Date post: | 20-Apr-2023 |
Category: |
Documents |
Upload: | independent |
View: | 0 times |
Download: | 0 times |
THE UNIVERSITY OF ADELAIDE
Discriminating and mapping soil variability with hyperspectral reflectance data
David Summers
B. Ag. Sci. (Hons), The University of Adelaide
B.A. Flinders University of South Australia
Thesis presented for the degree of Doctorate of Philosophy
Faculty of Sciences, School of Earth and Environmental Sciences
July 2009
i
Abstract
The classification and mapping of soils and soil variability is important for a variety of
environmental and agricultural applications. Advances in precision agriculture, better
understanding of environmental processes and improvements in mathematical models used
to predict and understand landscape phenomena all require detailed information about soils
at increasingly finer scales. The goal of this thesis was to address this need for fine scale
soil information by developing new mapping methodologies from hyperspectral remote
sensing and reflectance spectroscopy. The spatially continuous and rich spectral
information of hyperspectral data provides a powerful diagnostic tool for mapping and
monitoring the earth’s surface materials. Similarly, reflectance spectroscopy allows for
rapid and cost effective measurement of materials based on their spectral response. These
two technologies offer the potential to record information about soils and provide fine
scale or continuous surface information for natural resource management.
The research aimed to explore the extent to which variation in surface horizon soils could
be discriminated and mapped with hyperspectral reflectance data. The study examined the
prediction of soil properties and classes with spectroscopic measurements, the mapping of
surface soils through interpolation from sample sites and the analysis of hyperspectral
imagery. The influence of vegetative cover and soil type on the identification of soil class
and quantification of soil exposure was investigated using simulated imagery. Each of the
research components focused on the soil properties and range of variation typically
encountered in southern Australian agricultural regions.
Reflectance spectroscopy was used to discriminate select field soil survey classes and to
predict and quantify various laboratory derived soil properties. For both of these analyses
visible near-infrared reflectance spectra (350 – 2500 nm) were collected with an ASD
FieldSpec Pro using a hand held probe. The spectral separability of the commonly used
field survey classes texture, carbonate and Munsell colour (separated into hue, value and
chroma) was assessed using penalised discriminant analysis. Only Munsell chroma was
adequately discriminated; while other classes showed some separability, it was limited and
not sufficient for soil classification. Failure to adequately classify the soil property classes
Abstract ii
was attributed to the subjective nature of the field survey methods, as well as co-variance
between soil properties.
Quantitative prediction of laboratory-measured soil properties (clay, organic carbon, iron
oxide and carbonate) from reflectance spectra was conducted using partial least squares
regression. Clay and carbonate contents were the best predicted, although predictions of
iron oxide and organic carbon were also acceptable. The utility of reflectance spectroscopy
to provide inputs for soil mapping was assessed by comparison of kriged surfaces of soil
properties. This comparison indicated that the methodology captured the same variability
in the landscape over the same range in values for each of the soil properties.
Prediction of soil exposure and type through vegetation cover was assessed with two types
of simulated imagery which were created using spectra of soil, photosynthetic and non-
photosynthetic vegetation. Both simulated images had the same, known combinations of
soil and vegetation but the relative mixes were created differently. Soil and vegetation
cover fractions were retrieved from the images through linear spectral unmixing and
compared with the measured fractions. Soils were accurately identified and classified in
both image types. However, not all soil spectra were isolated from mixed pixels equally or
successfully to provide accurate abundance fractions: some spectral mixes of soil and
vegetation were incorrectly classified as different soils, highlighting potential sources of
error in unmixing procedures.
The mapping of surface soils was assessed using image derived soil endmembers and
HyMap hyperspectral image data. Endmembers were isolated from the imagery using a
pixel purity process before being used in a partial unmixing routine. Field estimates of soil
exposure and laboratory analysis of soil samples were correlated with unmixing
abundances and used to characterise areas mapped by the different soil endmembers. Only
a moderate correlation between the field and image derived soil exposure was found.
Furthermore, soil properties for the different endmembers showed little difference between
classes and the mean of all samples. However, more than 70% of the areas mapped by the
four endmembers were unique, indicating that they were spatially distinct. These results
imply that the spectral response of soils captured by the hyperspectral imagery is more
strongly influenced by land management and soil properties other than those determined
through laboratory analysis.
Abstract iii
Reflectance spectroscopy of surface samples offers the potential to quickly and reliably
predict soil properties. Results indicate that it can be applied successfully to local
geographic areas and interpolated with geostatistics to create maps. The mapping of soils
with hyperspectral data presents problems that stem both from issues of plant material
obscuring the soil surface and high variability in soil reflectance due to management and
landscape processes. The unmixing of soils and vegetation (photosynthetic and non-
photosynthetic) from simulated imagery was successful but showed the potential for mixed
pixels to be confused for non-target soils. Similarly, landscape and management process
are subject to high variability and are not necessarily related to soil properties relevant to
agricultural and environmental applications. To fully utilise remote sensing for mapping
soils in a natural environment further research is required.
iv
Declaration
This work contains no material which has been accepted for the award of any other degree
or diploma in any university or other tertiary institution to David Summers and, to the best
of my knowledge and belief, contains no material previously published or written by
another person except where due reference has been made in the text.
I give consent to this copy of my thesis when deposited in the University Library, being
made available for loan and photocopying, subject to the provisions of the Copyright Act
1968.
The author acknowledges that copyright of published works contained within this thesis (as
listed below) resides with the copyright holders(s) of those works.
I also give permission for the digital version of my thesis to be made available on the web,
via the University’s digital research repository, the Library catalogue, the Australasian
Digital Theses Program (ADTP) and also through web search engines unless permission
has been granted by the University to restrict access for a period of time.
David Summers
July 2009
v
Publications arising from this thesis
Refereed publications
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2005 Spectral discrimination of soil properties, In SSC 2005 Spatial Intelligence, Innovation and Praxis:
The National Biennial Conference of the Spatial Sciences Institute, Melbourne, Australia, September, 2005.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2009 Mapping soil variability with hyperspectral image data, In SSC 2009 Spatial diversity: The Biennial
International Conference of the Surveying and Spatial Sciences Institute, Adelaide, Australia, September-October, 2009.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. (In Press) Visible near-infrared reflectance spectroscopy as a predictive indicator of soil properties, Ecological
Indicators.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. (In Review) Unmixing of soil types and estimation of soil exposure with simulated hyperspectral imagery, International Journal of Remote Sensing.
Non-refereed publications
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2006 Spectral determination of soil properties under vegetation, In 13th Australasian Remote Sensing and
Photogrammetry Conference, Canberra, Australia, 18-22 October.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2007 Identification of soil properties under vegetation using hyperspectral imagery, In EcoSummit 2007, Beijing, China, 22-27 May 2007.
Publications vi
Proportion of contribution by author
This is a declaration of the extent of each author’s contributions to the refereed papers
arising from this thesis. The extent of each of author’s contribution is quantified for
conceptualisation, realisation and documentation. Each author gives permission for the
paper containing their contribution to be included in this thesis.
Percent contribution and permission to include paper in this thesis:
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2005 Spectral discrimination of soil properties, In SSC 2005 Spatial Intelligence, Innovation and Praxis:
The National Biennial Conference of the Spatial Sciences Institute, Melbourne, Australia, September, 2005.
Conceptualisation Realisation Documentation Signature
Summers, D. 80% 90% 85% ______________
Lewis, M. 10% 5% 10% ______________
Ostendorf, B. 5% 2.5% 2.5% ______________
Chittleborough, D. 5% 2.5% 2.5% ______________
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2009 Mapping soil variability with hyperspectral image data, In SSC 2009 Spatial diversity: The Biennial
International Conference of the Surveying and Spatial Sciences Institute, Adelaide, Australia, September-October, 2009.
Conceptualisation Realisation Documentation Signature
Summers, D. 80% 90% 85% ______________
Lewis, M. 10% 5% 10% ______________
Ostendorf, B. 5% 2.5% 2.5% ______________
Chittleborough, D. 5% 2.5% 2.5% ______________
Publications vii
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. (In Press) Visible near-infrared reflectance spectroscopy as a predictive indicator of soil properties, Ecological
Indicators.
Conceptualisation Realisation Documentation Signature
Summers, D. 80% 90% 85% ______________
Lewis, M. 10% 5% 10% ______________
Ostendorf, B. 5% 2.5% 2.5% ______________
Chittleborough, D. 5% 2.5% 2.5% ______________
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. (In Review) Unmixing of soil types and estimation of soil exposure with simulated hyperspectral imagery, International Journal of Remote Sensing.
Conceptualisation Realisation Documentation Signature
Summers, D. 80% 90% 85% ______________
Lewis, M. 10% 5% 10% ______________
Ostendorf, B. 5% 2.5% 2.5% ______________
Chittleborough, D. 5% 2.5% 2.5% ______________
viii
Acknowledgements
I need to thank all of my family and friends for their help and patience over the years.
Thanks to my parents John Summers and Deborah McCulloch for all of their support,
encouragement and advice. And thanks to Andi Sebastian, similarly for support,
encouragement and advice but also her general enthusiasm for everything. Thanks also to
Ella Sebastian for always being my little sister.
Friends like Kerry Ireland, Mathew Rice, Simon Krieg, Jo-Anne Krieg, Chris Iley, Simone
Iley, Amanda Whitford, Jacob Habner, Amy Roberts, Matthew Slade, Claire Sherman and
Steve Safralidis, were all invaluable in one way or another to getting through.
I would also like to thank Kirsty Baldock who is my wonderful partner in all things and
has supported me in this journey with encouragement, patience and good humour.
Thanks go to all of my supervisors Megan Lewis, Bertram Ostendorf, David
Chittleborough and David Maschmedt who provided advice and direction in what was
sometimes a tortuous path. Special thanks go to Megan, who was always available for
advice and prompt with responses, providing intelligent and insightful feedback, and able
to see the whole picture and the detail with seeming ease.
The students, researches and professional staff of the Spatial Information Group and Soil
and Land Systems who over the years have provided advice, assistance, friendship,
humour and distraction. In no particular order they are; Ramesh Raja Segaran, Greg Lyle,
Neville Crossman, Kenneth Clarke, Patrick O’Connor, Tonja Wright, Paul Bierman,
Melissa Fraser, Kate Langdon, Sjaan Davey, Mohsen Forouzangohar, Sean Mahoney,
Dorothy Turner, David Mitchell, Claire Trelibs, Allana Grech, Reza Jafari, David Gerner,
David Hart, Rowena Morris, Serhiy Marchuk, Anna Dutkiewicz, Victoria Marshall,
Davina White, Adam Kilpatrick, Troy Willats, Tom Ellis, Collin Rivers, Debbie Miller,
Cameron Grant, Ron Smernik. Extra special thanks goes to those who joined me for
morning tea nearly everyday and the occasional game of hacky sack.
I would like to thank Sean Mahoney, Anna Dutkiewicz and Amanda Whitford for their
help in the field and with collecting and recording samples. Kerry Ireland and Amanda
Whitford also provided invaluable help in the laboratory taking spectroscopic
measurements and being generally very good friends.
Acknowledgements ix
Thanks to the people who provided technical advice in all its many forms; Debbie Miller,
Colin Rivers and Alla Marchuk from the University of Adelaide; and Richard Merry and
Les Janik, formally of CSIRO but now just hanging around and generally knowledgeable.
This research was conducted with joint funding from the Cooperative Research Centre for
Future Farm Industries (CRC FFI) and The University of Adelaide. Funding from the CRC
FFI was part the project ‘Development and application of high resolution spatial diagnostic
tools to aid in deployment of perennial systems at a catchment scale’. Funding from The
University of Adelaide was as part of a Faculty of Sciences Divisional Scholarship. Thanks
to the Cooperative Research Centre for Future Farm Industries and to The University of
Adelaide for their financial support and training and the excellent community they created.
Special thanks to Daryll Richardson for all his help in many forms, and to all of the
students of the CRC who provided friendship and advice over the years.
x
Table of Contents
Abstract ............................................................................................................................ i
Declaration ..................................................................................................................... iv
Publications arising from this thesis ............................................................................... v
Acknowledgements .......................................................................................................viii
Table of Contents ............................................................................................................ x
List of Figures............................................................................................................... xvi
List of Tables ................................................................................................................ xix
Chapter 1 ......................................................................................................................... 1
Understanding Soil Variability ....................................................................................... 1
1.1 Introduction ....................................................................................................... 1
1.2 Scope ................................................................................................................. 4
1.3 Thesis Structure ................................................................................................. 6
1.4 References ......................................................................................................... 7
Chapter 2 ......................................................................................................................... 9
Identifying and Evaluating Remote Sensing Techniques and Methodologies for
Mapping Soils .................................................................................................................. 9
1.1 Introduction ....................................................................................................... 9
1.2 Scope of Review ................................................................................................ 9
1.3 Soil Formation and Mapping............................................................................ 10
1.3.1 Soil Formation ......................................................................................... 10
1.3.2 Soil Mapping in Australia......................................................................... 10
Table of contents xi
1.3.3 Traditional Soil Mapping Methodology.................................................... 11
1.4 Improving Soil Mapping .................................................................................. 12
1.4.1 Pedotransfer functions.............................................................................. 12
1.4.2 Geostatistical analysis .............................................................................. 13
1.4.3 Continuous Classification......................................................................... 14
1.4.4 Digital Elevation Models and Topographic Indices................................... 14
1.5 Remote Sensing and Reflectance Spectroscopy................................................ 16
1.5.1 Spectral Characteristics of Soils ............................................................... 17
1.5.2 Soil Reflectance Spectra........................................................................... 18
1.5.3 Limitations of Optical Remote Sensing for Soil Mapping......................... 26
1.5.4 Vegetation Discrimination and Mapping .................................................. 26
1.6 Summary ......................................................................................................... 28
1.7 References ....................................................................................................... 29
Chapter 3 ....................................................................................................................... 34
Spectral Discrimination of Soil Properties ................................................................... 34
3.1 Introduction ..................................................................................................... 34
3.1.1 Spectral Variation in Soils ........................................................................ 35
3.2 Methods........................................................................................................... 36
3.2.1 Sample collection ..................................................................................... 36
3.2.2 Physical sample analysis .......................................................................... 36
3.2.3 Reflectance spectra collection................................................................... 37
Table of contents xii
3.2.4 Statistical analysis .................................................................................... 38
3.3 Results and Discussion..................................................................................... 39
3.4 Conclusion....................................................................................................... 47
3.5 References ....................................................................................................... 47
Chapter 4 ....................................................................................................................... 49
Visible near-infrared reflectance spectroscopy as a predictive indicator of soil
properties....................................................................................................................... 49
4.1 Introduction ..................................................................................................... 49
4.1.1 Spectral Reflectance Variation in Soils..................................................... 51
4.2 Methods........................................................................................................... 53
4.2.1 Study site and sample collection............................................................... 53
4.2.2 Laboratory soil analysis............................................................................ 55
4.2.3 Reflectance spectra................................................................................... 55
4.2.4 Statistical analysis .................................................................................... 55
4.2.5 Spatial Prediction ..................................................................................... 57
4.3 Results and Discussion..................................................................................... 57
4.3.1 Soil Properties.......................................................................................... 57
4.3.2 Soil Spectral Characteristics ..................................................................... 58
4.3.3 Prediction of Soil Properties ..................................................................... 60
4.3.4 Mapping of Predicted Soil Properties ....................................................... 64
4.4 Conclusion....................................................................................................... 65
4.5 References ....................................................................................................... 67
Table of contents xiii
Chapter 5 ....................................................................................................................... 70
Unmixing of soil types and estimation of soil exposure with simulated hyperspectral
imagery .......................................................................................................................... 70
5.1 Introduction ..................................................................................................... 70
5.2 Materials and methods ..................................................................................... 73
5.2.1 Soil and vegetation samples...................................................................... 73
5.2.2 Collection of spectra and image creation .................................................. 74
5.2.3 Spectral unmixing .................................................................................... 77
5.3 Results ............................................................................................................. 78
5.3.1 Spectral characteristics ............................................................................. 78
5.3.2 Mixes of spectra ....................................................................................... 80
5.3.3 Unmixing ................................................................................................. 82
5.4 Discussion ....................................................................................................... 87
5.4.1 Unmixing ................................................................................................. 87
5.4.2 Discrimination of soils ............................................................................. 87
5.4.3 Discrimination of soil and vegetation ....................................................... 88
5.4.4 Unmixing errors ....................................................................................... 89
5.4.5 Virtual versus laboratory images .............................................................. 90
5.5 Conclusions ..................................................................................................... 91
5.6 References ....................................................................................................... 92
Table of contents xiv
Chapter 6 ....................................................................................................................... 95
Mapping soil variability with hyperspectral image data ............................................. 95
6.1 Introduction ..................................................................................................... 95
6.2 Methodology.................................................................................................... 97
6.2.1 Study site characterisation and sample collection...................................... 97
6.2.2 Laboratory soil analysis............................................................................ 98
6.2.3 Image acquisition and pre-processing ....................................................... 98
6.2.4 Endmember selection and partial unmixing ............................................ 100
6.2.5 Validation .............................................................................................. 100
6.3 Results and Discussion................................................................................... 101
6.3.1 Endmembers .......................................................................................... 101
6.3.2 Soil mapping .......................................................................................... 103
6.3.3 Validation .............................................................................................. 104
6.4 Conclusion..................................................................................................... 109
6.5 References ..................................................................................................... 110
Chapter 7 ..................................................................................................................... 113
Discussion and Conclusion.......................................................................................... 113
7.1 Introduction ................................................................................................... 113
7.2 Summary of specific contributions to knowledge ........................................... 114
7.2.1 Spectral discrimination of soil properties (Chapter 3) ............................. 114
7.2.2 Visible near-infrared reflectance spectroscopy as a predictive indicator of
soil properties (Chapter 4)...................................................................................... 115
Table of contents xv
7.2.3 Unmixing of soil types and estimation of soil exposure with simulated
hyperspectral imagery (Chapter 5) ......................................................................... 116
7.2.4 Mapping soil variability with hyperspectral image data (Chapter 6)........ 117
7.2.5 Overall assessment of thesis topic........................................................... 117
7.3 General discussion: wider significance and limitations................................... 118
7.3.1 Spectral discrimination of soil properties (Chapter 3) ............................. 118
7.3.2 Visible near-infrared reflectance spectroscopy as a predictive indicator of
soil properties (Chapter 4)...................................................................................... 118
7.3.3 Unmixing of soil types and estimation of soil exposure with simulated
hyperspectral imagery (Chapter 5) ......................................................................... 119
7.3.4 Mapping soil variability with hyperspectral image data (Chapter 6)........ 119
7.4 Recommendations for future research ............................................................ 120
7.5 Conclusion..................................................................................................... 121
7.6 References ..................................................................................................... 121
xvi
List of Figures
Figure 2.1: Representative reflectance spectra of soils collected in the U.S.A. and Brazil.
Curves a-e explained in text below (Stoner and Baumgardner 1981)................................ 18
Figure 2.2: Example of reflectance spectra Kaolinite minerals showing absorption bands (~
2200 µm) characteristic of clay minerals (Clark 1999)..................................................... 23
Figure 2.3: Showing spectral features of hematite, green grass and dry grass (Fraser 1991).
........................................................................................................................................ 25
Figure 3.1: Jamestown study site, 200km north of Adelaide............................................. 37
Figure 3.2: Mean reflectance spectra for soil properties measured: field texture, soil
carbonate and soil colour; hue, value and chroma ............................................................ 40
Figure 3.3: Plots of spectral discrimination of field texture and field soil carbonate
measurements showing first and second discriminant variables ....................................... 43
Figure 3.4: Plots of spectral discrimination of components of Munsell soil colour; hue,
value and chroma, showing first and second discriminant variables ................................. 44
Figure 3.5: Discriminant loading plots for field texture, soil carbonate, hue, value and
chroma indicate regions of the spectra most significant in the discrimination analysis. V1
and V2 indicate the first and second discriminant variable respectively. .......................... 46
Figure 4.1: Jamestown study site, 200 km north of Adelaide, South Australia. Polygons
show Common Soils from the Land and Soil Spatial Data for southern South Australia
(Soil and Land Program 2007), soil sample sites marked with black dots. The legend
describes the soil Order from the Australia Soil Classification (in bold) (Isbell 2002) as
well as the soil description from the Land and Soil Spatial Data for southern South
Australia.......................................................................................................................... 54
Figure 4.2: Mean spectra of quartiles for percent clay...................................................... 58
Figure 4.3: Mean spectra of quartiles for soil organic carbon. .......................................... 59
Figure 4.4: Mean spectra of quartiles for carbonate concentration.................................... 60
List of figures xvii
Figure 4.5: Mean spectra of quartiles for iron oxide content............................................. 60
Figure 4.6: Spectral loading weight graph for the prediction of clay content. ................... 62
Figure 4.7: Spectral loading weight graph for the prediction of soil organic carbon content.
........................................................................................................................................ 62
Figure 4.8: Spectral loading weight graph for the prediction of carbonate content............ 63
Figure 4.9: Spectral loading weight graph for the prediction of iron oxide content. .......... 63
Figure 4.10: Spatial distribution of measured and predicted soil properties following
Kriging............................................................................................................................ 64
Figure 5.1: (a) Demonstrates the configuration of the ASD high intensity reflectance probe
held in a clamp over the tray containing soil and leaves. (b) Demonstrates the incremental
movement of probe field of view over plant and soil interface. The solid lines indicate soil
where pure soil and vegetation spectra were collected. The dashed lines indicate the 10%
increments as the probe was moved. Not to scale............................................................. 76
Figure 5.2: The ‘laboratory image’ created from the measured spectra. Soil type is listed at
the bottom, vegetation cover type at the top and percent soil exposure on the left. ........... 77
Figure 5.3: Pure soil spectra (endmembers) from soils used in this experiment. ............... 79
Figure 5.4: Pure vegetation spectra (endmembers) from soils used in this experiment...... 80
Figure 5.5: Spectra collected from actual mixes of Sodic Clay and photosynthetic
Eucalyptus vegetation. Mixes of 0%, 20%, 40%, 60%, 80% and 100% soil are shown for
clarity. ............................................................................................................................. 81
Figure 5.6: Spectra collected from actual mixes of Sodic Clay and the non-photosynthetic
field pea. Mixes of 0%, 20%, 40%, 60%, 80% and 100% soil are shown for clarity......... 81
Figure 5.7: RMSE from ‘virtual’ (a: Eucalyptus, b: orange, c: pea straw) and ‘laboratory’
(d: Eucalyptus, e: orange, f: pea straw) images for each of the vegetation cover types. .... 82
List of figures xviii
Figure 5.8: Unmixing with the Sodic Clay endmember of the ‘virtual’ (a: Eucalyptus, b:
orange, c: pea straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for
each of the vegetation cover types. .................................................................................. 83
Figure 5.9: Unmixing with the Silty Loam endmember of the ‘virtual’ (a: Eucalyptus, b:
orange, c: pea straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for
each of the vegetation cover types. .................................................................................. 84
Figure 5.10: Unmixing with the Loam endmember of the ‘virtual’ (a: Eucalyptus, b:
orange, c: pea straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for
each of the vegetation cover types. .................................................................................. 85
Figure 5.11: Unmixing with the Clay Loam endmember of the ‘virtual’ (a: Eucalyptus, b:
orange, c: pea straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for
each of the vegetation cover types. .................................................................................. 86
Figure 6.1: HyMap image strip in true colour (bands 660.4, 557.9 and 468.9 nm)
superimposed on Landsat 7 panchromatic band of Jamestown-Belalie district. Ranges are
marked with arrows. ........................................................................................................ 99
Figure 6.2: Soil endmembers (EM 1, EM 2, EM 3 and EM 4) used in the Matched Filtering
analysis. Actual reflectance spectra are on the left and continuum removed spectra are on
the right......................................................................................................................... 102
Figure 6.3: Partial unmixing results of the four soil endmembers (EM 1, EM 2, EM 3 and
EM 4) isolated from the image. Bright areas indicate a high match with endmembers and
dark areas indicate a poor match.................................................................................... 104
Figure 6.4: Soil maps produced by the application of thresholds to partial unmixing
outputs. ......................................................................................................................... 105
xix
List of Tables
Table 3.1: Error matrices for five discriminant analysis made in this study. Texture: SCL =
Sandy Clay Loam, CL = Clay Loam, LMC = Light Medium Clay, MC = Medium Clay.
CO3: N = Nil, S = Slight, M = Moderate, H = High. ........................................................ 42
Table 3.2: First (V1), second (V2) and third (V3) variates from the analysis and attribution
error derived from the error matrices. .............................................................................. 45
Table 4.1: Summary of laboratory results from chemical and physical analysis. .............. 57
Table 4.2: Sample numbers, residual predictive deviation (RPD), root mean square error
(RMSE) and R2 results for data sets................................................................................. 61
Table 5.1: Laboratory measured soil properties of four soils used in the study. ................ 74
Table 6.1: Coefficient of determination (r2) for the relationship between field estimated soil
exposure and image derived soil exposure for each endmember..................................... 106
Table 6.2: Average soil laboratory results for the total soil samples and the sites
corresponding to each soil endmember. ......................................................................... 107
1
Chapter 1
Understanding Soil Variability
1.1 Introduction
In recent decades there has been a greater awareness of the need to better understand soil
variability. The impact of land degradation and the falling price of many agricultural
commodities have placed increasing pressure on policy makers and land managers to
improve management around Australia and the world (John et al. 2005, Kingwell and
Pannell 2005). Increasingly producers are attempting to improve productivity to maintain
profit margins while arable land becomes degraded from processes such as erosion, salinity
and acidity (Passioura 2002, John et al. 2005, Rengasamy 2006).
In order to remain profitable, farmers are turning to new technologies, such as precision
agriculture, to more efficiently manage assets and allocate resources (Passioura 2002). The
aim of precision agriculture is to refine management decision making through improved
understanding of spatial variables such as yield and soil properties (Bongiovanni and
Lowenberg-Deboer 2004, McBratney et al. 2005). While the current understanding of soil
variability is very advanced at a regional scale (≥ 1:50 000) there is much scope to improve
our understanding at finer scales. As an input for precision agriculture, accurate and
detailed information about soil variability at a farm scale is required. With farm scale soil
maps, farmers will be better able to relate yield variability to changing soil properties.
Improved understand of soil variability is also useful to help mitigate land degradation. It
is difficult and complex allocating resources to manage land degradation problems such as
dryland salinity, soil loss, soil acidity, water quality and biodiversity loss. A great deal of
information is required to understand the processes taking place. This includes
groundwater recharge, surface water flows, river salinity, water nutrient loadings, the
impact of loss of biodiversity and the cost of implementing land management change
(Beverly et al. 2003). Adding further to this complexity is the temporal and spatial
discontinuity between the implementation of management strategies and observed
outcomes. Impacts of these strategies often manifest themselves many kilometres from the
Chapter 1: Introduction 2
implementation site and are often only seen after years or even decades. One way to
mitigate the difficulty in linking the cause and effect is to estimate the likely outcomes of
landscape intervention from mathematical models of landscape or biological processes.
Models that would benefit from improved soil maps are those related to plant growth and
soil processes such as hydrology. Plant growth models estimate the suitability and growth
of plants under different conditions. They require inputs relating to the climate (e.g. rainfall
and temperature), soil (e.g. texture, depth, slope and available water holding capacity) and
the plants themselves. Information about the plant largely relates to how they interact with
the environment, for example, the pH range a plant can tolerate or the degree of aeration/
water-logging preferred by the plant (Hackett and Harris 1990). By contrast, soil hydrology
models involve the estimation of the movement of water through the soil. Thus they are
most concerned with the soil texture and structure as it affects water movement due to
saturated and unsaturated flow (Hatton 1998, Beverly and Croton 2001).
Early models were developed for use with point data, but with improving information
technology they are been being applied to regions and landscapes. As a result spatially
distributed input data on a range of variables including soil is required. Spatial data on soils
originally came from existing regional soil maps. While a unique and informative resource
for regional land managers and planners, the regional soil maps, with scales of 1:250 000
to 1:50 000, have substantial limitations when applied to finer local and farm scales of less
than 1:20 000 (Maschmedt 2000). The broad scale regional maps do not have the detail
required to portray within-paddock variability and thus inform decision making at the farm
level, or provide a suitable input for fine scale catchment modelling. To effectively map
these local land changes at a usable scale is difficult and expensive by conventional soil
mapping methodologies. Furthermore, the polygon-based unit representation of discrete
boundaries delineating homogeneous areas is not a realistic representation of the
continuous variability found in soils. While variation within a polygon may be
acknowledged by descriptors or attributes assigned to the mapping unit, that variation is
not spatially located.
A further limitation of traditional soil mapping is the reliance on laboratory analysis of
samples to quantify soil properties. These methods are generally time consuming and
expensive, requiring many consecutive steps and often involve toxic and corrosive
reagents. What is more, soils are not homogenous and mechanisms and interactions within
Chapter 1: Introduction 3
the soil matrix are difficult to understand. Conventional laboratory techniques do not
account for this complexity but instead rely on physical and chemical relationships
between limited components to explain observed interactions (Viscarra Rossel et al. 2006).
Consequently new methods such as mass spectroscopy, X-Ray diffraction, nuclear
magnetic resonance, and visible-near infrared and mid infrared spectroscopy are being
used to analyse soil composition. These methods are typically rapid and repeatable,
reducing the need for extractions and allowing for the analysis of the solid soil matrix
(Janik et al. 1998).
One approach widely employed to avoid expensive and time consuming laboratory
analysis is the use of soil field survey protocols (McDonald and Isbell 1990). Soil field
survey aims to derive as much information about soil as possible from a series of simple
protocols applicable in the field without the need for laboratory or ongoing analysis. These
methodologies are applied extensively in Australia around high value and irrigated
agriculture to determine soil properties and maximise irrigation efficiency. However,
despite the speed and relative affordability of soil field survey, it does suffer from
substantial limitations. Soil field survey is prohibitively expensive for all but the most
intensive land uses; broad acre agriculture does not generally provide sufficient returns to
make such expenditure affordable. Furthermore, most of the techniques used in soil field
survey are subjective, requiring extensive training to achieve acceptable accuracy.
Visible-near infrared and mid infrared spectroscopy (from here summarised as reflectance
spectroscopy) are particularly appealing because they are quick, require almost no sample
preparation and are relatively inexpensive. Most impressively some researchers claim to
achieve more accurate results with reflectance spectroscopy than with traditional
laboratory methods (Viscarra Rossel et al. 2006). Thus reflectance spectroscopy has the
potential to improve the speed and perhaps the accuracy of soil sampling. Such advances in
soil analysis could substantially improve the density of sampling and improve the spatial
resolution of mapping without prohibitive increases in cost. However, such techniques still
generally lend themselves to polygon-based unit representation, albeit facilitating
improved spatial resolution.
Remote sensing from satellites or aircraft is a technology that relies on similar principles as
reflectance spectroscopy, measuring light reflected from materials, but it provides these
measurements over a spatially continuous area in the form of images. Source materials of
Chapter 1: Introduction 4
the spectral response can be identified and the relative abundance of materials can be
mapped. Opposed to traditional mapping and monitoring methods that rely on point data
sources from which to project the properties of whole landscapes, remote sensing provides
information over the whole landscape with a ground resolution down to 2 or 3 metres.
Studies have shown this form of earth observation to be useful in mapping and monitoring
many surface features from geology and minerals, vegetation and ecology to soils and soil
properties. Soil mapping with remote sensing has been carried out largely in the northern
hemisphere (Drake et al. 1999). Generally these studies have examined one or two
particular soil properties (Galvao et al. 2001, Chabrillat et al. 2002) although there are
some exceptions to this where many soil properties have been examined simultaneously
(Ben-Dor et al. 2002). Techniques used in these studies to extract thematic information
from the imagery include spectral matching, mixture modelling (Drake et al. 1999), band
ratios (Ryan and Lewis 2000) and multivariate statistical classification (Palacios-Orueta
and Ustin 1996). Additionally, both in Australia and overseas there have been studies
aimed at mapping the expression of degradation such as salinity, mapping salt affected
soils and the indicator vegetation types (Sharma and Bhargava 1988, Hick and Russell
1990, Dutkiewicz et al. 2003). The most significant addition that remote sensing brings to
these applications is the spatial continuity of the data as opposed to the interpolation of
point data of traditional mapping and monitoring.
1.2 Scope
The research presented in this thesis addresses the need for improved information on soil
variation at scales appropriate for precision agriculture and landscape process modelling. It
assesses the potential for prediction of soil properties with visible-near infrared reflectance
spectroscopy and examines some of the limitations to quantification of soil properties
under plant cover. Spectroscopic analysis of soil samples is used to inform regional
mapping of surface soils with hyperspectral imagery. The research comprises four
components addressing these areas.
The first study aimed to discriminate samples into soil field survey classes from spectral
response curves measured under laboratory conditions. Field survey analysis is a common
method used for characterising soil samples to map soil properties. Texture and colour are
measured in almost all instances and carbonate is measured in most southern Australian
Chapter 1: Introduction 5
environments where it is prevalent in soil profiles. Despite protocols for these methods,
there is still some subjectivity and variation in measurement and although faster than lab
analyses can still be time consuming. Spectral analysis of these properties could provide a
new objective, rapid technique to assist soil survey. Furthermore, establishing a
methodology that can predict field survey classes would provide continuity between
spectroscopy techniques and traditional field survey.
A second laboratory study aimed to predict quantitative soil properties in order to
overcome the subjectivity inherent in the soil field survey. Rapid and relatively
inexpensive determination of soil properties through reflectance spectroscopy could
improve the resolution of existing maps and provide important inputs for modeling and
precision agriculture. The soil properties predicted from spectral response curves were clay
content, organic carbon content, iron oxide content and carbonate content. These were
chosen due to their requirement as inputs for current hydrological models and because of
their general importance in determining agricultural fertility.
The influence of photosynthetic and non-photosynthetic plant material on the detection and
quantification of soil types was examined in a third study. A pilot study into image analysis
over South Australia’s northern agricultural districts found direct sensing of soil properties
difficult due to crop residue obscuring the earth’s surface. As a result of that finding this
subsequent study aimed to determine realistic thresholds for the spectral determination of
soil type and soil exposure using imagery simulated from laboratory measured reflectance.
The simulated imagery allowed for specifically quantified abundance ratios of different
soil and cover materials. The land surface in agricultural districts in southern Australia is
typically obscured from imaging sensors by photosynthetic vegetation or crop residue. For
image remote sensing of soils to inform future soil mapping programs, the spectral
interaction of soil and plant cover must be understood.
The final study used airborne hyperspectral imagery in an attempt to map surface soils in a
broad acre agricultural district. This study aimed to examine the ability of hyperspectral
remote sensing to map soil variability and discriminate different soil types. Partial spectral
unmixing and image derived endmembers were used to minimise a priori knowledge and
examine the possibility of using hyperspectral imagery to inform subsequent soil sampling
and survey.
Chapter 1: Introduction 6
All of these studies were carried out using imagery and soils from two agricultural districts
of South Australia; Monarto, 50 km east of Adelaide and Jamestown, 200 km north of
Adelaide.
1.3 Thesis Structure
The thesis is structured with 8 chapters. This introductory chapter (Chapter 1) provides a
brief overview of the motivation behind the research and outlines the unifying research
theme. Chapter 2 provides a detailed review of remote sensing, reflectance spectroscopy
and soil mapping literature. The review examines the literature that was available until the
beginning of the research phase of the study. Specific knowledge gaps relating to each
component of the research are addressed in subsequent research chapters (Chapters 3-6)
and each of these chapters contains more recent literature relevant to their specific
objectives. Chapter 3 examines the discrimination of field survey soil classes using
laboratory collected reflectance spectra. This chapter has been peer reviewed and accepted
for publication in the proceedings of SSC 2005 Spatial Intelligence, Innovation and Praxis
Conference (Summers et al. 2005). Chapter 4 focuses on the prediction of quantitative soil
properties from laboratory-collected reflectance data. This Chapter has been peer reviewed
and accepted for publication in Ecological Indicators (Summers et al. In Press). Chapter 5
evaluates the spectral unmixing of soil and vegetation (photosynthetic and non-
photosynthetic) using simulated laboratory imagery. This chapter is currently in review for
publication in the International Journal of Remote Sensing (Summers et al. In Review).
Chapter 6 examines the unmixing and mapping of soils using HyMap hyperspectral
imagery in South Australia’s northern agricultural districts. This chapter has also been peer
reviewed and accepted for publication in the proceedings of SSC 2009 Spatial Diversity
(Summers et al. 2009). Therefore, the thesis is presented with Chapter 3, Chapter 4,
Chapter 5 and Chapter 6 as standalone articles for publication. Although they have been
reformatted to match the rest of the thesis, the content is unchanged from the submitted
articles. This style of presentation necessarily results in some areas of repetition,
particularly in the introductions, methods and reference lists. The discussion (Chapter 7)
provides an overview of the research findings under the unifying thesis topic and outlines
the acquired knowledge. Furthermore, it will discuss the limitation and significance of the
findings and outline to future research that arises from these findings.
Chapter 1: Introduction 7
1.4 References
Ben-Dor, E., Patkin, K., Banin, A. and Karnieli, A. 2002 Mapping of several soil properties using DAIS-7915 hyperspectral scanner data - a case study over clayey soils in Israel, International Journal of Remote
Sensing, 23, 1043-1062.
Beverly, C., Avery, A., Ridley, A. and Littleboy, M. 2003 Linking farm management with catchment response in modelling framework, In 11th Australian Agronomy Conference, Geelong,
Beverly, C. and Croton, J. T. 2001 Formulation and application of the unsaturated/saturated catchment models SUSCAT and WEC-C, Hydrological Processes, 15.
Bongiovanni, R. and Lowenberg-Deboer, J. 2004 Precision agriculture and sustainability, Precision
Agriculture, 5, 359-387.
Chabrillat, S., Goetz, A. F. H., Krosley, L. and Olsen, H. W. 2002 Use of hyperspectral images in the identification and mapping of expansive clay soils and the role of spatial resolution, Remote Sensing of
Environment, 82, 431-445.
Drake, N. A., Mackin, S. and Settle, J. J. 1999 Mapping vegetation, soils, and geology in semiarid shrublands using spectral matching and mixture modeling of SWIR AVIRIS imagery, Remote Sensing of Environment, 68, 12-25.
Dutkiewicz, A., Lewis, M. and Ostendorf, B. 2003 Evaluation of hyperspectral imagery for mapping the symptoms of dryland salinity, In Spatial Sciences Coalition 2003, Canberra,
Galvao, L. S., Pizarro, M. A. and Epiphanio, J. C. N. 2001 Variations in reflectance of tropical soils: Spectral-chemical composition relationships from AVIRIS data, Remote Sensing of Environment, 75, 245-255.
Hackett, C. and Harris, G. 1990 PLANTGRO: A software package for the prediction of plant growth, Griffith University, Melbourne.
Hatton, T. 1998 Catchment scale recharge modeling, In The basics of recharge and discharge (Ed, L. Zhang) CSIRO Publishing, Melbourne.
Hick, P. T. and Russell, W. G. R. 1990 Some spectral considerations for remote sensing of soil salinity, Australian Journal of Remote Sensing, 28, 417-431.
Janik, L. J., Merry, R. H. and Skjemstad, J. O. 1998 Can mid infrared diffuse reflectance analysis replace soil extractions?, Australian Journal of Experimental Agriculture, 38, 681-696.
John, M., Pannell, D. and Kingwell, R. 2005 Climate change and the economics of farm management in the face of land degradation: Dryland salinity in western Australia, Canadian Journal of Agricultural Economics, 53, 443-459.
Kingwell, R. and Pannell, D. 2005 Economic trends and drivers affecting the Wheatbelt of western Australia to 2030, Australian Journal of Agricultural Research, 56, 553-561.
Maschmedt, D. 2000 Assessing agricultural land: Agricultural land classification standards used in South Australia's land resource mapping program, Primary Industries and Resources South Australia, Adelaide,
McBratney, A., Whelan, B. M., Ancev, T. and Bouma, J. 2005 Future directions of precision agriculture, Precision Agriculture, 6, 7-23.
McDonald, R. C. and Isbell, R. F. 1990 Soil profile, In Australian soil and land survey: Field handbook (Eds, R. C. McDonald, R. F. Isbell, J. G. Speight, J. Walker and M. S. Hopkins) Inkata Press, Melbourne.
Palacios-Orueta, A. and Ustin, S. L. 1996 Multivariate statistical classification of soil spectra, Remote
Sensing of Environment, 57, 108-118.
Passioura, J. B. 2002 Environmental biology and crop improvement, Functional Plant Biology, 29, 537-546.
Rengasamy, P. 2006 World salinization with emphasis on Australia, Journal of Experimental Botany, 57, 1017-1023.
Ryan, S. and Lewis, M. 2000 Discrimination and mapping soils using HyMap hyperspectral imagery, Barossa valley, S.A., In 10th Australasian Remote Sensing and Photogrammetry Conference, Adelaide,
Chapter 1: Introduction 8
Sharma, R. C. and Bhargava, G. P. 1988 Landsat imagery for mapping saline soils and wet lands in north-west India, International Journal of Remote Sensing, 9, 39-44.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2005 Spectral discrimination of soil properties, In SSC 2005 Spatial Intelligence, Innovation and Praxis: The National Biennial Conference of the Spatial Sciences Institute, Melbourne, Australia, September, 2005.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2009 Mapping soil variability with hyperspectral image data, In SSC 2009 Spatial diversity: The Biennial International Conference of the Surveying and Spatial Sciences Institute, Adelaide, Australia, September-October, 2009.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. In Press Visible near-infrared reflectance spectroscopy as a predictive indicator of soil properties, Ecological Indicators.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. In Review Unmixing of soil types and estimation of soil exposure with simulated hyperspectral imagery, International Journal of Remote Sensing.
Viscarra Rossel, R. A., Walvoort, D. J. J., McBratney, A. B., Janik, L. J. and Skjemstad, J. O. 2006 Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties, Geoderma, 131, 59-75.
9
Chapter 2
Identifying and Evaluating Remote Sensing Techniques
and Methodologies for Mapping Soils
2.1 Introduction
This project investigates contemporary remote sensing and reflectance spectroscopy
technologies and how they can be used to effectively map soils at a resolution that provides
useful property-scale land management tools. The project was established to investigate the
possibilities provided by these new technologies to overcome some of the expense and
limitations of conventional soil mapping techniques.
2.2 Scope of Review
This review briefly examines the theory behind soil formation and current regional scale (≥
1:50 000) soil mapping methods to provide an understanding of what is available, the
benefits arising from current methodologies and current databases available, but also to
detail the relative shortcomings. The review details some of the methodologies used more
broadly in research and general soil mapping such as pedometrics and geostatistics and
explains how these fit into the broader context of understanding soil process and mapping
procedures. It also examines digital terrain data and highlights previous studies where the
different methods have been incorporated. The review then examines the use of remote
sensing and reflectance spectroscopy and discusses how these technologies have been used
for land monitoring and soil attribute mapping.
Chapter 2: Literature review 10
2.3 Soil Formation and Mapping
1.1.1 Soil Formation
Soil formation is generally attributed to five soil forming factors. These are parent
material, climate, topography, biological processes and time (Jenny 1941). These factors
all combine to effect soil composition.
The interaction of these factors was expressed by Jenny (1941) in the following equation:
S = ƒ (Cl, o, r, p, t)
Where S is soil, Cl is climate, o is organisms (including humans), r is topography, p is
parent material and t is time. This equation defines a relationship between landform
processes and soil formation and their resulting properties.
Since its inception this equation has been considered a qualitative approach to
understanding soil formation and many surveyors have used it as such. These surveyors
use it as part of their expert knowledge in understanding the factors that are important in
producing soil pattern (McBratney et al. 2003). Other researches have taken quantitative
approaches to the equation by trying to formalise it. These approaches generally involve
analysis where all but one function is kept constant and as such quantitative climofunctions
and topofunctions have been developed; however their use in soil mapping is limited
(McBratney et al. 2003). Nonetheless, these factors play an important role in soil
variability and as such should be acknowledged.
1.1.2 Soil Mapping in Australia
A number of major soil mapping programs have taken place in Australia. Many of these
were undertaken by the Commonwealth Scientific and Industrial Research Organisation
(CSIRO) while others were conducted by individual states. Early national maps include the
Atlas of Australian Soils prepared at 1:3,000,000 scale (Northcote et al. 1968). Although
some local scale mapping is dated as far back as the 1920s, modern techniques were not
applied until the 1940s (Taylor 1970). However, these surveys were generally broad scale
and driven by local catchment and rural area planning strategies. Still today the extent of
systematic soil mapping in some of Australia’s agricultural districts is limited. For
example, only 50% of soils in the Murray Darling Basin, Australia’s most important
Chapter 2: Literature review 11
agricultural area covering approximately 1,000,000 km2, are mapped at 1:250,000 and 3%
at 1:100,000 (McBratney et al. 2003). However, in Western Australia and South Australia
substantive efforts have been made since the 1980s to provide seamless mapping of the
agricultural districts across the states.
In Western Australia this mapping covers the south west agricultural districts and
combines surveys at various scales. A methodology has been developed to provide a nested
hierarchy of soil-landuse mapping units that allows for pre-existing and recent surveys to
be included into a seamless mapping database (Schoknecht et al. 2004). Thus, the resultant
database includes maps at various scales from 1:20,000 – 1:250,000, however, the vast
majority is at scales no smaller than 1:100 000. In South Australia soil mapping was
carried out at scales of 1:100 000 and 1:50 000 depending on the agricultural district. Soil
landscape units were developed to provide a means of determining the suitability of land
for different uses (Maschmedt 2000). The South Australian and Western Australian
databases provide a substantive and informative regional scale land assessment tool
(Maschmedt 2000, Schoknecht et al. 2004). However, the scale of most of the mapping
does not account for property scale soil variability that can have a significant effect on land
management decisions.
1.1.3 Traditional Soil Mapping Methodology
The traditional methodology used to make soil landscape maps involves expert knowledge
and significant expense in soil sampling and analysis. Aerial photographs are examined by
experts who delineate polygons of what appear to be different soils and combinations of
soil visible on the photos or inferred from position in the landscape. Soil surveyors then
collect soil samples from representative areas for each mapping unit to characterise the
soils within each polygon. The polygons are then assigned to soil classes based on the
composition of this analysis and depicted as such on seamless landscape unit maps (Gunn
et al. 1988, Schoknecht et al. 2004). While this is a popular and effective methodology to
map soil for various applications, there are limitations associated with it.
Continuous Variability and Polygons
A major limitation of traditional soil mapping is its dependence on polygon-based unit
representation. The continuous variability of soils in the landscape is portrayed by
homogeneous polygons with discrete boundaries. This results in class assignment
Chapter 2: Literature review 12
generalisation which involves grouping suites of soils to single mapping units using crisp
logic (Zhu 1997). While variation within each polygon may be accounted for through
descriptors or attributes assigned to the mapping unit, that variation is not spatially located
(Zhu 1997). This is rarely a realistic representation of soils. Variation in soils is
continuous, more often demonstrating a diffuse contrast from one soil to another rather
than an abrupt change.
Scale
While traditional regional scale soil maps provide useful and reliable information for some
purposes their low resolution does not account for property-scale soil variability that can
have a significant effect on land management decisions. Only soil attributes or objects
larger than a certain size (called the ‘minimum mapping size’) can be represented on maps
at a given scale. As a result, areas smaller than this minimum mapping size are either
incorporated into surrounding soil objects or entirely omitted: this is known as spatial
generalisation (Zhu 1997). Thus the resolution of traditional soil maps may be a limiting
factor when incorporated into environmental models using other ‘fine resolution’
environmental data (Zhu 1997). To effectively map local land changes at a usable sub-
catchment scale is difficult and prohibitively expensive by conventional soil mapping
methodologies. However, most data obtained from digital terrain analyses and remote
sensing provide for discrimination of areas less than one hectare and are thus capable of
describing small areas of the environment.
2.4 Improving Soil Mapping
Due to the expense and time consuming nature of traditional soil survey, recent decades
have seen the development of new methods to extend soil property prediction from
relatively sparse traditional data sets using secondary information (Bishop and McBratney
2001, McBratney et al. 2002). These methods include pedotransfer functions, geostatistics
and continuous classifications (McBratney et al. 2003).
1.1.4 Pedotransfer functions
Pedotransfer function (PTF) is a generic term for a soil prediction method that uses some
known soil property or properties to estimate another unknown property. PTFs came about
through a desire to predict difficult and expensive to measure soil properties from more
easily measurable, surrogate properties (Minasny et al. 1999, McBratney et al. 2002).
Chapter 2: Literature review 13
There are many soil properties that are prohibitive to measure over large areas and
especially at fine scales. PTFs are a method by which to estimate these properties from
more available data sets. The properties predicted from PTFs can be used in further
modelling at a field and regional scale (Mayr and Jarvis 1999, McBratney et al. 2002).
Most commonly PTFs are used to predict soil hydraulic properties, although this is not
their exclusive use (McBratney et al. 2002). Many studies have been undertaken in the
estimation of soil water retention curves based on properties such as texture, bulk density
and organic carbon (Mayr and Jarvis 1999, Minasny et al. 1999, McBratney et al. 2002).
Other uses include the estimation of pesticide leaching with relation to regional water flow
(Petach et al. 1991, Soutter and Pannatier 1996), modelling heavy metal movement and
accumulation (Tiktak et al. 1999) and yield estimation (Haskett et al. 1995, Timlin et al.
1996). However, even measuring surrogate properties for PTFs requires field sampling and
laboratory analysis. Therefore to improve the resolution of soil maps through the use of
pedotransfer functions will likely require increased sampling density and laboratory
analysis of discrete samples, all of which increases the cost of mapping at finer scales.
1.1.5 Geostatistical analysis
Geostatistical analysis has been used to aid in soil attribute prediction to improve soil
mapping (McBratney et al. 2003). Traditionally geostatistics provide a means to explain
variability between sample points in soil survey but it also offers a measure of uncertainty
in soil maps that is becoming increasingly important (Bishop and McBratney 2001, Bishop
et al. 2001). Geostatistical methods include kriging, co-kriging and regression kriging.
Kriging is a univariate approach to soil prediction that improves significantly on results
obtained by traditional methods such as multiple linear regression, but limits the inclusion
of other data sets such as remotely sensed data (Bishop and McBratney 2001). Co-kriging
on the other hand is a multivariate approach that allows the inclusion of ancillary variables
correlated with the primary data sets (McBratney et al. 2003). While initially these
ancillary data sets were other soil variables (McBratney et al. 2003), later studies
incorporated crop yield, terrain data and satellite remote sensing imagery (Bhatti et al.
1991, Ishida and Ando 1999, Bishop and McBratney 2001). Regression kriging on the
other hand involves kriging of the residuals of regression models such as multiple linear
regression or regression tree regression (Bishop and McBratney 2001). While geostatistics
offer significant tools for soil prediction between data points they generally allow only a
Chapter 2: Literature review 14
‘crisp’ allocation of membership to any one class. That is to say, a given data point on a
soil map can only belong to one class (Burrough et al. 1997). However, there are some
methods such as indicator kriging or stochastic simulation that allow the inclusion of
categorical data (Goovaerts 1997).
1.1.6 Continuous Classification
Continuous classification or fuzzy logic was developed out of the acknowledgement that
attributes in the landscape vary continuously across space (Burrough et al. 1997,
McBratney et al. 2000, Triantafilis et al. 2001). Continuous classification offers a means of
attributing partial (or fuzzy) membership of more than one class to a single data point.
Membership of a class to a data point is assigned a value between 0 and 1, with 0
indicating no membership and 1 indicating total membership (Burrough et al. 1997,
McBratney and Odeh 1997, Stein et al. 1998). In soil science fuzzy set theory is generally
used for classification, allowing continuous class membership across continuous space
(McBratney and Odeh 1997). Fuzzy k-means (also known as fuzzy c-means) are a means
by which to compute fuzzy membership to a class based on attribute data (Stein et al.
1998). This has been used in soil science for mapping of continuous classes and soil
attributes (McBratney et al. 2000, Triantafilis et al. 2001).
1.1.7 Digital Elevation Models and Topographic Indices
Digital elevation models (DEM) are becoming increasingly important in understanding
natural processes such as the formation of soils and the subsequent erosional and
depositional processes to which they are subject. It has long been acknowledged that
topography is an important factor in soil formation (Jenny 1941) and analysis of terrain
variables in the field or from air photos has also been used historically to aid in soil survey
(Boer et al. 1996, Burrough et al. 1997). In conventional soil survey, particularly at a local
scale, qualitative terrain variables are used to extrapolate point surveys out to broader
regions (McKenzie et al. 2000). Since the development of remote sensing and adequate
computer technology terrain data has been used more and more to improve the diagnostic
and predictive power of remote sensing and earth process modelling (Odeh et al. 1994,
McKenzie and Ryan 1999, Metternicht et al. 2002, Drysdale and Metternicht 2003).
Variation in soil properties such as texture, nutrient concentration and availability and
cation exchange capacity (CEC) have been correlated with variations in topography
Chapter 2: Literature review 15
(Brubaker et al. 1993). Statistical prediction methods have been used with landform
attributes to predict soil properties such as subsoil clay, depth to solum and depth to
bedrock (Odeh et al. 1994, 1995, Skidmore et al. 1997). Soil parameters such as
phosphorus and pH have been correlated with terrain position (Skidmore et al. 1997)
Terrain data and airborne multispectral imagery have been used to predict soil variability
to aid in soil sampling, significantly improving the effectiveness of sampling strategies for
soil survey (Drysdale and Metternicht 2003). Topographic and landform variables have
been incorporated with gamma ray spectroscopy to predict soil profile depth, total
phosphorus and total carbon with varying degrees of success (McKenzie and Ryan 1999).
Radiometrics and digital terrain data have been used to examine the relationship between
soil, landform attributes and proteoid plants (Verboom and Pate 2003).
There are generally considered to be two types of topographic indices: primary attributes
and secondary or compound attributes (Wilson and Gallant 2000, McBratney et al. 2003).
Primary attributes are those that are derived directly from DEMs. For example slope,
defined as the gradient, affects surface and subsurface water flow, precipitation,
vegetation, soil water content, and land capability class. Aspect, measured in degrees
clockwise from north, affects solar radiation, vegetation distribution and
evapotranspiration (Wilson and Gallant 2000). Secondary attributes involve a combination
of primary attributes and are used to characterise the spatial variability of landscape
processes. For example, the topographic wetness index, which is derived from catchment
area, slope gradient and soil transmissivity, predicts soil moisture as it is affected by
topography (Wilson and Gallant 2000).
Studies have used optical remote sensing and slope to assist in the determination of soil
variability for the purposes of soil sampling design for soil survey (Drysdale et al. 2002,
Drysdale and Metternicht 2003). Secondary and primary topographic indices are also used
in combination. Field morphology and soil depth have been predicted successfully using
indices such as slope, wetness index, stream power, curvature and upslope and downslope
area (Odeh et al. 1994, Gessler et al. 1995, Boer et al. 1996). Topographic wetness index,
slope, curvature and downslope slope were successfully used with radiometrics to predict
soil depth, total phosphorus and total carbon (McKenzie and Ryan 1999). Other studies
have used digital terrain data in combination with other explanatory variables such as
Chapter 2: Literature review 16
optical remote sensing and radiometrics (Cialella et al. 1997, Skidmore et al. 1997, Taylor
et al. 2002, Thwaites 2002b, 2002a, Verboom and Pate 2003).
2.5 Remote Sensing and Reflectance Spectroscopy
The fields of remote sensing and reflectance spectroscopy are based on the principle of
electromagnetic radiation being reflected from a material and then detected by a sensor.
Remote sensing records the reflectance over the earths surface, collected from airborne or
satellite sensors, creating a continuous image. The image is made up of pixels, each
recorded from different ground resolution units and with reflectance spectra characteristic
of the material within the field of view. Alternatively, reflectance spectroscopy collects
information from discrete samples, typically recorded in the field or laboratory. Each
sample provides one reflectance spectrum characteristic of the material being analysed.
Remote sensing and reflectance spectroscopy can be applied across many different
wavelengths of the electromagnetic spectrum, each with different strengths and
weaknesses. This thesis focuses on what is known as the optical range covering the visible,
near-infrared and shortwave-infrared1 regions of the spectrum (Vis-NIR-SWIR, 400 –
2500 nm). The advantages of the optical range are that it is a passive technology; there is a
range of airborne and satellite sensors available, and it offers a cost effectiveness and
repeatability not available from other technologies. This range is also advantageous
because it is here that solar irradiance is at a maximum and there is sufficient reflected
radiation to be recorded by passive sensors.
The progression of improved spectral and spatial resolution has allowed for continued
development in the application of remote sensing and reflectance spectroscopy in many
areas. The development from multispectral to hyperspectral remote sensing has given users
increased diagnostic power allowing for the detection and discrimination of more and more
of the earth’s surface features.
1 The SWIR is included in the NIR in some disciplines such as chemistry and reflectance spectroscopy.
Chapter 2: Literature review 17
1.1.8 Spectral Characteristics of Soils
Interest in the optical properties of soils coincided with the development of spectrometers
capable of measuring electromagnetic reflectance at fine resolutions and the development
of airborne and space-based remote sensing. In this spectral range many of the constituents
of soils are optically active, absorbing radiation at specific wavelengths. Reflectance
spectrometry and remote sensing record the radiation reflected from materials and so
provides information about the active absorption processes taking place.
Electromagnetic radiation interacts with matter at atomic, molecular and structural levels.
At an atomic and molecular level, translational, rotational and vibrational motion of the
nuclei determine the interaction (Ben-Dor et al. 1999). Most important in soil reflectance is
vibrational motion which can exist at several different energy levels in an atom or
molecule and results in the stretching of molecular bond lengths or the bending bond
angles (Ben-Dor et al. 1999). Transition between energy levels can occur due to emission
or absorbance of radiation at specific wavelengths or frequencies. The locations of these
wavelengths are called fundamental bands, overtone bands and combination bands
depending on the type of transition. Absorption at fundamental, overtone and combination
bands result in absorption features within the reflectance spectra. Overtone and
combination bands are common in soil reflectance spectra over the NIR-SWIR region
whereas fundamental bands do not occur in this range (Ben-Dor et al. 1999, Clark 1999).
Examples of overtone bands in soil reflectance spectra include the oxygen-hydrogen (OH)
stretch at about 1400 nm and that associated with the CO32+ ion at 2300 to 2350 nm. An
example of a combination band is the bending and stretching of aluminium-hydroxyl (Al-
OH) at 2200 nm (Clark 1999).
There are also bands associated with electron transitions. These occur due to changes in the
state of electrons attached to atoms or molecules caused by the absorption or emission of
radiation. The location of these bands are determined by the relative energy states of
electron shells around atoms and molecules but for the most part they occur in the
ultraviolet and visible portions of the spectrum (Clark 1999). For example, iron has a
feature in the Vis-NIR that results from electron transition between the ferrous ion (Fe2+)
and the ferric iron (Fe3+) (Ben-Dor 2002).
Electromagnetic radiation also interacts with matter at a physical or structural level. This
involves the reflection or scattering of radiation by a multitude materials that make up the
Chapter 2: Literature review 18
soil volume but does not cause changes in the position of absorption features or
chromophores (Ben-Dor 2002). The factors which affect this physical interaction include
particle size, viewing geometry, radiation intensity, incident angle, sample geometry and
azimuth angle (Clark 1999, Ben-Dor 2002). Spectral properties that are affected by
changes in these parameters are typically absorption feature intensity and spectral curve
through changes in baseline height (Clark 1999). While these factors are relatively easy to
control in laboratory experiments they are essentially uncontrollable in field and imaging
studies.
1.1.9 Soil Reflectance Spectra
Classification of soil reflectance spectra was initially carried out by Condit (1970). After
measuring 285 soil samples (both wet and dry) from the USA he found that they could be
represented by three distinct spectral curves. However, these curves were only in the range
300 to 1000nm and no attempt was made to relate the distinct spectral curve types to
chemical or physical properties of the soils.
Stoner and Baumgardner (1981) continued this work, conducting a study with 485
individual soil samples from the USA and Brazil. They discovered 5 distinct soil
reflectance curve forms that were identified by curve shape and the presence or absence of
absorption features (Figure 2.1).
Figure 2.1: Representative reflectance spectra of soils collected in the U.S.A. and Brazil. Curves a-e
explained in text below (Stoner and Baumgardner 1981).
Chapter 2: Literature review 19
The first three curves of Stoner and Baumgardner (1981) are considered the same as those
presented by Condit (1970). The organic dominated form (type a) shows low overall
reflectance characterised by a concave curve shape from 500 to 1300 nm. Strong water
absorption bands are present at 1450 and 1950 nm in ‘type a’ and most other curve forms.
The minimally altered form (type b) has overall high reflectance and a convex shape
between 500 to 1300 nm. It also has strong water absorption bands at 1450 and 1950 nm
with weaker bands at 1200 and 1770 nm. The iron-affected form (type c) is characterised
by a slight ferric iron absorption at 700 nm and a strong iron absorption band at 900 nm.
The organic affected form (type d) has an overall reflectance higher than the organic
dominated form with a concave shape from 500 to 750 nm and a convex shape from 750 to
1300 nm. The iron-dominated form (type e) has decreasing reflectance with increasing
wavelength beyond 750 nm.
Most of the research since 1980 has focused on understanding the relationship between soil
properties and soil reflectance, with the goal of using soil spectra to predict the physio-
chemical composition of the soils. Generally this research has studies northern hemisphere
soils in the U.S.A., Europe and the Middle East. However, while these regions often
present different soils and land management regimes, it is possible that techniques and
methodologies developed may be applicable to Australia.
Soil Colour
Soil colour is an important measurement made in the classification of soils. It relates to,
and influenced by, soil moisture, permeability, organic matter (OM) content, mineralogy
and texture (Murtha 1988, Metternicht et al. 2002). The Munsell soil colour chart (2000) is
usually used to determine soil colour. This consists of a three dimensional identification of
colour describing the hue, value and chroma of a soil. Hue is a measure of the dominant
wavelength of light reflected from soil and results from a combination of pigments present
(i.e. minerals and OM). Value is a measure of the lightness compared to absolute white
while chroma is a measure of the purity of the hue (Ben-Dor et al. 1999, Munsell Color
Company 2000). Spectral reflectance is a quantitative means of measuring soil colour. As
spectral reflectance of soils has a direct relationship with soil colour, it can provide
information on soil moisture, permeability, OM content and mineralogy.
Reflectance spectroscopy of soils in the visible region has been used to determine Munsell
soil colour with some success, although the accuracy of the conversion was affected by soil
Chapter 2: Literature review 20
texture (Fernandez and Schulze 1987). Soil colour has also been derived from reflectance
spectra and related to the hematite content of soils (Torrent et al. 1983). Studies have
established a significant relationship between the albedo (brightness) of soil and the
Munsell value (a measure of soil lightness) but no relationship with hue or chroma (Post et
al. 2000). This is probably because soil colour is a function of a range of attributes, for
example, quartz content, OM content, iron oxide and clay. Other studies have used
multispectral sensors to develop relationships between the imagery and soil colour. SPOT
imagery has been used to successfully identify variation in soil colour not represented in
soil field mapping units (Agbu et al. 1990). Another study found strong correlations
between Landsat image data to Munsell soil colour in semiarid rangelands in North
America (Post et al. 1994). Variations in soil colour were also used to map soil organic
carbon with digitised aerial photography, essentially using visible light for the predictions
(Chen et al. 2000). Some research has aimed to predict soil colour from simulated
hyperspectral sensors, and results are favourable when compared with similar multispectral
simulations. However, the increased complexity and variability of image data has limited
the application of these methods to hyperspectral images (Leone and Escadafal 2001).
Soil Moisture
It was generally accepted from early studies that as the moisture content of a soil increases
the spectral reflectance decreases (Baumgardner et al. 1985, Post et al. 2000, Galvao et al.
2001, Weidong et al. 2002). This decrease in reflectance with increasing moisture content
stems from two sources; soil particles covered with thin films of water and water on the
lattice sites of some minerals present in the soil. However, despite the changes in
reflectance intensity, the overall shape of the curve forms remain relatively unchanged
(Condit 1970, Stoner and Baumgardner 1981).
The findings of earlier investigations, while correct, have been modified somewhat by
subsequent studies. Later studies found that the decrease in reflectance with increasing
moisture content is more pronounced at longer wavelengths (>1450nm) (Weidong et al.
2002). Weidong et al. (2002) also found that at higher moisture contents the trend is
reversed and reflectance increases with increasing water content. They determined this
critical point’ of reversal to be somewhere around field capacity, although it varied for
different soils, and occurs before the point where water absorption is saturating the
reflectance signal.
Chapter 2: Literature review 21
Also important when considering the effect of moisture on soil spectra is the presence of
water absorption bands. These water absorption bands relate to the effects of vibrational
frequencies of water molecules beyond 2500 nm (Baumgardner et al. 1985). The
absorption bands occur strongly at 1450 and 1950 nm with sharp peaks that indicate well-
defined sites and broad bands that denote unordered sites. The broad unordered bands are
more common in naturally occurring soils (Baumgardner et al. 1985, Galvao et al. 2001).
There are also weak bands that appear at 970, 1200 and 1770 nm (Hunt 1977). It has also
been contested that soil moisture is the most important variable in determining the
reflectance differences in the 2080-2320 µm bands (as found on the middle IR bands of
Landsat 4 and 5) (Baumgardner et al. 1985). Studies have used reflectance spectroscopy
and remote sensing to develop reliable spectral models for soil moisture (Ben-Dor et al.
2002, Whiting et al. 2004).
pH
Studies have found no chromophoric properties for pH (Ben-Dor and Banin 1995, Ben-Dor
et al. 2002). Whereas Ben-Dor et al. (2002) found correlations often exist between
different soil properties that are spectrally featureless, allowing the use of prediction
equations to reliably map such properties, they were unable to determine such a
relationship for soil pH using hyperspectral image data. Other studies have successfully
used reflectance spectroscopy and advanced statistical methods (e.g. partial least squares
regression and multivariate adaptive regression splines) to predict pH (Reeves et al. 2002,
Shepherd and Walsh 2002). However, the results are generally less successful than for
other soil properties with distinctive chromophoric properties.
Soil Organic Matter
The amount of soil organic matter (SOM) and type of SOM can significantly influence soil
spectral characteristics. Increasing SOM content of soils results in an decrease in the
spectral reflectance over the visible to NIR wavelength range, especially if the SOM
content is greater than 2% (Stoner and Baumgardner 1981, Henderson et al. 1992). It has
been found that, over the range between 520-800 nm, soils with an OM content higher than
2% have a concave shape and those with less then 2% have convex shape (Stoner and
Baumgardner 1981).
Chapter 2: Literature review 22
Different types of SOM have varying effects on soil spectral reflectance. Humic acid
accounts for most of the dark pigment of SOM and has lower reflectance over the visible to
short-wave spectral range. Alternatively, fulvic acid has been found to have no significant
influence on soil reflectance (Henderson et al. 1992). This study also found that soil
reflectance decreased with increasing SOM and highlighted bands that respond best to
SOM differences to allow for analysis.
Reflectance spectroscopy of soils has been used to predict SOM content. A study of 10 soil
types from North America found no absorption band that could be attributed to organic
matter in the infrared region (Krishnan et al. 1980). However, they did find that the visible
region of the spectrum provided the most reliable predictor (R2 = 0.873) and that
increasing organic carbon increases the slope of the curve at 800 nm. A study of soils in
Thailand using artificial neural networks found Vis – NIR a reliable predictor of SOM (R2
= 0.86) (Daniel et al. 2003). Other studies have predicted organic carbon using similar
techniques with some success (Shepherd and Walsh 2002, Islam et al. 2003). A
hyperspectral image study found reliable features in the reflectance spectra of heavy clay
soils in Israel to map soil SOM using prediction (calibration) equations (R2m > 0.82) (Ben-
Dor et al. 2002). Another image study used digitised colour aerial photography was
successfully used to map SOM at a paddock scale (r2 = 0.997) (Chen et al. 2000). Both of
these image studies relied heavily on exposed soil and took place over largely cultivated
areas.
Mineralogy
The different minerals that make up the largest component of soils affect the spectral
reflectance of the soil through the presence of absorption bands and overall spectral
brightness. Quartz is the largest and most common component of soils; it displays no
unique absorption feature in the Vis-NIR-SWIR range although it does increase the overall
brightness.
Clay minerals do have distinctive absorption bands that are caused by unique vibrational
overtones, electronic and charge transfers, and conduction processes (for example Figure
2.2) (Clark 1999). These absorption bands provide a diagnostic tool and with reflectance
spectroscopy have been used to determine the specific mineralogy of soils (Clark et al.
1990). Spectral features characteristic of clay minerals (around 2200nm) were successfully
extracted from AVIRIS imagery and used to identify soil clay mineralogy (smectite,
Chapter 2: Literature review 23
kaolinite and illite) (Chabrillat et al. 2002). Similarly, absorption band position, depth and
asymmetry have been used to map alteration phases with AVIRIS imagery (van der Meer
2004). Mineralogical identification has been achieved when the target material is partially
obscured by vegetation due largely to the distinctive absorption features (Chabrillat et al.
2002).
Figure 2.2: Example of reflectance spectra Kaolinite minerals showing absorption bands (~ 2200 µµµµm)
characteristic of clay minerals (Clark 1999)
Texture and particle size
Soil texture is influenced by many factors including the amount, size and type of clay
mineralogy, organic matter, carbonates and soil structure. Particle size distributions refer
simply to the relative amounts of particles within the size classes of sand, silt and clay,
although they are probably the most determining factor of the soil texture (Murtha 1988).
There is a commonly observed relationship between soil composition and texture that
affects the determination of the contribution of soil texture to observed reflectance (Galvao
et al. 1997). For example, sandy soils have a higher reflectance, due to lower amounts of
OM, iron oxides and clay minerals, than heavy textured clay soils. These factors all
contribute to the spectral reflectance of the soils and it becomes unclear which property is
contributing to the spectral profile.
Decreases in particle size of a mineral can increase overall spectral reflectance
(Baumgardner et al., 1985). This is caused by more energy being reflected from the soil
mineral than is lost between coarser grained aggregates. Alternatively, clay (<0.002 mm)
generally has lower reflectance than soils with sand and/or silt (>0.002 mm), and finer
Chapter 2: Literature review 24
textured soils appear darker than coarse textured soils (Irons et al. 1989). This is possibly
due to the increased water holding capacity of clays. A study of tropical Brazilian soils
found that clay content can be more reliably measured in subsurface horizons because of
their lower OM. This is thought to be due to lignin and cellulose absorption near 2200 nm
masking clay absorption at the same location (Galvao et al. 1997, Galvao et al. 2001).
Despite these complications, reflectance spectroscopy in the Vis-NIR-SVIR has been used
to reliably predict clay content and other particle sizes in a number of studies (Chang et al.
2001, Shepherd and Walsh 2002, Cozzolino and Morón 2003).
In image studies surface crusts also affect reflectance spectra. They affect both albedo due
to particle size and also spectral absorption features due to changes in chemical
composition (Ben-Dor et al. 2003). This could have significant impacts on optical remote
sensing because it is the surface that is visible to the sensor. Surface crust may not be a
good predictor for what is under the surface because it is severely affected by management
practices as well as soil chemistry and physiology.
Iron Oxide
Iron oxide affects soil reflectance spectra with broad and shallow absorption features at
wavelengths lower than 1000 nm and overall lower albedo as iron oxide content increases
(Hunt 1977, White et al. 1997, Galvao et al. 2001). Reflectance spectroscopy has been
used to predict iron oxide content in a number of studies with ranging success. Some
studies have achieve relatively poor correlations (R2 = 0.5) (Islam et al. 2003) while others
have more successful (R2 = 0.64 and R2 = 0.9) (Chang et al. 2001, Cozzolino and Morón
2003). Iron oxide has also been correlated to surface soil reflectance within multispectral
and hyperspectral image studies (Stoner and Baumgardner 1981, Galvao et al. 2001,
Metternicht et al. 2002). Iron oxide abundance has been mapped using multispectral
imagery and changes in concentrations have been reliably predicted (r = 0.91) (White et al.
1997). Other studies have used principal components analysis to successfully map iron
oxide (Fraser 1991, Tangestani and Moore 2002). The study by White et al. (1997) was
carried out in a desert with little vegetation and a quartz-dominated desert soil. Other
studies have found that OM can interfere with the detection of iron oxide due to
interference of absorption features due to the different materials (Galvao et al. 1997).
Figure 2.3 shows the spectral characteristics of hematite with dry and green vegetation
demonstrating how they coincide in the spectral range. Other studies have also
Chapter 2: Literature review 25
demonstrated this interference between iron oxide and vegetation (Fraser 1991, Galvao et
al. 1997).
Figure 2.3: Showing spectral features of hematite, green grass and dry grass (Fraser 1991).
Salinity
Most of the salts responsible for soil salinity have no direct spectral features or
chromophores that allows for their discrimination. Despite this some studies have
successfully predicted salinity with reflectance spectroscopy (R2 = 0.65) (Shibusawa et al.
2001) although others have had less success (R2 = 0.1) (Islam et al. 2003). It has been
postulated that successful prediction of salinity is governed by inter-correlation between
other soil properties such as soil moisture (Ben-Dor et al. 2002).
For multispectral image studies the inclusion of topographic data is sometimes used to
mitigate the poor diagnostic power of the sensor and improve the classification. For
example a study used Landsat TM and DEM derived topographical indices to map salinity
in the Western Australian wheat belt (Caccetta et al. 2000). For these purposes the DEMs
were used to determine watershed parameters including ‘upslope area’, ‘upslope cleared
area’ and other factors effecting the formation and spread of salinity. A similar study used
topographic indices, Landsat TM imagery and conditional probability networks to monitor
increasing salinity in Western Australia also used Landsat TM imagery to predict areas at
risk of salinity using decision trees and DEMs to substantiate their data (Kiiveri and
Caccetta 1998). The use of other data sets improves the predictability of some landcovers
and provides an extra element to the diagnosis.
Chapter 2: Literature review 26
While hyperspectral sensors improve the diagnostic power of remotely sensed data and can
thus be used more independently of other data sets, the absence of spectral features in salt
still makes classification difficult. However, researchers have found that saline soils can be
mapped using other soil or vegetation properties as surrogates. Ben-Dor et al. (2002)
found, using DAIS-7915 hyperspectral scanner data, that soil salinity was correlated with
soil moisture (r = 0.58) in cultivated fields and was able to develop reliable prediction
equations. HyMap hyperspectral data has also been used to map salinity symptoms under
different agricultural environments. The characteristic features of samphire (Halosarcia
pergranulata) and gypsum (associated with salt scalds) were used as indicators of dryland
salinity at Point Sturt in Lake Alexandrina, South Australia (Dutkiewicz et al. 2009).
Similarly, samphire and other halophyte species such as Sea Blite (Sueda australis) and
Sea Barley Grass (Critesion marinun) have been used to map irrigation salinity with
HyMap hyperspectral imagery (Dehaan and Taylor 2002, 2003).
1.1.10 Limitations of Optical Remote Sensing for Soil Mapping
Optical remote sensing can only directly access the surface of materials covering the earth.
This presents a significant limitation in the mapping of soils. For most land uses the upper
surface of the soil is covered by material other than the soil for much of the year. This may
be photosynthetic crops, for example wheat or vines, or crop residues in the form of
stubble and loose material. These surface coverings significantly reduce the amount of
information received by the sensor that is directly about the soil itself. Furthermore, a
comprehensive soil map must consist of an analysis of the entire profile. Thus optical
remote sensing is not a tool to be used in isolation to map soil. The incorporation of other
techniques and technologies is warranted to provide comprehensive understanding of the
soils for the purposes of mapping.
1.1.11 Vegetation Discrimination and Mapping
Remote sensing is an important means by which to map vegetation and landcover on the
earth’s surface. Satellite sensors have long been used to determine the percentage cover of
vegetation and its converse, soil exposure (Bannari et al. 1995). These components of
landcover are important in understanding risk to natural resources of degradation such as
erosion and increasing salinity. Furthermore, in understanding vegetation distribution and
components it is sometimes possible to draw conclusions about the underlying soil
properties (Taylor et al. 2002). It is also important to understand vegetation reflectance and
Chapter 2: Literature review 27
how it interacts with that of other materials including soil. How vegetation reflectance
interacts with other materials depends on the state of the vegetation, dry or alive, and also
structural differences in vegetation, herbaceous or woody (Skidmore et al. 1997).
Multispectral imagery has long been used to discriminate and map vegetation variables
such as biomass, leaf area index and percent cover. Moreover, it has been used widely to
map vegetation condition by way of greenness indices. However, there are significant
limitations in the usefulness of multispectral imagery in discriminating variations in the
composition of vegetation (Elvidge 1990, Lewis 2000).
Hyperspectral imagery provides advantages over multispectral imagery for sub-pixel
discrimination and mapping of vegetation. The larger number of spectral bands in
hyperspectral data can potentially provide interpretation and discrimination of more sub-
pixel components. Moreover, the band placement more readily enables discrimination of
spectral features, further increasing diagnostic power of the data (Lewis et al. 2000).
There is some disagreement about the spectral regions of the EM spectrum that are useful
in the discrimination of vegetation. Some consider the VNIR provides the best spectral
information for vegetation due to water absorption features in the SWIR masking plant
spectral information (Elvidge 1990). However, studies have challenged this notion, finding
the SWIR valuable for semi-arid vegetation discrimination (Drake et al. 1999, Lewis
2000).
Using airborne multispectral sensor (AMS) hyperspectral imagery, functional components
of vegetation (i.e. trees versus shrubs), differences in species (i.e. Eucalyptus versus other
tree species) and different physiological conditions (i.e. actively growing versus dry litter)
have been adequately mapped (Lewis et al. 2000). Also multispectral satellite sensors such
as SPOT have been used to successfully map forest type, relying on vegetation structure
for discrimination (Xiao et al. 2002).
Furthermore, studies have used the normalised difference vegetation index (NDVI) and
other vegetation indices derived from multispectral airborne remote sensing to infer
variability in the underlying soil (Lamb 2000, Drysdale and Metternicht 2003). Plant
condition strongly influences vegetation reflectance spectra. Water stress, for example, has
been found to change plant spectral response and plant reflectance has been used to
estimate soil water content in cropping systems (Senay et al. 2000). However, this was
Chapter 2: Literature review 28
largely based on the plant biomass having a strong correlation with plant water which in
turn correlated positively with soil water. The above studies demonstrate that soil variation
and to some degree soil properties can be discriminated using multispectral remote
sensing.
Soil properties have been examined in eucalyptus forests of south-eastern Australia with
remote sensing, terrain data and GIS. Correlations were found between soil properties,
total phosphorus, exchangeable cations and electrical conductivity and spectral reflectance
(Skidmore et al. 1997). However, this study was undertaken in natural forests that were
relatively unaffected by modern agriculture. Thus there were not subject to fertiliser and
pesticide inputs which would significantly affect plant response in relation to soil
variability. Other studies have used remote sensing of vegetation as a surrogate for soil
properties in association with digital terrain data over mono-crop environments (Selige
1998). However, this is more difficult over heterogeneous cropping environments due to
variation in chemical and physical properties across different crop types affecting their
reflective properties (Skidmore et al. 1997).
2.6 Summary
Soil mapping in Australia is well advanced for various regional scale applications in some
parts of the country. However, there is much scope to improve upon the current scale of
mapping and provide a better resource for local applications. Furthermore there are large
areas of the continent where the soils are not well understood and new mapping programs
are likely in the future. Improving upon the scale of current soil maps and providing new
inventories of soils is expensive, labour intensive and time consuming by conventional
mapping methodologies. The application of remote sensing and reflectance spectroscopy
may provide a cost effective and rapid means by which to improve the resolution current
soil mapping and undertake new programs.
The spectral response of soils has been used to predict different properties in a variety of
applications. While some studies have applied reflectance spectroscopy to predicting soils
this has been with soil samples from large geographic extents and little effort has been
given to using the predictions for subsequent mapping. There is much scope to further
examine the use of reflectance spectroscopy and applying it to soil mapping. Furthermore,
the soils of southern Australia provide unique profile and landscape characteristics such as
Chapter 2: Literature review 29
low nutritional content and strong texture contrast resulting from extensive weathering,
low organic matter content and a high occurrence of salinity and sodicity.
2.7 References
Agbu, P. A., Fehrenbacher, D. J. and Jansen, I. J. 1990 Statistical comparison of spot spectral maps with field soil maps, Soil Science Society of America Journal, 54, 812-818.
Bannari, A., Morin, B., Bonn, F. and Huete, A. R. 1995 A review of vegetation indices, Remote Sensing
Review, 13, 95-120.
Baumgardner, M. F., Silva, L. R., Biehl, L. L. and Stoner, E. R. 1985 Reflectance properties of soils, Advances in Agronomy, 38, 1-44.
Ben-Dor, E. 2002 Quantitative remote sensing of soil properties, Advances in Agronomy, 75, 173-243.
Ben-Dor, E. and Banin, A. 1995 Near-infrared analysis as a rapid method to simultaneously evaluate several soil properties, Soil Science Society of America Journal, 59, 364-372.
Ben-Dor, E., Goldlshleger, N., Benyamini, Y., Agassi, M. and Blumburg, D. G. 2003 The spectral reflectance properties of soil structural crusts in the 1.2- to 2.5um spectral region, Soil Science Society of
America Journal, 67, 289-299.
Ben-Dor, E., Irons, J. R. and Epema, G. F. 1999 Soil reflectance, In Remote sensing of the earth sciences:
Manual of remote sensing, Vol. 3 (Ed, A. N. Rencz) John Wiley and Sons, New York, pp. 111-188.
Ben-Dor, E., Patkin, K., Banin, A. and Karnieli, A. 2002 Mapping of several soil properties using DAIS-7915 hyperspectral scanner data - a case study over clayey soils in Israel, International Journal of Remote
Sensing, 23, 1043-1062.
Bhatti, A. U., Mulla, D. J. and Frazier, B. E. 1991 Estimation of soil properties and wheat yields on complex eroded hills using geostatistics and thematic mapper images, Remote Sensing of Environment, 37.
Bishop, T. F. A. and McBratney, A. B. 2001 A comparison of prediction methods for the creation of field-extent soil property maps, Geoderma, 103, 149-160.
Bishop, T. F. A., McBratney, A. B. and Whelan, B. M. 2001 Measuring the quality of digital soil maps using information criteria, Geoderma, 103, 95-111.
Boer, M., Del Barrio, G. and Puigdefabregas, J. 1996 Mapping soil depth classes in dry Mediterranean areas using terrain attributes derived from a digital elevation model, Geoderma, 72, 99-118.
Brubaker, S. C., Jones, A. J., Lewis, D. T. and Frank, K. 1993 Soil properties associated with landscape position, Soil Sci Soc Am J, 57, 235-239.
Burrough, P. A., van Gaans, P. F. M. and Hootsmans, R. 1997 Continuous classification in soil survey: Spatial correlation, confusion and boundaries, Geoderma, 77, 115-135.
Caccetta, P. A., Allen, A. and Watson, I. 2000 The land monitor project, In 10th Australasian Remote Sensing and Photogrammetry Conference, Adelaide, Australia, 21-25 August 2000.
Chabrillat, S., Goetz, A. F. H., Krosley, L. and Olsen, H. W. 2002 Use of hyperspectral images in the identification and mapping of expansive clay soils and the role of spatial resolution, Remote Sensing of
Environment, 82, 431-445.
Chang, C., Laird, D. A., Mausbach, M. J. and Hurburgh, C. R., Jr. 2001 Near-infrared reflectance spectroscopy-principal components regression analyses of soil properties, Soil Sci Soc Am J, 65, 480-490.
Chen, F., Kissel, D. E., West, L. T. and Adkins, W. 2000 Field-scale mapping of surface soil organic carbon using remotely sensed imagery, Soil Sci Soc Am J, 64, 746-753.
Cialella, A. T., Dubayah, R., Lawrence, W. and Levine, E. 1997 Predicting soil drainage class using remotely sensed and digital elevation data, Photogrammetric Engineering and Remote Sensing, 63, 171-178.
Clark, R. N. 1999 Spectroscopy of rocks and minerals and principles of spectroscopy, In Remote sensing for
the earth sciences: Manual of remote sensing, Vol. 3 (Ed, A. N. Rencz) John Wiley and Sons, New York, pp. 3-58.
Chapter 2: Literature review 30
Clark, R. N., King, T. V. V., Klejwa, M., Swayze, G. and Vergo, N. 1990 High spectral resolution reflectance spectroscopy of minerals, Journal of Geophysics, 95, 12653-12680.
Condit, H. R. 1970 The spectral reflectance of American soils, Photogrammetric Engineering, 36, 955-966.
Cozzolino, D. and Morón, A. 2003 The potential of near-infrared reflectance spectroscopy to analyse soil chemical and physical characteristics, Journal of Agricultural Science, 140, 65-71.
Daniel, K. W., Tripathi, N. K. and Honda, K. 2003 Artificial neural network analysis of laboratory and in situ spectra for the estimation of macronutrients in soils of Lop Buri (Thailand), Australian Journal of Soil
Research, 41, 47-59.
Dehaan, R. and Taylor, G. R. 2002 Field-derived spectra of salinized soils and vegetation as indicators of irrigation-induced salinization, Remote Sensing of Environment, 80, 406-417.
Dehaan, R. and Taylor, G. R. 2003 Image-derived spectral endmembers as indicators of salinisation, International Journal of Remote Sensing, 24, 775-794.
Drake, N. A., Mackin, S. and Settle, J. J. 1999 Mapping vegetation, soils, and geology in semiarid shrublands using spectral matching and mixture modeling of SWIR AVIRIS imagery, Remote Sensing of Environment, 68, 12-25.
Drysdale, G., Delfos, J. and Metternicht, G. 2002 Low cost remote sensing approach for designing effective sampling strategies for soil properties for site-specific crop management, In 29th International symposium on Remote Sensing of Environment, Buenos Aires, Argentina, 8-12 April 2002.
Drysdale, G. and Metternicht, G. 2003 Utilising remote sensing and terrain data for designing multiscale sampling strategies for soil properties in agricultural fields, In Spatial Sciences Conference 2003, Canberra,
Dutkiewicz, A., Lewis, M. and Ostendorf, B. 2009, Vol. 30 Taylor & Francis, pp. 693 - 719.
Elvidge, C. D. 1990 Visible and near infrared reflectance characteristics of dry plant materials, International
Journal of Remote Sensing, 11, 1775-1795.
Fernandez, R. N. and Schulze, D. G. 1987 Calculation of soil colour from reflectance spectra, Soil Science
Society of America Journal, 51, 1277-1282.
Fraser, S. F. 1991 Discrimination and identification of ferric oxides using satellite thematic mapper data: A Newman case study, International Journal of Remote Sensing, 12, 635-641.
Galvao, L. S., Pizarro, M. A. and Epiphanio, J. C. N. 2001 Variations in reflectance of tropical soils: Spectral-chemical composition relationships from AVIRIS data, Remote Sensing of Environment, 75, 245-255.
Galvao, L. S., Vitorello, I. and Formaggio, A. R. 1997 Relationships of spectral reflectance and color among surface and subsurface horizons of tropical soil profiles, Remote Sensing of Environment, 61, 24-33.
Gessler, P. E., Moore, I. D., McKenzie, N. J. and Ryan, P. J. 1995 Soil-landscape modeling and spatial prediction of soil attributes, International Journal of Geographic Information Systems, 9, 421-432.
Goovaerts, P. 1997 Geostatistics for natural resources evaluation, Oxford University Press, New York.
Gunn, R. H., Beattie, J. A., Riddler, A. M. H. and Lawrie, R. A. 1988 Mapping, In Australian soil and land
survey handbook: Guidelines for conducting survey (Eds, R. H. Gunn, J. A. Beattie, R. E. Reid and R. H. M. van de Graaff) Inkata Press, Melbourne, pp. 90-112.
Haskett, J. D., Pachepsky, Y. A. and Acock, B. 1995 Estimation of soybean yields at county and state levels using GLYCIM: A case study for Iowa, Agronomy Journal, 87, 926-931.
Henderson, T. L., Baumgardner, M. F., Franzmeier, D. P., Stott, D. E. and Coster, D. C. 1992 High dimensional reflectance analysis of soil organic matter, Soil Science Society of America Journal, 56, 865-872.
Hunt, G. R. 1977 Spectral signatures of particulate minerals in the visible and near infrared, Geophysics, 42, 501-513.
Ishida, T. and Ando, H. 1999 Use of disjunctive cokrieging to estimate soil organic matter from Landsat thematic mapper image, International Journal of Remote Sensing, 20, 1549-1565.
Islam, K., Singh, B. and McBratney, A. 2003 Simultaneous estimation of several soil properties by ultra-violet, visible and near-infrared reflectance spectroscopy, Australian Journal of Soil Research, 41, 1101-1114.
Chapter 2: Literature review 31
Jenny, H. 1941 Factors of soil formation: A system of quantitative pedology, McGraw-Hill, New York.
Kiiveri, H. and Caccetta, P. A. 1998 Image fusion with conditional probability networks for monitoring the salinization of farmland, Digital Signal Processing, 8, 225-230.
Krishnan, P., Alexander, J. D., Butler, B. J. and Hummel, J. W. 1980 Reflectance technique for predicting soil organic matter, Soil Science Society of America Journal, 44, 1282-1285.
Lamb, D. W. 2000 The use of quantitative airborne multispectral imaging fro managing agricultural crops - a case study in south eastern Australia, Australian Journal of Experimental Agriculture, 40, 725-738.
Leone, A. P. and Escadafal, R. 2001 Statistical analysis of soil colour and spectroradiometric data for hyperspectral remote sensing of soil properties (example in a southern Italy Mediterranean ecosystem), International Journal of Remote Sensing, 22, 2311-2328.
Lewis, M. 2000 Discrimination of arid vegetation composition with high resolution CASI imagery, Rangeland Journal, 22, 141-167.
Lewis, M., Jooste, V. and deGasparis, A. A. 2000 Discrimination of arid vegetation with hyperspectral imagery, In 10th Australasian Remote Sensing and Photogrammetry Conference, Adelaide,
Maschmedt, D. 2000 Assessing agricultural land: Agricultural land classification standards used in South Australia's land resource mapping program, Primary Industries and Resources South Australia, Adelaide,
Mayr, T. and Jarvis, N. J. 1999 Pedotransfer functions to estimate soil water retention parameters for a modified Brooks - Corey type model, Geoderma, 91, 1-9.
McBratney, A. B., Mendonça Santos, M. L. and Minasny, B. 2003 On digital soil mapping, Geoderma, 117, 3-52.
McBratney, A. B., Minasny, B., Cattle, S. R. and Vervoort, R. W. 2002 From pedotransfer functions to soil inference systems, Geoderma, 109, 41-73.
McBratney, A. B. and Odeh, I. O. A. 1997 Application of fuzzy sets in soil science: Fuzzy logic, fuzzy measurements and fuzzy decisions, Geoderma, 77, 85-113.
McBratney, A. B., Odeh, I. O. A., Bishop, T. F. A., Dunbar, M. S. and Shatar, T. M. 2000 An overview of pedometric techniques for use in soil survey, Geoderma, 97, 293-327.
McKenzie, N. J., Gessler, P. E., Ryan, J. P. and O'Connell, D. A. 2000 The role of terrain analysis in soil mapping, In Terrain analysis: Principles and applications (Eds, J. P. Wilson and J. C. Gallant) John Wiley and Sons, New York.
McKenzie, N. J. and Ryan, P. J. 1999 Spatial prediction of soil properties using environmental correlation, Geoderma, 89, 67-94.
Metternicht, G., Newby, T., van der Berg, H., Paterson, G. and Booyens, B. 2002 Feasibility of using aster data for rapid farm scale soil mapping in South Africa, In 11th Australasian Remote Sensing and Photogrammetry Conference, Brisbane, Australia,
Minasny, B., McBratney, A. B. and Bristow, K. L. 1999 Comparison of different approaches to the development of pedotransfer functions for water-retention curves, Geoderma, 93, 225-253.
Munsell Color Company 2000 Gretag Macbeth, New York.
Murtha, G. G. 1988 Soil properties and soil performance, In Australian soil and land survey handbook:
Guidelines for conducting surveys (Eds, R. H. Gunn, J. A. Beattie, R. E. Reid and R. H. M. van de Graaff) Inkata Press, Melbourne, pp. 241-257.
Northcote, K. H., Beckmann, G. G., Bettenay, E., Churchward, H. M., van Dijk, D. C., Dimmock, G. M., Hubble, G. D., Isbell, R. F., McArthur, W. M., Murtha, G. G., Nicolls, K. D., Paton, T. R., Thompson, C. H., Webb, A. A. and Wright, M. J. 1968 Atlas of Australian soils, CSIRO, Melbourne.
Odeh, I. O. A., McBratney, A. B. and Chittleborough, D. J. 1994 Spatial prediction of soil properties from landform attributes derived from a digital elevation model, Geoderma, 63, 197-214.
Odeh, I. O. A., McBratney, A. B. and Chittleborough, D. J. 1995 Further results on prediction of soil properties from terrain attributes: Heterotopic cokriging and regression-kriging, Geoderma, 67, 215-226.
Petach, M. C., Wagenet, R. J. and DeGloria, S. D. 1991 Regional water flow and pesticide leaching using simulations with spatially distributed data, Geoderma, 48, 245-269.
Chapter 2: Literature review 32
Post, D. F., Fimbres, A., Matthias, A. D., Sano, E. E., Accioly, L., Batchily, A. K. and Ferreira, L. G. 2000 Predicting soil albedo from soil color and spectral reflectance data, Soil Science Society of America Journal, 64, 1027-1034.
Post, D. F., Lucas, W. M., White, S. A., Ehasz, M. J., Batchily, A. K. and Horvath, E. H. 1994 Relations between soil color and Landsat reflectance on semiarid rangelands, Soil Sci Soc Am J, 58, 1809-1816.
Reeves, J., McCarty, G. W. and Mimmo, T. 2002 The potential of diffuse reflectance spectroscopy for the determination of carbon inventories in soils, Environmental Pollution, 116, S277-S284.
Schoknecht, N., Tille, P. and Purdie, B. 2004 Resource management technical report 280: Soil-landscape mapping in South-western Australia, an overview of methodology and outputs, Department of Agriculture, Western Australia, November, 2004,
Selige, T. 1998 Spatial detection of soil properties for precision farming using remotely sensed imagery, terrain analysis and GIS, In 9th Australian Remote Sensing and Photogrammetry Conference, Sydney, Australia, 20-24 June.
Senay, G. B., Ward, A. D., Lyon, J. G., Fausey, N. R., Nokes, S. E. and Brown, L. C. 2000 The relations between spectral data and water in a crop production environment, International Journal of Remote Sensing, 21, 1897-1910.
Shepherd, K. D. and Walsh, M. G. 2002 Development of reflectance spectral libraries for characterization of soil properties, Soil Science Society of America Journal, 66, 988-998.
Shibusawa, S., I Made Anom, S. W., Sato, S., Sasao, A. and Hirako, S. 2001 Soil mapping using the real-time soil spectrophotometer, In Third European Conference on Precision Agriculture, Agro Montpellier,
Skidmore, A. K., Varakamp, C., Wilson, L., Knowles, E. and Delaney, J. 1997 Remote sensing of soils in a eucalypt forest environment, International Journal of Remote Sensing, 18, 39-56.
Soutter, M. and Pannatier, Y. 1996 Groundwater vulnerability to pesticide contamination on a regional scale, Journal of Environmental Quality, 25, 439-444.
Stein, A., Bastiaanssen, W. G. M., De Bruins, S., Crackness, A. P., Curran, P. J., Fabbri, A. G., Gorte, B. G. H., Van Groenigen, J. W., van der Meer, F. and Saldana, A. 1998 Integrating spatial statistics and remote sensing, International Journal of Remote Sensing, 19, 1793-1814.
Stoner, E. R. and Baumgardner, M. F. 1981 Characteristic variation in reflectance of surface soils, Soil
Science Society of American Journal, 45, 1161-1165.
Tangestani, M. H. and Moore, F. 2002 Porphyry copper alteration mapping at the Meiduk area, Iran, International Journal of Remote Sensing, 23, 4815-4825.
Taylor, J. K. 1970 The development of soil survey and Field pedology in Australia, 1927-67, CSIRO, Melbourne.
Taylor, M., J., Smettem, K., Pracilio, G. and Verboom, W. 2002 Relationships between soil properties and high-resolution radiometrics, central eastern Wheatbelt, western Australia, Exploration Geophysics, 33, 95-102.
Thwaites, R. 2002a Airborne gamma-ray spectrometry in soil landscape modelling for upland, erosional forestlands, In 11th Australasian Remote Sensing and Photogrammetry Conference, Brisbane, Australia, 2-6 September 2002.
Thwaites, R. 2002b Spatial terrain analysis for matching native tree species to sites: A methodology, New
Forests, 24, 81-95.
Tiktak, A., Leijnse, A. and Vissenburg, H. 1999 Uncertainty in a regional-scale assessment of cadmium accumulation in the Netherlands, Journal of Environmental Quality, 28, 461-470.
Timlin, D. J., Pachepsky, Y. A., Acock, B. and Whisler, F. 1996 Indirect estimation of soil hydraulic properties to predict soybean yield using GLYCIM, Agricultural Systems, 52, 331-353.
Torrent, J., Shwertmann, U., Fechter, H. and Alferez, F. 1983 Quantitative relationships between soil color and haematite content, Soil Science, 136, 354-358.
Triantafilis, J., Ward, A. D., Odeh, I. O. A. and McBratney, A. B. 2001 Creation and interpolation of continuous soil layer classes in the lower Namoi valley, Soil Science Society of America Journal, 65, 403-413.
Chapter 2: Literature review 33
van der Meer, F. 2004 Analysis of spectral absorption features in hyperspectral imagery, International
Journal of Applied Earth Observation and Geoinformation, 5, 55-68.
Verboom, W. and Pate, J. S. 2003 Relationships between cluster root-bearing taxa and laterite across landscapes in southwest western Australia: An approach using airborne radiometric and digital elevation models, Plant and Soil, 248, 321-333.
Weidong, L., Baret, F., Xingfa, G., Qingxi, T., Lanfen, Z. and Bing, Z. 2002 Relating soil surface moisture to reflectance, Remote Sensing of Environment, 81, 238-246.
White, K., Walden, J., Drake, N., Eckardt, F. and Settle, J. 1997 Mapping the iron oxide content of dune sands, Namib sand sea, Namibia, using Landsat thematic mapper data, Remote Sensing of Environment, 62, 30-39.
Whiting, M. L., Li, L. and Ustin, S. 2004 Predicting water content using Gaussian model on soil spectra, Remote Sensing of Environment, 89, 535-552.
Wilson, J. P. and Gallant, J. C. 2000 Digital terrain analysis, In Terrain analysis: Principles and applications (Eds, J. P. Wilson and J. C. Gallant) John Wiley and Sons, New York, pp. 1-27.
Xiao, X., Boles, S., Liu, J., Zhuang, D. and Liu, M. 2002 Characterization of forest types in Northeastern china, using multi-temporal spot-4 vegetation sensor data, Remote Sensing of Environment, 82, 335-348.
Zhu, A.-X. 1997 A similarity model for representing soil spatial information, Geoderma, 77, 217-242.
34
Chapter 3
Spectral Discrimination of Soil Properties
Published as refereed conference paper:
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2005 Spectral discrimination of soil properties, In SSC 2005 Spatial Intelligence, Innovation and Praxis: The National Biennial Conference of the Spatial Sciences Institute, Melbourne, Australia, September, 2005.
49
Chapter 4
Visible near-infrared reflectance spectroscopy as a
predictive indicator of soil properties
Published as journal article:
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. (In Press) Visible near-infrared reflectance spectroscopy as a predictive indicator of soil properties, Ecological Indicators.
4.1 Introduction
The classification, mapping and monitoring of soils is an important underpinning of
modern day natural resource management. Regional scale soil maps are traditionally
produced by dividing the landscape into mapping units through air-photo and landscape
interpretation from which sample sites are chosen to characterise the soils. For regional
planning these maps provide an excellent resource, but they do not provide sufficient detail
for localised soil and land management. Whereas soil variability within each of these
mapping units is often acknowledged in the map and accompanying report, it is not
depicted or quantified. Increasing concern over land degradation, agricultural productivity
and the loss of ecological services has led to a desire for greater understanding of land
resources and processes at scales larger than 1:50 000 scales.
Around the world governments are investing in programs to better understand soil
variability and create soil databases to better inform landscape planning and management
decisions. In South Australia the soils of the agricultural districts have been mapped and
information presented on maps at 1:50 000 and 1:100 000 scale (Soil and Land Program
2007). While these maps provide an excellent regional planning tool, finer spatial
resolution information is required to improve land management decisions at farm scale,
and to assist understanding and modelling of problems such as diminishing biodiversity
and dryland salinity. Unlike the agricultural districts, there is a paucity of data on the
nature and distribution of soils in South Australia’s pastoral zones. The pastoral districts
Chapter 4: Spectral prediction 50
cover large areas and contribute substantially to the State’s economic productivity. These
areas would benefit greatly from improved understanding of soil properties and their
variability as well as vegetation condition, ecology and biodiversity. Recent studies in
Australia’s arid region for example, have called for improved understanding of soil
heterogeneity as inputs for the monitoring of ecology and biodiversity, citing the lack of
spatial resolution as an impediment (Clarke 2008). Similar inputs have been used in other
parts of the world to predict vegetation community distributions (Miller et al. 2002).
Creating these maps and improving the spatial resolution of existing maps to provide
greater detail about soil variability can be prohibitively expensive by traditional soil survey
procedures (Sumfleth and Duttmann 2007) and can only be justified for the most intensive
agricultural systems. Pedotransfer functions have been employed to reduce the expense of
intensive soil mapping by using surrogates that are relatively inexpensive to measure, as
well as to predict less readily measured soil properties. Examples of this include using soil
colour to predict organic carbon content and using mechanical resistance as an indicator of
bulk density and clay content (McBratney et al. 2002). However effective these functions
are for some applications, pedotransfer functions do not provide a direct measurement of
soil properties nor are they provided for all soil properties of interest. Information or
indicators for a wider range of soil properties is needed.
In order to overcome the expense of traditional soil survey and the limitations of
pedotransfer functions, researchers are increasingly turning to remote sensing and, in
particular, reflectance spectroscopy. This form of earth observation can provide useful
indicators for mapping and monitoring many environmental features such as geology and
minerals (Bower and Rowan 1996, Clark 1999), vegetation and soil (Lewis 2000, Ben-Dor
et al. 2002, Sumfleth and Duttmann 2007) and even ecological habitats (Tiner 2004, Bock
et al. 2005). With field and imaging spectrometers becoming increasingly sophisticated,
there is potential for substantial improvement in the speed, reliability and resolution of soil
analysis. Spectral analysis of soil cores with field or laboratory spectrometers could
provide a new means of automated, rapid and objective profile evaluation, following the
approach now being developed for mineral characterisation of geological cores (Mauger et
al. 2004). In addition, new imaging spectrometers offer the prospect of detailed raster-
based mapping of surface soil properties with higher spatial resolution than is possible with
the current approaches.
Chapter 4: Spectral prediction 51
4.1.1 Spectral Reflectance Variation in Soils
Early studies of soil reflectance spectra over the visible (Vis, 400 – 700 nm), near-infrared
(NIR, 700 – 1300 nm) and shortwave-infrared2 (SWIR, 1300 – 2400 nm) region described
and classified different ‘curve forms’. For example, Condit (1970) identified three types of
curves amongst 285 soils from the United States, characterised by the overall shape of the
spectral response and changes in slope over the wavelength range. However, no attempt
was made to explain the spectral response in relation to physical or chemical properties of
the soils. A more comprehensive study by Stoner and Baumgardner (1981) described five
curve forms amongst 485 soils from the United States and Brazil, and also related specific
absorption features to soil organic carbon and iron oxide content in the soil. However, most
of the more recent research has investigated relationships between the soil properties and
soil reflectance with the aim of predicting the physio-chemical properties of the soil.
The clay mineralogy in soils has been distinguished in several studies using the short wave
infrared (SWIR) region of the spectrum (1300 – 2500 nm) (Islam et al. 2003), and
especially the 2200 nm absorption feature that is characteristic of clays (Ben-Dor 2002).
Soil texture and clay content have also been estimated from reflectance spectra, based on
the depth of specific clay absorption features (Ben-Dor and Banin 1995b) and statistical
analysis of the whole curve form (Brown et al. 2006, Viscarra Rossel et al. 2006). In a
limited study conducted in South Australia, relationships between soil texture and
laboratory and hyperspectral image spectra from the Barossa Valley region were described
(Ryan and Lewis 2000, 2001).
Early studies observed that increasing soil organic carbon (SOC) lowered albedo across the
whole visible, shortwave infrared and near infrared (Vis-NIR-SWIR) reflectance spectrum
(Stoner and Baumgardner 1981, Henderson et al. 1992). However, there appears to be a
threshold of 2% organic carbon below which the effect of SOC on soil reflectance is
greatly reduced (Baumgardner et al. 1985). SOC has been predicted from various portions
of the Vis-NIR-SWIR largely because it contains so many components. These components
include compounds such as lignin (e.g. 2050, 2351 nm), cellulose (e.g. 1370, 1725, 2347
nm), pectin (e.g. 1320, 1582, 1761, 2111 nm) and humus (e.g. 1929, 1932 nm), which are
2 The SWIR is included in the NIR in some disciplines such as chemistry and reflectance spectroscopy
Chapter 4: Spectral prediction 52
optically active across this spectral region and are thought to overlap in places (Elvidge
1990, Ben-Dor et al. 1997). SOC has been reliably predicted from both laboratory
reflectance spectroscopy and image spectroscopy (Ben-Dor et al. 2002, Daniel et al. 2004).
Some studies have focused on the VIS and NIR regions of the reflectance spectra,
(Krishnan et al. 1980, Vinogradov 1981, Daniel et al. 2003, Brown et al. 2006) whereas
others have used the SWIR region to predict SOC (Morra et al. 1991, Ben-Dor and Banin
1995b, Viscarra Rossel et al. 2006). An Australian study was able to predict SOC from
reflectance spectroscopy in the spectral range 1702 – 2052 nm in a simultaneous
determination of moisture, organic carbon and total nitrogen (Dalal and Henry 1986).
Iron oxide content of soils has been predicted from different spectral regions of the VIS-
NIR-SWIR, based on characteristic absorption features at 550 – 650 nm, 750 – 950 nm
(Ben-Dor and Banin 1995a) and 1406 and 2449 nm (Ben-Dor et al. 2006). The
concentration of iron oxide as wind blown dust on mangrove foliage has been predicted
using features at wavelengths: 518, 746, 927, 1261 and 1402 nm (Ong et al. 2003). Studies
have also found that SOC as low as 1.7% can severely decrease the influence of iron oxide
on the reflectance spectra in the VIS and NIR regions, and particularly decrease the
definition of the 900 nm absorption band (Galvao and Vitorello 1998).
The detection of soil carbonate in soils is complicated by its characteristic absorption
feature shifting to longer and shorter wavelengths depending on the impurities present
(Ben-Dor et al. 1999, Clark 1999). Furthermore, the depths of these spectral features are
dependent not only on the concentrations present but also on particle size and porosity (van
der Meer 1995). Despite this, correlations between absorption feature depth and carbonate
concentration have been established (Ben-Dor and Banin 1990). Correlations have also
been established between carbonate concentration in soil and reflectance spectra based on
changes in colour and albedo, (Ben-Dor and Banin 1995b, Ben-Dor et al. 1999).
The aim of this study was to determine the extent to which high-resolution reflectance
spectra in the visible, near infrared and shortwave infrared regions (400 – 2500 nm) could
be used as an indicator to predict selected surface soil properties. An increasing number of
studies have examined the reflectance properties of soils from temperate, Mediterranean
and tropical regions with moderate to high fertility properties but evidence from low
fertility soils is still sparse. In this study we examine soils from a South Australian region
that has a unique array of profile and landscape characteristics such as low nutritional
Chapter 4: Spectral prediction 53
content and strong texture contrast profile due to extensive weathering, low organic matter
content and a high occurrence of salinity and sodicity. The economic and environmental
importance of understanding variability in landscapes like these is becoming increasingly
accepted and has been highlighted by recent research (Lyle and Ostendorf In Review).
While some previous studies have applied mid-infrared spectroscopy (2500 – 25 000 nm)
to Australian soils (Janik and Skjemstad 1995, Dunn et al. 2002), we examine the optical
visible-near infrared range within which airborne and satellite-based imaging instruments
operate (400 – 2500 nm). The study is a precursor to hyperspectral image mapping of soils
in South Australian agricultural environments. For this reason, we focussed on properties
that are important determinants of soil agricultural capability and the extent to which they
can be simultaneously quantified and predicted from high-resolution reflectance spectra. In
addition, we aimed to identify the spectral regions or features that are most influential in
soil property discrimination, in order to guide future hyperspectral image enhancement and
feature-extraction methodologies. Many of the published spectral analyses of soils have
focussed on single soil properties. Here we address the combined spectral expression of
four key properties that are widely used to assess the agricultural and ecological capability
of soils. Moreover, we examine the proposal that reflectance spectroscopy could be used as
a cost effective means to improve the resolution of soil data for local and regional
inventories. Therefore, we sampled soils to encompass the range of types and variability in
properties that might be encountered in a regional mapping study. Most prior spectral
studies have assembled collections of soils from geographically disparate areas to provide
a wide range of characteristics for analysis. However, as an alternative (or complement) to
traditional soil survey, the methodology needs to be able to predict properties within a
limited region where variation is less pronounced. To further demonstrate the utility of
Vis-NIR reflectance spectroscopy for supplementing soil maps, kriging was used to create
continuous raster layers of the predicted soil properties.
4.2 Methods
4.2.1 Study site and sample collection
Soil samples were collected from the top 2 cm of 300 randomly selected sites in the
Jamestown-Belalie district, approximately 200 km north of Adelaide, South Australia
(Figure 4.1) (33.20611o S, 138.20611o E). The northern third of the study site is dominated
Chapter 4: Spectral prediction 54
by a north-south trending range of hills. A broad valley extends into the south-eastern part
of the study site and is interrupted by another, smaller north-south ridgeline. Several small
ephemeral creeks also traverse the study site, some originating in the hills to the north-east
and some outside the study area and running through the valley. Landuse in this area is
predominantly rain fed cereal cropping in the low lying areas and perennial pasture in the
hills.
Figure 4.1: Jamestown study site, 200 km north of Adelaide, South Australia. Polygons show Common
Soils from the Land and Soil Spatial Data for southern South Australia (Soil and Land Program 2007),
soil sample sites marked with black dots. The legend describes the soil Order from the Australia Soil
Classification (in bold) (Isbell 2002) as well as the soil description from the Land and Soil Spatial Data
for southern South Australia.
Soils have been mapped at 1:100 000 by the Department of Water, Land and Biodiversity
Conservation, South Australia (Soil and Land Program 2007) and are predominantly
Chromosols (Isbell 2002), the key profile characteristic being a strong texture contrast
between A and B horizons. These are described as Xeralfs within the Soil Taxonomy (Soil
Survey Staff 1999). Less widely distributed soils include Dermosols that have structure in
A and B horizons and a gradational texture profile, Calcarosols that have carbonate in the
profile and Rudosols which include shallow skeletal soils on rock. Textures of the B
horizon are often heavy clays that are almost invariably underlain by a carbonate-rich clay
horizon. In higher rainfall areas and some of the ranges there are isolated patches of
Chapter 4: Spectral prediction 55
Kurosols, which are acidic soils with a strong texture contrast between the A and B
horizons.
4.2.2 Laboratory soil analysis
Proportions of clay were calculated from particle size analyses using the hydrometer
method (Gee and Bauder 1986, Sheldrick and Wang 1993). It should be noted that this
methodology calculates clay fraction as determined by size (< 2 µm) and not mineralogy.
Therefore, other fine material (< 2 µm) such as iron oxides and silicates could be measured
in this fraction if it is present in the soil. The calcimeter method (Allison and Moodie 1982,
Nelson and Sommers 1986) was used to measure the carbonate concentration in the soil.
Organic carbon was determined by a modification of the Walkley and Black’s titration
method as outlined by Nelson and Sommers (1986). Iron oxide content was measured by
the sodium dithionate-citrate method (Olson and Roscoe 1986, Ross and Wang 1993).
4.2.3 Reflectance spectra
Prior to spectral measurement, soil samples were air dried in an oven at 60oC for 72 hours
and then passed through a 2mm sieve. Samples were placed in a Petri dish and screeded so
that the entire surface of the soil sample was level with the rim of the dish, thus
guaranteeing a uniform sample depth of 10mm and ensuring that reflectance measurements
recorded the soil surface and not the sample background. Soil spectra were collected using
a FieldSpec Pro spectrometer (Analytical Spectral Devices) that measures reflectance in 3
to 10 nm bandwidths over the range 350 – 2500 nm. A high-intensity contact probe was
used to optimise incidence and reflectance angles, minimise illumination differences and
atmospheric attenuation of the signal and allow for precise identification of the area
sampled. The quality of the spectral measurements was reviewed and noisy portions (350 –
400 nm) of the spectra were removed prior to analysis. The average of ten spectra for each
sample was used in subsequent statistical analysis.
4.2.4 Statistical analysis
The objective of the statistical analysis was to determine whether the reflectance spectra
could be used to predict the chosen soil properties, and to identify the spectral regions
contributing to the prediction. Multiple linear regression is a common multivariate tool
which, at its simplest level, forms a model that specifies the relationship between a
response variable (Y) and a set of dependent variables (X). However, multiple linear
Chapter 4: Spectral prediction 56
regression suffers from some significant limitations, the most important being the
overfitting of data when there are large numbers of highly correlated variables
(significantly more than the number of samples), as is often the case with hyperspectral
reflectance measurements. Partial least squares regression was developed in order to
overcome this limitation (Wold et al. 1983, Otto and Wegscheider 1985), through the
incorporation of aspects of principal components analysis and multiple linear regression.
More specifically, partial least squares regression finds a series of components or latent
vectors that provide a simultaneous reduction or decomposition of X and Y such that these
components explain, as much as is possible, the covariance between X and Y. This step
approximates principal components analysis, although in the latter the components only
explain variation in X and do not necessarily have any bearing on Y. This is then followed
by regression where Y is predicted from the reduction of X (Abdi 2003). The number of
latent vectors are chosen by a process of cross validation which outputs a root mean square
error (RMSE), with the aim of minimising both the number of latent vectors and the
RMSE. Partial least squares regression has been used previously over different spectral
ranges (Vis-NIR-SWIR-MIR) for the prediction of soil properties with varying degrees of
success (Janik et al. 1998, Walvoort and McBratney 2001, McCarty et al. 2002, Cozzolino
and Morón 2003, Ong et al. 2003).
Statistical analysis was carried out using The Unscrambler (Camo Software AS).
Calibration data was mean centred and cross-validation was used to determine the
minimum number of PLS factors required. A large proportion of the samples recorded no
carbonate in the laboratory analysis, with the result that the carbonate distribution amongst
the 290 samples was strongly skewed. To provide a range of values more suitable for
statistical analysis, the data set for carbonate analysis was reduced to 75 by randomly
selecting samples that returned zero carbonate in the laboratory analysis to include in the
statistical analysis along with all of the samples that contained higher carbonate levels.
Cross-validation was carried out using the ‘leave-out-one’ method where one sample is
systematically left out from each cycle of the regression until all the samples have been
excluded once. With different sample numbers for each of the soil properties examined,
this method of validation was chosen to provide for a uniform approach for all of the
analyses.
Chapter 4: Spectral prediction 57
The accuracy of the prediction models was tested with the residual predictive deviation
(RPD) which is the ratio of the standard error of performance to the standard deviation of
the reference data (Williams 2004). Interpretation of the RPD differs amongst authors and
applications. However, it is generally accepted that when applied to the prediction of soil
properties values below 1.5 indicate a poor predictive model, between 1.5 and 2.0 is
acceptable and greater than 2.0 is considered good (Chang et al. 2001, Dunn et al. 2002,
Cozzolino and Morón 2003, Janik et al. 2007). Values below one are considered
inadequate and indicate that the mean of the observed would be a better predictor
(Williams 2004).
4.2.5 Spatial Prediction
The kriging function within the spatial prediction program VESPER (Minasny et al. 2005)
was used to create raster surfaces of the measured and predicted surfaces. Local
variograms were used for clay content, organic matter content and iron oxide content while
low sample density required global variograms were used for carbonate content. Maps
were created with 100 m cell size.
4.3 Results and Discussion
4.3.1 Soil Properties
The percentage of clay in the samples ranged from 5% to 36% (Table 4.1), corresponding
to textural classes loamy sand, sandy loam, loam, silty loam, silty clay loam, clay loam and
clay (McDonald and Isbell 1990). Values for organic carbon were between 0.3 and 2.9%,
carbonate concentrations 0 to 26% and iron oxide concentrations in the range 0.8% to 3%.
Table 4.1: Summary of laboratory results from chemical and physical analysis.
CC
Clay content
(%)
OM
Organic Carbon (%)
IO
Iron oxide
(%)
CO3
Carbonate content
(%)
No. of samples 237 228 229 75
Mean 16.32 1.5 1.5 2.65
Std. deviation 5.42 0.53 0.37 5.37
Minimum 4.97 0.31 0.79 0.0
Maximum 35.98 2.9 3.05 25.67
Chapter 4: Spectral prediction 58
4.3.2 Soil Spectral Characteristics
The overall form of the spectra for all the soils was quite similar. Clay (2200 nm) and
water (1400 nm, 1900 nm) absorption features were present in all spectra while there were
differences in the albedo (intensity) and in the iron oxide (850 – 900 nm) and carbonate
(2300 nm) spectral features amongst the samples.
Figure 4.2 presents mean spectra for each of the quartiles from the laboratory analysis of
clay. The quartiles were determined by dividing the samples into four groups based on
their clay content, with each group containing 25% of the total range. Noteworthy is the
increasing depth of the absorption features at approximately 1400, 1900 and 2200 nm with
increasing clay content. These absorption features are caused by bending and stretching in
the O-H bonds of free water (1400 nm and 1900 nm) and the Al-OH lattice structure in
clay minerals (2200 nm) (Ben-Dor 2002, Viscarra Rossel et al. 2006). Illitic and
montmorillonitic clays dominate the study site area and the nature of the spectra supports
this, as the single symmetrical absorption at 2200 nm is diagnostic for these clays. Other
noticeable differences are evident in the VIS and NIR regions but are likely to be the result
of other factors, such as SOC or iron oxides.
Figure 4.2: Mean spectra of quartiles for percent clay.
Figure 4.3 shows the mean soil spectra of the quartiles from laboratory analysis for SOC.
There is a clear trend with increasing SOC: the spectra have increased slope around 800
nm and lower reflectance across the 400 – 2500 nm spectral range, shifts which have been
observed in other studies (Krishnan et al. 1980, Galvao and Vitorello 1998). In addition to
variation in SOC content, differences in albedo and the slope between 400 nm and 800 nm
Chapter 4: Spectral prediction 59
have been attributed also to the stage of organic carbon (OM) decomposition (Ben-Dor et
al. 1997). The spectra here show increased absorption depth at 2327 nm and 2357 nm,
features which have been attributed to differences in the OM composition (Ben-Dor et al.
1997). Although not investigated here, soils with a higher vegetative load will contain SOC
over a range of decomposition stages.
Figure 4.3: Mean spectra of quartiles for soil organic carbon.
Figure 4.4 depicts the mean spectra for the quartiles of carbonate content. The carbonate
absorption features were slight and limited to one spectral region (2325 nm). Although this
appears to be the only spectral expression of carbonate in our samples, previous studies
have used a range of wavelengths (1800 nm, 2350 nm and 2360 nm) to predict calcite in
soils (Ben-Dor and Banin 1990). Other spectral variations amongst our samples can only
be attributed to other soil properties. The iron oxide quartiles in Figure 4.5 demonstrate
increasing definition of the iron oxide features in the VIS-NIR. As the iron oxide
concentration increases, there is an increase in depth of absorption from 400 nm to 550 nm
and in the broad feature at 900 nm indicating that goethite dominates the samples rather
than hematite.
Chapter 4: Spectral prediction 60
Figure 4.4: Mean spectra of quartiles for carbonate concentration.
Figure 4.5: Mean spectra of quartiles for iron oxide content.
4.3.3 Prediction of Soil Properties
Table 4.2 presents the efficiency criterion (E), root mean square error (RMSE) and
regression coefficients (R2) obtained from each partial least squares analysis. The first two
PLS loading weights for each analysis in Figures 6, 7, 8 and 9 demonstrate the relative
importance of spectral regions in the prediction of each of the soil properties. Negative
peaks in the loading weight graphs indicate spectral regions that correlate positively with
the prediction and positive peaks are those areas that correlate negatively with the
prediction.
Chapter 4: Spectral prediction 61
Table 4.2: Sample numbers, residual predictive deviation (RPD), root mean square error (RMSE) and
R2 results for data sets.
Soil property Samples
(n) Factors RPD
RMSE
(%) R2
CC (%) 237 10 2.0 3.13 0.66
OC (%) 228 8 1.8 0.35 0.57
IO (%) 229 10 1.7 0.23 0.61
CO3 (%) 75 5 2.1 2.90 0.69
With ten prediction factors or latent vectors selected for the analysis (Table 4.2), 66% of
the variation in clay content was explained by the partial least squares regression model,
returning a RMSE of 3.13. An RPD of 2.0 indicates that the prediction was acceptable and
substantially better predictor than the mean of the observed clay contents. In Figure 4.6 the
first loading weight (PC1) is dominated by the clay absorption feature at 2200 nm and the
features at 1400 and 1900 nm. These three features are all related to the bending and
stretching of O-H bonds in the lattice minerals and water molecules, directly and indirectly
associated with the clay minerals. The 2200 nm region is specifically related to the
symmetric absorption feature that is diagnostic of the illite and montmorillonite that
dominate the clays in these soils. For all these spectral regions, increasing clay content
would result in more pronounced absorption features. These spectral regions were also
discriminants for field textural classes in soils from the same geographic region (Summers
et al. 2005). The second loading weight was dominated by the visible (400 – 700 nm) and
a portion of near infrared region (700 – 1300 nm) with some contribution from the same
regions as observed in the first loading weight. The importance of the visible spectral range
in this result indicates that there may be some co-variation between the clay content and
the colour of the soil. There is also a strong influence in the first and second loading
weights, starting at 2300 nm and increasing in contribution through to 2500 nm. This is the
initial stages of a water absorption feature that continues out of range to 2800 nm.
Chapter 4: Spectral prediction 62
Figure 4.6: Spectral loading weight graph for the prediction of clay content.
The analysis explained 57% of the variation (Table 4.2) in SOC using eight prediction
factors with an RMSE of 0.35. The RPD value of 1.8 is evidence of an acceptable model
although it could be improved with different calibration strategies (Chang et al. 2001). The
first loading weight was dominated by a relatively broad region extending from 550 nm in
the visible to 1000 nm in the NIR, with a maximum contribution near 700 nm (Figure 4.7).
Increased SOC generally produces visibly darker soils and it is likely that this contributed
to the prediction here. The second loading weight is dominated by a couple of peaks at
2100 and 2300 nm. Other studies have found these spectral regions to be associated with
lignin and humic acids and important in the prediction of SOC (Ben-Dor et al. 1997).
Figure 4.7: Spectral loading weight graph for the prediction of soil organic carbon content.
Substantially fewer samples were available for the carbonate analysis than for the other
soil properties (Table 4.1), but the coefficient of determination was the highest for all the
soil properties in the study (0.69) using 5 prediction factors (Table 4.2). The analysis also
returned an acceptable RPD value (2.1) and a reasonable RMSE (2.9) (Table 4.2). The first
loading weight (Figure 4.8) is dominated by a peak at 2300 nm which is directly associated
with carbonate in reflectance spectra. There is also some influence from a peak at 1900 nm
Chapter 4: Spectral prediction 63
that extends into a ‘plateau’ to around 2100 nm. The second loading weight shows a broad
peak from 600 to 1100 nm indicating some influence from red visible wavelengths to the
near infrared range. A previous study of a similar set of soils also found that the visible
region contributed to discrimination of carbonate classes but that the discrimination was
dominated by absorption features associated with water (1900 nm), clay (2200 nm) and
carbonate (2300 nm) (Summers et al. 2005).
Figure 4.8: Spectral loading weight graph for the prediction of carbonate content.
The prediction of iron oxide explained 61% of the variability in the samples using ten
prediction factors with an RMSE of 0.23 (Table 4.2). The RPD value was 1.7, which is the
lowest of all the soil properties examined in this study although still within the acceptable
range. The first loading weight (Figure 4.9) shows the range from 400 to 1100 nm to be
most influential in the prediction. Within this range there are two maximum ‘peaks’ one at
550 nm and one at 900 nm, both regions associated with spectral characteristics of iron
oxide species. The second loading weight is dominated by a portion (400 – 550 nm) of the
visible range, associated with the blue and green, and peaks at 1900, 2200 and 2300 nm,
associated with water, clay and carbonate respectively.
Figure 4.9: Spectral loading weight graph for the prediction of iron oxide content.
Chapter 4: Spectral prediction 64
4.3.4 Mapping of Predicted Soil Properties
The kriged geographic distributions of the measured and predicted soil properties are
displayed in Figure 4.10. Comparison between the measured and predicted soil properties
demonstrate similar patterns and value ranges for the soil properties examined. The two
maps of clay content show lower values in the hills towards the north-east and in the sandy
soils of the south-west although overall the area has limited variability. The most
substantial difference between the two maps is in the centre where the predicted map
demonstrates less variability. Organic carbon shows the greatest variation between the
measured and predicted maps of all of the soil properties examined. However despite that,
the overall pattern between the two maps is consistent. In both maps organic carbon
content is lower in the valley areas which are dominated by cropping, while in the hills,
which are predominantly pasture, there is a build up of organic carbon. The small band of
sandy soils in the south-east has unexpectedly high organic carbon contents although this
too could be the product of the pasture and forestry landuse in that area.
Figure 4.10: Spatial distribution of measured and predicted soil properties following Kriging.
Chapter 4: Spectral prediction 65
There are some minor differences between the measured and predicted carbonate maps,
however the same trend is evident in both. The central portion of each map shows higher
carbonate contents, particularly along the southern edge. Both the measured and predicted
carbonate maps correspond well with polygons classed as calcareous within the Land and
Soil Spatial Data (Figure 4.1). The iron oxide maps show a good match between the
measured and predicted soil properties, each demonstrating the same pattern and with a
few small differences in the centre of the map. The area is dominated by red-brown earths
and predicted iron oxide content reflects this with a relatively high and even distribution
across the study site. The lower iron oxide levels in the south-west corner are associated
with the small band of sandy soils found there.
4.4 Conclusion
Visible–near infrared reflectance spectra collected under controlled laboratory conditions
were employed as an indicator for the prediction of selected soil properties. Partial least
squares regression overcame the collinearity problems associated with large numbers of
highly correlated variables and relatively small sample numbers. We have shown that it is
possible to predict clay content, soil organic carbon, iron oxide content and carbonate
content from reflectance data produced with a high-resolution laboratory spectrometer.
Furthermore, all of the samples were collected from the same geographical area in order to
test prediction of soil properties over a naturally occurring range and provide a prediction
that can be related to a regional image analysis. The predicted soil properties have also
been examined geographically in relation to existing soil maps with some discussion of
how they relate to the landscape and the usefulness of the method in future soil mapping
projects. However, it should be noted that recalibration of PLS predictive functions would
be required for different soil types and mineralogy.
Carbonate and clay content were best predicted followed by iron oxide and organic carbon.
Validation R2 for all analyses was above 0.5 and the RPD was acceptable for all soil
properties. We showed the utility of particular regions of the 400 – 2500 nm spectrum for
prediction of clay content (1900 and 2200 nm), SOC (600 – 900 nm), iron oxides (400 –
1100 nm) and carbonate (1900 – 2300 nm). This demonstrates the ability to use this
methodology as an indicator for rapid and reliable soil mapping. Laboratory analyses of
soil samples in support of traditional survey methods are expensive and time consuming.
Field and laboratory measurement potentially offers a rapid, cost effective method for
Chapter 4: Spectral prediction 66
prediction of soil properties. Such studies could also be expanded to include the analysis of
whole profiles and provide a more comprehensive understanding of the solum. Moreover,
the results from this study can inform subsequent image studies which would allow the
application of similar and related methodologies to spatially continuous remotely sensed
imagery. However, further studies on different soils are required to confirm the efficiency
of these predictors as indicators of soil properties and variability.
The Land and Soil Spatial Data in this region is produced at a relatively broad scale (1:100
000) and soil units are depicted with discrete polygons units. This provides a valuable
regional planning tool but lacks the spatial resolution for finer scale applications. For
example, the soil properties represented in any one polygon are, in some cases, only 50%
reliable (Soil and Land Program 2007). This is largely the result of scale and the absence
of soil variability depiction within polygons. The predictions of soil properties show that
reflectance spectroscopy could be used to improve the spatial resolution of soil inventories
such as these. Furthermore, we have demonstrated how simple kriging can be used to
create a raster maps of the predicted soil properties and that these maps are comparable to
the measured soil properties. It should be noted that there is room to improve the prediction
accuracy of the reflectance spectroscopy in this study and achieving higher accuracy would
benefit any soil survey carried out with these techniques. However, the improved spatial
resolution available from greater sampling density at reduced costs could counteract some
of the expected error.
While this study examines only surface soils, the spectral methodology would need to be
extended to the profile to fully supplement traditional soil survey. Vis-NIR reflectance
spectroscopy has been successfully used to catalogue and classify geological cores and in
situ soil profiles (Mauger et al. 2004, Ben-Dor et al. 2008) and a combination of those
techniques with the ones used here could provide a new methodology for complete
description of the soil profile. These combined methodologies could be used to supplement
traditional soil survey with the aim of improving the resolution of current soil mapping
programs and to expand soil mapping to areas that are currently excluded due to economic
imperatives such as arid and pastoral zones. It is also possible that the sampling density
could be increased to the point where raster based maps could be produced at reliably fine
scales.
Chapter 4: Spectral prediction 67
4.5 References
Abdi, H. 2003 Partial least squares (PLS) regression, In Encyclopaedia for research methods for the social
sciences (Eds, M. Lewis-Beck, A. Bryman and T. Futing) Sage Publications, Thousand Oaks (CA).
Allison, L. E. and Moodie, C. D. 1982 Carbonate, In Methods of soil analysis (Ed, A. L. Page) Soil Science Society of America, Madison, pp. 1379-1396.
Analytical Spectral Devices 2001 ASD Field Spec. Pro., Colorado.
Baumgardner, M. F., Silva, L. R., Biehl, L. L. and Stoner, E. R. 1985 Reflectance properties of soils, Advances in Agronomy, 38, 1-44.
Ben-Dor, E. 2002 Quantitative remote sensing of soil properties, Advances in Agronomy, 75, 173-243.
Ben-Dor, E. and Banin, A. 1990 Near-infrared reflectance analysis of carbonate concentration in soils, Society for Applied Spectroscopy, 44, 1064-1069.
Ben-Dor, E. and Banin, A. 1995a Near-infrared analysis (NIRA) as a method to simultaneously evaluate spectral featureless constituents in soils, Soil Science, 159, 259-270.
Ben-Dor, E. and Banin, A. 1995b Near-infrared analysis as a rapid method to simultaneously evaluate several soil properties, Soil Science Society of America Journal, 59, 364-372.
Ben-Dor, E., Carmina, K., Heller, D. and Chudnovsky, S. 2008 A novel combined optical method for (sic) objectively map soil in a near real time domain, In The 21st Congress of the International Society for Photogrammetry and Remote Sensing, Beijing, China, 3-11 July 2008.
Ben-Dor, E., Inbar, Y. and Chen, Y. 1997 The reflectance spectra of organic matter in the visible near-infrared and short wave infrared region (400-2500 nm) during a controlled decomposition process, Remote
Sensing of Environment, 61, 1-15.
Ben-Dor, E., Irons, J. R. and Epema, G. F. 1999 Soil reflectance, In Remote sensing of the earth sciences:
Manual of remote sensing, Vol. 3 (Ed, A. N. Rencz) John Wiley and Sons, New York, pp. 111-188.
Ben-Dor, E., Levin, N., Singer, A., Karnieli, A., Braun, O. and Kidron, G. J. 2006 Quantitative mapping of the soil rubification process on sand dunes using an airborne hyperspectral sensor, Geoderma, 131, 1-21.
Ben-Dor, E., Patkin, K., Banin, A. and Karnieli, A. 2002 Mapping of several soil properties using DAIS-7915 hyperspectral scanner data - a case study over clayey soils in Israel, International Journal of Remote
Sensing, 23, 1043-1062.
Bock, M., Rossner, G., Wissen, M., Remm, K., Langanke, T., Lang, S., Klug, H., Blaschke, T. and Vrscaj, B. 2005 Spatial indicators for nature conservation from European to local scale, Ecological Indicators, 5, 322-338.
Bower, T. L. and Rowan, L. C. 1996 Remote mineralogic and lithological mapping of the Ice River alkaline complex, British Columbia, Canada, using AVIRIS data, Photogrammetric Engineering and Remote Sensing, 62, 1376-1143.
Brown, D. J., Shepherd, K. D., Walsh, M. G., Dewayne Mays, M. and Reinsch, T. G. 2006 Global soil characterization with VNIR diffuse reflectance spectroscopy, Geoderma, 132, 273-290.
Camo Software AS 2007 The Unscrambler 9.7, Oslo, Norway.
Chang, C., Laird, D. A., Mausbach, M. J. and Hurburgh, C. R., Jr. 2001 Near-infrared reflectance spectroscopy-principal components regression analyses of soil properties, Soil Sci Soc Am J, 65, 480-490.
Clark, R. N. 1999 Spectroscopy of rocks and minerals and principles of spectroscopy, In Remote sensing for
the earth sciences: Manual of remote sensing, Vol. 3 (Ed, A. N. Rencz) John Wiley and Sons, New York, pp. 3-58.
Clarke, K. D. 2008 Landscape-scale measurement and monitoring of biodiversity in the Australian
rangelands, PhD Thesis, University of Adelaide, Adelaide, South Australia.
Condit, H. R. 1970 The spectral reflectance of American soils, Photogrammetric Engineering, 36, 955-966.
Cozzolino, D. and Morón, A. 2003 The potential of near-infrared reflectance spectroscopy to analyse soil chemical and physical characteristics, Journal of Agricultural Science, 140, 65-71.
Chapter 4: Spectral prediction 68
Dalal, R. C. and Henry, R. J. 1986 Simultaneous determination of moisture, organic carbon and total nitrogen by near infrared reflectance spectroscopy, Soil Science Society of America Journal, 50, 120-123.
Daniel, K. W., Tripathi, N. K. and Honda, K. 2003 Artificial neural network analysis of laboratory and in situ spectra for the estimation of macronutrients in soils of Lop Buri (Thailand), Australian Journal of Soil
Research, 41, 47-59.
Daniel, K. W., Tripathi, N. K., Honda, K. and Apisit, E. 2004 Analysis of VNIR (400-1100nm) spectral signatures for estimation of soil organic matter in tropical soils of Thailand, International Journal of Remote
Sensing, 25, 643-652.
Dunn, B. W., Beecher, H. G., Batten, G. D. and Ciavarella, S. 2002 The potential for near-reflectance spectroscopy for soil analysis - a case study from the riverine plain of south-eastern Australia, Australian
Journal of Experimental Agriculture, 42, 607-614.
Elvidge, C. D. 1990 Visible and near infrared reflectance characteristics of dry plant materials, International
Journal of Remote Sensing, 11, 1775-1795.
Galvao, L. S. and Vitorello, I. 1998 Role of organic matter in obliterating the effects of iron on spectral reflectance and colour of Brazilian tropical soils, International Journal of Remote Sensing, 19, 1969-1979.
Gee, G. W. and Bauder, J. W. 1986 Particle-size analysis, In Methods of soil analysis, part 1. Physical and
mineralogical methods, Vol. 2 (Ed, A. Klute) American Society of Agronomy - Soil Science Society of America, Madison, Wisconsin, USA, pp. 383-409.
Henderson, T. L., Baumgardner, M. F., Franzmeier, D. P., Stott, D. E. and Coster, D. C. 1992 High dimensional reflectance analysis of soil organic matter, Soil Science Society of America Journal, 56, 865-872.
Isbell, R. F. 2002 The Australian soil classification, CSIRO Australia, Melbourne.
Islam, K., Singh, B. and McBratney, A. 2003 Simultaneous estimation of several soil properties by ultra-violet, visible and near-infrared reflectance spectroscopy, Australian Journal of Soil Research, 41, 1101-1114.
Janik, L. J., Merry, R. H., Forrester, S. T., Lanyon, D. M. and Rawson, A. 2007 Rapid prediction of soil water retention using mid infrared spectroscopy, Soil Science Society of America Journal, 71, 507-514.
Janik, L. J., Merry, R. H. and Skjemstad, J. O. 1998 Can mid infrared diffuse reflectance analysis replace soil extractions?, Australian Journal of Experimental Agriculture, 38, 681-696.
Janik, L. J. and Skjemstad, J. O. 1995 Characterisation and analysis of soils using mid-infrared partial least squares: II. Correlations with laboratory data, Australian Journal of Soil Research, 33, 637-650.
Krishnan, P., Alexander, J. D., Butler, B. J. and Hummel, J. W. 1980 Reflectance technique for predicting soil organic matter, Soil Science Society of America Journal, 44, 1282-1285.
Lewis, M. 2000 Discrimination of arid vegetation composition with high resolution CASI imagery, Rangeland Journal, 22, 141-167.
Lyle, G. and Ostendorf, B. In Review A high-resolution spatial indicator of economic performance in the grain growing regions of Australia, Ecological Indicators.
Mauger, A. J., Keeling, J. L. and Huntington, J. F. 2004 Bringing remote sensing down to earth: CSIRO Hylogger as applied to the Tarcoola goldfield, South Australia, In 12 Australasian Remote Sensing and Photogrammetry Conference, Fremantle, Western Australia,
McBratney, A. B., Minasny, B., Cattle, S. R. and Vervoort, R. W. 2002 From pedotransfer functions to soil inference systems, Geoderma, 109, 41-73.
McCarty, G. W., Reeves, J. B., III, Reeves, V. B., Follett, R. F. and Kimble, J. M. 2002 Mid-infrared and near-infrared diffuse reflectance spectroscopy for soil carbon measurement, Soil Science Society of America
Journal, 66, 640-646.
McDonald, R. C. and Isbell, R. F. 1990 Soil profile, In Australian soil and land survey: Field handbook (Eds, R. C. McDonald, R. F. Isbell, J. G. Speight, J. Walker and M. S. Hopkins) Inkata Press, Melbourne.
Miller, T. F., Mladenoff, D. J. and Clayton, M. K. 2002 Old-growth northern hardwood forests: Spatial autocorrelation and patterns of understory vegetation, Ecological Monographs, 72, 487-503.
Chapter 4: Spectral prediction 69
Minasny, B., McBratney, A. and Whelan, B. M. 2005 Vesper version 1.62, Australian Centre for Precision Agriculture, University of Sydney, Sydney, Australia.
Morra, M. J., Hall, M. H. and Freeborn, L. L. 1991 Carbon and nitrogen analysis of soil fractions using near-infrared reflectance spectroscopy, Soil Science Society of America Journal, 55, 288-291.
Nelson, D. W. and Sommers, L. E. 1986 Total carbon, organic carbon and organic matter, In Methods of soil
analysis, Vol. 2 (Ed, A. L. Page) Soil Science Society of America, Madison, pp. 53-579.
Olson, R. V. and Roscoe, E. 1986 Iron, In Methods of soil analysis, Vol. 2 (Ed, A. L. Page) Soil Science Society of America, Madison, pp. 301-312.
Ong, C. C. H., Cudahy, T. J., Caccetta, M. S. and Piggott, M. S. 2003 Deriving quantitative dust measurements related to iron ore handling from airborne hyperspectral data, Mining Technology, 112, 158-163.
Otto, M. and Wegscheider, W. 1985 Spectrophotometric multicomponent analysis applied to trace metal determinations, Analytical Chemistry, 57, 63-69.
Ross, G. L. and Wang, C. 1993 Extractable Al, Fe, Mn, and Si, In Soil sampling and methods of analysis (Ed, M. R. Carter) Lewis Publishers, Boca Raton, pp. 239-246.
Ryan, S. and Lewis, M. 2000 Discrimination and mapping soils using HyMap hyperspectral imagery, Barossa valley, S.A., In 10th Australasian Remote Sensing and Photogrammetry Conference, Adelaide,
Ryan, S. and Lewis, M. 2001 Mapping soils using high resolution airborne imagery, Barossa valley, S.A., In Inaugural Australian Geospatial Information and Agriculture Conference Incorporating Precision Agriculture in Australasia 5th Annual Symposium, Sydney, NSW, 17-19 July.
Sheldrick, B. H. and Wang, C. 1993 Particle size distribution, In Soil sampling and methods of analysis (Ed, M. R. Carter) Lewis Publishers, Boca Raton.
Soil and Land Program 2007 Land and soil spatial data for southern South Australia - GIS format, Department of Water, Land and Biodiversity Conservation, South Australia, Accessed
Soil Survey Staff 1999 Soil taxonomy: A basic system of classification for making and interpreting soil
surveys, U.S. Government Print Office, Washington D.C.
Stoner, E. R. and Baumgardner, M. F. 1981 Characteristic variation in reflectance of surface soils, Soil
Science Society of American Journal, 45, 1161-1165.
Sumfleth, K. and Duttmann, R. 2007 Prediction of soil property distribution in paddy soil landscapes using terrain data and satellite information as indicators, Ecological Indicators, 8, 485-501.
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2005 Spectral discrimination of soil properties, In SSC 2005 Spatial Intelligence, Innovation and Praxis: The National Biennial Conference of the Spatial Sciences Institute, Melbourne, Australia, September, 2005.
Tiner, R. W. 2004 Remotely-sensed indicators for monitoring the general condition of "Natural habitat" In watersheds: An application for Delaware's Nanticoke river watershed, Ecological Indicators, 4, 227-243.
van der Meer, F. 1995 Spectral reflectance of carbonate mineral mixtures and bidirectional reflectance theory: Quantitative analysis for application in remote sensing, Remote Sensing Review, 13, 67-94.
Vinogradov, B. V. 1981 Remote sensing of the humus content of soils, Soviet Soil Science, 11, 114-123.
Viscarra Rossel, R. A., Walvoort, D. J. J., McBratney, A. B., Janik, L. J. and Skjemstad, J. O. 2006 Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties, Geoderma, 131, 59-75.
Walvoort, D. J. J. and McBratney, A. 2001 Diffuse reflectance spectrometry as a proximal sensing tool for precision agriculture., In Third European Conference on Precision Agriculture, Montpellier,
Williams, P. 2004 Implementation of near-infrared technology, In Near-infrared technology in the
agricultural and food industries (Eds, P. Williams and K. Norris) American Association of Cereal Chemists, Inc., St. Paul, Minnesota.
Wold, S., Albano, C., Dunn, W. J., Esbensen, K., Hellberg, S., Johansson, E. and Sjöström 1983 Pattern recognition: Finding and using regularities in multivariate data, In Food research and data analysis (Eds, H. Martens and H. Russwurm) Applied Science Publishers, Essex, England, pp. 147-188.
70
Chapter 5
Unmixing of soil types and estimation of soil exposure
with simulated hyperspectral imagery
Submitted for journal publication:
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. (In Review) Unmixing of soil types and estimation of soil exposure with simulated hyperspectral imagery, International Journal of Remote Sensing.
5.1 Introduction
Remote sensing is useful for mapping and monitoring many environmental features
including geology, minerals (Bower and Rowan 1996, Clark 1999), vegetation and soil
(Lewis 2000, Ben-Dor et al. 2002, Sumfleth and Duttmann 2007) and even ecological
habitats (Tiner 2004, Bock et al. 2005). With increasing sophistication of field and imaging
spectrometers, there is potential for substantial improvement in the speed, reliability and
resolution of inventory and monitoring of natural and agricultural systems. New sensors
offer the prospect of detailed raster-based mapping of land surface characteristics with
higher spatial resolution and variation than is possible with the current approaches.
Some hyperspectral image studies to discriminate and map soils in agricultural regions
have been conducted (Ben-Dor et al. 2002, Dehaan and Taylor 2003, Dutkiewicz et al.
2003, Taylor 2004), but many of them suffer from a common limitation. Unless the land is
fallow or recently ploughed, some degree of vegetation, either actively growing or as crop
residue, obscures the soil from the imaging instrument (Metternicht and Zinck 2003). In
Australia the current best practice in croplands employs a minimum tillage regime that,
where possible, minimises soil disturbance. Thus, under a well-managed agricultural
system there is little exposure of soil to allow for unobscured remote sensing.
Studies aiming to map soil types in situations with partial vegetation cover typically use
spectral unmixing methods to identify materials in mixed pixels (Asner and Heidebrecht
2002, Alemie 2005, Lu and Weng 2007, Zhang et al. 2007). Linear mixture analysis is
based on the assumption that the spectrum of a pixel is a weighted linear combination of
Chapter 5: Spectral unmixing with simulated data 71
the spectra of materials within the instantaneous field of view; spectral contributions for
the different materials are proportional to their abundances (Settle and Drake 1993,
Dennison and Roberts 2003). Endmembers, or spectra of ‘pure’ materials in the image,
theoretically representative of all materials in a scene, are used as inputs into the unmixing
process. Such methods result in estimations of fractional abundance in the form of a
greyscale image for each input spectrum. Errors in the unmixing occur when the number of
endmembers approach the spectral dimensions of the image, when endmembers are poorly
selected and not sufficiently distinct from one another, or are not sufficiently representative
of materials in the image (Malenovsky et al. 2007).
Differentiating soils and non-photosynthetic plant residue is difficult because of the
spectral similarity of the two materials (Daughtry 2001, Nagler et al. 2003). Photosynthetic
vegetation has a unique spectral signature in the visible and near-infrared (400 – 1000 nm)
that is not present in non-photosynthetic vegetation, making it much easier to differentiate
(Daughtry et al. 2005, Daughtry et al. 2006). Furthermore, studies have found variable
responses from the mixtures of different soils with the same cover type (Nagler et al.
2003). Photosynthetic vegetation cover under 30%, as typically found in arid and semiarid
regions, appears to have little effect on the determination of soil type from hyperspectral
data, but increasing plant cover severely limits the ability to accurately model soil and its
exposure (Okin et al. 2001). In addition, spectral confusion may occur when mixtures of
soil and vegetation cover mimic the spectral characteristics of some soil types with no
vegetation cover, or where the same level of plant cover on different soils produces
variable spectral responses (Okin et al. 2001).
To fully utilise hyperspectral imagery for soil studies there is a need to understand the
combined reflectance of both the soil and the cover materials as well as the pure
endmember spectra. The spectral response of a soil, even with a well defined spectral
expression, is a function of the physical constituents as well as the exposure of the soil to
the sensor. In situations with partial soil exposure variations in spectral response of, for
example, the depth of the clay absorption feature can be attributed to varying clay contents
as well as differing proportions of soil exposure. Research is needed to clarify the
influences of variable plant cover on spectral sensing of different soil types, and to identify
limits to the detection different soil types under these conditions.
Chapter 5: Spectral unmixing with simulated data 72
Laboratory-based reflectance spectroscopy is an increasingly popular method for soil
sample analysis and has the potential to greatly improve speed of measurement. Most
successful prediction of soil properties, has been based upon high spectral resolution
reflectance spectroscopy of prepared samples in the laboratory or exposed soils in the field
(Viscarra Rossel et al. 2006b). These studies have included the spectral ranges of ultra
violet (UV) (200 – 400 nm), visible (vis) (400 – 700 nm), near infrared (NIR) (700 – 1300
nm), short wave infrared (SWIR) (1300 – 2500 nm) and mid infrared (2500 – 25 000 nm)
and in some cases different combinations of these ranges (McCarty et al. 2002, Cozzolino
and Morón 2003, Islam et al. 2003, Viscarra Rossel et al. 2006a). It should be noted that
SWIR is a remote sensing term and this range is typically included in the NIR in studies
relating to reflectance spectroscopy.
Unlike image based remote sensing conducted from airborne and satellite platforms,
samples are generally small, discrete units that are examined in the laboratory, often after
some form of preparation. The illumination of samples is achieved with an active source;
for visible near-infrared analysis this is typically a halogen light. Like remote sensing,
reflectance spectroscopy allows for the rapid examination of materials, and in the case of
soil analysis, eliminates much of the laboratory work usually associated with conventional
measurement. Reflectance spectroscopy also eliminates many of the complications
associated with remote sensing, such as atmospheric attenuation. Studies have found it
useful for the determination of soil properties including clay content, carbonate, organic
matter, iron oxide, cation exchange, pH and many more (Janik and Skjemstad 1995,
McCarty et al. 2002, Cozzolino and Morón 2003, Viscarra Rossel et al. 2006a).
Reflectance spectroscopy analysis is typically performed on isolated samples, and the
applicability of findings to imagery is, in some cases, limited. However, some studies have
attempted to apply laboratory spectra to problems encountered with image based remote
sensing. Differentiation and quantification of soil, vegetation and crop residue has been
carried out using laboratory reflectance spectra of material combinations within controlled
experiments. These studies have used wavelength-specific vegetation indices, including the
normalised difference vegetation index (NDVI), soil adjusted vegetation index (SAVI) and
the cellulose absorption index (Nagler et al. 2000, Daughtry 2001, Nagler et al. 2003).
Furthermore, laboratory-measured spectra have also been used to create artificial images
for testing and retrieval of spectral mineral components using image analyses such as
Chapter 5: Spectral unmixing with simulated data 73
spectral mixture analysis (Hussey 1998). These laboratory based techniques and artificial
image analyses have been useful for evaluating and comparing analytical techniques,
allowing determination of pixel composition and accuracy of results without the
requirements of extensive field work.
This study examined the vis-NIR-SWIR (400 – 2400 nm) spectral expression of different
mixes of vegetation cover and surface soils from a southern Australian agricultural region
and the ability to distinguish material abundances and soil types with spectral unmixing. In
particular the study aimed to examine the extent to which soil exposure could be reliably
quantified from variable mixes of soils with photosynthetic and non-photosynthetic
vegetation cover. Furthermore, we aimed to examine the influence of soil spectral
characteristics on the estimation of abundance and the degree to which different soil types
can be isolated from mixed and pure pixels using linear mixture analysis.
The ability to accurately estimate soil exposure and identify soil types was evaluated
through linear unmixing of spectra derived from controlled mixes of four different soils
and three plant cover types. The linear mixture analysis was applied to two types of
artificial hyperspectral imagery: a ‘laboratory image’, created from physical mixes of
various soils with different vegetation, and a ‘virtual image’ created by weighted linear
combinations of pure soil and vegetation spectra. The virtual image was seen as a control
for the spectra of physical mixtures of soil and vegetation, in that the mixing proportions
were precisely known and the mixture of spectra was strictly linear. Further to this,
comparison of the virtual and laboratory image results was included to determine the utility
of ‘virtual’ images in future investigations.
5.2 Materials and methods
5.2.1 Soil and vegetation samples
Four soils, two photosynthetic vegetation types and a non-photosynthetic crop residue were
chosen to simulate the range of soils and cover types commonly found in natural and
agricultural settings in southern Australia.
Soils for the study, each with differing physical and chemical properties, were
representative of surface horizons from the Monarto agricultural region, 50 km east of
Adelaide, South Australia. This region consists mostly of Chromosols and Calcarosols in
Chapter 5: Spectral unmixing with simulated data 74
the Australian Soil Classification (Isbell 2002) which translate roughly as Xeralfs and
Calciargids or Calciorthids respectively in the Soil Taxonomy (Soil Survey Staff 1999).
The soils were a sodic clay, loam, silty loam and a clay loam. Samples were analysed for
particle size fractions (Gee and Bauder 1986) carbonate content (Allison and Moodie
1982), organic carbon content (Nelson and Sommers 1986) and free iron oxide content
(Ross and Wang 1993) (Table 5.1).
Foliage from a native Australian Eucalyptus tree (Spotted Gum, Eucalyptus maculata H.)
and a perennial horticultural tree (orange, Citrus sinensis L.) was used to provide
photosynthetic vegetation cover for the experiment, while dry crop residues of agricultural
field pea (Pisum sativum L.) provided samples of non-photosynthetic vegetation. These
vegetation types were chosen to represent native vegetation (eucalyptus), irrigated
horticulture (orange) and the most prevalent non-photosynthetic material in southern
Australian broad acre agricultural landscapes (crop residue).
Table 5.1: Laboratory measured soil properties of four soils used in the study.
Soil Clay
(%)
Carbonate (%)
Iron Oxide
(%)
Organic Carbon (%)
Munsell Soil Colour
Sodic Clay 32.2 0.2 1.1 0.2 5 YR 5/6
Loam 11.4 10.7 0.6 1.6 10 YR 5/3
Silty Loam 18.8 23.3 0.7 0.6 2.5 Y 7/2
Clay Loam 29.2 0.0 0.7 1.5 10 YR 3/3
5.2.2 Collection of spectra and image creation
Prior to spectral collection, soils were air-dried and sieved to 2 mm. Soil samples were
placed in a 200 x 100 x 20 mm tray and the soil surface was screeded level to the rim of
the tray to provide a uniform soil depth of 20 mm. Air drying and sieving the soil removed
some of the complexities that would be encountered in a traditional image scene analysis
enabling the study to focus on the spectral distinction of endmembers. Furthermore,
replicating the micro-variability of soil properties such as micro topography and surface
crusting was deemed unrealistic for a laboratory experiment such as this. Also, most soil
image studies carried out in southern Australia would focus on data collection at times of
peak soil exposure when the soil is very dry.
Spectra of the orange and Eucalyptus foliage were measured within one hour of collection,
while the dry field pea was collected several days prior. Eucalyptus and orange leaves were
Chapter 5: Spectral unmixing with simulated data 75
placed three layers deep and overlapping the soil surface to cover an area of 100 cm2. This
was found sufficient to prevent light transmission through the leaves and thus deemed
adequate to ensure no spectral interference from the soil below. Field pea stalks were glued
together on a thin piece of plastic to cover an area of 100 cm2: both the plastic and the glue
were tested and found to be spectrally featureless. The stalks were placed in layers to
provide complete coverage and prevent the transmission of light through the stalks to
ensure spectral sampling was not influenced by the underlying soil. Samples of plant and
residue cover were placed in small troughs within the soil to ensure a level sampling
surface between the different materials. Replicating variations in leaf orientation was not
attempted, partly to focus on the spectral characteristics but also as a practical measure.
The orientation of field pea stalks did mimic that of crop residue in a typical agricultural
environment in southern Australia at the end of summer.
Spectral measurements were made with an Analytical Spectral Devices FieldSpec Pro
Spectrometer, a 2150 band sensor which collects data between 350-2500 nm. It has a
sampling interval of 1.4 nm in the 350 -1100 nm range (FWHM = 3 nm) and of 2 nm in the
1000 – 2500 nm range (FWHM = 10 -12 nm). The spectrometer was calibrated against a
white Spectralon reference panel to prevent drift and ensure consistency across
measurements. Wavebands below 400 nm were considered noisy and removed, reducing
the number of bands to 2101, spanning the range 400-2500 nm. Each spectrum used in the
analysis was the averaged combination of 10 measurements collected with the
spectrometer. An ASD high intensity reflectance probe (A122000) with internal halogen
lamp was used for data collection. This probe is configured with the optical fibre
approximately 20o off nadir and 60 mm above the sample, and the halogen lamp in the
nadir position. The probe was fitted with a field of view (FOV) limiter which provided a
sample spot size of 30 mm diameter.
To collect spectra for the laboratory image the probe was placed in a clamp so that the field
of view was filled by the plant cover in the tray. The tray was then moved incrementally,
so that the soil exposure in each field of view increased in 10% increments from 0% to
100% (Figure 5.1). As the FOV is circular, the linear distances moved for each 10%
increase in area are not the same. The distance the tray was moved for each 10% increase
in area was determined using Newton’s Method (Kelley 2003). Thin pieces of wood, each
a specific width corresponding to the different distances required for 10% increases in area,
Chapter 5: Spectral unmixing with simulated data 76
were placed between the tray and a solid stop at the appropriate intervals. Through this
collection method it was expected that spectral mixing would be purely linear. This was
repeated for each of the three vegetation types over the four soils. The small field of view
of the probe and the very small distance that the probe was moved for each increment
provides some possibility of errors in the collection of the different mixes. While every
effort was taken to measure each increment and ensure that mixes were accurate this is one
possible source of error in creating the physical mixes. Pure soil and vegetation spectra
were collected for creation of the virtual image in the same manner as for the laboratory
image. However, the mixed spectra of the virtual image were created as weighted linear
combinations of the pure spectra. Ten percent increments were again used to create an
image of the same mixes as the laboratory image.
a
b
Figure 5.1: (a) Demonstrates the configuration of the ASD high intensity reflectance probe held in a
clamp over the tray containing soil and leaves. (b) Demonstrates the incremental movement of probe
field of view over plant and soil interface. The solid lines indicate soil where pure soil and vegetation
spectra were collected. The dashed lines indicate the 10% increments as the probe was moved. Not to
scale.
The measured reflectance spectra and the calculated mixed spectra were incorporated into
a spectral library and then converted into artificial hyperspectral images. Spectra from the
four soils, the three different vegetation cover types and 11 cover fractions were
represented in each image of 12 samples, 11 lines and 2101 spectral bands (Figure 5.2).
The images created in this way allowed for the examination of the unmixing process with
known endmembers as inputs into the algorithm and measured or known fractions of each
of the soil and vegetation mixes.
100% vegetation within field of view
100% soil within field of view
Chapter 5: Spectral unmixing with simulated data 77
Figure 5.2: The ‘laboratory image’ created from the measured spectra. Soil type is listed at the bottom,
vegetation cover type at the top and percent soil exposure on the left.
5.2.3 Spectral unmixing
The images were analysed using linear spectral unmixing in order to determine the relative
abundance of soil from mixed spectra for comparison with the known fractions measured
during spectral collection. The unmixing process is based on the principle that the
reflectance spectrum of a given pixel is the weighted linear combination of spectra in the
field of view. The procedure assumes that the photons interact with only one material and
that ‘non-linear’ mixing does not occur (as when photons have multiple interactions with
materials)(Ray and Murray 1996, Zhu 2005). Input reference spectra for the unmixing
were the four pure soil spectra (100% soil exposure) and the three pure plant cover spectra
(0% soil exposure) for each of the images.
Linear unmixing is summarised in Equation 1:
iij
n
j
ji erfR +=∑=1
(1)
where Ri is the reflectance of a pixel in band i, ƒj is the fractional abundance of endmember
j in band i, rij is the reflectance of the pure endmember j in band i, ei is the residual error
Chapter 5: Spectral unmixing with simulated data 78
associated with band i and n is the number of endmembers. Equation 1 is constrained by
the assumption that the sum of the spectral components in each pixel should equate to 1.0
as defined by Equation 2:
11
=∑=
n
j
jf (2)
Spectral unmixing may be constrained or unconstrained. Constrained unmixing forces the
algorithm to assign fractional abundances that sum to one for each pixel and where no
negative abundances are permitted, whereas unconstrained allows unlimited negative and
positive abundances. Unconstrained unmixing has the advantage that the algorithm is not
being forced to unity and erroneous output abundances (less than zero or greater than one)
indicate a poor unmixing solution but do provide an avenue to improve the analysis
(Malenovsky et al. 2007). Erroneous output abundances theoretically arise from
incomplete assessment of the ‘pure’ materials within the image, i.e. an improper number of
endmembers, inadequate selection of endmembers to represent those materials, or a high
degree of collinearity between endmembers. Image noise and atmospheric attenuation are
also known to also affect the unmixing process (Settle and Drake 1993). The reality is that
if the outputs are negative or do not sum to unity the abundance fractions become
unrealistic and lose their meaning in the physical world and forcing them to do so will not
improve the analysis (van der Meer and De Jong 2000, Graña and D'Anjou 2004,
Malenovsky et al. 2007). In this study we used unconstrained spectral unmixing. The
unmixing was carried out using the ENVI 4.4 software package (RSI 2007).
Linear mixture analysis produces an image with estimates of each endmember fraction
within each pixel, and an estimate of the root mean squared error (RMSE) associated with
the unmixing. Endmember fractions for each pixel in the images were tabulated and
compared with the input measured or calculated fractions.
5.3 Results
5.3.1 Spectral characteristics
Reflectance spectra of all the soils (Figure 5.3) show pronounced water absorption features
(1400 and 1900 nm) and a clay absorption feature (2200 nm) with differences in symmetry
and depth indicating different clay species present. The sodic clay and clay loam show
Chapter 5: Spectral unmixing with simulated data 79
largely symmetrical clay absorption features, indicating likely domination by illite and
smectite minerals. The loam and silty loam have asymmetrical absorption features at 2200
nm indicating they contain higher proportion of kaolinite group minerals but are still likely
dominated by illite and smectite. Also present in the loam and silty loam is an absorption
feature at 2316 nm, characteristic of carbonates and reflecting the high content determined
from the laboratory analysis (Table 5.1). The sodic clay spectra contains a distinctive iron
oxide response over the visible-near infrared range (400-900 nm), as expected from the
high iron oxide content (Table 5.1). A subtle iron oxide spectral feature is also present in
the silty loam but not in the loam and clay loam, despite very similar concentrations found
in laboratory analysis (Table 5.1). This is likely the result of the higher organic matter
content in the loam and silty loam which has been shown to mask the spectral response of
iron oxide (Galvao and Vitorello 1998). The silty loam and loam also have an absorption
feature at 2388 nm that is possibly the result of organic matter in the soil (Henderson et al.
1992, Ben-Dor et al. 1997) despite very different concentrations found in laboratory
analysis (Table 5.1).
Figure 5.3: Pure soil spectra (endmembers) from soils used in this experiment.
Figure 5.4 shows the spectra of the vegetation used as cover in the experiment. Although
both the Eucalyptus and orange foliage showed overall characteristics typical of actively
photosynthetic vegetation, they differ in particular spectral regions. The orange has a more
pronounced chlorophyll green reflectance maximum at 550 nm, as well as more
pronounced water absorption features at 1400 nm and 1900 nm. The field pea residue
showed spectral characteristics typical of dry senescent organic matter. Most noticeably
there is a broad absorption at 2100 nm with two smaller absorption at 2261 and 2327 nm,
Chapter 5: Spectral unmixing with simulated data 80
which result from cellulose and lignin in the plant residue. Unlike the photosynthetic
vegetation spectra, the pea straw shows some similarity in overall form and albedo to the
soil spectra. There is increasing reflectance through the visible and near infrared and
specific absorption features in the 2000-2300 nm region.
Figure 5.4: Pure vegetation spectra (endmembers) from soils used in this experiment.
5.3.2 Mixes of spectra
Both methods of spectral mixing, the physical mixes (laboratory image) and the linear
weighted combinations (virtual image), created spectra that showed an even progression
from pure vegetation to soil. Examples of spectra from the sequences of physical mixes of
soil and plant material can be seen in Figure 5.5 and Figure 5.6. Because the soil and
photosynthetic vegetation differ markedly in albedo across most of the measured spectral
range, the sequence of mixed spectra shows pronounced gradients in reflectance intensity
from 500-2500nm (Figure 5.5). In addition, the increased influence of soil spectra in the
reflectance data can be seen with the reduction of the chlorophyll absorption at 650 nm,
along with a reduction in the red-edge and overall albedo in the near-infrared range (700-
1300 nm). There is also a change in shape of water absorption features at 1400 and 1900
nm and the appearance of a clay absorption feature at 2200 nm. In Figure 5.6 the changes
in reflectance characteristics are less evident as the soil fraction increases in the mix with
dry plant residue. There is little difference in albedo between the soil and non-
photosynthetic plant spectra, other than in the near infrared (800-1400 nm). However, the
clay absorption feature at 900 nm and the water absorption features at 1400 and 1900 nm
become more pronounced as the fraction of soil increases.
Chapter 5: Spectral unmixing with simulated data 81
Figure 5.5: Spectra collected from actual mixes of Sodic Clay and photosynthetic Eucalyptus
vegetation. Mixes of 0%, 20%, 40%, 60%, 80% and 100% soil are shown for clarity.
Figure 5.6: Spectra collected from actual mixes of Sodic Clay and the non-photosynthetic field pea.
Mixes of 0%, 20%, 40%, 60%, 80% and 100% soil are shown for clarity.
Chapter 5: Spectral unmixing with simulated data 82
5.3.3 Unmixing
The unmixing process results in a grey scale image for each of the input endmembers and a
root mean squared error (RMSE) image that indicates the level of error associated with the
unmixing of each pixel. For the virtual image, the RMSE (Figure 5.7) for all pixels was
low. Higher RMSE was found in pixels dominated by vegetation for all soils except the
loam which had higher errors in soil pixels. The clay loam had double the RMSE of the
other soils except under Eucalyptus where the loam errors were highest. For the laboratory
image, the RMSE (Figure 5.7) of unmixing was generally higher than that from the virtual
image. Under Eucalyptus all soils between 20% and 60% exposure had increased error.
Under the orange and the pea straw the clay loam had substantially higher RMSE, more
than twice that of the other soils. All pixels with high soil content (>60%) returned low
RMSE under each cover type.
Figure 5.7: RMSE from ‘virtual’ (a: Eucalyptus, b: orange, c: pea straw) and ‘laboratory’ (d:
Eucalyptus, e: orange, f: pea straw) images for each of the vegetation cover types.
Chapter 5: Spectral unmixing with simulated data 83
Sodic clay
Abundance fractions for the sodic clay were comparable to the known soil exposure for
both the synthetic and measured spectral mixes (Figure 5.8). The sodic clay was
recognised as the input endmember and the fractional combinations of soil and the
vegetation were unmixed accordingly. Furthermore, the non-target soil spectra were not
confused with the sodic clay spectra. The unmixing fractions for the virtual image retrieved
the calculated soil exposure more accurately than the fractions from the measured
mixtures. The errors in unmixing the laboratory image were restricted to over estimation
between 50% and 100% soil exposure. However, in all cases, the errors in estimation of
soil fraction were less than 0.1 (10%).
Figure 5.8: Unmixing with the Sodic Clay endmember of the ‘virtual’ (a: Eucalyptus, b: orange, c: pea
straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for each of the vegetation cover
types.
Chapter 5: Spectral unmixing with simulated data 84
Silty loam
Fractions of the silty loam endmember were retrieved correctly from unmixing of the
virtual image (Figure 5.9) in most instances. However, when mixed with the orange cover,
the soil fraction was underestimated by at least 10% at all exposures, with under-estimation
greater at lower soil fractions. The laboratory image (Figure 5.9) demonstrated the same
over estimation of soil exposure between 50% and 100% that was evident in sodic clay
(Figure 5.8). Soil abundance under the orange was also underestimated to a similar extent
as in the virtual image. Not present in the laboratory image unmixing is the
misclassification of the loam as silty loam as seen in the virtual image.
Figure 5.9: Unmixing with the Silty Loam endmember of the ‘virtual’ (a: Eucalyptus, b: orange, c: pea
straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for each of the vegetation cover
types.
Chapter 5: Spectral unmixing with simulated data 85
Loam
The loam soil fraction was unmixed less successfully than previous soils from both the
virtual and laboratory images. Although the loam fraction unmixed from the virtual image
(Figure 5.10) showed a linear increase from 0-100% exposure, the magnitude of the soil
fraction was underestimated by up to 50%. This error was substantially worse under the
Eucalyptus cover type compared to the orange and pea straw. The unmixing pattern from
the laboratory image (Figure 5.10) with the loam endmember is quite different from the
pattern with the virtual image. The unmixing fractions generally followed a sigmoidal
trend with increasing soil exposure. Under orange and pea straw the range of estimated
fractions was feasible (0-1), but negative soil abundances were recorded below 50%
exposure under Eucalyptus cover. Errors in estimation of the soil fraction were greatest
below 60% soil exposure under Eucalyptus and orange but over 40% exposure under pea
straw.
Figure 5.10: Unmixing with the Loam endmember of the ‘virtual’ (a: Eucalyptus, b: orange, c: pea
straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for each of the vegetation cover
types.
Chapter 5: Spectral unmixing with simulated data 86
Clay loam
The clay loam endmember unmixing of the virtual image (Figure 5.11) returned accurate
fractional abundances for the target soil under Eucalyptus but misclassified the silty loam
at low exposures and the loam at high exposures. Under the orange the target soil showed a
one to one increase but underestimated the soil fraction by up to 20%. Unmixing of the
laboratory image (Figure 5.11) showed substantial misclassification of up to 0.6 (60%)
with all the non-target soils under Eucalyptus. The fractional abundance of the target soil
was also overestimated at most soil exposures. Under the orange there was again a
negative abundance below 30% exposure and an overestimation above 40%. The pea straw
unmixing returned low fractional abundances at all exposures above 20%.
Figure 5.11: Unmixing with the Clay Loam endmember of the ‘virtual’ (a: Eucalyptus, b: orange, c:
pea straw) and ‘laboratory’ (d: Eucalyptus, e: orange, f: pea straw) images for each of the vegetation
cover types.
Chapter 5: Spectral unmixing with simulated data 87
5.4 Discussion
5.4.1 Unmixing
Of the soils examined, the sodic clay was unmixed most accurately. This is likely due to
the very distinct spectral features such as iron oxide (400-900 nm), water (1400 and 1900
nm) and clay (2200 nm). In addition it has a moderately high albedo over most of the
wavelength range examined. The silty loam was also unmixed well and has an iron oxide
absorption feature (400-900 nm), distinctive clay feature (2200 nm) and carbonate
absorption feature (2316 nm). This soil also has the highest albedo but relatively small and
uncharacteristic water absorption features (1400 and 1900 nm). Unmixing of the loam and
clay loam was the least accurate; these soils have the least distinct spectra. Loam had a
carbonate absorption feature (2316 nm) and moderate albedo while the clay loam had the
lowest albedo.
5.4.2 Discrimination of soils
There were few cases of non-target soils being classified as target spectra but they occurred
largely in unmixing the virtual image. Pure loam was classified as up to 0.33 (33%) of the
target soil silty loam (Figure 5.9) and as up to 0.4 (40%) of the clay loam (Figure 5.11).
Despite a substantial difference in albedo, the clay loam and loam were very spectrally
similar; water (1400 and 1900 nm) and clay (2200 nm) absorption features are of
comparable intensity and shape with little else to differentiate them other than
chromophores at 2316 nm and 2388 nm in the loam which are not present in the clay loam
spectra. The silty loam and loam are also spectrally similar, again despite contrasting
reflectance intensity, differing mostly in the mild iron oxide absorption feature present in
the silty loam and not in the loam. It is unexplained why this occurred only in the virtual
image but not in the laboratory image. Given the spectral similarity of the soils some
degree of misclassification such as this was expected from both images.
The observed misclassification of mixed pixels as a different soil-vegetation combination
potentially undermines the unmixing of airborne and satellite imagery in resource
management and mapping applications. It appears from these results that the unique
spectra of different soils can affect the ability of unmixing algorithms to correctly estimate
mixed abundances. Contrary to expectations (Asner and Heidebrecht 2002, Bannari et al.
2006, Daughtry et al. 2006), the greatest misclassification here was observed with
Chapter 5: Spectral unmixing with simulated data 88
photosynthetic vegetation and soil mixes, rather than non-photosynthetic soil mixtures.
Nonetheless, despite these complications, for the soils examined in this study, few
difficulties were encountered in isolating pure soil pixels.
5.4.3 Discrimination of soil and vegetation
Typically the separation of soil and non-photosynthetic vegetation with unmixing is more
difficult than with soil and photosynthetic vegetation (Bannari et al. 2006). In our
experiments however, the soil and field pea residue pixels were unmixed better than the
soil-photosynthetic vegetation mixes. In both the virtual and laboratory images there was
evidence of confusion between the target soil endmembers and pure photosynthetic
vegetation. This was particularly unexpected given the spectral characteristics of the soils
and photosynthetic vegetation are so distinct. Pure vegetation spectra, with significant
water absorption features and red edge, bear little resemblance to the soil spectra examined
in this study.
In Figure 5.10a and Figure 5.10d pure Eucalyptus spectra over sodic clay were unmixed as
having up to 0.12 (12%) fractional abundance of loam spectra. Comparing the spectra of
these materials (Figure 5.3 and Figure 5.4) there were spectral characteristics, such as the
iron oxide feature (400-900 nm), that bears some resemblance to the characteristic red-
edge of photosynthetic vegetation. This misclassification only occurred with the
Eucalyptus and not the orange which has a higher overall albedo and more pronounced
red-edge. Similarly, pure Eucalyptus and orange over silty loam were incorrectly given
nearly 0.2 (20%) fractional abundance of the target spectra clay loam (Figure 5.11).
However, unlike the sodic clay there is no obvious spectral similarity between soil (Figure
5.3) and the photosynthetic vegetation (Figure 5.4).
There was some misclassification of mixes of soils and vegetation as the pure soil
endmember or target soil. The unmixing with sodic clay (Figure 5.8) and to a lesser extent
silty loam (Figure 5.9), gave expected results where there was a clear recognition of pixels
containing target and non-target soils mixed with vegetation. Alternatively, for loam
(Figure 5.10) fractional abundances of soil and vegetation are substantially incorrect
despite successful separation of different soil and soil-vegetation mixes. For the clay loam
(Figure 5.11) mixtures of soil and vegetation were incorrectly classified as soil. Similar
observations have been made in a previous study (Okin et al. 2001) where some vegetation
and soil mixes were confused with pure soil spectra in the unmixing process. However, in
Chapter 5: Spectral unmixing with simulated data 89
that field study, undertaken in a semiarid environment, the vegetation was observed to
have minimal water absorption features and a very small red edge, which is not the case
here.
5.4.4 Unmixing errors
There were many instances where the unmixing resulted in a negative abundance (e.g.
Figure 5.11) and some with abundances greater than unity (e.g. Figure 5.10). In traditional
image unmixing a negative result would be common and due to the difficulty in isolating
all endmembers within the scene and ensuring that those spectra chosen as inputs are pure
and not mixtures of different materials. However, in this experiment all the endmembers in
the ‘scene’ were known and they were all known to be pure. Nonetheless errors in our
experiment would be expected and would prevent the algorithm from perfectly inverting
the mixed spectra. Firstly some correlation between the endmembers should be expected.
In both images there are similarities between the soil spectra and between the
photosynthetic vegetation spectra. Secondly, there is some variation between each of the
‘pure’ spectra used in the image and the reference spectra used in the unmixing because
each was collected separately by the spectrometer. The algorithm cannot account for this
variation and errors are unavoidably introduced to the unmixing.
The disparity between the known soil exposure and the fractional abundance coincides
with the areas of high RMSE. For the laboratory image the areas of largest RMSE
corresponded with the 20% to 60% soil-vegetation mixes that were incorrectly unmixed as
some fraction of the target soil. These inaccuracies were evident with the clay loam (Figure
5.11a) where soil and vegetation mixes were given a high reading (up to 60%) and with the
loam (Figure 5.11a) where a substantial negative fractional abundance was returned. For
the virtual image there is substantially lower RMSE; however, the areas of largest RMSE
still correspond with the poor estimation of soil exposure. For example, the
misclassification of loam as clay loam (Figure 5.11) under all cover types is reflected in the
relatively high error given loam in the RMSE graph (Figure 5.7).
Chapter 5: Spectral unmixing with simulated data 90
5.4.5 Virtual versus laboratory images
The virtual image served as a control to determine the accuracy of measured spectra for the
laboratory image. As spectral mixes in the virtual image were created from weighted linear
combinations they contained exact proportions of the pure spectral endmembers. If the
fractions of soil and vegetation in the laboratory image were accurately measured, and the
radiance recorded by the sensor was a linear mixture of the reflectance from these
components, then retrieval of the endmember spectra should be similar. There are two
potential reasons for the substantial differences between the unmixing results from the
laboratory and virtual images. Firstly, fractions of soil and vegetation could have been
inaccurately measured and secondly the radiance measured by the spectrometer is not a
perfect weighted linear mix of the fractional constituents.
Small errors in measurement of cover fractions in the laboratory spectra may account of
some apparent errors in unmixing. The one to one relationships evident in some virtual
image outputs (Figure 5.8 and Figure 5.9) but not the outputs from the laboratory image
(Figure 5.8 and Figure 5.9) are to some degree a reflection of this. However, the difference
in other unmixing results between the two image types far exceeded the expected error
from the collection of the laboratory image spectra (e.g. Figure 5.10 and Figure 5.11).
Nonetheless, the errors evident in Figure 5.10 (although negative) and Figure 5.11 appear
consistent across all the soils despite each soil and vegetation combination being measured
independently. For example, the misclassification of the soil-vegetation combinations
evident in the unmixing of the laboratory image (Figure 5.11d) were not present in the
unmixing of the virtual image (Figure 5.11a) but the misclassification of the non-target
soils as the target soil (clay loam) in the laboratory image was relatively uniform.
Therefore it appears that much of the discrepancy between the laboratory and virtual image
outputs resulted from non-linear mixing of the material constituents in the laboratory
image. The relatively accurate retrieval of soil fractions from the virtual image compared
with the laboratory image suggest that real world mixing is not linear. Previous studies
have used a similar methodology to create virtual mixes of spectra and found the results
adequate for measuring constituent material abundance. One such study (Daughtry 2001)
used ratio indices such as the cellulose absorption index to quantify crop residue on
different soils but did not attempt to identify the soils themselves. Another study (Hussey
Chapter 5: Spectral unmixing with simulated data 91
1998) examined spectral mixture analysis of different mineral combinations made through
weighted linear combinations and compared them to the spectra of physical mixes. Some
differences were observed between the virtual and physical mixes but these were attributed
to minor discrepancies in mineral purity, and while successful, the study aimed to
determine the physical limits for unmixing based on band numbers and signal noise rather
than material constituents.
Our comparison of virtual and physical mixing suggest that radiance measurement by the
instrument is not a consistently linear mix of component spectra. Many remote sensing
studies using spectral mixture analysis assume linear mixing yet suggest ‘non-linear
mixing’ as a cause for unmixing errors. Generally accepted causes of the non-linear mixing
include transmission of light through the vegetative cover and the scattering of light off
multiple surfaces before reaching the sensor. This experimental design sought to minimise
these factors by using multiple layers of vegetation to reduce transmission of light through
the leaf and collecting spectral mixes from a flat surface to reduce scattering. Thus these
two parallel experiments have tested, and provided stronger evidence that, mixing of soil
and cover types is not perfectly linear. While this is generally well accepted within the
remote sensing community , there is little systematic experimentation to quantify and
explore it under controlled laboratory conditions.
5.5 Conclusions
This study used a technique combining laboratory reflectance spectra and spectral mixture
analysis to identify soil fractions from mixed pixels containing soil, photosynthetic
vegetation and non-photosynthetic vegetation. The methodology provided images that
could be analysed by standard hyperspectral feature extraction algorithms. It should also be
emphasised that this study was conducted in a controlled laboratory. The effects of
atmospheric attenuation, soil surface roughness, soil moisture content and leaf orientation
are not considered.
Results also show the unmixing process successfully recognised and classified the different
soils within both image types. However, not all soil spectra were isolated from mixed
pixels equally or successfully to provide accurate abundance fractions. This highlights
potential problems of techniques like linear spectral mixture analysis with evidence of
confusion between pixels of mixed constituents (soil and vegetation) and other materials
Chapter 5: Spectral unmixing with simulated data 92
(different soil types). While other studies have suggested this possibility (Okin et al. 2001),
this research shows conclusively that spectral confusion occurs even in images with limited
and well understood endmembers. The comparison between virtual and laboratory images
cast some doubt on assumptions regarding the combination of pure spectra in mixed pixels
with further evidence that it is not consistently linear. This is largely accepted for image
studies conducted in heterogeneous terrain with rough and undulating soil surfaces, and
differing plant geometry with variable leaf orientation (Ray and Murray 1996, Malenovsky
et al. 2007). These results demonstrate the limitations of this technique even carried out in
an essentially linear environment. Furthermore, spectra for the two image types were
collected under identical conditions and as such the comparison is that between the mixing
processes alone.
Visible and near-infrared remote sensing provides enormous scope in monitoring and land
management to improve our understanding and practices in agricultural and environmental
applications. However, the power of these techniques is limited by our understanding of
processes on the ground. Confusion of mixed pixels with pure pixels and the non-linear
spectral mixing evident in this study has impacts on real world image studies undertaken to
monitor ground cover or soil. Further research is required to better understand the process
at work here.
5.6 References
Alemie, B. K. 2005 Spectral unmixing of hyperspectral and multispectral images for predictive mapping of
surface soil organic matter, Master of Science Thesis, International Institute for Geo-Information Science and Earth Observation, Enschede, The Netherlands.
Allison, L. E. and Moodie, C. D. 1982 Carbonate, In Methods of soil analysis (Ed, A. L. Page) Soil Science Society of America, Madison, pp. 1379-1396.
Asner, G. P. and Heidebrecht, K. B. 2002 Spectral unmixing of vegetation, soil and dry carbon cover in arid regions: Comparing multispectral and hyperspectral observations, International Journal of Remote Sensing, 23, 3939-3958.
Bannari, A., Pacheco, A., Staenz, K., McNairn, H. and Omari, K. 2006 Estimating and mapping crop residues cover on agricultural lands using hyperspectral and IKONOS data, Remote Sensing of Environment, 104, 447-459.
Ben-Dor, E., Inbar, Y. and Chen, Y. 1997 The reflectance spectra of organic matter in the visible near-infrared and short wave infrared region (400-2500 nm) during a controlled decomposition process, Remote
Sensing of Environment, 61, 1-15.
Ben-Dor, E., Patkin, K., Banin, A. and Karnieli, A. 2002 Mapping of several soil properties using DAIS-7915 hyperspectral scanner data - a case study over clayey soils in Israel, International Journal of Remote
Sensing, 23, 1043-1062.
Bock, M., Rossner, G., Wissen, M., Remm, K., Langanke, T., Lang, S., Klug, H., Blaschke, T. and Vrscaj, B. 2005 Spatial indicators for nature conservation from European to local scale, Ecological Indicators, 5, 322-338.
Chapter 5: Spectral unmixing with simulated data 93
Bower, T. L. and Rowan, L. C. 1996 Remote mineralogic and lithological mapping of the Ice River alkaline complex, British Columbia, Canada, using AVIRIS data, Photogrammetric Engineering and Remote Sensing, 62, 1376-1143.
Clark, R. N. 1999 Spectroscopy of rocks and minerals and principles of spectroscopy, In Remote sensing for
the earth sciences: Manual of remote sensing, Vol. 3 (Ed, A. N. Rencz) John Wiley and Sons, New York, pp. 3-58.
Cozzolino, D. and Morón, A. 2003 The potential of near-infrared reflectance spectroscopy to analyse soil chemical and physical characteristics, Journal of Agricultural Science, 140, 65-71.
Daughtry, C. S. T. 2001 Discriminating crop residues from soil by shortwave infrared reflectance, Agronomy
Journal, 93, 125-131.
Daughtry, C. S. T., Doraiswamy, P. C., Hunt, J., E.R., Stern, A. J., McMurtrey III, J. E. and Prueger, J. H. 2006 Remote sensing of crop residue cover and soil tillage intensity, Soil and Tillage Research, 91, 101-108.
Daughtry, C. S. T., Hunt, E. R., Jr., Doraiswamy, P. C. and McMurtrey, J. E., III 2005 Remote sensing the spatial distribution of crop residues, Agronomy Journal, 97, 864-871.
Dehaan, R. and Taylor, G. R. 2003 Image-derived spectral endmembers as indicators of salinisation, International Journal of Remote Sensing, 24, 775-794.
Dennison, P. E. and Roberts, D. A. 2003 Endmember selection for multiple endmember spectral mixture analysis using endmember average RMSE, Remote Sensing of Environment, 87, 123-135.
Dutkiewicz, A., Lewis, M. and Ostendorf, B. 2003 Evaluation of hyperspectral imagery for mapping the symptoms of dryland salinity, In Spatial Sciences Coalition 2003, Canberra,
Galvao, L. S. and Vitorello, I. 1998 Role of organic matter in obliterating the effects of iron on spectral reflectance and colour of Brazilian tropical soils, International Journal of Remote Sensing, 19, 1969-1979.
Gee, G. W. and Bauder, J. W. 1986 Particle-size analysis, In Methods of soil analysis, part 1. Physical and
mineralogical methods, Vol. 2 (Ed, A. Klute) American Society of Agronomy - Soil Science Society of America, Madison, Wisconsin, USA, pp. 383-409.
Graña, M. and D'Anjou, A. 2004 Feature extraction by linear spectral unmixing, In Knowledge-based
intelligent information and engineering systems, pp. 692-698.
Henderson, T. L., Baumgardner, M. F., Franzmeier, D. P., Stott, D. E. and Coster, D. C. 1992 High dimensional reflectance analysis of soil organic matter, Soil Science Society of America Journal, 56, 865-872.
Hussey, M. C. 1998 Surface detection of alkaline ultramafic rocks in semi-arid and arid terrains using
spectral geological techniques, PhD Thesis, Open University, Southampton, United Kingdom.
Isbell, R. F. 2002 The Australian soil classification, CSIRO Australia, Melbourne.
Islam, K., Singh, B. and McBratney, A. 2003 Simultaneous estimation of several soil properties by ultra-violet, visible and near-infrared reflectance spectroscopy, Australian Journal of Soil Research, 41, 1101-1114.
Janik, L. J. and Skjemstad, J. O. 1995 Characterisation and analysis of soils using mid-infrared partial least squares: II. Correlations with laboratory data, Australian Journal of Soil Research, 33, 637-650.
Kelley, C. T. 2003 Solving nonlinear equations with Newton's methodology, Society for Industrial Applied Mathematics, Philadelphia.
Lewis, M. 2000 Discrimination of arid vegetation composition with high resolution CASI imagery, Rangeland Journal, 22, 141-167.
Lu, D. and Weng, Q. 2007 A survey of image classification methods and techniques for improving classification performance, International Journal of Remote Sensing, 28, 823-870.
Malenovsky, Z., Bartholomeus, H. M., Acerbi-Junior, F. W., Schopfer, J. T., Painter, T. H., Epema, G. F. and Bregt, A. K. 2007 Scaling dimensions in spectroscopy of soil and vegetation, International Journal of
Applied Earth Observation and Geoinformation: Advances in airborne electromagnetics and remote sensing
of agro-ecosystems, 9, 137-164.
Chapter 5: Spectral unmixing with simulated data 94
McCarty, G. W., Reeves, J. B., III, Reeves, V. B., Follett, R. F. and Kimble, J. M. 2002 Mid-infrared and near-infrared diffuse reflectance spectroscopy for soil carbon measurement, Soil Science Society of America
Journal, 66, 640-646.
Metternicht, G. I. and Zinck, J. A. 2003 Remote sensing of soil salinity: Potentials and constraints, Remote
Sensing of Environment, 85, 1-20.
Nagler, P. L., Daughtry, C. S. T. and Goward, S. N. 2000 Plant litter and soil reflectance, Remote Sensing of
Environment, 71, 207-215.
Nagler, P. L., Inoue, Y., Glenn, E. P., Russ, A. L. and Daughtry, C. S. T. 2003 Cellulose absorption index (CAI) to quantify mixed soil-plant litter scenes, Remote Sensing of Environment, 87, 310-325.
Nelson, D. W. and Sommers, L. E. 1986 Total carbon, organic carbon and organic matter, In Methods of soil
analysis, Vol. 2 (Ed, A. L. Page) Soil Science Society of America, Madison, pp. 53-579.
Okin, G. S., Roberts, D. A., Murray, B. and Okin, W. J. 2001 Practical limits on hyperspectral vegetation discrimination in arid and semiarid environments, Remote Sensing of Environment, 77, 212-225.
Ray, W. T. and Murray, B. C. 1996 Nonlinear spectral mixing in desert vegetation, Remote Sensing of
Environment, 55, 59-64.
Ross, G. L. and Wang, C. 1993 Extractable Al, Fe, Mn, and Si, In Soil sampling and methods of analysis (Ed, M. R. Carter) Lewis Publishers, Boca Raton, pp. 239-246.
RSI 2007 ENVI version 4.4, Research Systems Inc., Boulder, Colorado.
Settle, J. J. and Drake, N. A. 1993 Linear mixing and the estimation of ground cover proportions, International Journal of Remote Sensing, 14, 1159 - 1177.
Soil Survey Staff 1999 Soil taxonomy: A basic system of classification for making and interpreting soil
surveys, U.S. Government Print Office, Washington D.C.
Sumfleth, K. and Duttmann, R. 2007 Prediction of soil property distribution in paddy soil landscapes using terrain data and satellite information as indicators, Ecological Indicators, 8, 485-501.
Taylor, G. R. 2004 Field and image spectrometry for soil mapping, In 12th Australasian Photogrammetry and Remote Sensing Conference, Fremantle, Western Australia,
Tiner, R. W. 2004 Remotely-sensed indicators for monitoring the general condition of "Natural habitat" In watersheds: An application for Delaware's Nanticoke river watershed, Ecological Indicators, 4, 227-243.
van der Meer, F. and De Jong, S. M. 2000 Improving the results of spectral unmixing of Landsat thematic mapper imagery by enhancing the orthogonality of end-members, International Journal of Remote Sensing, 21, 2781-2797.
Viscarra Rossel, R. A., McGlynn, R. N. and McBratney, A. B. 2006a Determining the composition of mineral-organic mixes using UV-VIS-NIR diffuse reflectance spectroscopy, Geoderma, 137, 70-82.
Viscarra Rossel, R. A., Walvoort, D. J. J., McBratney, A. B., Janik, L. J. and Skjemstad, J. O. 2006b Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties, Geoderma, 131, 59-75.
Zhang, L., Wu, B., Huang, B. and Li, P. 2007 Nonlinear estimation of subpixel proportion via kernel least square regression, International Journal of Remote Sensing, 28, 4157-4172.
Zhu, H. 2005 Linear spectral unmixing assisted by probability guided and minimum residual exhaustive search for subpixel classification, International Journal of Remote Sensing, 26, 5585-5601.
95
Chapter 6
Mapping soil variability with hyperspectral image data
Published as a refereed conference paper:
Summers, D., Lewis, M., Ostendorf, B. and Chittleborough, D. J. 2009 Mapping soil variability with hyperspectral image data, In SSC 2009 Spatial diversity: The Biennial International Conference of the Surveying and Spatial Sciences Institute, Adelaide, Australia, September-October, 2009.
113
Chapter 7
Discussion and Conclusion
7.1 Introduction
The overall goal of this thesis was to contribute to the development of tools to understand
and map soils in southern Australia. This was identified as a gap in knowledge from the
point of view of precision agriculture, where improved understanding of soil variability is
an important input in improving farming efficiency and productivity. Furthermore,
improved understanding of soil variability in the landscape is seen as vital to improve the
accuracy and precision of models to better understand landscape processes for applications
including ecology, biodiversity and soil hydrology.
The thesis has contributed to this goal by examining spectral reflectance methodologies
that have the potential to improve the efficiency of soil sample analysis, allowing for
sampling densities greater than is typical for regional soil analysis and mapping.
Additionally, this work examined the spectral unmixing of hyperspectral image data to
map surface soil variability, exploiting the continuous nature of remotely sensed images
and the high diagnostic power of hyperspectral reflectance data.
Chapters 3 and 4 examined the use of hyperspectral reflectance spectroscopy to
discriminate select soil field survey classes and predict laboratory measured soil properties
respectively. The discrimination of soil field survey classes (Chapter 3) provided some
insight into the complex relationships and collinearity of soil properties such as clay
content, carbonate content and soil colour. However, this study also highlighted the
inherent problems of soil field survey, a relatively subjective measure of soil properties, for
quantitative research. Alternatively, Chapter 4 examined the prediction of quantitative soil
properties determined from laboratory analysis using partial least squares regression and
achieved substantially better results. Following the regression analysis, kriging of the
measured and predicted data was used to create soil raster layers. Comparison of the
measured and predicted raster layers found they mapped similar variability in the
landscape over comparable ranges in soil properties.
Chapter 7: Conclusion 114
Chapters 5 and 6 examined the possibility of using hyperspectral image data to identify soil
variability in the landscape. Vegetative cover was identified as a major problem in
achieving this aim as it obscures the soil surface from the sensor. In order to examine the
complexities of this problem, two types of simulated imagery were developed which
provided known mixes of the various constituents for subsequent analysis and comparison
(Chapter 5). Finally, HyMap airborne hyperspectral imagery was used to map soil types in
the landscape. Endmembers were isolated from the imagery and were used in partial
unmixing algorithms in an attempt to identify soil variation (Chapter 6).
7.2 Summary of specific contributions to knowledge
7.2.1 Spectral discrimination of soil properties (Chapter 3)
The major aim of Chapter 3 was to investigate the ability of visible, near infrared and
shortwave infrared reflectance spectroscopy to predict various field survey soil properties
in a localised geographical region in order to supplement soil survey. These were clay
content, carbonate content and the components of Munsell colour (hue, value and chroma).
The primary motivation behind this study was to determine the compatibility of reflectance
spectroscopy to complement soil field survey. While soil field survey is conducted
extensively in southern Australia’s intensive agricultural areas it is prohibitively expensive
in broadacre and dryland agricultural areas. Reflectance spectroscopy was investigated as a
means to expand the areas mapped using field survey methodologies more cost effectively
while maintaining some continuity between methodologies.
The study involved the collection of 293 soil samples from the Jamestown – Belalie
district, a northern agricultural region in South Australian. Samples were analysed using
conventional field survey methodologies and reflectance spectra were collected before the
development of penalised discriminant analysis models to discriminate classes. The
chroma component of Munsell colour was the only soil property that was adequately
discriminated using the hyperspectral reflectance data. All the other properties examined
were well discriminated in one or two of their classes but overall accuracy was poor.
Findings from Chapter 3 also indicate that there was substantial co-variation in the spectral
properties of the soil properties examined. Consideration has been given that this co-
variation substantially limited discrimination of soil properties.
Chapter 7: Conclusion 115
However, it is also possible that the subjective nature of field survey classification
introduced a considerable source of variation into the classification. For example, while
field texture analysis provides a useful and repeatable assessment of the physical behaviour
of soil in the field, it is nonetheless subjective and prone to user error (McDonald and
Isbell 1990). Similarly, soil colour is also subjective; individuals can perceive colour
differently, but also, the soil colour classification involves matching soil to the closest
colour chip and there is rarely a perfect match (McDonald and Isbell 1990). In this study
efforts were made to minimise error, firstly through the analysis of replicates and secondly
by using a single trained and competent soil scientist to carry out the analysis.
7.2.2 Visible near-infrared reflectance spectroscopy as a predictive indicator of soil
properties (Chapter 4)
The major aim of Chapter 4 was to investigate the use of visible near-infrared spectroscopy
as a predictor of laboratory measured soil properties in a localised geographical region in
order to supplement soil survey. The motivation behind this was to overcome the
subjective nature of the field survey used in Chapter 3, provide a more rigorous test of
spectroscopic prediction and an objective cost-effective methodology to improve the
spatial resolution of soil mapping in dryland agricultural regions. The soil properties used
in the analysis were clay, carbonate, organic carbon and iron oxide contents. These were
chosen because they are important determinants of soil agricultural capability and also
because they are considered important as inputs into soil hydrology models.
This study involved the analytical determination of soil properties in the laboratory, the
collection of further reflectance spectra of soil samples using the ASD Field Spec Pro. and
the development of prediction models using partial least squares regression (PLSR).
Following the prediction of soil properties kriging was employed to model surface soil
properties across the landscape from both the measured and predicted datasets. These
layers were then compared to determine the utility of the predicted data as a supplement
for soil survey.
The results show that all soil properties were adequately predicted. The model explained
more than 65% of the variability in clay and carbonate content and residual predictive
deviations (RPD) of 2.0 and 2.1 respectively indicate substantially better prediction than
the mean of the observed values. Soil organic carbon and iron oxide were less successfully
predicted but still achieved r2 values of 0.57 and 0.61 respectively and acceptable RPD
Chapter 7: Conclusion 116
values of greater than 1.5. The different measured and predicted layers produced following
kriging of the point sample sites show similar patterns indicating that soil spatial variability
was similarly represented in both approaches. The PLSR results and comparison of the
surface layers produced demonstrates that the PLS prediction from spectroscopic
measurements provides a suitable method to efficiently supplement and enhance traditional
soil survey.
7.2.3 Unmixing of soil types and estimation of soil exposure with simulated
hyperspectral imagery (Chapter 5)
The major aim of Chapter 5 was to examine the complex interaction of soil and vegetative
cover of different types. The goal was to better understand how soil surface properties can
be measured with remote sensing imagery in environments where soil exposure is limited.
The assessment involved the collection of four distinct soil types from the Monarto region
in South Australia, chosen to provide physical and chemical diversity as well as spectral
variability. Vegetation types typical of common landuses in southern Australia were also
collected, namely native trees (eucalyptus), horticulture (orange) and dryland agriculture
(crop residue). These materials were used to create two types of simulated imagery: one,
called the laboratory image, created from real mixes of the soil and vegetation,
incrementally increasing the amount of vegetation of in the field of view during spectral
collection. The other, called the virtual image, created by weighted linear combinations of
pure soil and vegetation spectra. The pure soil endmembers were then used as inputs into
linear unmixing algorithms. The classification of soils types in mixed pixels and the
determination of fractional soil exposure were then assessed from the output images.
Results show that the soils were successfully recognised and classified within both image
types. However, not all soil spectra were isolated from mixed pixels equally or
successfully to provide accurate abundance fractions. For example, the only soil that
showed accurate unmixing abundances at most exposures was the sodic clay. Importantly,
in some cases, such as with the unmixing of loam and clay loam, mixed pixels were
classified as non-target soils indicating that the unmixing process interpreted mixed pixels
as a different soil endmember. This presents a complication when attempting to unmix
soils and vegetation in the landscape for the purposes of mapping soil variability.
Chapter 7: Conclusion 117
7.2.4 Mapping soil variability with hyperspectral image data (Chapter 6)
The major aim of this research was to map soil types from airborne hyperspectral image
data without using a priori knowledge of soil variability or composition in the landscape to
inform unmixing endmembers. HyMap hyperspectral imagery was collected over the
Jamestown – Belalie district, a dryland agricultural region 200 km north of Adelaide. The
image, which is dominated by crop residue, covers an area of broad valleys and small hills
where the landuse is mostly cropping and grazing. Soil endmembers were determined
through a process where pure pixels were isolated statistically in n-dimensional space.
These endmembers were then used in the unmixing to map soil variability and the results
were compared with quantitative soil properties determined from sample sites within the
mapped areas. Further to this, the endmember abundance was compared to visual field
assessment of soil exposure made during sample collection.
Four distinct endmembers were isolated in the pixel purity process and each mapped
different areas in the landscape using the partial unmixing algorithm. However, the
laboratory analysis of soil samples was unable to characterise any difference between the
areas mapped. Furthermore, the coefficients of determination between the image derived
soil abundance and the field estimated soil exposure indicate that little of the variance was
captured through the image analysis. While the use of partial unmixing to identify surface
soil variation in the landscape may provide a useful tool to inform soil survey, the results
here were limited. Possible explanations for this include poor endmember selection
through the pixel purity process and the lack of variation in the surface soils within the
hyperspectral image. However, it is also appears that the dominant influences on the soil
response as recorded by the airborne hyperspectral sensor are related to land management
(e.g. tillage), or properties such as moisture and colour not quantified by the laboratory
measurements.
7.2.5 Overall assessment of thesis topic
The research summarised above represents a substantial contribution to the use of soil
reflectance and hyperspectral remote sensing to better understand soil variability and map
soil properties with these technologies. This thesis strengthens existing knowledge by
testing the prediction of soil properties from reflectance spectroscopy over a limited
geographical area. The research also provided an assessment of how that prediction can be
used to generate soil maps. Simulated hyperspectral imagery was used to assess the
Chapter 7: Conclusion 118
spectral unmixing of soil and vegetation to assess the use of hyperspectral image data in
mapping soil variability through vegetative cover of different types. The mapping of soil
variability in a dryland agricultural region dominated by crop residue was also examined
using HyMap airborne hyperspectral data.
7.3 General discussion: wider significance and limitations
The work conducted and presented in this thesis has made some important contributions to
knowledge. The significance and limitations of the research specific to the aims of the
different studies has been discussed within the relevant chapters. The following section
covers the wider significance and the limitations to generalisation of the research.
7.3.1 Spectral discrimination of soil properties (Chapter 3)
The spectral discrimination of soil properties presented in Chapter 3 was conducted using
field survey methodologies. These field survey techniques provide a cost effective and
useful assessment of the soil properties for land managers, largely targeted at improving
irrigation efficiency and environmental sustainability. It was proposed that spectral
discrimination of field survey soil properties may provide a cost effective means to apply
similar classifications in dryland areas. Furthermore, the spectral discrimination of field
classes may provide a more quantitative, objective means by which to discern field classes.
However, results from this study indicate limited success in this regard. The limited results
may stem from inaccurate assessment of the field classes, which may be overcome through
the incorporation of multiple individuals undertaking the field classification. Alternatively,
poor classification of the spectral data through the penalised discriminant analysis may
have been a factor. This could be caused by co-variance between the different soil classes
and subsequent analyses may be improved through stratification. However, given the
successful prediction of soil properties using partial least squares regression, no further
attempt to improve the classification of field survey classes was made for this research.
7.3.2 Visible near-infrared reflectance spectroscopy as a predictive indicator of soil
properties (Chapter 4)
The use of visible near-infrared reflectance spectroscopy as a predictive indicator of soil
properties in this study was successful. One of the main goals of the research was to
predict properties over a range of soil variation encountered in a limited geographical area
that would normally be the subject of soil survey. Because of this, the application of the
Chapter 7: Conclusion 119
model developed here to different geographical areas is limited. In order to apply a similar
methodology to different geographical areas new prediction models would need to be
developed. Although previous research (Janik et al. 1995, Viscarra Rossel et al. 2006) has
had some success applying similar methodologies to soils from broad geographical extents,
little attempt has been made predict values over local areas. Furthermore, the broad
geographical extents of these studies limits the utility of reflectance spectroscopy
methodologies to supplement soil survey to improve the resolution of regional soil maps.
Nonetheless, the results of the model developed here provide clear evidence that the
methodology can be applied to areas of limited variability with relative success. Moreover,
this research shows that the prediction of soil properties from reflectance spectroscopy can
be used with geostatistical methods such as kriging in order to develop soil maps.
7.3.3 Unmixing of soil types and estimation of soil exposure with simulated
hyperspectral imagery (Chapter 5)
The classification of soils and determination of exposure from mixed pixels of various soil
and vegetation types was examined using simulated imagery. Limitations of this
methodology arise from the small number of soils and vegetation cover types used in the
simulated imagery. It is quite likely that greater variation in soil and vegetation would be
encountered in some image studies. Nonetheless, there can be little doubt that the results
presented in this study demonstrate spectral confusion in the unmixing. Further limitations
include the simulated data itself providing a near perfectly linear environment and the
absence of topographic variation, which is unlikely to be encountered in image studies.
However, while these factors are a limitation, they also provide for quantitative assessment
of the unmixing itself by restricting the number of variables. The wider implications of this
study are that combined soil and vegetation mixes can be confused for different soil types
and this must be considered in future work.
7.3.4 Mapping soil variability with hyperspectral image data (Chapter 6)
The mapping of soil variability with hyperspectral imagery provided limited success in
identifying surface soils of measurably different properties. Four spectrally distinct soil
endmembers were extracted from the image and used to map distinctly different areas of
agricultural landscape. However, the methodology failed to identify measurably different
soils. Causes of this may be similar to those outlined in Section 7.3.3 where linear spectral
unmixing was found to confuse soil endmembers with mixed pixels of vegetation and soil.
Chapter 7: Conclusion 120
However, the complexity of the image data does not allow for such a definitive conclusion
to be drawn. Other possible causes include poor identification of endmembers, insufficient
sampling points to measure differences in the unmixing outputs and lack of soil variation
in the landscape. Time and funding restrictions have prevented return field visits that
would allow for further soil sampling and the opportunity to refine conclusions. While
these results are disappointing they do raise many questions for further avenues of
research. The scope for using methodologies like this to map soil variability is ever
increasing with new satellites planned that would allow for finer temporal resolution
through repeat visits. Greater image data availability and improved field work may provide
definitive results that were not achieved in this study.
7.4 Recommendations for future research
The following areas of necessary research were identified through the work presented in
this thesis.
• Further assessment of the utility of spectral discrimination of soil field survey
classes may provide more useful results. The work presented in this thesis could be
improved through the introduction of quality tests of field survey classifications.
The easiest way to do this is to utilise multiple operators and compare field survey
results before the applying the discriminant analysis.
• The next step in the prediction of soil properties using reflectance spectroscopy is
to incorporate sub-surface soils analysis. This could be achieved through the
analysis of soil cores similar to that currently done with geological cores (Mauger
et al. 2004, Ben-Dor et al. 2008). Such an analysis would provide a three
dimensional understanding of soil variability crucial for complete landscape
management.
• Follow up investigations should be made into the spectral unmixing of soil and
vegetation. This would require a more exhaustive physical soil analysis to provide
validation data. The inclusion of properties such as soil water, soil colour and
surface crusting in the analysis may improve the results.
• The introduction of new satellite hyperspectral image sensors will provide the
ability to repeat sample areas of the landscape at a high spectral and spatial
Chapter 7: Conclusion 121
resolution. One possible advantage of such instruments is that complete soil
coverage within regions could be built up over time as seasonal changes in land
cover provide for direct soil exposure to the sensor.
7.5 Conclusion
This thesis has contributed significantly to improving the use of reflectance spectroscopy
and remote sensing in mapping and understanding soil variability in the landscape. The
prediction of soil properties using reflectance spectroscopy is a powerful tool and could
certainly aid in improving the resolution of soil maps. This technique could be applied to
other regions with the development of new prediction models and could also be expanded
to include sub-surface soil properties, thus providing a three dimensional soil map.
Understanding the spectral unmixing of soil and vegetative cover is an important
component for successful image mapping of surface soil variability. The simulated
imagery provided a useful tool to demonstrate some of the problems encountered when
using unmixing algorithms with hyperspectral imagery. While less successful, the partial
unmixing of image derived soil endmembers form hyperspectral image data may yet
provide a useful tool in understanding soil variability at relatively fine scales and over
large extents. However, further research in this area with improved datasets is required to
develop a useful tool for this application.
7.6 References
Ben-Dor, E., Carmina, K., Heller, D. and Chudnovsky, S. 2008 A novel combined optical method for (sic) objectively map soil in a near real time domain, In The 21st Congress of the International Society for Photogrammetry and Remote Sensing, Beijing, China, 3-11 July 2008.
Janik, L. J., Skjemstad, J. O. and Raven, M. D. 1995 Characterization and analysis of soils using mid infrared partial least-squares, I. Correlations with XRF-determined major-element composition, Australian Journal of
Soil Research, 33, 621-636.
Mauger, A. J., Keeling, J. L. and Huntington, J. F. 2004 Bringing remote sensing down to earth: CSIRO Hylogger as applied to the Tarcoola goldfield, South Australia, In 12 Australasian Remote Sensing and Photogrammetry Conference, Fremantle, Western Australia,
McDonald, R. C. and Isbell, R. F. 1990 Soil profile, In Australian soil and land survey: Field handbook (Eds, R. C. McDonald, R. F. Isbell, J. G. Speight, J. Walker and M. S. Hopkins) Inkata Press, Melbourne.
Viscarra Rossel, R. A., Walvoort, D. J. J., McBratney, A. B., Janik, L. J. and Skjemstad, J. O. 2006 Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties, Geoderma, 131, 59-75.