

extraction tasks (Hill and Thompson, 2005; Leckie et al., 2003).

The objective of this chapter therefore, is to investigate whether the same approach of fusing

multispectral images and LiDAR data allows us to classify the constituent Eucalyptus species that

make up the scattered trees in our ‘farmscape’. In this chapter we will separately use the LiDAR and

high resolution multispectral images as well as a combination of the two.


9.2.1 Study Area

The study area used in this work is the same as that described earlier in Chapter 5 (Section 5.3.1), mix

of forested area (Mount Duval), open woodland and mixed pastures. Much of the farm area is

dominated by 5 Eucalypt species of varying age and stem and canopy density. The major species


121

occurring in the area are Apple Box (AB, Eucalyptus bridgesiana) Stringy Bark (SB, Eucalyptus

caliginosa), Red Gum (RG, Eucalyptus blakelyi), White Gum (WG, Eucalyptus viminalis), and

Yellow Box (YB, Eucalyptus melliodora). A total of 88 trees comprising these 5 species (SB-34,

WG-18, YB-31, AB-15 and RG-5) were used for the training and validation process.

9.2.2 Remote sensing datasets

Multispectral imagery of the study area was acquired at approximately 1045 hrs (AEST) on 3

November 2011 using an ADS40 airborne digital scanner described earlier in Chapter 5 (Section

5.3.2). Flown at an altitude of 1920 m above ground level (AGL), the 24-bit images were acquired

with a spatial resolution of approximately 50 cm in five spectral bands: Band 1 (NIR 0.7-1 μ , Band

2 (Red 0.6-0.7 μ ), and Band 3 (Green 0.5-0.6 μ ), Band 4 (Blue 0.4-0.5 μ ). The image transects

were mosaiced and the complete image geo-referenced using ground control (Section 5.3.2).

The LiDAR data, previously described in Section 7.2.3, was acquired on June 1, 2013 using a Trimble

Harrier 68i/G1 system. The LiDAR data encompassed an area of approximately 200 ha (previously

described in Section 7.2.3), which included the 5 Eucalyptus species. Using the raw point dataset, a

DTM was created from the last return data, a digital surface model (DSM) from the first return data

and, by subtracting the DTM from the DSM, a canopy Height Model (CHM) with a vertical resolution

of 1m was also created.

9.3 Methodology

Object based segmentation and classification can be grouped broadly as a three step process. 1)

Segmentation which is grouping of features called objects. 2) Defining the object based metrics and 3)

Classification based on these defined metrics. This study investigated each of these steps to improve

forest species classification through integration of multispectral ADS40 imagery and 3-dimensional

LiDAR data. eCognition Developer 8.64 software (formerly Definiens) of Trimble Germany GmbH

(München, Germany), which was specifically created as a powerful instrument for object-oriented

image analysis (Benz et al., 2004), was chosen for the purposes of this study.

9.3.1 Image Segmentation and Classification

Image segmentation was carried out on three different image types: (1) the individual bands of the

ADS40 (spectral-based); (2) LiDAR derived layers (LiDAR-based); and (3) both the spectral and

LiDAR derived layers (Spectral/LiDAR based). For the image segmentation process, the spectral and

shape homogeneity criterion based on color/shape ratio, compactness/smoothness ratio for object

shape, and a scale parameter for resultant object size and the input layer weighting were stipulated.

Since the three segmentation processes, namely that applied to the multispectral data only, the LiDAR


122

data only, and spectral/LiDAR-combination, were different in their characteristics, hence the user

specified parameter settings varied between the data types. Here the optimum parameter values were

determined by trial and error, with a visual assessment of the final canopy boundaries used to assess

the veracity of the settings. The image layer weights were tailored for the best differentiation of

tree/cluster areas, with greater weights assigned to the green and NIR bands (e.g., Gitelson et al.,

1996). For LiDAR-based segmentation, the three layers namely height (return differences), CHM and

intensity were used. Higher weights were assigned to CHM and height and the lowest weight to the

intensity image as the characteristics of the intensity image depends on various environmental factors

(e.g., Im et al., 2008). In the case of the spectral/LiDAR based segmentation involving a combination

of the above two datasets, the same weighting was applied to the collective image and LiDAR

datasets, respectively. Table 9.1 summarizes the values assigned to each parameter in each scheme.

Though segmentation largely depends on the scale parameter, as it affects the granularity of the

objects formed Chapter 3 (Section 3.3.1), changing the shape and compactness also led to an increase

in the quality of segmentation results, as observed, for example in Machala and Zejdova (2014). A

range of segmentation with different scale parameters were carried out and tested before reaching to

the optimal scale for each dataset.

A number of object based metrics were calculated based on the spectral, the topographic (ie CHM,

DTM, DSM) and the intensity information. Together with mean and standard deviation of

segmentation layers, higher-order texture measurements such as GLCM (Grey Level Co-occurrence

Matrix) and GLDV (Grey-Level Difference Vector) (Haralick, 1986) were derived for the green and

near infrared bands because of their usefulness in species discrimination (e.g., Gitelson et al., 1996).

Other geometric metrics of objects (e.g., shape etc.) were not computed as they are generally not

found useful in vegetation classification (e.g., Yu et al., 2006). Overall, a total of 26 metrics were

generated, 20 metrics derived from ADS40 multispectral layers, 6 from the LiDAR-derived

topographic layers (Table 9.1).

Spectral features such as mean, brightness and standard deviation of the spectral bands were

calculated using the segmentation bands. Spectral indices like normalized difference vegetation index

(NDVI) and the simple ratio (SR) were calculated using Red and IR bands, as was the area of each

segmented region and its relationship to neighboring objects. Statistically significant features were

then tested with the feature space optimization tool within eCognition. When defining the feature

space, some textural features were also tested (Textures after Haralick - GLCM Mean, Standard

Deviation, Homogeneity and Contrast).


123

Table 9.1 : Metrics defined for rule based classification

Datasets Data Layers Object Metrics

Multispectral ADS40 Blue Mean and standard deviation of each layers

Green Brightness of each layer

Red GLCM mean and standard deviation for green and IR

Infra Red GLDV mean and contrast for green and IR

LiDAR CHM Mean and standard deviation of each layer

Height

Intensity

Rule-based classification was used in this study for classification of segmented objects into five

different Eucalypt tree species, namely Stringy Bark (SB), White Gum (WG), Yellow Box (YB),

Apple Box (AB) and Red Gum (RG). Training point features for each of the species were collected

from each of the data types (ie image only, LiDAR only or integrated image-LiDAR) and used to

classify the respective segmented objects. The approach for classifying individual tree species was to

define rules based upon metrics best describing the training features. Classification of the LiDAR

data was based on seven tree height ranges, namely 5-10 m, 10-15 m, 15-20 m, 20-25 m, 25-30 m, 30-

35 m, 35-40 m and > 40 m, based on the nature of the species observed in the field. Based on the

intensity values and height, trees were classified in five different categories, taking reference from the

field based measurements. The classification of multispectral ADS40 which can be thought of as a

combination of supervised and unsupervised classification approach where the rules were able to

detect the species class which was later given a specific species name with the help from field based

measurements.

9.3.2 Accuracy Assessment

Field sampling was conducted in the month of September through December 2012. Field

measurements of vegetation structural properties along with species information were collected during

this period. Depending upon tree species available in the study area, all 88 samples pertaining to the

five different Eucalypt types (SB-34, WG-18, YB-31, AB-15 and RG-5) were used to assess the

classification accuracy. The field based tree locations were converted into point coverage along with

the tree attributes using ARCGIS ver. 10. These field sampling locations which represented the

sample size for each class was used to evaluate the accuracy with which the trees have been classified.

An error matrix was constructed to estimate the classification accuracy which consisted of producer’s


124

accuracy (PA), user’s accuracy (UA), overall accuracy and Kappa co-efficient. The error matrix was

generated for all the three classification results and was compared. Comparisons between Kappa

coefficients were performed to evaluate the effect of (1) the integration of spectral data and LiDAR

data in image segmentation, (2) the integration of spectral and LiDAR data sources in classification.


An example of coincident LiDAR and multispectral image data (without radiometric scales) for a

small subset of the overall study area is given in Figure 9.1.

Figure 9.1 (a) ADS40 multispectral image of the study area rendered in false colour, and (b) grey-

scale CHM of 1m vertical resolution derived from the LiDAR returns (no radiometric scale for

brevity)

As expected, a visual examination the objects generated from the various segmentation processes

showed the outputs to be highly dependent on the scale parameters (Table 9.2).


125

Table 9.2 : Applied data segmentation parameters

Datasets Data Layers Weight

assigned

Scale Color and

Shape

parameter

Compactness

smoothness

Multispectral ADS40 Blue 1 70 0.2 0.5

Green 5

Red 1

Infra Red 5

LiDAR CHM 5 70 0.7 0.5

Height 5

Intensity 0.5

LiDAR/Multispectral ADS40

combined

CHM 2 270 0.7 0.9

Height 5

Intensity 0.5

Blue 1

Green 5

Red 1

Infra Red 5

A Scale of 70 was found to be effective in delineating tree crowns in the LiDAR and multispectral

data datasets (Figure 9.2(a) and 9(b)) whereas a larger Scale value (270) worked very well with the

combined multispectral/LiDAR data, effectively delineating the tree crowns (Figure 9.2(c)). A close

look at the segmentation results of the LiDAR CHM showed elongated segments (Figure 9.2(b)) but

only on the low value (of CHM) areas. This is expected given the vertical resolution of the derived

CHM, when there is low understory and the ground surface is relatively heterogeneous. However, this

did not affect the delineation of tree crowns given their considerable heights (~ 10 m) above the

ground surface. The segmentation process was more challenged for the multispectral image data;

segmentation at lower scale values appeared fragmented while larger values of Scale resulted in more

generalization with some of the understory cover found included within the crown segment (as

encountered and discussed earlier in Chapter 3, (Section 3.3.1). However, this could have been

overcome by the use of stepwise segmentation, but did not try here. In addition to the value of Scale,


126

the relative weights assigned to the image layer, the color/shape ratios and smoothness/compactness

ratios also had an influence on segmentation.

Figure 9.2 Examples of the result of data segmentation (a) ADS40 multispectral image (b) 1 m

resolution CHM derived from the LiDAR data, and (c) combined LiDAR/multispectral data.

The confusion matrix summarizing the classification performance of the three datasets is given in

Table 9.3. A classification based on the combined use of LiDAR and multispectral imagery yielded

the highest classification accuracy with an overall accuracy of 60.5%. Unlike the multispectral

imagery, the shadows were well segmented from the canopy envelopes in the combined datasets,

owing to the LiDAR derived height information.

The lowest accuracy achieved by the multispectral data layers echoes previously identified challenges

(Section 3.3.1) that spectral metrics such as mean and brightness did help in delineating tree crown

from exposed understory and while the textural properties helped in separating species, these metrics

were found not very effective in discriminating between the crown envelop and understory vegetation.

The NDVI layer did not significantly improve the image-based segmentation.

The species Red Gum and Apple Box were the most difficult to classify using the multispectral

dataset due to their occurrence in lesser numbers in the study area. Stringy bark was found to coexist

with the Apple Box and to a lesser extent the White Gum, while Red Gum the least accurately

classified was mixed in all of the other categories. The highest classification accuracy achieved, based

on spectral and texture metrics, was for the Stringy bark species.


127

Table 9.3 Comparison of species classification accuracy (%)

Multispectral LiDAR Combined

Species Name PA UA PA UA PA UA

Yellow box 39.02 64.42 61.44 66.73 63.93 76.14

Stringy Bark 74.84 40.15 62.19 32.88 74.9 43.54

White Gum 32.25 49.92 42.16 82.91 42.95 93.21

Red Gum 21.28 5.48 89.1 28.49 73.76 42.04

Apple Box 2.75 5.33 39.06 26.62 60.54 50.29

Overall Accuracy 39.06% 48.56% 60.99%

Kappa 0.22 0.43 0.51

PA = producer’s accuracy, UA = user’s accuracy

Clearly the ancillary LiDAR derived layers helped in classifying trees based on different height

ranges. However, when species-wise classification was performed using these height ranges only, a

substantial amount of intermixing was observed especially between String Bark and White Gum. It

was however easier to classify species like Apple Box and Red gum as they behaved differently so far

as tree height and tree structure is concerned. The topographic information helped in increasing the

classification accuracy, but intensity data did not contribute much and therefore the classification rely

only on height information in this case. A preliminary classification results based on seven different

height ranges were found satisfactory with respect to field based measurements. This helped in

identifying species that can be classified on the basis of height ranges. The following observations

were made with respect to height based classification: (a) White gum was classified well because of

its huge crown and tree height characteristics; (b) the same logic can be applied for Apple Box due to

its typical height characteristics, however, a substantial amount of intermixing of Apple Box with

other species was observed throughout the study area and hence height information alone is not

working very well in this case; (c) trees with intermingled crowns were difficult to classify and edge

of the tree crown were misclassified in all case; (d) Yellow Box and Stringy bark showed maximum

mixing probably due to overlapping height ranges. To overcome this, tree crown area were taken into

consideration which helped in refining the accuracy as Yellow Box had larger crown compared to

Stringy bark. The confusion in case of tree edges could be due to LiDAR returns. Figure 9.3 shows an

example of the height based classification results.


128

Figure 9.3 Classification output based on height ranges of LiDAR

Figure 9.4 The species classifications generated using (a) ADS40 multispectral image, (b) 1 m

resolution CHM derived from the LiDAR data, and (c) the combined LiDAR/multispectral datasets.

The combined use of LiDAR and multispectral data allowed accurate identification of tree species

with the overall classification accuracy of 61 % (kappa 0.51) at scale 270. The inclusion of

‘topographic’ attributes along with (image) spectral information in the classification increasing the

overall classification accuracy by ~ 20 %. Similar findings were reported by others (e.g., Ke at al.,

2010). Fig 9.4 shows the classified outputs for all the three datasets. Apple Box compared to all other


129

tree species performed better in LiDAR and good with the combined use of the two. Stringy bark was

classified well by the single and combined use of the datasets firstly due to abundance and secondly

due to a distinct crown shape.

9.5 Conclusions

The study evaluated the accuracy of classifying 5 different species of Eucalypt trees scattered

(individually and in small clusters) across a farmscape using LiDAR and multispectral imagery

(ADS40), both individually and in combination. The spectral based classification approach using

multispectral data; a mix of supervised and unsupervised classification system, proved unable to

differentiate between the species because of the heterogeneity associated with the image. Of the 4

wavebands available in the image, the green and infrared bands helped the most. The relatively poor

classification accuracy (39.06 %) of the segmentation suggests that a multi-level hierarchical

classification approach would more likely increase the accuracy and overcome the heterogeneity (Ke

at al., 2010).

The height based discrimination available to the LiDAR data, while only slightly improving the

classification accuracy compared to the image-based dataset, illustrated the limitations of relying only

on height groupings as a means of discriminating tree species, especially when single tree (and even

small clusters) is not growth limited due to competition. Other metrics like tree canopy shape, spectral

information within the tree crown, the intensity of laser returns and type of laser returns could also be

taken into account for tree species discrimination, similar to the study carried by Holmgren et al,

2008. However, this is a matter of further research. The overall accuracy (48.56 %) achieved was

higher by around 10 % than the multispectral alone. The classification accuracy largely depends on

the accuracy of the segmentation achieved. The contrast associated with the LiDAR dataset helped in

achieving better segmentation results than multispectral.

The integration of LiDAR and multispectral data resulted in more accurate species classification than

using either of the dataset independently. The inclusion of height information helped overcome at

places where the spectral and textural attributes failed resulting in higher overall accuracy (60.5 %).

Chapter 10 Conclusions

130

Chapter 10

Conclusions

10.1 Summary

Image-based remote sensing systems have evolved rapidly with metre resolution satellite systems

such as WorldView2, and sub-metre airborne systems now widely available. The main objective of

this thesis was to investigate the potential of very high resolution, image-based remote sensing data

for estimating a key parameter, namely the diameter at breast height (DBH) of scattered Eucalyptus

trees in typical grazing farmland in south eastern Australia. Whilst more sophisticated airborne

scanning techniques such as LiDAR are gaining prominence for their ability to provide detailed

surface and structural measurements of tree canopies, imaging systems are currently simpler to deploy

and operate, and the data, per unit area of acquisition remain (at least to date) considerably cheaper to

acquire. There are however attributes of trees that only LiDAR is capable of directly measuring, such

as vertical canopy dimensions and porosity. Consequently LiDAR data have also featured in this work

as an adjunct to the image datasets; the integration of which was used in delineating, for example the

individual species of Eucalyptus that featured in this work.

Owing to the importance of Eucalypt species in an Australian landscape, and in particularly in the

New England region of south eastern Australia where this work was conducted, five different species;

Eucalyptus bridgesiana, Eucalyptus caliginosa, Eucalyptus blakelyi, Eucalyptus viminalis, and

Eucalyptus melliodora of genus eucalyptus were studied. Simple regression models were developed

linking the crown projection area and height of both isolated trees and for tree clusters of up to 27

stems, to DBH. Based solely on ground measurements, the model explained 67% and 68%,

respectively, of the variance in stem DBH in two cases. A single model involving both single trees

and the tree clusters to predict average stem DBH had similar explanatory power (R2 = 0.71) and

yielded a mean prediction error in average DBH per stem of ±13 cm. The results also indicated that it

was sufficient to use crown projection area for DBH prediction, however, it was found that the

inclusion of tree height as a parameter in the equations increased, slightly, the overall accuracy of

DBH estimations. The results of this study established the fact that DBH of the scattered Eucalyptus

trees could be estimated using crown projection area, which itself is a measured variable from remote

sensing data. Very high spatial resolution (15 cm) aerial imagery, and high spatial resolution (50 cm)

airborne (Airborne Digital Sensor, ADS40), and satellite (World View 2, WV2) imagery were then

evaluated for their use in identificating and delineating tree canopies, from which crown projection

area and crown diameter could subsequently be extracted. Owing to the complexity of high resolution


131

remote sensing data both pixel based and object based image classification schemes were tested with a

very high spatial resolution CIR image (15 cm). This study served to highlight the advantages and

disadvanteges associated with pixel and object based image classification, and in the latter, the

important role played by setting the appropriate feature parameters in delineating the desired objects

(ie tree crowns). The results verified that the object-based classification of individual tree crowns (or

crown clusters) had an improvement in overall classification accuracy compared to the pixel-based

classification, . In addition to this, an attempt was also made to quantify the variation in tree cover

area estimates taking manual method of vectorization as a reference. For this study WV2 data set with

a spatial resolution of 50 cm was used. This study showed that both the object based and supervised

pixel-based method of classification performed equally well and that the spatial resolution, within the

range of this work, did not overly effect the accuracy of tree area estimations. This research explored

the potential of using remote sensing data of tree crowns as a possible alternative for field based

measurements for estimating the diameter at breast height (DBH). Considering the limitations

associated with airborne images in terms of their availability, temporal resolution and the often

significant image acquisition costs, an attempt was made to replace airborne image data with that

from space borne platforms of similar spatial resolution. The relative performance of three sensors in

terms of crown area extraction was investigated to determinate an appropriate spatial resolution of

image datasets necessary to extract tree crown descriptors in scattered trees and tree clusters in a

typical farmscape. The remote sensing data: ADS40 digital airborne imaging (50 cm), spaceborne

WorldView2 (50 cm) and Color Infrared (CIR) imagery (15 cm) were tested for their ability to infer

crown characteristics. \

Even though crown projection area proved to be the most accurate parameter from which to infer

DBH, the results also indicated that inclusion of a tree height measurement could increase the

predictive performance. A shadow-based method for estimating the height of single eucalyptus trees

from the very high spatial resolution imagery was proposed and tested. The method used the projected

tree shadows on the ground, taking into account ground slope and aspect and solar illumination angles

(elevation and azimuth). The accuracy of the height estimated in this work (MPE/RMSE ±5.6 m)

demonstrated a possible pathway to inferring the height of individual trees from imagery alone.

Possibly the least expected of the results in this work was the fact that the canopy projected area-DBH

relationship for single eucalyptus trees (of 5 species) and the average canopy projected area-average

DBH relationship for tree clusters containing between 2 and 25 stems, were statistically

indistinguishable. Although high resolution remote sensing data was found useful in extracting

projected crown dimensions for tree clusters, from which the average DBH within the cluster can be

derived, calculating the total DBH within the cluster requires knowledge of the number of stems

within the cluster. This study explored the utility of LiDAR data to estimate tree height of single trees


132

and also to ascertain stem numbers within a tree cluster. The LiDAR based measurement method

estimated tree height with a MPE of 1.44 m (6.5 % error). Similarly LiDAR based stem density

measurements were in agreement with the field based measurements. The results led to the conclusion

that the algorithm TreeVaw came up with the best estimate with MPE of 2 trees; the downside being

that TreeVaw failed with smaller trees which were less than 12m high.

The data also helped in quantifying canopy volume of our candidate eucalypts, an important factor in

inferring tree yield, with achieved estimation accuracies very close to the field based measurements.

Both the image and LiDAR datasets were also combined in an attempt to delineate the individual

Eucalyptus species within the area of study. Here the image and LiDAR datasets were used separately

and also in combination to assess their ability in generating species descriptions for this study. The

combined image-LiDAR datasets proved most effective in delineating between the species.

Overall, the study suggests that even though both LiDAR and multispectral imagery could effectively

be used to estimate tree characteristic estimates like tree height, canopy dimensions, canopy volume,

species, LiDAR achieved better estimation accuracies than multispectral datasets. However owing to

the cost associated with the LiDAR datasets, it would not be thought to be fit for larger areas in terms

of cost and time.

Overall, the research demonstrated the potential of using image-based methods for estimating DBH in

our candidate trees and tree clusters with an accuracy that may equal that of the significantly more

expensive and complex LiDAR systems. As with other technologies, the cost of LiDAR systems and

data acquisition will likely decrease in time and such data will become more widely available. The

research demonstrated the potential of using image-based remote sensing data as a plausible

alternative not only for field based measurements of DBH and other tree parameters estimations at

farmscape level studies, but as an alternative to LiDAR-type systems. However, where 3D tree

descriptors are necessary, for example to quantify canopy volume, overall tree biomassand possibly

for tree species identification, LiDAR along with image data are an effective combination. In other

words, remotely sensed imagery for DBH and biomass assessment in scattered trees and tree clusters

in farmscapes will continue to play an important role in the future .

10.2 Scope for further work

The study developed and applied a combination of field and remote sensing based tree measurements

to five different Eucalyptus species in north eastern New South Wales, Australia. Such allomteric

models are often site and species specific; hence the performance of these models needs to be tested

for Eucalyptus species in other regions as well. While encouraging as the cluster versus single tree


133

model comparison was, it is of considerable interest to test this assertion, again over a larger range of

species and in different regions.

Of course, the landscape investigated in this work was only 662 ha, and while encapsulating

considerable variation in soils, elevation and aspect (for example as reported in Garraway and Lamb,

2011), it would be expected that the robustness and precision of any model could potentially be

enhanced by including other landscape parameters. While this study was supported by sophisticated

remote sensing datasets like LiDAR, which offers the advantage of large area coverage, the cost of

acquiring the LiDAR data restricted the analysis to a very small area for the stem density study with

few sample points and this means the results are less likely to be applicable to other areas. Terrestrial

scanning system which has widely been used in Australia could also seem as an alternative, as these

provide data in much finer details. Airborne laser scanning provides digital terrain model at 10-50 cm

precision digital height model at about 1m precision (Mass 2005), whereas terrestrial laser scanning

provides data with precise stem geometry information which can be relied upon.

Remote sensing and conventional field based methods have been extensively used in forest based

studies but not so in farmscapes which contain scattered trees. Our native eucalyptus trees, both

individual and in clusters are an important feature of our farmscapes and they contribute significantly

to above and below-ground carbon stocks in these landscapes. There will be a growing need to assess

carbon and biomass stocks across our farmscapes in order to fully quantify carbon storage change in

response to management and provide evidence-based support for carbon inventory and ultimately

carbon trading. Such large scale assessments are likely to only ever be feasible using remote sensing

techniques.

The method for tree crown identification undertaken in the above study was Object based image

classification using k-NN technique. As an initiation to further research the author would like to test

other non parametric classification techniques like SVM and Decision tree and their performance in

forest landscape. It is also a matter of further research to have an insight into other tree metrics like

percentage plan cover, crown porosity and percentage foliage cover, which have not been undertaken

in the present research.

134

References

Agresti, A., 1996. An Introduction to Categorical Data Analysis, Wiley, New York, 312 p.

Albrigo, L. G., C. A. Anderson, and G. J. Edwards. 1975. Yield estimation of ‘Valencia’ orange

research plots and groves. Proc. Fla. State Hort. Soc. 88: 44–49

Andersen, H. 2009. Using airborne light detection and ranging (LIDAR) to characterize forest stand

condition on the Kenai peninsula of Alaska. Western Journal of Applied Forestry. 24, 95–102.

Apan A.A. (1997). Land cover mapping for tropical forest rehabilitation planning using remotely-

sensed data. International Journal of Remote Sensing, 18(5): 1029–1049.

Aplin, P., P. Atkinson, P.Curran, 1997. Fine spatial resolution satellite sensors for the next decade. International Journal of Remote Sensing 18, 3873–3881.

Arzai, A.H., Aliyu, B.S., 2010. The relationship between canopy width, height and trunk size in some

tree species growing in the savana zone of Nigeria. Bajopas. 3(1), 260–263

Asner, G.P., Palace, M., Keller, M., Pereira Jr., R., Silva, J.N. M. and Zweede, J.C. (2002). Estimating

canopy structure in an Amazon forest from laser range finder and IKONOS satellite observations.

Biotropica, 34(4): 483–492.

Asner, G.P., and Warner, A.S., 2003. Canopy shadow in IKONOS satellite observations of tropical

forests and savannas. Remote Sensing of Environment, 87(4): 521–533, doi:10.1016/j.rse.2003.08.006

Attiwill, P.M., Adams, M.A., (Eds) 1996. Nutrition of Eucalypts (CSIRO Publishing) 448 pages.

Baatz, M. and A. Schape, 2000, Multiresolution segmentation—an optimization approach for high

quality multi-scale image segmentation. In Angewandte Geographische Informationsverarbeitung XII,

Beitra¨ge zum AGIT-Symposium Salzburg 2000 (Karlsruhe: Herbert Wichmann Verlag), pp. 12–23.

Baatz, M., U. Benz, S. Dehghani, M. Heynen, A. H¨ oltje, P. Hofmann, I. Lingenfelder, et al. 2004.

eCognition professional user guide, version 4.0. Definiens Imaging GmbH. Munchen, Germany:

Definiens.

Bacher, U., & Mayer, H. (2000). Automatic extraction of trees in urban areas from aerial imagery.

International Archives of Photogrammetry and Remote Sensing, 33(Part B 3/1), 51 – 57.

Baffetta, F., Corona, P., Fattorini, L., 2011. Assessing the attributes of scattered trees outside the

forest by a multi-phase sampling strategy. Forest. 84 (3), 315–325

Barbosa, P.M., Stroppiana, D., Grégoire, J.M., Pereira, J.M.C., 1999. An assessment of vegetation fire

in Africa (1981–1991): Burned areas, burned biomass, and atmospheric emissions. Global

Biogeochemical Cycles 13, 933-950.

Barnes, P., Wilson, B.R., Trotter, M.G., Lamb, D.W., Reid, N., Koen, T., 2011a. The patterns of

grazed pasture associated with scattered trees across an Australian temperate landscape: an

investigation of pasture quantity and quality. Rangel. J. 33, 121–130.

Barnes, P., Wilson, B.R., Reid, N., Koen, T.B., Lockwood, P., Lamb, D.W., 2011b. Litterfall and

associated nutrient pools extend beyond the canopy of scattered eucalypt trees in temperate pastures.

Plant Soil. 345, 339–352

135

Beets, P.N., Kimberley, M.O., Oliver, G.R., Pearce, S.H., Graham, J.D., Brandon, A., 2012.

Allometric Equations for Estimating Carbon Stocks in Natural Forest in New Zealand. Forests. 3,

818–839.

Bella, I.E., 1971. A new competition model for individual trees. Forest. Science. 17, 364–372

Benz, U. C., P., Hoffmann, G., Willhauck, I., Lingenfelder, and M., Heynen, 2004. Multi-resolution

Object-oriented Fuzzy analysis of Remote Sensing Data for GIS-ready Information. ISPRS Journal of

Photogrammetry and Remote Sensing, 58, pp. 239–258.

Benediktsson, J.A., Pesaresi, M. and Arnason, K. (2003). Classification and feature extraction for

remote sensing images from urban areas based on morphological transformations. IEEE Transactions

on Geoscience and Remote Sensing, 41 (9): 1940–1949.

Bertolette, D. R., & Spotskey, D. B (1999). Fuel model and forest type mapping for FARSITE input.

In G. E. Gollberg (Ed.), The joint fire science conference and workshop. Boise, ID: University of

Idaho and International Association of Wildland Fire.

Bergen, K., Colwell, J. and Sapio, F. “Remote sensing and forestry: Collaborative implementation for

a new century of forest information solutions”, Journal of Forestry, vol. 98, no. 6, pp. 4–9, 2000

Biging, G.S., Dobbertin, M., 1992. A Comparison of Distance-Dependent Competition Measures for

Height and Basal Area Growth of Individual Conifer Trees. Forest Science. 38(3), 695–720.

Biging, G.S., Dobbertin, M., 1995. Evaluation of Competition Indices in Individual Tree Growth

Models. Forest Science. 41 (2), 360–377.

Bird, P., Bicknell, D., Bulman, P., Burke, S., Leys, J., Parker, J., 1992. The role of shelter in Australia

for protecting soils, plants and livestock. Agroforest. Sys. 20, 59–86.

Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS Journal of

Photogrammetry and Remote Sensing, 65(1): 2–16.

Blaschke, T., and J. Strobl, 2001. What’s wrong with pixels? Some recent developments interfacing

remote sensing and GIS. GeoBIT/GIS, 6: 12–17.

BoM., 2014. Bureau of Meteorology Climate Statistics for Australian Locations: Armidale airport.

http://www.bom.gov.au/climate/averages/tables/cw_056238.shtml. Accessed 3rd March 2014.

Brambley, R.G.V., P.A. Hill, P.J. Thorburn, F.J. Kroon, and K. Panten., 2008. Precision Agriculture

for improved environmental outcomes: Some Australian perspectives. Landbauforschung. 58:161–

178.

Brandtberg, T., Warner, T., Landenberg, R., & McGraw, J. (2003). Detection and analysis of

individual leaf-off tree crowns in small footprint, high sampling density lidar data from the eastern

deciduous forest in North America. Remote Sensing of Environment,85, 290−303.

Breiman, L., J. Friedman, R. Olshen, and C.J. Stone, 1984. Classification and Regression Trees,

Chapman and Hall, New York, 146–150 p

Brieman, L. 2001. Random Forests. Machine Learning, 45(1), 5–32

Brown, S., T. Pearson, D. Slaymaker, S. Ambagis, N. Moore, D. Novelo and W. Sabido, 2005.

Creating a Virtual Tropical Forest from three-dimensional Aerial Imagery to Estimate Carbon Stocks,

Ecological Applications, 15(3):1083–1095.

136

Brown de Colstoun, E.C., M.H. Story, C. Thompson, K. Commisso, T.G. Smith, and J.R. Irons, 2003.

National Park vegetation mapping using multitemporal Landsat/data and a decision tree classifier,

Remote Sensing of Environment, 85(3):316–327.

Budreski, K.A., R. H. Wynne, J. O. Browder, and J. B. Campbell, 2007. Comparison of Segment and

Pixel-based Non-parametric Land Cover Classification in the Brazilian Amazon Using Multi-

temporal Landsat TM/ETM+ Imagery. Photogrammetric Engineering and Remote Sensing,73(7):

813.

Cade, B.S., 1997. Comparison of tree basal area and canopy cover in habitat models: Subalpine forest.

J. Wildlife Management. 61(2), 326–335.

Cao, C., Bao, Y., Xu ,M. , Chen,W., Zhang, H., He,Q. , Li, Z., Guo, H., Li, J., Li X., & Li,G. (2012)

Retrieval of forest canopy attributes based on a geometric-optical model using airborne LiDAR and

optical remote-sensing data. International Journal of Remote Sensing Vol. 33, No. 3, 10 February

2012, 692–709

Case, B.S. and Hall, R.J. 2008. Assessing prediction errors of generalized tree biomass and volume

equations for the boreal forest region of West Central Canada. Canadian Journal of Forest Research,

38, 878–889

Chastain, R.A., & Townsend, P.A. (2007). Use of landsat ETM and topographic data to characterize

evergreen understory communities in appalachian deciduous forests. Photogrammetric Engineering &

Remote Sensing, 73, 563–575

Chave, J., Andalo, C., Brown, S., Cairns, M., Chambers, J., Eamus, D., Fölster, H., Fromard, F.,

Higuchi, N., Kira, T., 2005. Tree allometry and improved estimation of carbon stocks and balance in

tropical forests. Oecologia 145, 87–99.

Chubey, M. S., Franklin, S. E. and Wulder, M. A. (2006). Object-based analysis of Ikonos-2

imagery for extraction of forest inventory parameters. Photogrammetric Engineering and Remote

Sensing 72(4):383–394.

Clark, M. L., Roberts, D. A., & Clark, D. B. (2005). Hyperspectral discrimination of tropical

rain forest tree species at leaf to crown scales. Remote Sensing of Environment, 96, 375−398

Coder, D. K. D. 2000. Crown Shape Factors & Volumes. Tree Biomechanics Series. University of

Georgia: University of Georgia.

Cohen, W. B., & Spies, T. A. (1992). Estimating structural attributes of Douglas-fir/western hemlock

forest stands from Landsat and SPOT imagery. Remote Sensing of Environment, 41, 1 – 17.

Cook, S.E., and R.G.V. Bramley, 2000. Coping with variability in agricultural production-

Implications for soil testing and fertilizer management. Communications in Soil Science and Plant

Analysis, 31: 1531–1551.

Commonwealth of Australia, 1999. Comprehensive Regional Assessment, World Heritage Sub-

theme: Eucalypt-dominated vegetation. Report of the Expert Workshop, Canberra, 8 & 9 March 1999.

(Commonwealth of Australia), 105 pages.

Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data.

Remote Sensing of Environment, 37(1): 35–46.

Congalton, R.G. and K. Green, 2009. Assessing the accuracy of remotely sensed data: principles and

137

practices, CRC/Taylor & Francis, Boca Raton.

Cutini A, Chianucci F, Manetti MC, 2013. Allometric relationships for volume and biomass for stone

pine (Pinus pinea L.) in Italian coastal stands. iForest 6: 331–337 [online 2013-08-29] URL:

http://www.sisef. it/iforest/contents/?id=ifor0941-006

Dash, J., E. Steinle, R.P. Singh, and H.P. Bähr, 2004. Automatic building extraction from laser

scanning data: An input tool for disaster management, Advances in Space Research, 33:317–322.

Dare, P. M.. 2005. Shadow Analysis in High-Resolution Satellite Imagery of Urban Areas,

Photogrammetric Engineering & Remote Sensing, 2005, Vol. 71, No. 2, pp. 169–177.

Dassot, M., Barbacci , A., Colin A., Fournier, M.,and Constant, T. Tree architecture and biomass

assessment from terrestrial LiDAR measurements: a case study for some Beech trees (Fagus

sylvatica). SilviLaser 2010.

Definiens, 2004, eCognition 4 User Guide, Definiens Imaging GmbH Munich.

Dehvari A., and Heck R. J., 2009. Comparison of object-based and pixel based infrared airborne

image classification methods using DEM thematic layer. Journal of Geography and Regional

Planning, 2 (4): 086–096.

Drake, J.B.; Dubayah, R.O.; Clark, D.; Knox, R.; Blair, J.B.; Hofton, M.; Chazdon, R.L.; eishampel,

J.F.; Prince, S. 2002a. Estimation of tropical forest structural characteristics using large-footprint

LIDAR. Remote Sensing of Environment. 79: 305–319.

Duveiller, G., P. Defourny, B. Desclee, and, P. Mayaux, 2008. Deforestation in Central Africa:

Estimates at regional, national and landscape levels by advanced processing of systematically-

distributed Landsat extracts. Remote Sensing of Environment, 112(5): 1969–1981.

Eamus, D., McGuinness, K., and Burrows, W (2000) Review of Allometric Relationships for

Estimating Woody Biomass for Queensland, the Northern Territory and Western Australia. National

Carbon Accounting System Technical Report No. 5A.

Edson, C., and Wing, M.G. (2011). Airborne Light Detection and Ranging (LiDAR) for individual

Tree Stem Location, Height and Biomass Measurements. (2011). Remote Sensing (Open Access), doi:

10.3390/rs3112494. Vol. 3, pp 2494–2528

Elliott, H.J., Bashford, R., Greener, A., 1993. Effects of defoliation by the leaf beetle, Chrysophtharta

bimaculata, on growth of Eucalyptus regnans plantations in Tasmania. Aust. Forest. 56, 22–26.

Fang, D.Q., C.T. Federici, and M.L. Roose. 1998. A high resolution linkage map of the citrus tristeza

virus resistance gene region in Poncirus trifoliata (L.) Raf. Genetics 150:883–890.

Falcucci, A., L. Maiorano, and L. Boitani, 2007. Changes in land-use/land-cover patterns in Italy and

their implications for biodiversity conservation. Landscape Ecology, 22(4): 617–631.

Fisher, P., 1997. The pixel: a snare and a delusion. International Journal of Remote Sensing, 18(3):

679–685.

Flanders, D., M. Hall-Beyer, and J. Pereverzoff, 2003. Preliminary evaluation of eCognition object-

based software for cut block delineation and feature extraction. Canadian Journal of Remote Sensing,

29(4): 441–452.

Fleck, S. Mölder, I., Jacob, M., Gebauer, T., Jungkunst, H.F., Leuschner, C., 2011. Comparison of

conventional eight-point crown projections with LIDAR-based virtual crown projections in a

temperate old-growth forest. Annals Forest Sci. 68, 1173–1185.

http://www.sisef/

138

Foody, G.M., Boyd, D.S., Cutler, M.E.J., 2003. Predictive relations of tropical forest biomass from

Landsat TM data and their transferability between regions. Remote Sensing of Environment 85, 463-

474.

Foody, G.M., 2004. Thematic map comparison: evaluating the statistical significance of differences in

classification accuracy. Photogrammetric Engineering and Remote Sensing, 70(5): 627–634.

Franklin, J., Strahler, A.H., 1988. Invertible canopy reflectance modeling of vegetation structure in

semi-arid woodland. IEEE Trans. Geoscience and Remote Sensing 26, 809–825.

François A. Gougeon and Donald G. Leckie. 2006 The Individual Tree Crown Approach Applied to

Ikonos Images of a Coniferous Plantation Area. Photogrammetric Engineering & Remote Sensing

Vol. 72, No. 11, November 2006, pp. 1287–1297.

Fu, K.S. and Mui, J.K. (1981). A survey of image segmentation, Patter Recognition, vol 13, pp 3-16

Garraway, E., Lamb, D.W., 2011. Delineating spatial variations in soil carbon using remotely sensed

data: A New England case study, Proceedings of the 2011 Regional Convention, Newcastle Division,

Engineers Australia (R. Patterson Ed.), (Newcastle Division Engineers Australia: Newcastle West,

Australia), University of New England, Armidale. NSW 16th -18th September 2011. ISBN 978-0-

85825-870-9, 45-52.

García, M., Riaño, D., Chuvieco, E., Danson, F.M., 2010. Estimating biomass carbon stocks for a

Mediterranean forest in central Spain using LiDAR height and intensity data. Remote Sensing of

Environment 114, 816–830.

Gering, L.R., May, D.M., 1995. The Relationship of Diameter at Breast Height and Crown Diameter

for Four Species Groups in Hardin County, Tennessee. Southern J. Appl. Forest.. 19 (4), 177–181.

Gertner, G.Z. 1991. The sensitivity of measurement error in stand volume estimation. In Canadian

Journal of Forest Research, vol. 20, pp. 800–804.

Gerylo, G., R. J. Hall, S. E. Franklin, A. Roberts, and E. J. Milton, 1998. Hierarchical image

classification and extraction of forest species composition and crown closure from airborne

multispectral images, Can. J. Rem. Sen., 24, (3), 219–232.

Gill, S. J., Biging, G.S., Murphy, E.C., 2000. Modeling conifer tree crown radius and estimating

canopy cover. Forest Ecology and Management 126, 405–416.

Gitelson, A. A., Kaufmann, Y. J., & Merzlyak, M. N. (1996). Use of Green channel in remote

sensing of global vegetation from EOS-MODIS. Remote Sensing of Environment, 58, 289−298

Goh, K. M., Mansur, I., Mead, D. J., Sweet, G. B., 1996. Biological nitrogen fixing capacity and

biomass production of different understorey pastures in a Pinus radiata-pasture agroforestry system in

New Zealand. Agroforest. Sys. 34, 33–49.

Goodchild, M.F., M. Yuan, and T.J. Cova, 2007. Towards a general theory of geographic

representation in GIS. International Journal of Geographical Information Science, 21(3): 239– 260.

Goodenough, D. G., Chen, H., Han, D. T., & Li, J. Y. (2005). Multisensor Data fusion for

aboveground carbon estimation. Proceedings of XXVIIIth General Assembly of the International

Union of Radio Science (URSI), October, 23–29, 2005, New Delhi, India (pp. 1−4)

Gougeon, F.A. 1995. A crown-following approach to the automatic delineation of individual tree

crowns in high spatial resolution aerial images. Can. J. Rem. Sensing 21(3):274–284.

139

Graham, S., Wilson, B. R., Reid, N. and Jones,H. (2004). Scattered paddock trees, litter chemistry,

and surface soil properties in pastures of the New England Tablelands, New South Wales. Soil

Research, vol. 42, no. 8, pp. 905–912, 2004.

Greenberg, J.A., Dobrowski, S.Z., Ustin, S.L. (2005) “Shadow allometry: Estimating tree structural

parameters using hyperspatial image analysis” Remote Sensing of Environment, 97:15–25.

Grote, R., 2003. Estimation of crown radii and crown projection area from stem size and tree position.

Annals Forest Sci. 60, 393–402.

Hall, R.J., Morton, R.T., Nesby, R.N., 1989. A Comparison of Existing Models for DBH Estimation

for Large –scale photos. Forestry Chronicle. 114–116.

Hall, R., Skakun, R., Arsenault, E., Case, B., 2006. Modeling forest stand structure attributes using

Landsat ETM+ data: Application to mapping of aboveground biomass and stand volume. Forest

Ecology and Management 225, 378–390.

Hamburg, S.P., Zamolodchikov, D.G., Korovin, G.N., Nefedjev, V.V., Utkin, A.I., Gulbe, J.I., Gulbe,

T.A., 1997. Estimating the carbon content of Russian forests; a comparison of phytomass/volume and

allometric projections. Mitigation and Adaptation Strategies for Global Change 2, 247–265.

Haralick, R. and Shapiro, L. (1985). Survey: image segmentation techniques. Computer Vision,

Graphics, and Image Processing, 29, 100–132.

Haralick, R. M., K. Shanmugam, and I. Dinstein. 1973. Textural features for image classification.

IEEE TransActions on Systems, Man, and Cybernetics 3:610–21

Haralick, R. M. (1986). Statistical image texture analysis. In T. Y. Young, & K. S. Fu (Eds.),

Handbook of pattern recognition and image processing (pp. 247–279). New York: New York.

Hardin, P.J., 1994. Parametric and nearest-neighbor methods for hybrid classification: a

comparison of pixel assignment accuracy. Photogrammetric Engineering and Remote Sensing, 60,

1439– 1448.

Hay, G.J., G. Castilla, M.A.Wulder, J.R. Ruiz, 2005. An automated object-based approach for the

multiscale image segmentation of forest scenes. International Journal of Applied Earth

Observation and Geoinformation, 7: 339–359.

Heikkonen, J., and A. Varfis, 1998. Land-cover land-use classification of urban areas: A remote

sensing approach, International Journal of Pattern Recognition and Artificial Intelligence, 12(4):475–

489.

Heinzel, J.N., Weinacker, H., and Koch, B., 2008. Full automatic detection of tree species based on

delineated single tree crowns – a data fusion approach for airborne laser scanning data and aerial

photographs, Proceedings of the SilviLaser 8th international conference on LiDAR applications in

forest assessment and inventory, (pp. 76–85), September 18-19, 2008, Edinburgh, UK.

Hester D.B., S.A.C. Nelson, H.I. Cakir, S. Khorram, and H. Cheshire., 2010. High – resolution

landcover change detection based on fuzzy uncertainty analysisand change reasoning. International

Journal of Remote Sensing, 31(2): 455–475.

Hemery, G.E., Savill, P.S., Pryor, S.N., 2005. Applications of the crown diameter-stem diameter

relationship for different species of broadleaved trees. Forest Ecoogy and Management.. 215, 285–

294.

140

Hill, R. A., & Thomson, A. G. (2005). Mapping woodland species composition and structure using

airborne spectral and LIDAR data. International Journal of Remote Sensing, 26, 3763−3779

Hirschmugl, M., Ofner, M., Raggam, J., Schardt, M., 2007. Single tree detection in very high

resolution remote sensing data. Remote Sensing of Environment, 110, pp. 533–544.

Houghton, R.A., 2005. Aboveground forest biomass and the global carbon balance. Global Change

Biol. 11, 945–958.

Hollaus, M., Wagner, W., Schadauer, K., Maier, B., Gabler, K., 2009. Growing stock estimation for

alpine forests in Austria: a robust lidar-based approach. Canadian Journal of Forest Research 39(7),

1387–1400.

Holmgren, J., & Persson, A. (2004). Identifying species of individual trees using airborne laser

scanner. Remote Sensing of Environment, 90, 415−423

Huertas, A., and Nevatia, R. (1988). “Detecting Buildings in Aerial Images,” Computer Vision,

Graphics and Image Processing, 41(2):131–152.

Hudak, A.T., Crookston, N.L., Evans, J.S., Falkowski, M.J., Smith, A.M.S., Gessler, P.E., Morgan,

P., 2006. Regression modeling and mapping of coniferous forest basal area and tree density from

discrete-return lidar and multispectral satellite data. Canadian Journal of Remote Sensing 32, 126–

138.

Hudak, A.T.; Evans, J.S.; Stuart Smith, A.M. LiDAR Utility for Natural Resource Managers. Remote

Sening. 2009, 1, 934–951

Hunter, M.O., Keller, M., Victoria, D. and Morton, D.C. (2013). Tree height and tropical forest

biomass estimation. Biogeosciences, 10: 8385–8399

Huiyan,G.U., Limin, D., Gang, W., Dong, X., Shunzhong, W

& Hui, W. Estimation of forest volumes

by integrating Landsat TM imagery and forest inventory data Science in China: Series E

Technological Sciences 2006 Vol.49 Supp. I 54—62 DOI: 10.1007/s11431-006-8107-z

Hyyppa, J., Hyyppa, H., Leckie, D., Gougeon, F., Yu, X., and Maltamo, M. (2008) Review of

methods of small footprint airborne laser scanning for extracting forest inventory data in boreal

forests. International Journal of remote Sensing, 29(5), 1339–1366.

Im, J., Jensen, J. R., & Hodgson, M. E. (2008). Object-based land cover classification using high-

posting-density LIDAR data. GIScience & Remote Sensing, 45, 209−228.

Irving, R., and McKeown, D. (1989). “Methods for exploiting the Relationship Between Buildings

and their Shadows in Aerial Imagery,” IEEE Transactions on Systems, Man and Cybernetics, SMC

19(6): 1564–1575.

Jensen, J.R., 2009. Remote sensing of the environment, Pearson Education India.

Jennings, S.B., Brown, N.D., Sheil, D., 1999. Assessing forest Canopies and Understorey

illumination: canopy closure, canopy cover and other measures. Forestry. 72 (1), 59–73.

Jobin, B., S. Labrecque, M. Grenier, and G. Falardeau, 2008. Object-based classification as an

alternative approach to the traditional pixel-based classification to identify potential habitat of the

Grasshopper Sparrow. Environmental Management, 41(1): 20–31.

Jonson, J. H., and Freudenberger, D. 2011. Restore and sequester: estimating biomass in native

Australian woodland ecosystems for their carbon-funded restoration. CSIRO PUBLISHING

http://www.ncbi.nlm.nih.gov/pubmed?term=Labrecque%20S%5BAuthor%5D&cauthor=true&cauthor_uid=17985180

http://www.ncbi.nlm.nih.gov/pubmed?term=Grenier%20M%5BAuthor%5D&cauthor=true&cauthor_uid=17985180

http://www.ncbi.nlm.nih.gov/pubmed?term=Falardeau%20G%5BAuthor%5D&cauthor=true&cauthor_uid=17985180

141

Australian Journal of Botany, 59, 640–653 http://dx.doi.org/10.1071/BT11018

Johansen, K., N. C. Coops, S.E. Gergel, and Y. Stange, 2007. Application of high spatial resolution

satellite imagery for riparian and forest ecosystem classification. Remote Sensing of Environment,

110(1): 29–44.

Kamagata N., Hara K., Mori M., Akamatsu Y., Li Y., Hoshino Y. (2006) - A new method of

vegetation mapping by object-based classification using high resolution satellite data. In: First

International Conference on Object-based Image Analysis (OBIA 2006), XXXVI (4/C42) Salzburg,

Austria.

Ke, Y., Quackenbush, L.J.L., 2011. A review of methods for automatic individual tree-crown

detection and delineation from passive remote sensing. International Journal of Remote Sensing 32,

4725–4747.

Ke, Y. and L.J. Quackenbush, 2008. Comparison of individual tree crown detection and delineation

methods, In Proceedings of 2008 ASPRS Annual Conference (American Society of Photogrammetry

and Remote Sensing, Bethesda, Maryland), April 28-May 2, 2008, Portland, Oregon.

Ke, Y., Quackenbush, L.J.L., and Im. J.,2010. Synergistic use of Quickbird imagery and lIDAr data

for object based forest species classification. Remote Sensing of Environment.114, 1141–1154.

Keith, H., Barrett, D., and Keenan, R (2000) Review of Allometric Relationships for Estimating

Woody Biomass for New South Wales, The Australian Capital Territory, Victoria, Tasmania and

South Australia. National Carbon Accounting System Technical report No. 5B

Kini, A.U., and Popescu, S.C. ( 2004). TreeVaW : A versatile tool for Analysing Forest Canopy

LiDAR Data – A Preview with an Eye towards Future. In Proceedings of the ASPRS 2004 Fall

Conference, Kansas City, MO, USA, 12–16 September 2004; pp1–10

Kim HY, Lieffering M, Kobayashi K, Okada M, Miura S. 2003. Seasonal changes in the effects of

elevated CO2 on rice at three levels of nitrogen supply: a free air CO2 enrichment (FACE)

experiment. Global Change Biology 9: 826–837.

Kim M., Madden M. (2006) - Determination of optimal scale parameter for alliance-level

forest classification of multispectral IKONOS image. Commission IV, WG IV/4 on Proceeding of 1st

OBIA Conference, 4-5 July, Salzburg, Austria.

Kim S., 2010. LiDAR based species classification using multivariate cluster analysis. ASPRS Annual

Conference. San Diego, California, April 26-30.

Köstner, B., Falge, E., Tenhunen, J. D., 2002. Age-related effects on leaf area/sapwood area

relationships, canopy transpiration and carbon gain of Norway spruce stands (Picea abies) in the

Fichtelgebirge, Germany. Tree Physiol.. 22, 567–574.

Krosgad, F., and Schiess, P. 2004 – The Allure and pitfalls of using LIDAR topography in harvest and

road design. Proceedings of 12th International Mountain Logging Conference. June 13-16, 2004,

Vancouver, B.C.

Kuyah, S., Dietz, J., Muthuri, C., Jamnadass, R., Mwangi, P., Coe, R., Neufeldt, H., 2012. Allometric

equations for estimating biomass in agricultural landscapes: I. Aboveground biomass. Ag. Ecosys.

Enviro. 158, 216–224.

Landsberg, J., Ohmart, C., 1989. Levels of insect defoliation in forests: patterns and concepts. Trends

Res. Ecol. Evol.. 4, 96–100.

142

Lamb, D.W., 2000. The use of qualitative airborne multispectral imaging for managing agricultural

crops – A case study in south eastern Australia. Australian Journal of Experimental Agriculture,

40(5): 725–738.

Lamb, D.W., M. Trotter, A. Murphy, and I. Young, 2013. “UNE SMART Farm: Showcasing the

value of broadband connectivity in the business snd lifestyle of farming”. In. Proceedings of the

Digital Rural Futures Conference. D.W.Lamb (Editor), 26-28th June 2013 (University of New

England: Australia), 89.

Lambert, M., Ung, C., Raulier, F., 2005. Canadian national tree aboveground biomass equations.

Canadian Journal of Forest Research 35, 1996–2018.

Lara A. Arroyo, , Kasper Johansen , John Armston , Stuart Phinn Integration of LiDAR and

QuickBird imagery for mapping riparian biophysical parameters and land cover types in Australian

tropical savannas. Forest Ecology and Management 259 (2010) 598–606

Laymon, 2005. Landcover classification comparison with different resolution

imagery. http://www.ncrste.msstate.edu. Accessed May 2012.

Leblon B., Gallant L. and Granberg H. (1996). Effects of shadowing types on ground-measured

visible and near-infrared shadow reflectances. Remote Sensing of Environment. 58(3): 322–328.

Leckie, D.G., Gougeon, F., Walsworth, N., and Paradine, D. (2003). Stand delineation and

composition estimation using semi automated individual tree+ crown analysis. Remote Sensing of

Environment, 85, 355–369

Leckie, D., Gougeon, F., Hill, D., Quinn, R., Armstrong, L., & Shreenan, R. (2003). Combined high

density LiDAR and multispectral imagery for individual tree crown analysis. Canadian Journal of

Remote Sensing, 29, 633−649

Leckie, D.G., and Gillis, M.D., “Forest inventory in Canada with emphasis on map production”. The

Forestry Chronicle., vol. 71, no. 1, pp. 74–88, 1995

Lee, A.C., Lucas, R.M., 2007. A Lidar-derived canopy density model for tree stem and

crownmapping in Australian forests. Rem. Sens. Env. 111(4), 493–518.

Lefsky, M.A., Cohen, W.B., Parker, G.G., and Harding, D.J., 2002a. Lidar remote sensing for

ecosystem studies, Bioscience, Vol. 52, No. 1, pp.19–30.

Lefsky, M.A., Cohen, W.B., Harding, D.J., Parker, G.G., Acker, S.A., and Gower, S.T., 2002b. Lidar

remote sensing of above-ground biomass in three biomes, Global Ecology & Biogeography, No. 11,

pp. 393–399.

Lefsky, M.A., Harding, D., Cohen, W.B., Parker, G., and Shugart, H.H., 1999a. Surface lidar remote

sensing of basal area and biomass in deciduous rorests of Eastern Maryland, USA, Remote Sensing of

Environment, No. 67, pp. 83–98. ASPRS 2008 Annual Conference Portland, Oregon ♦ April 28 –

May 2, 2008

Lefsky, M.A., Cohen, W.B., Acker, S.A., Parker, G.G., Spies, T.A., and Harding, D., 1999b. Lidar

remote sensing of the canopy structure and biophysical properties of Douglas-fir western hemlock

forests, Remote Sensing of Environment, (70) pp. 339–361.

Leverett, R. (2010). Measuring Tree Height by Tape and Clinometer Scenarios. Bulletin of the

Eastern Native Tree Society, 5(3-4): 3–12.

Liang, X., Hyyppa, J., & Matikainen, L. (2007). First-last pulse signatures of airborne laser

143

scanning for tree species classification, Deciduous–coniferous tree classification using difference

between first and last pulse laser signatures. Proceedings of the ISPRS workshop, Laser Scanning

2007 and Silvilaser 2007 (pp. 253−257).

Lim, K., Hopkinson, C., and Treitz, P.(2008) Examining the effects of sampling point densities on

laser canopy height and density metrics. Forestry Cronicle, 84(6), pp. 876–885

Lim, K.S., Treitz, P.M., 2004. Estimation of above ground forest biomass from airborne discrete

return laser scanner data using canopy-based quantile estimators. Scandinavian Journal of Forest

Research 19, 558–570.

Lindeberg ,T. “Feature detection with automatic scale selection”, International Journal of Computer

Vision, vol. 30, no. 2, pp. 77–116, 1998

Liow, Y., and T. Pavlidis, T. (1990). “Use of Shadows for Extracting Buildings in Aerial Images,”

Computer Vision, Graphics and Image Processing, 49: 242-277 Liu W., Yamazaki F., 2012. Object-

based shadow

Liu, H., D. Sherman, and S. Gu, 2007. Automated extraction of shorelines from airborne light

detection and ranging data and accuracy assessment based on Monte Carlo simulation, Journal of

Coastal Research, 23:1359–1369.

Liu, X., Zhang, Z., Peterson, J. and Chandra, S. 2007. LiDAR-derived high quality ground control

information and DEM for image orthorectification. GeoInformatica 11(1), 37–53.

Loos, R., Niemann, O., 2006. Identification of individual trees and canopy shapes using LiDAR data

for fire management. IEEE, pp. 3755–3757.

Loos, R., Niemann, O., Visintini, F., 2007. Identification of partial canopies using first and last return

lidar data. Our Common Borders – Safety, Security, and the Environment Through Remote Sensing

October 28 – November 1, 2007, * Ottawa, Ontario, Canada

Lu, D. and Weng, Q. 2007. A survey of image classification methods and techniques for improving

classification performance, International Journal of Remote Sensing, 28:5, 823-870.

Lu, D., 2006. The potential and challenge of remote sensing‐based biomass estimation. International

Journal of Remote Sensing 27, 1297–1328.

Machala, M. and Zejdová, L.2014. Forest Mapping Through Object-based Image Analysis of

Multispectral and LiDAR Aerial Data. European Journal of Remote Sensing - 2014, 47: 117–131 doi:

10.5721/EuJRS20144708

Maclean, G.A., and Krabill, W. B. (1986), Gross merchantable timber volum estimation using

airborne Lidar system, Canadian Journal of Remote Sensing, Vol. 12, pp 7–18

Makungwa, S.D., Chittock, A., Skole, D.L., Kanyama-Phiri, G.Y., Woodhouse, I.H., 2013. Allometry

for Biomass Estimation in Jatropha Trees Planted as Boundary Hedge in Farmers’ Fields. Forests. 4,

218–233.

Malimbwi, R., Solberg, B., Luoga, E., 1994. Estimation of biomass and volume in miombo woodland

at Kitulangalo Forest Reserve, Tanzania. Journal of Tropical Forest Science 7, 230-242.

Maltamoa, M., , Eerikaïnenb, K., Pitka¨nena, J., Hyyppa¨ J., and Vehmasa., M. (2004) Estimation of

timber volume and stem density based on scanning laser altimetry and expected tree size distribution

functions. Remote Sensing of Environment 90 (2004) 319 – 330

144

Manakos, I., T. Schneider, U. Ammer, 2000. A comparison between the ISODATA and the eCognition

classification methods on basis of field data. IAPRS 33, 133–139.

Manning, A.D., Fischer, J. and Lidenmayer, D.B. 2006. Scattered trees are keystone structures.

Implications for conservation. Biological. Conservation. 132, 311–321

Mathieu R., Aryal,J. and Chong A. K., , “Object-Based Classification of IKONOS Imagery for

Mapping Large-Scale Vegetation Communities in Urban Areas”, Sensors, vol.7, pp. 2860–2880, 2007

McComb, J. W., Roberts, S. D., & Evans, D. L. (2003). Influence of fusing lidar and multispectral

imagery on remotely sensed estimates of stand density and mean tree height in a managed loblolly

pine plantation. Forest Science, 49, 457 – 466.

Mcinerney, D.O., Suarez-Minguez, J., Valbuena, R., and Nieuwenhuis, 649 M., 2010, Forest canopy

height retrieval using Lidar data, medium-resolution 650 satellite imagery and kNN estimation in

Aberfoyle, Scotland. Forestry, 83, pp. 651 195−206.

McGaughey, R. (2010). FUSION/LDV: Software for LIDAR Data Analysis and Visualization. United

States Department of Agriculture Forest Service Pacific Northwest Research Station

FUSION/LDV: Version 2.90

Medhurst, J.L. and Beadle, C.L. (2001) Crown structure and leaf area index development in thinned

and unthinned Eucalyptus nitens plantations. Tree Physiology, 21: 989–999.

Minho, K., 2009. Object-based spatial classification of forest vegetation with ikonos imagery,

Ph.D. Dissertation, University of Georgia, Georgia, 133p.

Moeur, M., 1997. Spatial models of competition and gap dynamics in old growth Tsuga

heterophylla/Thuja plicata forests. Forest Ecol. Man. 94, 175–186.

Mundt, J.T., N.F. Glenn, K.T. Weber, and J. Pettingill, 2006. Determining target detection limits and

accuracy delineation using an incremental technique. Remote Sensing of Environment, 105, 34– 40.

Myint S.W., Gober P., Brazel A., Grossman-Clarke S., Weng Q. (2011) - Per-pixel vs. object-based

classification of urban land cover extraction using high spatial resolution imagery. Remote Sensing of

Environment, 115 (5): 1145–1161. doi: http://dx.doi. org/10.1016/j.rse.2010.12.017

Næsset, E., & Bjerknes, K.-O. 2001. Estimating tree heights and number of stems in young

forest stands using airborne laser scanner data. Remote Sensing of Environment, 78,

328–340

National Parks and Wildlife Service, 2003. Duval Nature Reserve, Plan of Management.

(Commonwealth of Australia) ISBN 0 7313 6600 X, 17 pages.

Næsset, E., Økland, T., 2002. Estimating tree height and tree crown properties using airbo

rne scanning laser in a boreal nature reserve. Remote Sensing of Environment 79, 105–115.

Nelson, R., Krabill, W. and Maclean, G. 1984: Determining forest canopy characteristics using

airborne laser data. Remote Sensing of Environment 15, 201–12.

Nelson, R., Krabill, W. & Tonelli, J. 1988a. Estimating forest biomass and volume using airborne

laser data. Remote Sens. Environ. 24: 247–267.

Nelson R., 1997. Modelling forest canopy heights. The effects of canopy shape. Rem. Sens. Env. 60,

327–334.

http://dx.doi/

145

Neumann, F.G., 1993. Insect pest problems of eucalypt plantations in Australia. 3. Victoria.

Australian Forest. 56, 370–374.

O’Brien, S.T., Hubbell, S.B.Spiro, P., Condit, R., Foster, R.B., 1995. Diameter, height, crown and age

relationships in eight neotropical tree species. Ecology. 76 (6), 1926–1939.

Oliver, I., Pearce, S., Greenslade, P. J. M., Britton, D. R., 2006. Contribution of paddock trees to the

conservation of terrestrial invertebrate biodiversity within grazed native pastures. Australian Ecol. 31,

1–12.

Orka, H.O., Nasset, E and Bollandsas, O.M. 2007. Utilizing airborne laser intensity for tree

speciesclassification. Onternational Archives of the Photogrammetry, remote Sensing and Spatial

Information Sciences 36. Part3/W52:300–304

Ozdemir , I. (2008) Estimating stem volume by tree crown area and tree shadow area extracted from

pan-sharpened Quickbird imagery in open Crimean juniper forests. International Journal of Remote

Sensing. Vol 29. Pages 5643–5655

Palace, M., M. Keller, G. P. Asner, J. N. M. Silva, and C. Passos. 2007. Necromass in undisturbed and

logged forests in the Brazilian Amazon. Forest Ecology and Management 238: 309–318.

Paul, K. I., Roxburg, S. H., Ritson , P., Brooksbank, K., England, J. R., Larmour, J. S., Raison, J.R.,

Peck, A., Wildy, D. T., Sudmeyer, R. A., Giles, R., Carter, J., Bennett, R., Mendham, D. S.,

Huxtable, D., and Bartle, J. R. 2013. Forest Ecology and Management 310, 1005–1015

Persson, A., Holmgren, J. and Soderman, U. 2002. Detecting and measuring individual trees using

and airborne alser scanner. Photogrammetric Engineering and remote Sensing 68:925–932

Phua, M.H., Saito, H., 2003. Estimation of biomass of a mountainous tropical forest using Landsat

TM data. Canadian Journal of Remote Sensing 29, 429-440.

Platt, R.V., and L. Rapoza, 2008. An evaluation of an Object-Oriented Paradigm for Land Use/Land

Cover Classification. The Professional Geographer, 60(1): 87–100.

Popescu, S.C., Wynne, R.H., 2004a. Seeing the trees in the forest: using lidar and multispectral data

fusion with local filtering and variable window size for estimating tree height. PERS. 70 (5), 589–604.

Popescu, S.C., Wynne, R.H., Scrivani, J.A., 2004b. Fusion of small-footprint Lidar and multispectral

data to estimate plot-level volume and biomass in deciduous and pine forests in Virginia, USA. Forest

Sci. 50 (4), 551–564.

Popescu, S.C., et al (2002). Estimating plot level tree heights with LiDAR: local filtering with a

canopy height based variable window size. Computers and Electronics in Agriculture. 37(103),pp.

71–95.

Popescu, S.C., Wynne, R.H., and Nelson, R.F. (2003) Measuring individual tree crown diameter with

LiDAR and assessing its influence on estimating forest volume and biomass. Canadian journal of

Remote Sensing, Vol. 29, No. 5, pp. 564–577

Popescu, S. C. (2007). Estimating biomass of individual pine trees using airborne lidar. Biomass &

Bioenergy, 31, 646–655.

Pouliot, D.A., D.J. King, F.W. Bell, and D.G. Pitt, 2002. Automated tree crown detection and

delineation in high-resolution digital camera imagery of coniferous forest regeneration, Remote

Sensing of Environment, 82:322–334

Pretzsch, H., 2009. Forest dynamics, growth and yield: From measurement to model. Springer. 664

pages.

146

Puissant, A., J. Hirsch, and C. Weber, 2005. The utility of texture analysis to improve per‐pixel

classification for high to very high spatial resolution imagery. International Journal of Remote

Sensing, 26(4): 733–745.

Ravindranath, N.H., Ostwald, M., 2008. Carbon inventory methods: handbook for greenhouse gas

inventory, carbon mitigation and roundwood production projects. Springer Verlag.

Reitberger, J., Krzystek, P., Stilla, U. 2006. Analysis of fill waveform LIDAR data for tree species

classification. Symposium of ISPRS commission III Photogrammetric Computer Vision PCV’06.

Bonn, Germany. 20-22 September 2006. IAPRS, XXXVI, Part 3: 228–233

Robertson, L.D. and D.J. King, 2011. Comparison of pixel-and object-based classification in land

cover change mapping. International Journal of Remote Sensing, 32(6): 1505–1529.

Röhle H., Huber W., 1985. Untersuchungen zur Methode der Ablotung von Kronenradien und der

Berechnung von Kronengrundflächen, Forstarchiv. 56, 238–243.

Roy, P., Ravan, S.A., 1996. Biomass estimation using satellite remote sensing data—an investigation

on possible approaches for natural forest. Journal of Biosciences 21, 535-561.

Ryherd, S., C. Woodcock, 1996. Combining spectral and texture data in the segmentation of

remotely sensed images. Photogrammetric Engineering and Remote Sensing, 62, 181–194.

Sanquetta, C.R., Dalla Corte, A.P., Silva. F.S., 2011a. Biomass expansion factor and root-to-shoot

ratio for Pinus in Brazil. Carbon Bal. Man. 6(6), 1–8.

Sanquetta, C.R., Dalla Corte, A.P., Jacon. A.D., 2011b. Crown area and trunk diameter relationship

for tree species at a mixed-araucaria natural forest in the mid-southern Parana State, Brazil. Floresta.

41 (1), 63–72.

Scurlock, J., Johnson, K., Olson, R., 2002. Estimating net primary productivity from grassland

biomass dynamics measurements. Global Change Biology 8, 736–753.

Shackelford, A.K., and C.H. Davis, 2003. A combined fuzzy pixel-based and object-based approach

for classification of high-resolution multispectral data over urban areas. IEEE Transactions on

Geoscience and Remote Sensing, 41(10): 2354–2363.

Shahtahmassebi, A., Yang N., Wang K., Moore, N. and Shen Z. (2013). Review of shadow detection

and de-shadowing methods in remote sensing. Chinese Geographical Science, 23(4): 403–420.

Shettigara, V.K., and Sumerling, G.M. (1998) Height determination of extended objects using

shadows in SPOT images. Photogrammetric Engineering and Remote Sensing, 64 (1): 35–44.

Smith, M.W., 2008. Relationship of trunk size to selected canopy size parameters for native pecan

trees. HortSci. 43(3), 784–786.

Smith, J.L. and J. J. Halvorson. 2011. Field Scale Studies on the spatial Variability of soil quality

indicators in Washington State, USA. Applied and Environmental Soil science, Volume 2011, Article

ID 198737, 7 pages, doi:10.1155/2011/198737(2011).

Snowdon, P., Raison, J., Keith, Heather., Ritson, P., Grierson, P., Adams, M., Montagu, K., Bi, Hui-

quan., Burrows, W., Eamus, D., 2002. Protocol for Sampling Tree and Stand Biomass. National

Carbon Accounting System, Technical Report No. 31 (Australian Greenhouse Office).

Sobrino, J., and N. Raissouni, 2000. Toward remote sensing methods for land cover dynamic

monitoring: application to Morocco. International Journal of Remote Sensing, 21(2): 353–366.

147

Soto-Pinto, L., Anzueto, M., Mendoza, J., Ferrer, G.J., de Jong, B., 2010. Carbon sequestration

through agroforestry in indigenous communities of Chiapas, Mexico. Agroforest Sys. 78, 39–51.

Song, C., Dickinson, M. B., Su, L., Zhang, S., and Yaussey, D. 2010. Estimating average tree crown

size using spatial information from Ikonos and QuickBird images: Across-sensor and across-site

comparisons. Remote Sensing of Environment 114 (2010) 1099–1107

Steininger, M., 2000. Satellite estimation of tropical secondary forest above-ground biomass: data

from Brazil and Bolivia. International Journal of Remote Sensing 21, 1139-1157.

Stromgaard, P., 1985. Biomass estimation equations for miombo woodland, Zambia. Agroforestry

systems 3, 3-13.

Sumida, A., Komiyama, A. 1997. Crown spread patterns for five deciduous broad-leaved woody

species: Ecological significance of the retention patterns of larger branches. Annals of Biol. 80, 759–

766.

Tadesse, W., Coleman, T.L. and T.D. Tsegaye (2003). Improvement of Landuse and Land-cover

Classification of an Urban Area Using Image Segmentation from Landsat ETM+ Data. Proceedings of

the 30th International Symposium on Remote Sensing Environment, November 10-14, 2003,

Honolulu, Hawaii.

Ter-Mikaelian, M.T., Korzukhin, M.D., 1997. Biomass equations for sixty-five North American tree

species. Forest Ecol. Man. 97, 1–24.

Thenkabail, P.S., Stucky, N., Griscom, B.W., Ashton, M.S., Diels, J., Van Der Meer, B., Enclona, E.,

2004. Biomass estimations and carbon stock calculations in the oil palm plantations of African

derived savannas using IKONOS data. International Journal of Remote Sensing 25, 5447-5472.

Thomas, N., C. Hendrix, and R.G. Congalton, 2003. A comparison of urban mapping methods using

high-resolution digital imagery. Photogrammetric Engineering and Remote Sensing, 69(9): 963– 972.

Tonolli, S., Dalponte, M., Vescovo, L., Rodeghiero, M., Bruzzone, L., Gianelle, D., 2011. Mapping

and modeling forest tree volume using forest inventory and airborne laser scanning. European

Journal of Forest Research, 1–9.

Todd, K.W., Csillag, F., and Atkinson, P.M., 2003. Three-dimensional mapping of light transmittance

and foliage distribution using lidar, Canadian Journal of Remote Sensing, Vol. 29, No. 5, pp. 544–

555.

Townshend J.R.G., Huang C., Kalluri S.N.V., Defries R.S., Liang S., Yang K. (2000) . Beware of Per-

pixel Characterization of Land Cover. International Journal of Remote Sensing, 2000, vol. 21, no. 4,

839–843

Tumbo, S. D., Salyani, M., Whitney, J. D., Wheaton ,T. A., Miller ,W. M. 2002. Investigation of

laser and ultrasonic ranging sensors for measurements of citrus canopy volume. Applied Engineering

in Agriculture Vol. 18(3): 367–372

Unger,D.R., Hung, I-Kuai, Brooks, R. & Williams, H. GIScience & Remote Sensing Volume 51, Issue

3, 2014.

van Ewijk, K. Y., Randin, C.F., Treitz, P.M., and Scott, N. A. 2014. Predicting fine-scale tree

species abundance patterns using biotic variables derived from LiDAR and high spatial resolution

imagery. Remote Sensing of Environment 150 (2014) 120–131

148

Verma, N.K., Lamb, D.W, Reid, N. and Wilson, B. (2014a) “An allometric model for estimating

DBH of isolated and clustered Eucalyptus trees from measurements of crown projection area”, Forest

Ecology and Management. 326: 125–132.

Verma, N.K., Lamb, D.W., Reid, N. and Wilson, B. (2013) “Tree cover extraction from 50 cm

worldview imagery: A comparison of image processing techniques”. IEEE International Geoscience

& Remote Sensing Symposium, IGARSS 2013, July 21-26, 2013, Melbourne, Australia. 192–195.

Verma, N., Lamb, D.W., Reid, N. and Wilson, B. (2014b) “A comparative study of land cover

classification techniques for ‘farmscapes’ using very high-resolution remotely sensed data”,

Photogrammetric Engineering and Remote Sensing. 80 (5), 461–470

Vieira ICG, de Almeida AS, Davidson EA, Stone TA, de Carvalho CJR, Guerrero JB. Classifying

successional forests using Landsat spectral properties and ecological characteristics in eastern

Amazonia.Remote Sensing of Environment. 2003;87:470–481.

Wang, L., W. Sousa, and P. Gong, 2004. Integration of object-based and pixel-based classification for

mapping mangroves with IKONOS imagery. International Journal of Remote Sensing, 25(24): 5655–

5668.

Wang, L., Gong, Peng, Biging, G.S.G., 2004. Individual tree-crown delineation and treetop detection

in high-spatial-resolution aerial imagery. Photogrammetric Engineering and Remote 3114, 351–358.

Walraven, R. (1978). Calculating the Position of the Sun. Solar Energy, 30: 393–397.

Weber, K.T., J. Théau, K. Serr, 2008. Effect of coregistration error on patchy target detection

using high-resolution imagery. Remote Sensing of Environment, 112, 845–850.

West, G. and Glasby, T. (2010). A new system for high resolution mapping of seagrass. Proc 19th

NSW Coastal Conference ("Coastal Management – all aboard, making it work!"), 10 – 12 November

2010, Batemans Bay, NSW, Australia.

Wheaton, T. A., J. D. Whitney, W. S. Castle, R. P. Muraro, H. W. Browning, and D. P. H.

Tucker.1995. Citrus scion and rootstock, topping height, and tree spacing affect tree size, yield, fruit

quality and economic return. J. Amer. Soc. Hort. Sci. 120(5): 861–870

White, J., 1981. The allometric interpretation of the self-thinning rule. J. Theoret. Biol. 89, 475–500.

Wilson, B. R., Growns,I. and Lemon, J., “Scattered native trees and soil patterns in grazing land on

the Northern Tablelands of New South Wales, Australia”, Soil Research, vol. 45, no. 3, pp. 199–205

Williams, R.J., Zerihun, A., Montagu, K.D., Hoffman, M., Hutley, L.B., and Chen, X. 2005.

Allometry for estimating aboveground tree biomass in tropical and subtropical eucalypt woodlands.

Australian Journal of Botany, 53, 607–619

Willhauck, G., 2000. Comparison of object oriented classification techniques and standard image

analysis for the use of change detection between SPOT multispectral satellite images and aerial

photos. International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3.

Amsterdam.

Wilson, B. R., 2002. Influence of scattered paddock trees on surface soil properties: a case study of

the Northern Tablelands of NSW. Ecol. Man. Restor. 3, 211–219.

Wong, T.h., Mansor, S.B., Mispan, M.R., Ahmad, N. and W.N.A. Sulaiman (2003). Feature

extraction based on object oriented analysis. Proceedings of ATC 2003 Conference, 20-21 May 2003,

Malaysia.

149

Woodcock, C.E. and Strahler, A.H. 1987. The factor of scale in Remote sensing. Remote Sensing of

Environment, 21, 311–332

Wulder, M. A., & Seemann, D. (2003). Forest inventory height update through the integration of lidar

data with segmented Landsat imagery. Canadian Journal of Remote Sensing, 29, 536−543.

Wulder, M.A., Hall, R.J., Coops, N.C., Franklin, S.E., 2004. High spatial resolution remotely sensed

data for ecosystem characterization. Bioscience 54, 511–521

Xiuwan, C., 2002. Using remote sensing and GIS to analyse land cover change and its impacts on

regional sustainable development. International Journal of Remote Sensing, 23(1): 107–124.

Yan, G., J.F. Mas, B.H.P. Maathuis, Z. Xiangmin, and P. M. Van Dijk, 2006. Comparison of

pixel‐based and object‐oriented image classification approaches—a case study in a coal fire area,

Wuda, Inner Mongolia, China. International Journal of Remote Sensing, 27(18): 4039–4055.

Yemefack, M., D.G. Rossiter, R. Njomgang, 2005. Multiscale characterization of soilvariability

within an agricultural landscape mosaic system in southern Cameroon. Geoderma, Volume 125, Issue

1-2: 117–143.

Yinghai Ke, Lindi J. Quackenbush ⁎, Jungho Im. Synergistic use of QuickBird multispectral imagery

and LIDAR data for object-based forest species classification. Remote Sensing of Environment 114

(2010) 1141–1154

Yu, Q., P. Gong, N. Clinton, G. Biging, M. Kelly, and D. Schirokauer, 2006. Object-based detailed

vegetation classification with airborne high spatial resolution remote sensing imagery.

Photogrammetric Engineering and Remote Sensing, 72(7): 799–811.

Zhang Q. and J. Wang (2003). A rule-based urban landuse inferring method for fine-resolution

multispectral imagery. Canadian Journal of Remote Sensing, Col. 29, No. 1 pp. 1–13.

Zheng, D., Rademacher, J., Chen, J., Crow, T., Bresee, M., Le Moine, J., Ryu, S.R., 2004. Estimating

aboveground biomass using Landsat 7 ETM+ data across a managed landscape in northern

Wisconsin, USA. Remote Sensing of Environment 93, 402-411.

Date post:	03-Oct-2020
Category:	Documents
Upload:	others
View:	9 times
Download:	0 times

Estimating trunk diameter at breast height for scattered ... · tree trunk diameter, more formally...

Documents