ANALYSIS OF LAND USE AND LAND COVER CHANGE IN KISKATINAW RIVER WATERSHED: A REMOTE SENSING, GIS & MODELING APPROACH
by
Siddhartho Shekhar Paul
B.Sc. (Honors), University of Dhaka, Bangladesh, 2008 M.Sc., University of Dhaka, Bangladesh, 2009
THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF SCIENCE IN
NATURAL RESOURCES AND ENVIRONMENTAL STUDIES (ENVIRONMENTAL SCIENCE)
UNIVERSITY OF NORTHERN BRITISH COLUMBIA
August, 2013
© Siddhartho Shekhar Paul, 2013
UMI Number: 1525672
All rights reserved
INFORMATION TO ALL USERS The quality of this reproduction is dependent upon the quality of the copy submitted.
In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.
UMIDissertation PiiblishMig
UMI 1525672Published by ProQuest LLC 2014. Copyright in the Dissertation held by the Author.
Microform Edition © ProQuest LLC.All rights reserved. This work is protected against
unauthorized copying under Title 17, United States Code.
ProQuest LLC 789 East Eisenhower Parkway
P.O. Box 1346 Ann Arbor, Ml 48106-1346
ABSTRACT
This thesis study was conducted to capture the land use and land cover
(LULC) change dynamics in Kiskatinaw River Watershed, BC, Canada. A
combination of remote sensing, GIS and modeling approach was utilized for this
purpose. Landsat TM and ETM+ satellite images of the years 1984, 1999 and 2010
were analyzed using object oriented image classification technique to produce LULC
maps and detect the associated changes. The dynamic nature of different forest
types, increase in built-up area and significant depletion of wetlands were found to
be notable among the detected LULC changes. Thereafter, a multi-layer perception
neural network technique was used to model transition potentials of various LULC
types, which was later realized with a Markov Chain land use model to predict
future changes. The integration of advanced satellite remote sensing tools and
neural network aided Markov Chain modeling was illustrated to be an effective
means for LULC change detection and prediction in Kiskatinaw River Watershed.
Table of Contents
ABSTRACT.............................................................................................................................. ii
Table of C ontents...................................................................................................................iii
List of Figures.......................................................................................................................... v
List of Tables..........................................................................................................................vii
ACKNOWLEDGEMENT................................................................................................... viii
DEDICATION........................................................................................................................ ix
CHAPTER 1: INTRODUCTION........................................................................................... 1
1.1 Background.................................................................................................................... 1
1.2 Purpose of study............................................................................................................ 4
1.3 Organization of the thesis............................................................................................ 5
CHAPTER 2: LITERATURE REVIEW................................................................................. 6
2.1 Land use and land cover (LULC) changes................................................................6
2.1.1 Concepts of LULC change..................................................................................... 6
2.1.2 Remote sensing (RS) and GIS techniques in LULC change analysis............. 8
2.2 Land use m odeling...................................................................................................... 15
2.2.1 Artificial neural network (ANN) m odels.......................................................... 16
2.2.2 Spatial statistical models...................................................................................... 16
2.2.3 Cellular Automata (CA) models......................................................................... 17
2.2.4 Application of land use m odels.......................................................................... 17
CHAPTER 3: DATA AND METHODS.............................................................................. 20
3.1 Overview of Kiskatinaw River W atershed............................................................20
3.1.1 Location and extent.............................................................................................. 20
3.1.2 Kiskatinaw R iver.................................................................................................. 22
3.1.3 Study area sub-basin............................................................................................ 24
3.1.4 Surficial Geology, Soil & Biophysical characteristics..................................... 25
iii
3.1.5 Water use values in KRW.................................................................................... 27
3.2 Methodology................................................................................................................ 28
3.2.1 Reconnaissance su rvey ........................................................................................ 28
3.2.2 Data selection and collection............................................................................... 28
3.2.3 Description of the land use and land cover classes.........................................32
3.2.4 Image pre-processing and analysis....................................................................40
CHAPTER 4: RESULTS AND DISCUSSION.................................................................... 71
4.1 Results of RS & GIS analysis of satellite im ages.....................................................71
4.2 Accuracy assessm ent.................................................................................................. 78
4.3 Discussion on LULC changes.................................................................................... 81
4.3.1 LULC change in individual sub-basin...............................................................85
4.3.2 Wetland depletion.............................................................................................. 108
4.3.3 Natural gas development infrastructure......................................................... I l l
4.4 Land use m odeling.................................................................................................... 113
4.4.1 Results of modeling analysis............................................................................. 113
4.4.2 Discussion on modeled outcomes.................................................................... 117
CHAPTER 5: CONCLUSION AND RECOMMENDATIONS.....................................126
5.1 Sum m ary.................................................................................................................... 126
5.2 Limitations and recommendations for future w ork.............................................130
Bibliography.........................................................................................................................132
List of Figures
Figure 1: Variation in spectral reflectance for different LULC types............................12Figure 2: Difference in PBC and OOIC.............................................................................. 13Figure 3: Study area- Kiskatinaw River W atershed.........................................................21Figure 4: Channel network within Kiskatinaw River W atershed..................................23Figure 5: Kiskatinaw River Watershed sub-basins..........................................................24Figure 6: Surficial geology of KRW.................................................................................... 26Figure 7: Different wavelengths of Landsat b ands..........................................................32Figure 8: Cropland sampled during ground truth survey..............................................33Figure 9: Coniferous forest sampled during ground truth survey................................33Figure 10: Deciduous forest sampled during ground truth survey..............................34Figure 11: Mixed forest sampled during ground truth survey......................................35Figure 12: Planted or re-growth forest sampled during ground truth su rvey ........... 36Figure 13: Forest fire area sampled during ground truth survey..................................36Figure 14: Cut block sampled during ground truth su rvey ...........................................37Figure 15: Pasture land sampled during ground truth survey......................................38Figure 16: View of One Island lake sampled during ground truth survey................. 38Figure 17: Wetland sampled during ground truth survey .............................................39Figure 18: A gas development infrastructure sampled during ground truth survey 40Figure 19: Band- 4 of Landsat 2010 image before processing.........................................42Figure 20: Mosaiced 2010 image with 5-4-3 band combination ............................43Figure 21:1984 Landsat satellite image used in this study.............................................45Figure 22:1999 Landsat satellite image used in this study.............................................46Figure 23:2010 Landsat satellite image used in this study.............................................47Figure 24: Number of segments and fineness varying with ST value...........................51Figure 25:1984 image classification; A) MAXLIKE, B) SEGCLASS..............................54Figure 26:1999 image classification; A) MAXLIKE, B) SEGCLASS..............................55Figure 27: 2010 image classification; A) MAXLIKE, B) SEGCLASS..............................56Figure 28: Landsat image analysis fram ework.................................................................60Figure 29: MLP neural network (after Eastman, 2012)....................................................64Figure 30: Land use modeling fram ew ork........................................................................70Figure 31: Area covered by each LULC type in 1984.......................................................72Figure 32: KRW LULC map of 1984...................................................................................73
Figure 33: Area covered by each LULC type in 1999.......................................................74Figure 34: KRW LULC map of 1999................................................................................... 75Figure 35: Area covered by each LULC type in 2010.......................................................76Figure 36: KRW LULC map of 2010................................................................................... 77Figure 37:1984-1999 change analysis for LULC types....................................................83Figure 38:1999-2010 change analysis for LULC types....................................................84Figure 39: LULC map of Mainstem sub-basin in 1984.................................................. 87Figure 40: LULC map of Mainstem sub-basin in 1999.................................................. 88Figure 41: LULC map of Mainstem sub-basin in 2010.................................................. 89Figure 42: Change in LULC within Mainstem sub-basin..............................................90Figure 43: LULC map of Brassey sub-basin in 1984....................................................... 91Figure 44: LULC map of Brassey sub-basin in 1999....................................................... 92Figure 45: LULC map of Brassey sub-basin in 2010....................................................... 93Figure 46: Change in LULC within Brassey sub-basin....................................................93Figure 47: LULC map of Flalfmoon-Oetata sub-basin in 1984..................................... 95Figure 48: LULC map of Halfmoon-Oetata sub-basin in 1999..................................... 96Figure 49: LULC map of Halfmoon-Oetata sub-basin in 2010..................................... 97Figure 50: Change in LULC within Halfmoon-Oetata sub-basin..................................97Figure 51: LULC map of East KRW sub-basin in 1984....................................................99Figure 52: LULC map of East KRW sub-basin in 1999..................................................100Figure 53: LULC map of East KRW sub-basin in 2010..................................................101Figure 54: Change in LULC within East KRW sub-basin.............................................102Figure 55: LULC map of West KRW sub-basin in 1984.................................................104Figure 56: LULC map of West KRW sub-basin in 1999.................................................105Figure 57: LULC map of West KRW sub-basin in 2010.................................................106Figure 58: Change in LULC within West KRW sub-basin............................................107Figure 59: Wetlands converted to other land-use type (1984-2010).............................109Figure 60: Gains and losses in wetland area (1984-2010)..............................................110Figure 61: Changes in built-up area and cut block (1984-2010)...................................I l lFigure 62: Transition probabilities from planted/re-growth forest to deciduous forest 116Figure 63: Intermediate stage LULC m ap from hard prediction.................................120Figure 64: Final stage LULC map from hard prediction...............................................121
Figure 65: Difference in area for each LULC type between existing and predictedm aps................................................................................................................................... 123Figure 66: Soft prediction output for 2020.......................................................................125
List o f Tables
Table 1: Description of satellite imageries used in LULC change detection............... 31Table 2: Spectral features of Landsat bands......................................................................31Table 3: Driver variables for transition potential m odeling...........................................67Table 4: Surface area covered by each LULC type in a particular y e a r ........................78Table 5: Accuracy assessment error matrix for 2010 image classification................... 79Table 6: Accuracy assessment sum m ary ...........................................................................80Table 7: Sensitivity of transition model for forcing independent variables...............115Table 8: Transition probability m atrix ............................................................................. 117Table 9: Expected transition of pixels...............................................................................119Table 10: Area calculated for predicted LULC m aps.....................................................123
ACKNOWLEDGEMENT
When you leave home for an uncertain period of time and travel half way of
the world to achieve something, you m ust know that you are starting a difficult
journey to reach that 'special achievement'. Pursuing the graduate study at UNBC
was that special to me. Not to mention, this precious journey would not have been
possible without the gracious support from some great people.
First and foremost, I would like to express my sincere gratitude to my
supervisor Dr. Jianbing Li whose unreserved support and guidance at every stage of
doing this research mentored me to reach the goal. His elaborate efforts and patience
for my intellectual development are truly appreciated. I would also like to offer my
indebtedness to my committee members Dr. Roger Wheate and Dr. Liang Chen for
their guidance throughout this research. Despite busy schedules, they were more
than willing to meet me when I needed advice.
I am grateful to Scott Emmons, Senior Instructor, GIS lab for his various sorts
of technical support. My sincere thankfulness goes to my research team member,
Gopal Saha and others for being beside me for every single need. I would like to
offer my appreciation to my family and friends who are with me to share my laughs
and cries and who I can feel every moment. The research was funded by Peace River
Regional District, City of Dawson Creek, Geoscience BC, Encana, and BP Canada.
DEDICATION
I dedicate this work to the people who feel and fight for a partition free world and
who care for unbiased goodwill towards humankind
ix
CHAPTER 1: INTRODUCTION
1.1 Background
Human beings, since the earliest stage of settlement, are dependent on land
for their food production and various sorts of economic development which have
been constantly modifying the global landscape. The relentless pressure to meet the
needs of burgeoning population and demand driven development activities have
amplified the stress on earth's land (Foley et al., 2011; Weinzettel et al., 2013). In this
context, anthropogenic activity and its concomitant land use and land cover (LULC)
changes have become an inevitable issue for the present time and accentuating the
risks of environmental degradation around the globe (Paiboonvorachat, 2008;
Stabile, 2012).
Over half of the world's landscape is influenced by hum an activities or under
some sort of anthropogenic development and since the historic past, many natural
resources have been heavily used or even depleted in the worst cases (Foley, et al.,
2005, Goldewijk, et al., 2011). The impacts of this widespread LULC change on the
natural environment are multi faceted, including climate change, alteration of
hydrological cycle, increased water extraction, impairment of water quality,
degradation of soil nutrients, amplified surface erosion, and loss of biodiversity
(Turner, et al., 2007, Paiboonvorachat, 2008). Therefore, information on land use and
1
land cover, changing trends and optimal use of the land resources have become
predestined criteria for land use planning and effective natural resources
management of an area.
Watersheds in the north-eastern part of British Columbia (BC), Canada have
been experiencing widespread LULC changes over the past few years due to the
convergence of various industrial interests, for example, logging, mining, oil and gas
development, large scale hydro development etc. (Lee & Hanneman, 2012). Among
the north-eastern BC's watersheds, Kiskatinaw River watershed (KRW) which is the
study area of this research, features a significant portion of industrial development
activities and associated LULC changes. Being the only source of drinking water
supply to the City of the Dawson Creek (DC) and neighbouring village of Pouce
Coupe, KRW plays a dominant role in north-eastern BC's life and environment
(Saha, et al., 2013), but unfortunately, information on LULC changes within the
watershed is scanty. Therefore, there is a pressing need to understand the LULC
system within the watershed to assess its impact on the overall watershed dynamics.
Researchers around the globe have long been enjoying the effectiveness of
remote sensing (RS) technology for extracting current and previous land use and
land cover (LULC) information and for providing robust inventory of LULC
changes (Ridd and Liu, 1998; Mas, 1999; Paul et al., 2012; Chen et al., 2013). Recent
2
advancement of RS tools and combination of Geographic Information System (GIS)
with RS makes this technique more successful and introduces a wider scope of
research including LULC change detection, LULC modeling and prediction (Araya,
2009, Paul et al., 2012). Every change in land cover can be reflected by the alteration
of radiance value captured by the remote sensor, e.g. satellite image sensor (Mas,
1999). Later, this variation in radiance value is gauged by comparing multi-temporal
satellite images or aerial photographs (Chen et al., 2013; Saha et al., 2013) and LULC
maps are produced for change detection. The information gathered from the change
detection analysis can be further realized by a land use modeling approach. The
simulation of land use models has recently proven valuable in land use planning,
environmental impact assessment, policy making etc. (Schulp, et al., 2008; Kline, et
al., 2007) and thus, being widely used globally. Land use models are capable of
exploring the transition potentials of various LULC types for a given set of driver
variables (Kamusoko, et al., 2009). This information can then be used for predicting
future LULC information for a study area.
It is anticipated that this thesis study will provide a comprehensive insight
about the land use-land cover system within the Kiskatinaw River watershed. The
LULC information congregated by RS, GIS and modeling analysis of this research
will update decision makers and development practitioners about the magnitude
3
and nature of long term LULC change in KRW. As a result, informed environmental
policies and management strategies could be implemented and practiced.
1.2 Purpose of study
This thesis study aims to capture the LULC change in Kiskatinaw River
watershed (KRW) to compare scenarios before and after the economic growth in this
vicinity using RS, GIS and modeling techniques. So, the specific objectives
encompassed by this study are:
• To assess the changes in land use and land cover occurring within KRW
based on the analysis of remotely sensed satellite imagery
• To model the transition potentials for each LULC type
• To predict future LULC scenarios
4
1.3 Organization of the thesis
Overview on the study area
Detail of the data used
Purpose of the study
Structure of the thesis
Background of the study
Describes the data analysis process
Chapter 1: Introduction
Introduces the research
Chapter 3: Data and Methods
Provides details about the chosen study methods
Chapter 2: Literature Review
Explains existing literatures and context of the study, introduces available methods of analysis
Chapter 4: Results and Discussion
Describes the analyzed results
Presents the results Explains the deliverables
__________ i____________Chapter 5: Conclusion
Summarizes the research outputs,
explains limitations and provides directions for future research
5
CHAPTER 2: LITERATURE REVIEW
2.1 Land use and land cover (LULC) changes
2.1.1 Concepts of LULC change
Although the terms 'land use' and 'land cover' are sometimes used
interchangeably, each has a distinct meaning. Land cover is the bio-physical layer
covering the earth surface, while land use represents the human utilization of the
land cover. Land cover includes earth's land surface distribution of vegetation,
water, desert and ice as well as the biota, soil, topography etc. in the immediate
subsurface, and it also includes hum an activity areas, such as settlement, mine
exposure etc. (Lambin, et al., 2003; Oumer, 2009). On the other hand, land use is
attributed to how humans exploit the land cover to serve their own purposes and
includes features such as residential zones, agricultural farms, logging areas etc.
(Zubair, 2006; Oumer, 2009). In this context, land use influences the changes in land
cover; therefore, LULC change can be defined as the modification of surface features
on earth's landscape which is realized by the difference in their surface appearance
assessed at two different times (Ayele, 2011).
The current global condition of mass environmental change and
sustainability issues elucidates the gravity of LULC change detection research in
different parts of the world. Though LULC changes entail both natural (e.g. weather,
6
flooding, earthquake etc.) and anthropogenic causes, the ever-increasing demand of
the mushrooming population has designated the anthropogenic influences as the
most dramatic (Turner, et al., 2007; Foley, et al., 2011; Weinzettel, et al., 2013).
At present, the undisturbed pristine areas cover less than 50% of the total
earth's landscape; forest cover is only 30% which was around 50% some 8000 years
ago (Oumer, 2009). Diverse and intense anthropogenic activities around the world
are attributed for most of these LULC changes. For this reason, research is
conducted around the world to study this dynamic LULC alteration and devoted
efforts are underway to explore its connection with the disturbances happening in
the earth system.
7
2.1.2 Remote sensing (RS) and GIS techniques in LULC change analysis
2.1.2.1 Definition of RS and GIS:
Remote sensing is defined in various ways in the literature, but with similar
meaning. For example, United Nations (1986) defined remote sensing as
'the sensing of the Earth’s surface from space by making use o f the properties o f
electromagnetic waves emitted, reflected or diffracted by the sensed objects, for the purpose of
improving natural resources management, land use and the protection of the environment'
Lillesand, et al. (2008), defined remote sensing as (p. 1):
'the science and art o f obtaining information about an object, area, or phenomenon
through the analysis o f data acquired by a device that is not in contact with the object, area,
or phenomenon under investigation'
In short, remote sensing is the study of satellite images or aerial photographs which
are capable of differentiating earth's land use and land cover types by variation in
their electromagnetic signature. On the other hand, geographic information systems
(GIS) refer to any scientific effort that incorporates geographical data to visualize,
analyze, and explore geographically referenced information of a location.
8
The United States Geological Survey (USGS, 2007) defined GIS as:
'In the strictest sense, a GIS is a computer system capable of assembling, storing,
manipulating, and displaying geographically referenced information, that is data identified
according to their locations'
2.1.2.2 Application in LULC mapping
In combination, RS and GIS serve efficiently for earth observations and
associated information analysis (Paiboonvorachat, 2008; Araya, 2009; Paul, et al.,
2012). Viewing earth from space enables us to comprehend the cumulative influence
of hum an activities on earth surface's natural state. Capturing and analyzing this
information by the RS and GIS tools provides a cost effective record of LULC in an
accurate and timely m anner (Ridd & Liu, 1998; Chen, et al., 2013).
Availability of multiple satellite sensors offering image data with fine spatial
resolution, high geometric precision and short revisit intervals has m ade satellite
remote sensing more appealing than aerial photography and manual data collection
methods for LULC change detection and modeling (Aplin, et al., 1997; Stabile, 2012).
With the advancement of satellite image analysis tools and ease-of-access to various
off-the-shelf image processing software, satellite remote sensing has been gaining
9
wide popularity for investigating LULC change. For example, Musaoglu, et al.,
(2005) analyzed Landsat and SPOT satellite images for land use change monitoring
(1975-2001) in the Beykoz region of Istanbul. Land cover change dynamics were
monitored in Africa using high spatial resolution satellite data by Brink & Eva
(2009). Supervised classification of Landsat image was performed by El-Kawy, et al.
(2011) to provide recent and historical LULC conditions for the western Nile delta.
Satellite remote sensing was also employed in New Zealand for estimating change
in forest cover, i.e. area of afforestation and deforestaion to meet the reporting
obligation under Kyoto protocol (Dymond, et al., 2012). Land use change derived by
shrub cover growth in northern slope of Alaska was mapped by Beck et al. (2011)
using IKONOS and SPOT satellite data. Thus, satellite remote sensing is being vastly
utilized at different parts of the world for diverse LULC change detection
approaches.
2.1.2.3 Satellite image analysis
Satellite image analysis entails digital image processing which involves
manipulation and interpretation of the digital image data by specific computer
programs to display and extract meaningful information about the surface of the
earth. Digital image classification which is among the basic image analysis processes
10
governs most of the LULC change detection study (Matinfar, et al., 2007). Image
classification which is normally performed on multispectral images, i.e. images with
more than one spectral band, automatically categorizes all the image pixels into
various land cover classes based on their similar digital num ber (DN) values
(Lillesand, et al., 2008).
The satellite sensors record the variation in the electromagnetic radiation
from each part of the earth surface and assign it with a distinct DN value for each
spectral band (Oumer, 2009). The range of digital number varies from sensor to
sensor and depends on the radiometric resolution, which is attributed to the sensor's
sensitivity to various level of incoming energy (Ayele, 2011). For example, Landsat
Multispectral Scanner (MSS) satellite sensor detects radiation in the range from 0 to
63 DN; while Landsat Thematic Mapper (TM) sensor's DN value ranges from 0 to
255 (NASA, 2011). The variation in the spectral reflectance of a particular LULC type
is captured during the digital image classification process and thus the LULC map
of an area is constructed. For instance, the spectral signature of water is different
from that of vegetation for each band of a multi-spectral imagery and vice-versa.
Figure 1 explains the variation in spectral reflectance for three LULC types: water,
vegetation and soil in Landsat TM imagery.
11
"mm fit m L«̂ TMb«di n i50 -
Sail
LVegetation
30
Watero04 O ui U 14 2.0 23
Figure 1: Variation in spectral reflectance for different LULC types
(adapted from Richards & Jia, 1999)
Various classification approaches and algorithms have been adopted by
researchers around the globe for classifying satellite imagery (Gao & Mas, 2008). The
conventional method is a pixel based classification (PBC) technique which classifies
the image based on each single image pixel (Dean & Smith, 2003) (Figure 2). The
remotely sensed satellite imagery comprises rows and columns of pixels whose
spectral similarity and dissimilarity work as the basis of PBC. The classification
process groups the like-pixels under distinct LULC types (Casals-Carrasco, et al.,
2000). Though the PBC technique has been well developed and successfully applied
in many cases, it has some limitations essentially because spatial photo-interpretive
elements namely texture, context and shape are disregarded during PBC and the
12
pixels do not represent true geographical objects (Hay & Gastilla, 2006; Blaschke,
2010). These issues contribute to lower classification accuracy in many studies.
a) Pixel based classification b) Object oriented classification
Green = Vegetation, Blue = Water
a) Individual pixels have been identified for vegetation and water classes based on spectral reflectance i.e. DN values; b) Image objects or segments comprising several pixels have been identified for vegetation and water classes based on homogeneities of spectral, spatial and
other characteristics.
Figure 2: Difference in PBC and OOIC
PBC classification techniques which employ the supervised method,
unsupervised method or some combinations (Enderle & Weih, 2005), do not
consider the spatial and contexual information of the pixels of interest; but this
information could be used to produce more accurate LULC classification output,
13
particularly when using high resolution satellite data (De Jong, et al., 2001; Benz, et
al., 2003; Dwivedi, et al., 2004; Matinfar et al., 2007; MacLean et al., 2013). In this
context, the object oriented image classification (OOIC) technique m ade a timely
arrival in the research era of remote sensing. Although the concept of OOIC was
introduced in 1970s, OOIC started to attract the demand of researcher community
after mid-1990s with the advancement of remote sensing data processing software
and hardware as well as with the increased availability of high spatial and spectral
resolution imagery (De Kok, et al., 1999).
Over time, many faceted issues regarding PBC have increased dissatisfaction
among the remote sening users which has been compensated by the object oriented
image classification (OOIC) gaining wide popularity over the last few years
(Blaschke, 2010; Chen et al., 2013). Unlike per-pixel classification, OOIC classifies the
imagery by image segments which comprise groups of spectrally homogeneous
pixels. Segments are building blocks in OOIC (Hay and Castilla, 2008; Blaschke,
2010) and govern the classification process by their own characteristics (such as:
segment size, shape, texture, zonal statistics etc.) instead of the individual
characteristics of each pixel (MacLean et al., 2013). Image segmentation which is the
basic step in object oriented classification, divides the imagery into homogeneous,
continuous and contiguous objects (Gao & Mas, 2008). Figure 2 demonstrates the
difference between PBC and OOIC. In Figure 2a vegetation and water were
14
classified pixel by pixel, but in Figure 2b, vegetation and water objects were
identified by segment based classification.
The image segments or objects provide OOIC the leverage of using spatial,
spectral, textural and contextual information during LULC classification.
2.2 Land use modeling
Land use modeling, at the present time, plays a pivotal role in many natural
resources management and decision making processes. Land use models are
effective tools to analyze the causes and consequences of land use-land cover change
and create an enhanced understanding of the land use system in an area (Verburg, et
al., 2004; Stabile, 2012). The use of land change models is multi-dimensional. For
example, they were used in biodiversity monitoring (Verburg, et al., 2008), for
estimating loss of vegetation cover (Echeverria, et al., 2008), for forest management
(Kamusoko, et al., 2013), in urban expansion and planning (Sun, et al., 2007) etc..
Researchers around the globe have been devising and utilizing a wide variety
of land use models, all of which are diverse in their formulations, objectives and
capabilities. There are whole landscape models, distributional landscape models as
well as spatial landscape models (Baker, 1989; Singh, 2003). Since the spatial details
including natural and human processes have greater impacts on land use change
15
system, spatial modeling has taken over other modeling methods in m any studies.
Progress in remote sensing and GIS research has made significant contributions in
these spatial landscape modeling methods (Singh, 2003).
Spatial land use modeling research has employed various approaches, a few
of which has been explained below.
2.2.1 Artificial neural network (ANN) models
ANN serves as a machine learning tool which is capable of quantifying and
forecasting complex behaviour and patterns of LULC change (Pijanowski, et al.,
2002). ANN imitates the interconnected neural system in the hum an brain. The basic
element or nodes, called neurons are connected in layers and perform the processing
in an ANN model. The neuron's output is derived from multiplying the input signal
by specific weights which are determined by using training algorithms of which
back propagation is the most popular approach (Pijanowski, et al., 2002; Singh, 2003;
Pijanowski, et al., 2005). ANN is used to identify the pattern of land use change and
hence, transition from one land use type to another can be predicted.
2.2.2 Spatial statistical models
Spatially explicit statistical modeling of LULC changes is a widely used
approach for understanding processes related to LULC change and quantifying their
16
influences on the change dynamics (Semeels & Lambin, 2001). Various spatial
statistical modeling techniques have been adopted by researchers for understanding
current LULC change and projecting the future change scenarios. Multiple linear
regression, Markov Chain methods, Multivariate modeling tools etc. are just a few.
When spatial information is aggregated with statistical analysis, the land use
modeling becomes more realistic and effective (Veldkamp & Lambin, 2001).
2.2.3 Cellular Automata (CA) models
The cellular automata (CA) is a popular spatially explicit land use modeling
tool. The output of the CA model emerges from interactions between individual
cells which are the fundamental modeling unit (Batty, 2005); this is why, CA is
frequently considered as powerful technique for modeling complex land use change
(Hasbani, 2008). CA is enhanced by its natural affinity with GIS and remotely sensed
data use (Torrens & O'Sullivan, 2001).
2.2.4 Application of land use models
Specific land use change process was focused in many land use modeling
approaches (Mas, et al., 2004; He, et al., 2008), whereas some other models integrated
multiple change dynamics (Dietzel & Clarke, 2007; Overmars, et al., 2007). A
Regressional statistical land use model was used by Aspinall (2004); a cellular
17
automata mechanistic model was applied by Walsh et al. (2008) for modeling
agricultural expansion and deforestation; a combination of various models has also
been used in many studies as in Kamusoko et al. (2013) for modeling multiple land
use processes. Though complex models are capable of generating robust output,
simulation of these models entails rigorous and difficult parameterization as well as
large cost and time (Benito et al., 2010). However, Markov Chain as a simple land
use model, is a useful and popular tool in this context and covers large spatial
extent (Weng, 2002). The Markov model calculates transition matrix for various land
use features based on current driving factors and predicts the future land use change
pattern if the driving forces continue in future (Mubea et al., 2010). Markov Chain
has been successfully used in many studies with few reported issues in varied
settings; for example Weng (2002) employed Markov model along with remote
sensing and GIS analysis to model land use dynamics in a coastal region of China;
Islam & Ahmed (2011) modeled urban sprawl in Dhaka city using GIS aided
Markovian modeling, while Freier, et al. (2011) used Markov Chain for modeling
rangelands under climate change scenarios in semi-arid environment of Morocco.
The transition matrix which comprises all the estimated transition potentials
serves as one of the basics in Markov Chain land use modeling. The non-parametric
multi-layer perception (MLP) technique which is an artificial neural network has the
merit to fit complex non-linear relationships between driving variables and land use
18
for producing accurate transition potential estimation (Sangermano, et al., 2010;
Eastman, 2012). Multi-layer perception tool has the ability to perform efficiently
even with less training data which has made it a convenient and preferred technique
(Civco, 1993; Chan, et al., 2001; Martinuzzi, et al., 2007). The augmentation of
computing power and performance as well as availability of user friendly software
packages have supported and substantially increased the use of MLP neural
network techniques in land use studies (Li & Yeh, 2002; Pijanowski, et al., 2002). A
number of studies have reported the effectiveness of artificial neural networks in
land use modeling (Pijanowski, et al., 2005; Almeida, 2008; Lin, et al., 2011). It is
claimed in various studies that a MLP network with 3 layers - input, hidden and
output is capable of estimating any polynomial function and its ability is almost
unequivocal for solving very complex regression and land use classification and
modeling problems (Eastman, 2012). So the integration of MLP neural network and
Markov Chain model could reinforce the land use modeling study by aggregating
statistical and spatial characteristics of LULC variations.
19
CHAPTER 3: DATA AND METHODS
3.1 Overview of Kiskatinaw River Watershed
3.1.1 Location and extent
Kiskatinaw River Watershed (KRW) is located on the Alberta Plateau of
north-eastern British Columbia, near the British Columbia-Alberta border (Figure 3).
KRW lies between longitude 119° 59' W to 121° 7' W and latitude 54° 58' N to 56° 5'
N. The total area of the watershed is 4097 km2, although this study focuses on the
upper Kiskatinaw watershed which is 2836 km2 (Figure 3). The City of Dawson
Creek and the village of Pouce Coupe are located on the north-east of the study area.
The municipality of Tumbler Ridge is near the south-western periphery of KRW.
Arras, located at the northernmost edge of the study area serves as Dawson Creek's
water intake station. Steep slopes of the Rocky Mountain Foothills characterize the
western portion of the watershed, while undulating plains projecting into BC from
Alberta delineate the eastern portion (Kiskatinaw River IWMP, 1991).
20
55°0
’0"N
55°2
0'0"N
55
°40,
0"N
56°0
*0MN
120°50'0MW 120°0'0"W
Poucacoupe
A Major locationMajor roads______________________ L120°50,0,,W 1 2 0 W W
Figure 3: Study area- Kiskatinaw River Watershed
55°0
'0HN
55°2
0'0"N
55
°40'0
"N
56°0
'0"N
3.1.2 Kiskatinaw River
The Kiskatinaw River is a tributary of the Peace River. The River originates in
the foothills of the Rocky Mountains, near Tumbler Ridge, and flows approximately
200 km north before joining the Peace River at the Alberta border with BC (Peace
Forest District, 2010) (Figure 4). The water supply area rises from an elevation of
680m at Arras in the northernmost edge to 1,300m at Bear Hole Lake in the
southernmost boundary (Dobson Engineering Ltd., 2007). The eastern and western
confluences meet the main confluence of the river almost at the middle of the
watershed near its eastern border (Figure 4). The average annual flow rate is 10 m3/s,
but in January it drops to 0.052 m3/s which makes it more complicated to establish
an effective water resources management for the watershed (City of Dawson Creek,
2009). The watershed receives an average annual precipitation of 499 mm,
comprising 320 mm of rain and 179 mm of snow.
22
LegendMain ch an n e l o f K iskatinaw River
Tributaries of K iskatinaw River
Elevation (m)
Figure 4: Channel network within Kiskatinaw River Watershed
23
3.1.3 Study area sub-basin
For study purposes, the KRW study area has been divided into five sub
basins which are Mainstem, Brassey, Halfmoon-Oetata, East KRW and West KRW
(Figure 5). Among these, West KRW covers the largest area of 1005 km2, followed by
East KRW (996 km2), Mainstem (433 km2), Brassey (208 km2) and Halfmoon-Oetata
(194 km2).
Figure 5: Kiskatinaw River Watershed sub-basins
24
3.1.4 Surficial Geology, Soil & Biophysical characteristics
Surficial Geology map of the KRW area is displayed in Figure 6. Seven types
of surficial deposits are identified in this watershed, namely alluvial, colluvial,
eolian, glaciofluvial, lacustrine, morainal, and organic. Morainal deposits
predominate the area whereas alluvial deposits exist mostly along the confluences of
the river. Eolian deposits are observed in the central-eastern areas.
Lower slopes and valley bottoms are predominately covered in thick
sequences of fine-grained lacustrine deposits. Glaciofluvial, colluvial and organic
deposits are not widespread and can be found locally. Colluvial deposits which are
derived mainly from mass-wasting processes are found on mid-slopes of the
watershed (Dobson Engineering Ltd., 2007). Clay and silt loams are the dominant
soil type in this area although sandy loam is also found.
Most of the watershed belongs to Boreal White and Black Spruce
biogeoclimatic zones with a minor component of Engelmann Spruce Sub-alpine Fir
(Dobson Engineering Ltd., 2007).
25
Surficial DepositAlluvial
Colluvial
Eolian
Glaciofluvial
Lacustrine
Morainal
Organic
I20 km
Figure 6: Surficial geology of KRW
26
3.1.5 Water use values in KRW
Kiskatinaw River watershed in northern-eastern BC serves as a dynamic and
crucial water resource for the Peace River Regional District (PRRD) area. It is the
primary drinking water source for the City of Dawson Creek, village of Pouce
Coupe and thousands of rural inhabitants in PRRD. Over decades, the dominant
land use activity in this area has been agriculture including grain production,
livestock and mixed farming (Jacklin, et al., 2003). Forestry is also a major land use
activity in this densely forested watershed. But over the past few years, natural gas
development has been dominating the scene by its intensity and rapid expansion
(Forest Practices Board , 2011). KRW is included in the Montney shale gas play
which is one of the major shale basins in North America. Recent gas development
within the watershed is dominating the land use and land cover dynamics in this
area. As a result, the water demand in the watershed is increasing at an average rate
of about 3.2% per year (Saha, et al., 2013).
27
3.2 Methodology
3.2.1 Reconnaissance survey
The first important step in this research was the reconnaissance survey in the
study area. Before actual data collection, this survey provided overall information
on the study watershed and allowed to comprehend the gravity of land use and
land cover change within it. Field trips were performed at a num ber of sites in the
watershed and general discussion was conducted with the Dawson Creek city
authority, individuals performing diverse research in this watershed and
community representatives who belong to various stakeholders of the watershed.
All of these created a deep insight about the study area and its ongoing activities
which whetted the overall research enthusiasm and strategy.
3.2.2 Data selection and collection
The research is based on the analysis of satellite imagery of the study
watershed. Therefore, the first and foremost task was the selection of satellite sensor
and associated images. During this process, the prior considerations were the
purpose of the study, objects to be identified, and the availability of images.
Based on literature review and previous experiences, Landsat satellite images
were selected for this study for a number of reasons.
28
a) For long term change detection, Landsat data are available since 1972. This
robust and continuous data inventory stores satellite data for every part of the world
from 1972 till today. Since this study aims to detect the LULC changes in KRW from
1984 to 2010, Landsat data was the best available option.
b) Landsat satellite has a repeat interval of 16 days. This property of this
sensor has increased the flexibility of data selection, especially when cloud cover is a
major limitation in satellite data selection.
c) Last but not least, Landsat data are freely available and it was a huge
support to this graduate research.
After selecting the satellite sensor, the next task was to collect necessary
Landsat imagery for this study. This step was also governed by these factors:
i) The objective of the study which was to capture development activity
driven LULC change within KRW from early 1980's to 2010,
ii) Image quality: the main hindrance was to obtain cloud free analyzable
imagery
iii) Image acquisition time: to obtain images captured at more or less the same
time of the year is important, because seasonal variability changes the appearance of
land use features and this can impact the quality of analysis.
29
After considering all of these factors, three Landsat images were selected for
analysis and downloaded from the data warehouse of USGS Earth Resources
Observation and Science Center (EROS):
1) 1984 Landsat 4/5 TM imagery: represents early stage of industrial
development in this forested watershed
2) 1999 Landsat ETM+ imagery: represents the status at the beginning of the
gas industry booming within the watershed
3) 2010 Landsat 4/5 TM imagery: represents current status within KRW
All the images downloaded were either from late July or early August within
a span of 10-18 days to keep the analysis free from seasonal variability impact. Table
1 summarizes the data description. Two separate scenes for each year had to be
downloaded to cover the whole watershed, one from path 48 row 21 and another
from path 48 row 22.
30
Table 1: Description of satellite imageries used in LULC change detection
Year Day Satellite imagery Spectral resolution1 Spatial resolution
1984 July 17 Landsat 4/5 TM Band 1 to 5 & 7 30 m
1999 August 4 Landsat ETM+ Band 1 to 5 & 7 30 m
2010 July 25 Landsat 4/5 TM Band 1 to 5 & 7 30 m
1 thermal band 6 was excluded in this analysis
Landsat data are multi-spectral with seven different color bands although the
Landsat Enhanced Thematic M apper Plus (ETM+) sensor has an additional
panchromatic band 8 which was not used in this study, along with thermal band 6
for both TM and ETM+ sensors. Table 2 and Figure 7 describe in detail the spectral
features of different Landsat bands. Landsat images have spatial resolution of 30 m
i.e. each pixel of the image covers 30m X 30m area on land.
Table 2: Spectral features of Landsat bands
Band Number W avelength Interval Spectral Response
1 0.45-0.52 p m Blue-Green
2 0.52-0.60 p m Green
3 0.63-0.69 p m Red
4 0.76-0.90 p m Near IR
5 1.55-1.75 p m Mid-IR
6 10.40-12.50 p m Thermal IR
7 2.08-2.35 p m Mid-IR
31
NEAR-IR SOLAR ■IVISIBLEl 1 I REFLECTED P I
ULTRAVIOLET
M1DIR ! FAR-IR-
1.0 2.0 3j0 5.0 10 15 20 30WAVELENGTH, Jim____________
Figure 7: Different wavelengths of Landsat bands (after NASA, 2011)
3.2.3 Description of the land use and land cover classes
Based on the reconnaissance field observation and local information, it was
decided to concentrate on 11 land use and land cover (LULC) classes during the
satellite data analysis. The selected LULC classes are cropland, coniferous forest,
deciduous forest, mixed forest, planted or re-growth forest, forest fire, cut block,
pasture, water, wetland, built-up area. Each of the classes is described below.
Cropland: This class includes all the cultivated lands used for crop
production. This comprises mostly flat areas and also some steep slopes where
various crops are grown (Figure 8).
32
Figure 8: Cropland sampled during ground truth survey
Coniferous Forest: The forested lands that predominantly comprise
evergreen trees throughout the year are defined as coniferous forest in this
classification scheme (Figure 9).
Figure 9: Coniferous forest sampled during ground truth survey
33
Deciduous Forest: The forested lands with predominance of broadleaf trees
that lose their leaves seasonally, particularly at the end of the frost-free season, are
classified as deciduous forest in this study (Figure 10).
Figure 10: Deciduous forest sampled during ground truth survey
Mixed Forest: The forested lands which have both evergreen coniferous and
broadleaf deciduous trees and no predominance of one category are defined as
mixed forest in this classification scheme (Figure 11).
34
Coniferous tree
Deciduous tree
Figure 11: Mixed forest sampled during ground truth survey
Planted or Re-growth Forest: The forested lands that comprise young
coniferous and/or deciduous plants which have re-grown or been planted after
forest fire or clear cutting or any other decay event are classified in this class. Some
herb-shrub may be included in this class since these are hard to differentiate during
digital image classification with imagery of 30 m resolution. Planted and re-growth
forests are very common in this watershed (Figure 12).
35
Figure 12: Planted or re-growth forest sampled during ground truth survey
Forest Fire: This class comprises fire affected forest land with burnt, dead
trees. A fire event occurred in the Hourglass area of this watershed in 2006. So this
class only appeared in 2010 image classification (Figure 13).
Figure 13: Forest fire area sampled during ground truth survey
36
Cut block: This class represents the forest clear cut area which was removed
for industrial (mostly) or other purposes. Cut blocks are copious in this watershed.
Figure 14 shows a cut block in KRW which was cleared by the gas development
industry.
Figure 14: Cut block sampled during ground truth survey
Pasture: The lands which are maintained for livestock production as well as
used for perennial hay or forage cultivation are defined as pasture in this
classification scheme. Figure 15 shows a typical pasture land in the study area.
Pasture and croplands may be interchanged after several years.
37
Figure 15: Pasture land sampled during ground truth survey
Water: This class comprises open water bodies with more than 95% water
surfaces. It includes river channels, lakes etc. (Figure 16).
Figure 16: View of One Island lake sampled during ground truth survey
38
Wetland: In this classification scheme, wetlands are non-forested and/or
slightly forested marshes, swamps etc. where the groundwater table is at, near or
above the surface for significant part of the year. W etlands are one of the common
features in KRW (Figure 17).
Figure 17: Wetland sampled during ground truth survey
Built-up Area: This class includes lands covered with human-built structures
like: houses, roads, industrial infra-structures etc. (Figure 18).
39
Figure 18: A gas development infrastructure sampled during ground truth survey
3.2.4 Image pre-processing and analysis
The image pre-processing and analysis entail a number of steps before
generating the final output. During this process, several computer software
packages were used. Pre-processing was mostly performed with PCI Geomatica
10.2, image analysis was conducted with IDRISI Selva 17.0 while ArcGIS 10.1 and
Quantum GIS 1.7 were used at different phases of the analysis and map generation.
40
3.2.4.1 Pre-processing of Landsat image
The downloaded Landsat images for each band needed to undergo several
pre-processing steps. The pre-processing of images prepares them for the
classification analysis. All the bands of the two Landsat scenes were downloaded as
separate image files (.tiff) which were layer stacked together for classification
analysis.
Figure 19 shows an image of an individual Landsat band downloaded from
the USGS EROS database. These individual bands were then stacked sequentially
from 1 to 7 using the 'transfer' function in PCI. Stacked bands were then translated
to PCI default image format (.pix). When layer stacking of all the bands from each
scene into two separate image files was executed, they were ready for the next step
of 'mosaicing'. During mosaicing, both of the new image files were joined together
to form a single image file which was later clipped to get the full extent of the study
area. Figure 20 displays mosaiced output of 2010 Landsat image with 5-4-3 band
combination.
The downloaded Landsat images have been already georeferenced to
projection system Universal Traverse Mercator (UTM), zone 10 with datum WGS 84
which was utilized throughout the analysis.
41
Sheene frdn* 25• ■ <► . ’ . . .
Scene from path 48 & row, 22
Figure 19: Band- 4 of Landsat 2010 image before processing
Figure 20: Mosaiced 2010 image with 5-4-3 band combination (KRW area in red boundary)
Each band of the Landsat image has its own characteristics as discussed
above and contains distinct signatures for the associated LULC features. Each LULC
type absorbs or reflects a particular range of wavelength. This phenomenon is
recorded in each Landsat band with a particular wavelength range. This is why, the
43
nature of these different Landsat bands had to be thoroughly studied to make a
decision as to which combination of three bands would be most interpretive during
classification analysis and visual elucidation. After literature surveying and lab
examination, 5-4-3 band combination was selected for RGB color composite, i.e.
band 5 in the red, band 4 in the green and band 3 in the blue. This combination
provides the user with the greatest amount of information and color contrast which
makes it easier to differentiate different LULC features (NASA, 2011). Healthy
vegetation and forest cover appear with strong green color, while soil is mauve. It
clearly contrasts water bodies with distinct blue color where range of blue color
varies with depth and turbidity of water. It also provides robust agricultural
information. Built-up areas appear in dark purple or pink. The 5-4-3 band
combination was used for all three Landsat images for the years 1984,1999, 2010 and
color composite images were produced accordingly for classification. Figure 21,
Figure 22 and Figure 23 display the Landsat images for the respective years in 5-4-3
band combination and clipped to the study area extent.
44
RGBRed: Band 5Green: Band 4 Blue: Band 3
V20 km
Figure 21:1984 Landsat satellite image used in this study
45
3.2.4.2 Image classification
Digital image classification in remote sensing is the detection and clustering
of similar image pixels into the same information categories which are produced
from several spectral bands of a satellite image (Campbell, 2002; Paiboonvorachat,
2008). The object oriented image classification (OOIC) method was used in this
analysis. OOIC, as described in the literature review section, has the ability to
generate more accurate and meaningful result than conventional pixel based
methods.
Image classification in this study was conducted using the segment classifier
in IDRISI Selva which comprises three distinct and mutually dependent modules,
namely SEGMENTATION, SEGTRAIN and SEGCLASS. SEGMENTATION is the
process by which spectrally similar, homogeneous pixels are grouped into
individual image segments or polygons. SEGTRAIN generates training and
signature files for the final classification step. Finally, SEGCLASS is a majority rule
classifier which uses segmentation, training files and a pixel based classification
output for its performance.
48
3.2.4.2.1 First step - segmentation of the image
During the segmentation process in IDRISI, the spectral similarity of the
image pixels are quantified using variance of pixel values within a moving window
which is a user defined filter. Both spectral and spatial characteristics of the imagery
are critical for properly delineating image segments. Accordingly, all six Landsat
bands (1 to 5 & 7) were used for image segmentation in this study.
In its first step, the SEGMENTATION module uses the moving window to
generate a variance image. The homogeneity of pixels controls the variance value;
the more the homogeneity, the lower the variance value. The homogeneous pixels
are grouped into image segments based on their variance values and assigned with
discrete IDs. The smaller the threshold value, the smaller the size of the segments as
smaller threshold value searches for more homogeneity.
In this analysis, all three Landsat images for the years 1984, 1999 and 2010
were segmented using these parameters: window width and height 3 x 3 , weight
mean factor 0.5, weight variance factor 0.5, similarity tolerance 10. The width and
height of the moving window assists IDRISI to derive variance image for each layer;
the weights for the mean and the variance factors evaluate the similarity between
neighbouring segments; similarity tolerance (ST) is used to control the
generalization level during the segmentation process given that the smaller the
49
tolerance value, the higher the num ber of image segments and the finer the
segmentation output (Egberth and Nilsson, 2010). Figure 24 shows how the number
of segments and fineness vary with similarity tolerance value of 10 and 50.
50
Figure 24: Number of segments and fineness varying with ST value
A) Original Image (5-4-3 band combination), B) Segments generated with ST=50 (red outline), C) Segments generated with ST=10 (red outline)
51
3.2A.2.2 Second step - generating training profile
After performing segmentation of all the images, the SEGTRAIN module was
used to create training profiles to be used for classification. The segmented images
were overlaid on the respective RGB composite images and segments were
identified for its particular LULC classes.
The training class generation entailed a rigorous use of field sampling data,
higher resolution aerial photograph for some parts as well as personal knowledge
about this watershed since many parts of this remote densely forested watershed are
physically inaccessible. Several factors were considered while selecting the sampling
locations by onscreen viewing of the imagery - a) at least 30 sampling polygons for
each LULC type, b) spatial distribution of the data polygons, c) accessibility to
sample location etc.. At the end of this process and based on the sampled data,
SEGTRAIN module generated a training file which was forwarded for further
analysis in the next step.
52
3.2A2.3 Third step - classification of images
This is the final step in the digital OOIC procedure. The SEGCLASS module
on IDRISI requires a supervised classification image for its final analysis.
Accordingly, supervised classification was performed for each of the years using
maximum likelihood (MAXLIKE) algorithm where training classes from the
previous step were utilized. In the MAXLIKE process, pixels are assigned to the
most likely class based on a comparison of the subsequent probability that it belongs
to each of the signatures being considered. Then, the SEGCLASS module executed
the final classification using the MAXLIKE output, segmented image from the first
step and segment-based training and signature files from the second step. The
SEGCLASS resulted in less noisy, smoother and improved classified output
compared to the MAXLIKE output. Figures 25 - 27 demonstrate the classification
outputs generated for the 1984,1999 and 2010 images respectively.
These outputs were then clipped to the study area extent and input into
ArcGIS for map creation and were analyzed on IDRISI for accuracy assessment,
change detection and land use modeling.
53
■ CroplandI I C oniferous Forest■ Deciduous Forest■ Mixed Forest■ Ptaitfed_Regrowti Forest■ Cut Bloc*I I Pasture■ Water■ Wetend H i But-upArea
B
»
Hi Cropland I I Coniferous Forest■ D eciduous F orest^ i Mixed forest Hi Ptartted_Regrowth Forest■ Cut Block I I P astu re■ W ater■ Wetland H i Buft-upArea
Figure 25:1984 image classification; A) MAXLIKE, B) SEGCLASS
54
■ CroplandI I coniferous Forest■ Deciduous Forest H Mixed Forest■ Ptarted.Regroifrth Forest■ Cut Block I I Fasfcve■ Water■ Wetland■ Buift-upArea
R
" i* 1
■ CroplandI I Coniferous Forest H I Deciduous Forest■ Mixed ForestH i Planted Regrowrih Forest■ CutBtodc
Pasture Water Wetland BuitapAiea
Figure 26:1999 image classification; A) MAXLIKE, B) SEGCLASS
55
■ C roplandI I CorifcrouB F orest■ D eciduous F orest■ M ixed F orest■ Ptantedjtegrowth Forest■ F o rest fire I I Cut Block■ Pasture
□ B u t-q p A rea
B ■ C roplandI I C oniferous Forest B D eciduous Forest■ M ixed F o rest■ Pfaflted.R egraw lti F o rest■ F o rest fire | | cut Block I P astu re
■ w etland I 1 B u t-u p A ree
Figure 27: 2010 image classification; A) MAXLIKE, B) SEGCLASS
56
3.2.4.3 Accuracy assessment
Assessing accuracy of digital image classification output is very important.
Accuracy assessment is usually performed either using a new set of ground truth
data or by comparing with a previously classified reference map for selected
sampling points.
In this study, accuracy assessment was an intricate task as there was no
reference LULC map available for the study area for the years 1984 and 1999. Thus,
there were no available ground truth data for those years. But it was possible to
perform the accuracy assessment using ground tru th data for the 2010 classification
output. Since the same signature and training information were used for classifying
all three images, the accuracy assessment of 2010 confidently confirmed the accuracy
of the other two, assuming that the land cover was consistent over the years.
The design of the sampling program is critical. In this study, stratified
random sampling method was applied for the ground data collection. The stratified
random scheme works by dividing the area into a rectangular matrix of cells and
then chooses a random location within each cell for sampling. Spatially distributed
20 sample points were selected for each class in LULC classification scheme. Each of
these points was checked in the field or with higher resolution images (Google
earth), where locations were inaccessible. Every match between classified LULC
57
map and ground truth information was counted as 1 and for mis-match, it resulted
in 0. All of these information were summarized in an error matrix which is the most
common method used by researchers for classification accuracy assessment (Oumer,
2009). The overall accuracy, user's accuracy, producer's accuracy as well as Kappa
coefficient were calculated.
Overall accuracy is defined as the ratio between the total num ber of samples
which are correctly classified and the total number of samples considered for the
accuracy assessment. User's accuracy corresponds to error of commission. It refers
to the measurement of how many of the samples of a particular class matched
correctly. It is defined by the following ratio:
Total number of samples that are correctly classified in a given categoryUser's accuracy =----------------- —— :-------;------:------;— :— ;-------------------------------
Total number of samples in that category
On the other hand, producer's accuracy corresponds to errors of omission. It
is a measure of how much of land in each LULC category was classified correctly. It
is calculated as:
58
Total number o f samples which are correctly classified for a given categoryProducer's accuracy =---------------------------------------------------------------------————— (2)
Total number o f samples that are classified to that particular category
The kappa coefficient estimates the agreement between a modeled scenario
and reality (Congalton R. G., 1991). It determines if the results displayed in an error
matrix are significantly better than random (Lillesand, et al., 2008). For an error
matrix with num ber of rows and column, kappa coefficient is computed as:
K = (NA - B ) / (N2 - B) (3)
Where, N = total number of observations included in the error matrix
A =the sum of correct classifications contained in the diagonal elements
B = the sum of the products of row total and column total for each LULC type
in the error matrix
Figure 28 summarizes all the Landsat image analysis steps in a flow chart:
59
Pre-processing
Image classification
Post-classificationanalysis
OOIC ClassificationMAXLIKE classification
Segmentation of the image
Accuracy assessment
Generation of training profile
LULC maps generation for 1984,1999 & 2010
Mosaicing of scenes, Clipping to KRW area
Layer stacking of bands (1-5 & 7): Transfer & Translate
Manual editing of the classified output to include missing features
Landsat imagery download for 1984,1999 & 2010 (scene from Path 48, Row 21 & Path 48, Row 22)
Figure 28: Landsat image analysis framework
60
3.2AA Land use modeling
RS and GIS along with computer based modeling tools have become a
popular and efficient means of simulating current and future land use and land
cover change and hence, assist in land use planning and natural resources
management (Herold, et al., 2003; Araya, 2009).
Various tools exist in the era of land use modeling which are diverse in
design, spatial scale, temporal dimension, data availability etc. The present study
utilizes Markov Chain model for simulating LULC changes in Kiskatinaw River
Watershed.
The Markov Chain model is a unique and widely used tool in land use
modeling which demonstrates the LULC changes as a stochastic process (Weng,
2002). In the Markovian system, the future state of a land use system is modeled on
the basis of the immediate proceeding state (Araya, 2009). The Markov Chain
analysis describes the probability of LULC changes from one period to another by
constructing a transition probability matrix between period-1 and period-2. The
basic hypothesis of Markov Chain prediction is that future land use at time (t+1), Xm
is a function of current land use at time t, Xr, i.e.
X m = f (Xt) (4)
61
The transition probability Pm* is represented by the probability that a cell of
land cover type Um alters into land cover type m in between the model period.
Therefore, if the transition probabilities are congregated in a transition Matrix, P,
then Xm can be derived from the following equation (Benito, et al., 2010):
X w = X t.P (5)
The whole land use modeling task was performed in the 'Land Change Modeler
(LCM)' on IDRISI Selva.
3.2.4.4.1 Land Change Modeler (LCM)
LCM is a powerful application in IDRISI which integrates a set of tool to
understand the dynamics of land use-land cover conversion and its associated
impacts. In this study, LCM provided support for LULC change analysis, transition
potential modeling and LULC change prediction. For all these tasks, LCM used the
LULC maps generated for the years 1984, 1999 and 2010. The change analysis was
performed for two separate periods, one from 1984 to 1999 and another from 1999 to
2010. But the transition potential modeling and change prediction were carried out
for the whole period from 1984 to 2010.
62
The 'Change Analysis' on LCM assesses changes between time Ti and T2 . In
the first set, changes were evaluated for T1 = 1984 and T2 = 1999; in the second set, it
was for Ti = 1999 and T2 = 2010. As a result, an in-depth comprehension could be
gained about the LULC change dynamic of the study area. The changes that are
identified are transitions from one state of land use-land cover to another. The gain
and loss in area for each LULC type as well as the net change were calculated and
corresponding graphics were generated. Outside of LCM, changes in individual sub
watershed were also evaluated and maps were created using ArcGIS.
After the change analysis, the next task on LCM was to model the potential of
land transitions. At first, the LCM project was set for Ti = 1984, T2 = 2010 and
corresponding LULC maps were loaded to the project. Then, LCM calculated the
transitions occurring between 1984 and 2010. At this stage, transition maps were
created which basically shows the LULC changes from one type to another. The
transition maps were organized within empirically evaluated transition sub-models
that were driven by the same underlying variables. These driver variables were used
to model the chronological change processes.
LCM modeled the transition potentials for each identified transition between
different LULC types by using a multi-layer perception (MLP) neural network
technique. The non-linear neural networks can be regarded as a complex
63
mathematical function which has the ability to convert multi-variant input data to a
desired output. MLP in LCM uses a back propagation algorithm for transition
prediction. A typical MLP network contains one input layer, one output layer and
one or more hidden layers, each containing multiple nodes which is akin to the
neural network of brains (Lin, et al., 2011) (Figure 29). Each node in the three layers
is connected to other with varying weights and plays critical role during the
modeling
Input:Imagedata
Figure 29: MLP neural network (after Eastman, 2012)
► Transition
Hidden•aver Output
Layer
procedure (Eastman, 2012). The hidden layer nodes are crucial to the execution of
MLP such that it loses its ability to learn and make use of interaction effects without
64
them (Chan, et al., 2001). At the beginning of the MLP process, sampling size was
selected for each transition which is the num ber of change pixels that would be used
in the modeling. Half of the samples were used for training and half were utilized
for validation of the transition model. As mentioned before, each transition was
modeled under a set of input variables which are basically various GIS layers and
likely to affect the land use change within the watershed. Table 3 explains the input
variables used in this analysis. The selection of these variables entailed careful
consideration of KRW's land use activities and data availability, and they were
presented as different GIS layers in the modeling process. All of the driving
variables were tested for their effects on modeling accuracy and skill statistics
calculated by using the following equations 6 and 7 (Eastman, 2012). The variables
which resulted in lower accuracy (i.e. below 50%) and skill statistics were removed
from analysis, and then the remaining variables were considered to govern the
transition process for each transition sub-model. Modeling parameters, e.g. hidden
layer nodes, learning rates, momentum factor, sigmoid constant etc. were
interactively estimated by the MLP modeling tool and produced the best result.
MLP provides the measure of accuracy (in %) and skill statistics (value: -1 to +1) as
an appraisal of efficiency of the prediction process.
65
The expected accuracy can be determined by the following equation
(Eastman, 2012):
EA = 1 / (T+P) (6)
Where,
EA = expected accuracy,
T = the number of transitions in the sub-model
P = the num ber of persistence classes
And the measure of model skill is then expressed as:
S = ( A - E A ) / (1 -E A ) (7)
Where,
A = measured accuracy from the analysis
EA = expected accuracy
66
Table 3: Driver variables for transition potential modeling
Driver variable layer RoleDistance to gas development infrastructure
Shale gas development industry: responsible for forest clear cutting, road development, high am ount of water extraction from the river etc.
Forest cut blocks planned for future harvesting
Forestry industry: Cut blocks planned and mapped for future harvesting dictates the plantation and regrowth process, hence the shifting of forest types.
Cumulative kill by mountain pine beetle infestation
Active management of the mountain pine beetle attack is on action in this watershed since its detection in 2004 which includes aggressive forest harvesting. This driver comprises location and number of cumulative kill by pine beetle infestation.
Distance to major channel network The channel network controls the general hydrology, wetlands dynamics, gas development activities etc. in this watershed.
Digital elevation model (DEM) and topographic wetness index (TWI)
These two determine the hydrological flow path i.e. overall hydrological process within a watershed; hence these control the wetland dynamics in the watershed. TWI is defined as Ln(A/tanB) (Sorensen, et al., 2005)where,A = local upslope area draining through a certain point per unit contour length, tanB = the local slope
67
In the final step of LULC modeling, the transition probabilities estimated
from MLP neural network modeling were fed into the "Change Prediction" module
on LCM to generate future LULC scenario in 2020 by using Markov Chain (MC)
modeling. The MC model in LCM could generate a transition probability matrix and
a transition area matrix based on which the prediction process was performed. The
probability of change from one LULC type to another was contained in the
transition probability matrix. The transition area matrix records the num ber of pixels
that are expected to convert from one LULC type to another, and it was created
through the multiplication of each column in the transition probability matrix by the
num ber of pixels in a corresponding LULC type (Myint & Wang, 2006). Based on
these transition matrices, MC model was used to generate both hard and soft
predictions of LULC within the study area. The hard prediction produces a LULC
map derived by a multi-objective land allocation algorithm contained in LCM which
considers all of the calculated transitions for creating lists of host classes (i.e. losing
land area) and claimant classes (i.e. gaining land area) (Eastman, 2012). On the other
hand, soft prediction identifies the vulnerability of LULC change for a set of
transitions based on a method of logical "OR" aggregation. This method is based on
the principle that a location is more vulnerable to LULC change if it is subject to
several transitions than if it is only subject to a single transition. The output of
logical "OR" aggregation for a pixel is equal to (a+b-ab), where 'a ' represents the
68
probability of that pixel transition to one LULC type and 'b ' represents its transition
probability to another LULC type. For example, if a particular pixel has a probability
of 0.40 to be changed to one LULC type and 0.30 to another LULC type, the logical
"OR" operation would evaluate the LULC change vulnerability as (0.40 + 0.30 - 0.40
x 0.30 = 0.58). More detailed descriptions of transition probability estimation and
LULC change vulnerability assessment can be found in Eastman (2012). Both of the
hard and soft predictions in this study were performed with an intermediate stage at
2015 and a final stage at 2020.
The land use modeling process has been summarized in the following flow
chart (Figure 30):
69
r
LULC map of Ti
Inputs
LULC map of T2
1—.......................................................... ... ... I
Driver variables
Change Analysis
Change map/gain & loss in area/net
change (T1-T2)
Figure 30: Land use m odeling framework
Transition Model
MLP neural network
Transition potentials: map / matrix
Change Prediction
Markov Chain model
Hard prediction:
projected LULC m ap of T3
Soft prediction: projected
vulnerability of
70
CHAPTER 4: RESULTS AND DISCUSSION
4.1 Results of RS & GIS analysis of satellite images
The land use-land cover maps produced by integration of remotely sensed
image classification and corresponding GIS editing have provided im portant LULC
information for this study area. Analysis of 1984 imagery has been summarized in
Figure 31 and Figure 32. It is shown that the major share of the land area was
covered by different types of forest among which coniferous forest comprised the
maximum of 37.35%, followed by deciduous forest (28.09%) and mixed forest
(12.41%). In total, all of the forest types including planted and re-growth forest
occupied 80% of the total land area of the study watershed in 1984. Wetlands
covered a significant portion of the area (16.02%). Other LULC features, namely
cropland, cut block, pasture, water, and built-up area comprised 0.82%, 1.58%,
0.23%, 0.76% and 0.64% respectively of the study watershed.
71
Built-up area
W etland
W ater
Pasture
C ut block
Planted or regrow th forest
M ixed forest
Deciduous forest
Coniferous forest
Cropland
200 400 1000 1200
Figure 31: Area covered by each LULC type in 1984
72
Built-up area Coniferous forest Cropland Cut block Deciduous forest
Mixed forest PasturePlanted or regrowth forestWaterWetland
N20 km
Figure 32: KRW LULC map of 1984
73
Figure 33 and Figure 34 display the output generated from the analysis of
1999 Landsat imagery. As in the 1984 analysis, forest types comprised the dominant
portion (85%) of the land area among which coniferous forest covered the maximum
of 41.45%, followed by deciduous forest 23.30%, mixed forest 15.92%, and planted
and re-growth forest 4.94%. Wetlands occupied only 7.79% of the watershed area in
1999. Other LULC features, namely cropland, cut block, pasture, water, and built-up
area comprised 1.12 %, 1.54%, 1.82%, 0.75% and 1.39% respectively of the study
watershed.
Built-up area
Wetland
Water
Pasture
Cut block
' Wanted or regrowth forest
Mixed forest
Deciduous forest
Coniferous forest
Cropland
0 200 400 600 800 1000 1200
Figure 33: Area covered by each LULC type in 1999
74
Built-up area Coniferous forest Cropland Cut block Deciduous forest
Mixed forest PasturePlanted or regrowth forestWaterWetland
fj20 km
Figure 34: KRW LULC map of 1999
75
Analysis of 2010 Landsat imagery is displayed in Figure 35 and Figure 36.
Kiskatinaw watershed maintained its LULC nature in 2010 with the principal
portion covered by various forest types. As before, the dominant forest type is
coniferous forest comprising 39.05% of the study area, followed by deciduous forest
28.83%, mixed forest 12.87% and planted and re-growth forest 5.53%. Wetlands
coverage continued to decline with a total of 6.45% of the watershed. The Hour
Glass forest fire event in this area in 2006 represents 1.17% of the watershed area.
The fire affected forest cover is located near the south western boundary of the area.
Finally, 0.66%, 0.93%, 2.12%, 0.72% and 1.66% of the study watershed are occupied
by cropland, cut block, pasture, water, and built-up area respectively.
Forest fire
Wetland
Pasture
Wanted or regrowth forest
Deciduous forest
Arei (km2)Cropland
0 200 400 600 800 1000 1200
Figure 35: Area covered by each LULC type in 2010
76
H i Built-up area Hi Coniferous forest
Cropland H Cut block
Hil Deciduous forest Hi Forest fire Hi Mixed forest
PastureM B Planted or regrowth forest HI Water
Wetland
20 km
Figure 36: KRW LULC map of 2010
77
Table 4 represents a compendium of the total area and percent of area
covered by individual LULC types.
Table 4: Surface area covered by each LULC type in a particular year
Total study area 2836 km 2
1984 1999 2010LULC type km2 % of total km2 % of total km2 % of totalCropland (CL) 23.27 0.82 31.70 1.12 18.82 0.66Coniferous forest (CF) 1059.06 37.35 1175.45 41.45 1107.84 39.05Deciduous forest (DF) 796.65 28.09 660.79 23.30 815.34 28.83Mixed forest (MF) 351.97 12.41 451.57 15.91 365.88 12.87Planted or regrowth forest (P/RF) 59.94 2.10 140.08 4.94 157.23 5.53Cut block (CB) 44.70 1.58 43.46 1.54 26.38 0.93Pasture (PS) 6.53 0.23 51.63 1.82 60.30 2.12Water (WT) 21.49 0.76 21.18 0.75 20.48 0.72Wetland (WL) 454.22 16.02 220.82 7.79 183.30 6.45Built-up area (BA) 18.17 0.64 39.32 1.39 47.24 1.66Forest fire (FF) 0.00 0.00 0.00 0.00 33.19 1.17
4.2 Accuracy assessment
The accuracy of satellite image classification could be constrained by the
resolution of images used and lack of fine details as well as unavoidable
generalization impacts (Oumer, 2009) and therefore, errors are always expected.
This is why, to ensure prudent utilization of the produced LULC maps and their
associated statistical results, the errors and accuracy of the analyzed outputs should
be quantitatively explained.
78
As clarified in the 'Data and Methods' section, the accuracy tests for the 1984
and 1999 image classification were not possible due to unavailability of reference
data, the test was only performed for 2010 image analysis. Table 5 shows the
corresponding accuracy assessment error matrix for 2010 analysis. The randomly
generated sample points were tested with ground truth information as well as with
higher resolution imagery, where the point was inaccessible. The numbers
highlighted in grey are matching samples for each LULC type, others are mis-match.
Table 5: Accuracy assessment error matrix for 2010 image classification
LULC typeGround Truth
CL CF DF MF P/RF CB PS WT WL BA FF Row Total
CL 16 2 1 1 20
i , CF 20 20
DF 19 1 20
MF 2 18 20
P/RF 1 19 20
CB 1 1 18 20
PS 2 17 1 20
WT 20 20
WL 1 2 2 15 20
BA 1 1 18 20
FF 1 19 20
Column Total 19 20 21 21 21 21 22 20 16 19 20 199
CL=Cropland, CF=Coniferous forest, DF=Deciduous forest, MF=Mixed forest, P/RF=Planted or regrowth forest, CB=Cut block, PS=Pasture, WT=Water, WL=Wetland, BA=Built-up area, FF=Forest fire
79
Based on the information from Table 5, the overall accuracy, user's accuracy,
producer's accuracy and overall kappa coefficient were calculated using the formula
stated in previous chapter. Table 6 recapitulates the calculated results.
Table 6: Accuracy assessm ent summary
LULC type User's Accuracy (%) Producer's Accuracy (%)
Overall Accuracy 90.45%
CL 80.00 84.21CF 1 0 0 . 0 0 1 0 0 . 0 0
DF 95.00 90.48MF 90.00 85.71
P/RF 95.00 90.48CB 90.00 85.71PS 85.00 77.27
WT 1 0 0 . 0 0 1 0 0 . 0 0 Overall KappaWL 75.00 93.75 coefficient 0.89BA 90.00 94.74FF 95.00 95.00
The overall accuracy and kappa coefficient for 2010 image classification were
90.45% and 0.89 respectively. The producer's and user's accuracy for all classes
except cropland, pasture and wetland were higher than overall accuracy. Lower
accuracy for these classes may be partly attributed to the medium spatial resolution
of Landsat data and general smaller size of the feature areas.
80
4.3 D iscussion on LULC changes
Forestry and agriculture have long been shaping the land use and land cover
activities in the Kiskatinaw river watershed. But KRW's location within the Montney
shale gas play and its associated natural gas development activity over the past few
years have been determining the LULC dynamics within the watershed. The role of
these dom inant hum an interactions has been demonstrated in the present land use
change analysis.
The LULC maps described above were input into a change analysis module
on IDRISI 'Land Change Modeler (LCM)'. The analysis operated by LCM revealed
some significant LULC change dynamics within the watershed. The study area
which is a forested watershed maintained its large mature forest cover (around 80%)
within the study period from 1984 to 2010. Subtle changes have been observed for
the mature coniferous, deciduous and mixed forest types. The noticeable change of
the planted or re-growth forest type and cut blocks indicates the impacts of the
forestry industry in this area, although cut blocks in recent images may be attributed
to the gas development industry as well. In many cases, cropland and pasture were
hard to differentiate during the digital image classification process of Landsat data
since these land use types are more or less similar in spectral signature. But
combinedly, cropland and pasture show a considerable change between 1984 and
2 0 1 0 which highlights the amplified agricultural and farming activities within the
81
watershed. This analysis identified a striking change in the extent of wetlands, while
most of the wetland depleted between 1984 and 1999, estimated as a loss of 233.22
km2. This significant change in wetland area needs further investigation to
understand the depletion dynamics. An increase of 29 km 2 of built-up area indicates
the recent industrial booming in this area, particularly shale gas development
activity. The forest fire affected 33.39 km 2 represents the Hour Glass fire event in
2006 within and around the study area.
For comprehensive spatial analysis, LCM was used to account for changes
from 1984 to 1999 and from 1999 to 2010 in separate sets. Gain and loss for all land
use types as well as their contribution to net change were quantified for each time
period. Figure 37 shows the change analysis from 1984 to 1999 and Figure 38
exhibits the change analysis from 1999 to 2010. In these figures, the area gain for a
particular LULC type includes the converted land cover area which previously
belonged to another LULC type. Accordingly, the lost area of any LULC type was
changed to some other types. From 1984 to 1999, there is a striking negative change
observed for the deciduous forest and wetlands which indicates higher loss than
gain of area for these LULC types. In contrast, the sharp positive net change for
LULC features such as: cropland, coniferous forest, mixed forest, planted and re
growth forest, pasture and built-up area, refer to the higher gain in area than loss.
There are also some interesting changes identified from 1999 to 2010. The wetlands
82
continued the decreasing trend during this time period and total depletion
accounted for 270.92 km2 of area from 1984 to 2010 which needs further rigorous
investigation to explore the reasons behind this widespread depletion.
400■ Area gained (km2)
-400
Net change (km2) ̂^CF
MF
P R F
CB
PS
WT
-300 -250 -200 -150 -100 -50 0 50 100 150
Figure 37:1984-1999 change analysis for LULC types
83
4 0 0
300
200
100
-100
-200
-300
Area ga ine d (km 2) Area lost (km 2)
•<fiQ
-100
DF
P/RF
PS
W
BA
FF
Net change (km2)
-50 -25 50 100 125 150 1 "5 200
Figure 38:1999-2010 change analysis for LULC types
During 1999 to 2010, the deciduous forest showed a marked increase whereas
other forest types between 1984 and 1999 exhibited a sharp decrease. Forest fire also
contributed positively to the net change which was already discussed above. The
84
small but constantly increasing built-up area underscores the ongoing development
activities in this area.
4.3.1 LULC change in individual sub-basin
As mentioned in the previous chapter, KRW is subdivided into five sub
basins, namely Mainstem, Brassey, Halfmoon-Oetata, East KRW and West KRW.
The nature and amount of LULC change at sub-basin level depends on its LULC
type and anthropogenic signature. Therefore, it is valuable to study the LULC
changes at sub-basin scale.
4.3.1.1 Mainstem sub-basin
Mainstem sub-basin covers the north-eastern portion of the study area and
holds the main course of the Kiskatinaw River. The City of Dawson Creek's water
intake station, Arras is located just at the northern most edge of this sub-basin. This
is why Mainstem sub-basin's LULC is considered crucial for this w atershed's overall
water resource management.
Figures 39 - 41 show the LULC maps of Mainstem area in 1984, 1999 and
2010. The change in LULC is depicted in Figure 42. From the maps and graph, it is
85
evident that the major forest type in this sub-basin is deciduous which remained
more or less consistent throughout the time period from 1984 to 2010 though a slight
depletion is observed. Mixed forest covers a greater area than coniferous forest in
Mainstem. Increasing planted and re-growth forest area is correlated with the
decreasing area of forest cut blocks. Also, fluctuations of the croplands and pasture
extent are interrelated. The modest spectral resolution of the satellite data used in
this study has contributed to some confusion while classifying these two LULC
types. The shrinking of wetland area is observed in this sub-basin as well, however,
the area was depleted to its lowest in 1999 and started to regain thereafter.
86
" I - a.- V
Built-up area Coniferous fore6t Cropland Cut bloc*
Deciduous forest mMixed forest mPasturePlanted or regrowth forest
WaterWetland
k
N
10 km
Figure 39: LULC map of Mainstem sub-basin in 1984
87
1999
Buit-up area Coniferous forest Cropland Cut block
Deciduous forest mMixed forest mPasturePlanted or regrowtn forest
WaterWetland
10km
Figure 40: LULC map of Mainstem sub-basin in 1999
88
I Built-up area I Coniferous forest Cropland
I Cut block
Deciduous forest m Mixed forest mPastureRanted or regrowth forest
WaterWetland I
10 km
Figure 41: LULC map of Mainstem sub-basin in 2010
89
300 Mainstem LULC Change
250
200
150
100
50
0CL CF DF MF
1198419992010
Figure 42: Change in LULC w ithin M ainstem sub-basin
A slow, but steady increase in the built-up area is captured from this analysis
which may be attributed to various infrastructure developments in this sub-basin.
Among all the sub-watersheds, Mainstem is the closest to the City of Dawson Creek
and contains high amounts of agricultural and farming activities which defend the
enhanced infrastructure development within this sub-basin.
4.3.1.2 Brassey sub-basin
Brassey sub-basin is located in the north-western corner of the study
watershed. Figures 43 - 46 describe the LULC changes within Brassey. Similar to
Mainstem, this sub-watershed's forest area is also dominated by deciduous forest,
90
followed by mixed and coniferous forest. The significant and dynamic extent of
planted and re-growth forest during the study period refers to the forest industry
interest within Brassey.
| Built-up area | Coniferous forest Cropland
I Cut block
Deciduous forest Mixed forest PasturePlanted or regrowth forest
WaterWetland
5 km
Figure 43: LULC map of Brassey sub-basin in 1984
91
Built-up area Coniferous forest Cropland Cut block
Deciduous forest mMixed forest mPasturePlanted or regrowth forest
WaterWetland t
5 km
Figure 44: LULC map of Brassey sub-basin in 1999
92
| Built-up area I Coniferous forest Cropland
I Cut block
| Deciduous forest Mixed forest PasturePlanted or regrowth forest
WaterWetland I
5 km
Figure 45: LULC map of Brassey sub-basin in 2010
Brassey LULC Change
198419992010
n i i i i i i i r
CL CF DF MF P/RF CB PS WT WL BA
Figure 46: Change in LULC w ithin Brassey sub-basin
93
Wetlands comprise a notable portion of Brassey. Interestingly, the wetland
area has increased slightly between 1984 and 2010 in this sub-watershed even
though the wetlands faced massive depletion on the whole. Consistent increase of
the built-up area underlines the anthropogenic modification in this sub-basin.
4.3.1.3 Halfmoon-Oetata sub-basin
Halfmoon-Oetata is the smallest sub-basin within this study watershed. This
is another deciduous forest dominant sub-basin. But during the study period,
deciduous forest area was highest in 1984, then dropped in 1999 and started
increasing thereafter. On the other hand, unlike the two sub-basins explained above,
coniferous forest area in Halfmoon-Oetata was constantly higher than mixed forest
during these three study years. Planted and re-growth forest elevated continuously
which signifies the planned forest harvesting within this sub-basin. Cropland and
pasture have a minor share in this sub-watershed. Figures 47 - 50 show the LULC
changes within Halfmoon-Oetata.
94
Built-up area Coniferous forest Cropland Cut block
Deciduous forestMixed forest _PasturePlanted or regrowth forest
WaterWetland i
5 km
Figure 47: LULC map of Halfmoon-Oetata sub-basin in 1984
95
Built-up area Coniferous forest Cropland Cut block
Deciduous forestMixed forest mPasturePlanted or regrowth forest
WaterWetland t
5 km
Figure 48: LULC map of Halfmoon-Oetata sub-basin in 1999
96
Built-up area Coniferous forest Cropland
I Cut block
Deciduous forest Mixed forest PasturePlanted or regrowth forest
WaterWetland I
5 km
Figure 49: LULC map of Halfmoon-Oetata sub-basin in 2010
120
100
80 -\
60
40
20
0
Halfmoon-Oetata LULC Change
198419992010
I PI I r
T 1----------1----------r 1----------1----------1----------r
CL CF DF MF P/RF CB PS WT WL BA
Figure 50: Change in LULC within Halfmoon-Oetata sub-basin
97
Wetland extent was minimum in 1999, however, it was similar in 1984 and
2010. Thus, wetlands were depleted to a larger amount in between 1984 and 1999,
but increased between 1999 and 2010. The consistent increase in the built-up area is
significant in this sub-basin as it is mostly due to the recent shale gas development
infrastructures.
4.3.1.4 East KRW sub-basin
East KRW sub-basin covers the south-eastern portion of the watershed and is
home to the east confluence of the Kiskatinaw River. Figures 51 - 54 display the
LULC changes within East KRW. Over the last few years, this sub-watershed
remains very busy for gas development activities which have been reflected by the
continuous increase of the built-up area.
98
Built-up area Coniferous forest Cropland Cut block
Deciduous forest mMixed forestPasturePlanted or regrowth forest
WaterWetland
I
10 km
Figure 51: LULC map of East KRW sub-basin in 1984
99
Built-up area I Coniferous forest Cropland
I Cut block
Deciduous forest mMixed forestPasturePlanted or regrowth forest
WaterWetland I
10 km
Figure 52: LULC map of East KRW sub-basin in 1999
100
Built-up area I Coniferous forest Cropland Cut block
| Deciduous forest m Mixed forest PasturePlanted or regrowth forest
WaterWetland t
10 km
Figure 53: LULC map of East KRW sub-basin in 2010
101
600East KRW LULC Change
500
400
300
200
100
0CL CF DF MF P/RF CB PS WT WL BA
1198419992010
Figure 54: Change in LULC within East KRW sub-basin
Unlike the other three northern most sub-basins, the landscape of East KRW
is significantly coniferous forest dominant with a gradual increase in extent.
Deciduous and mixed forest shared mostly the same amount of land area in 1984,
1999 and 2010 although minor changes are observed. A smaller amount of planted
and re-growth forest and cut block is also noteworthy. Wetlands exhibited a
continuous and striking diminution during the study period within this sub
watershed. The area covered by cropland and pasture is minimal.
102
4.3.1.5 West KRW sub-basin
West KRW occupies the south-western portion of the study area and holds
the western confluence of the Kiskatinaw River. The most conspicuous LULC
feature of this sub-basin is the fire affected forest area in the central-western border.
Similar to East KRW, this is another coniferous forest dominant sub-watershed
while the coniferous area reached its maximum in 1999. Other mature forest types,
e.g. deciduous and mixed forest showed minor changes during the study period.
Gradual increment of the planted and re-growth forest area may be supported by
the declining forest cut block area. Figures 55 - 58 show the LULC changes within
East KRW.
103
Built-up area Coniferous forest Cropland Cut block
I Deciduous forest
Forest fireMixed forest MPasturePlanted or regrowth forest
WaterWetland t
10 km
Figure 55: LULC map of West KRW sub-basin in 1984
104
1999
Built-up area Coniferous forest Cropland Cut block
I Deciduous forest
Forest fire H I! Mixed forest H IPasturePlanted or regrowth forest
WaterWetland I
10 km
Figure 56: LULC map of West KRW sub-basin in 1999
105
2010
I Built-up area i Coniferous forest Cropland Cut block
I Deciduous forest
Forest fireMixed forest ■ ■PasturePlanted or regrowth forest
WaterWetland t
10 km
Figure 57: LULC map of West KRW sub-basin in 2010
106
Wetland areas dropped to their lowest coverage in 1999 and then showed a
very minor increase in 2010. Built-up area is also escalating, but remained mostly the
same from 1999 to 2010. Cropland and pasture covers insignificant land area in this
sub-basin indicating minimal agricultural and farming activities.
600West KRW LULC Change
500
400 OJ
300
200
100
0A r &
11984
1999
2010
Figure 58: Change in LULC w ithin West KRW sub-basin
107
4.3.2 Wetland depletion
A striking depletion of wetlands in KRW has been captured in this satellite
image analysis. Wetlands are a vital component for regional ecosystems. Wetlands
play a dominant role in the global carbon cycle, containing about 12% of the global
carbon pool (Sahagian & Melack, 1998). Thus, this prevalent depletion of wetlands
has become a vital concern, given the current state of global warming and climate
change as well as ongoing high demand of water use (groundwater-surface water
interaction). Most of this extensive wetland depletion occurred between 1984 and
1999. Figure 59 shows contribution of wetlands to the net land use change within the
watershed and clearly exhibits that most of the lost wetland area has been converted
to various forest categories. Contribution to pasture land, built-up area and forest
fire also need to be noted as these conversions may have important implications to
the ecosystem of the watershed.
108
FF
BA
WT
PS
CB
P/RF
MF
DF
CF
CL
20 40 60 80 100 1200
Figure 59: W etlands converted to other land-use type (1984-2010)
Figure 60 shows gained, lost and persistent wetland area within the
watershed from 1984 to 2010. It is clear that a very minor part of wetlands remains
unchanged within the watershed during this period. The major change has occurred
in the southern portion which is dominated by coniferous forest. The depletion in
wetland extent may cause multi-dimensional impacts on the watershed, such as:
change in carbon storage, loss of biodiversity, increase in flooding, decrease in water
quality etc. Therefore, further rigorous research should be designed and performed
to investigate the reason and effects of this extensive depletion of wetlands.
109
4.3.3 Natural gas development infrastructure
The unconventional shale gas development within Kiskatinaw watershed has
gained tremendous interest over the past few years, particularly in the late 1990s. A
great deal of oil and gas activity is under intense operation within this watershed for
shale gas development. As a result, oil and gas infrastructures, such as: drilling
wells, drilling pads, petroleum development roads etc. are becoming widespread
features in this watershed, influencing the land use change dynamics within this
area.
r \ ,
1984 1999 2010
Built-up area
20 KmCut block
Figure 61: Changes in built-up area and cut block (1984-2010)
111
Figure 61 explains the conditions in 1984,1999 and 2010 for built-up area and
cut block features which were changed mainly due to gas development industry. In
the figure, the built-up area comprises permanent major roads, petroleum access
roads and petroleum development roads; conversely, cut block feature includes
forest clear cut for installing drilling pads and wells as well as forest harvesting by
the forestry industry. Thus, both these contain footprints of natural gas development
in this watershed. It is evident that there was almost no or very little gas activity in
1984, but it intensified by 1999. The road network was mostly compartmentalized in
East KRW sub-basin in 1999 which underscores amplified development activity
within this sub-basin. There is some activity observed in the Brassey and Halfmoon-
Oetata as well. The larger cut blocks are mostly due to forest industry harvesting,
but the relatively small clear cuts are a gas development signature.
In 2010, the East KRW gas development continued at a high pace.
Additionally, Brassey, Halfmoon-Oetata and north-western part of Mainstem drew
significant industry interest; however, West KRW and the southern most edge of
East KRW remained unaffected even in 2010. This intensified gas development
activity has multi-dimensional impacts on KRW's natural landscape as it is
attracting a tremendous inflow of population from outside and their concomitant
multi-faceted demands. Therefore, a proper planning on this regard needs to be
strictly framed and implemented.
112
4.4 Land use m odeling
A series of transition sub-models were identified by the IDRISI Land Change
Modeler (LCM) based on the change analysis between two input land use maps of
1984 and 2010. All of these empirically evaluated sub-models were considered
individually for modeling except for those with forest fire class since prediction on
forest fire is out of the scope of this study.
4.4.1 Results of modeling analysis
At the initial stage of modeling, driver variables as described in Table 3 were
evaluated for their potential explanatory power in LULC change projection. The
selection and processing of driver variables entailed careful consideration of KRW's
ongoing land use activities and data availability. The variables played both static
and dynamic role during the modeling process.
The multi-layer perception (MLP) neural network tool on Land Change
Modeler (LCM) has a strong evaluation procedure for measuring the efficiency of
driver variables. Input variables of each transition sub-models were evaluated under
MLP. For example, modeling efficiency of transition from wetlands to pasture was
evaluated as shown in Table 7. Four governing driving variables were identified for
113
the sub-model of wetland transition to pasture, including "distance to gas
development infrastructure", "distance to major channel network", "topographic
wetness index (TWI)", and "digital elevation model (DEM)". A num ber of scenarios
were considered during the evaluation process considering one or multiple variables
to be constant each time. This way, the impact of the constant variables in the
modeling process was determined. In the example, the variable (V4) digital
elevation model (DEM) was identified as the most influential variable for modeling
the transition from wetlands to pasture. The modeling accuracy and skill m easure of
this sub-model were found to be 78.55% and 0.5709, respectively.
The MLP neural network modeling process generated transition potential
maps for each evaluated transition sub-models. For example, Figure 62 exhibits the
transition potential map for the conversion from planted or re-growth forest to
deciduous forest. The maximum possibility of this transition is observed as 50%
within the watershed, particularly in the northern half of the watershed. The
transition potential maps generated from MLP modeling were used in Markov
Chain (MC) model for calculating the amount of change to be expected for each
transition and predicting future scenarios. MC process recorded the modeled
transition probabilities in a matrix which contains information on exactly how much
land would be expected to alter from the later date (2010) to the prediction date
(2020).
114
Table 7: Sensitivity of transition model for forcing independent variables
V I Distance to gas development infrastructure
Independent VariablesV 2 Distance to major channel networkV 3 Topographic wetness index (TWI)
V 4 Digital elevation model (DEM)
Forcing a Single Inde jendent Variable to be ConstantAccuracy (%) Skill measure Influence order
Sensitivity of model
With all variables 78.55 0.5709 N/AV 1 constant 78.55 0.5709 4 (least influential)V 2 constant 78.37 0.5674 2
V 3 constant 78.45 0.569 3V 4 constant 49.81 -0.0038 1 (most influential)
Forcing All Independent Variables Except One to be ConstantAccuracy (%) Skill measure
With all variables 78.55 0.5709All constant but V 1 49.79 -0.0042
Sensitivity of m odel All constant but V 2 49.79 -0.0042All constant but V 3 49.8 -0.004All constant but V 4 78.27 0.5654
Backwards Stepwise Constant Forcing
Variables included Accuracy (%) Skill measureWith all variables All variables 78.55 0.5709
Sensitivity of modelStep 1: V [1] constant, [2,3,4] 78.55 0.5709
Step 2: V [1,3] constant [2,4] 78.45 0.5690Step 3: V [1,3,2] constant [4] 78.27 0.5654
115
ir' ■ / / .. j i iV
v * ' V ^ ,
0.48 - 0.49
0.49 - 0.500 - 0.46 20 km0.46 - 0.48
Figure 62: Transition probabilities from planted/re-growth forest to deciduous forest
116
4.4.2 Discussion on modeled outcomes
The transition probability matrix (Table 8) elucidates that coniferous (0.92)
and deciduous (0.82) forest have high probability of maintaining most of their
current extent in 2020; however, this probability is 54% for mixed forest. Chance of
mixed to coniferous forest transition is 0.19. The likelihood of converting planted or
re-growth forest to deciduous is 0.33 while there is a 40% chance that the planted or
re-growth forest will remain the same. These outcomes represent the planned and
healthy forestry practice in this area since there is no major depleting trend for any
forest type observed. The matrix exhibits that transition from cropland to pasture
and vice-versa will be continuing during the model period. There is a 92% chance
that there will be very minor change in open water extent.
Table 8: Transition probability matrix
Given • l ii: j__~j C -1______
LULC 1riUDdDliliy Ul tlld llg lltg IU
type CL CF DF MF P/RF CB PS WT WL BACL 0.16 0.00 0.05 0.00 0.10 0.01 0.51 0.00 0.13 0.04CF 0.00 0.92 0.01 0.02 0.01 0.01 0.00 0.00 0.03 0.02DF 0.00 0.00 0.82 0.06 0.10 0.00 0.00 0.00 0.01 0.01MF 0.00 0.19 0.14 0.54 0.03 0.01 0.01 0.00 0.06 0.01
P/RF 0.04 0.00 0.33 0.00 0.40 0.00 0.10 0.00 0.13 0.00CB 0.04 0.04 0.05 0.38 0.06 0.02 0.14 0.00 0.26 0.03PS 0.14 0.00 0.03 0.00 0.03 0.05 0.59 0.00 0.15 0.01
WT 0.00 0.00 0.00 0.01 0.00 0.00 0.04 0.92 0.03 0.00WL 0.00 0.22 0.14 r 0.30 0.00 0.00 0.02 0.00 0.31 0.00BA 0.05 0.01 0.02 0.00 0.00 0.04 0.03 0.00 0.03 0.84
117
Similarly, there is a 84% chance that the built-up area will remain unaltered.
At the same time, probabilities of gain in built-up area extent from other LULC types
are also insignificant; but this projection dem ands further investigation since it
would make more sense if the chances of built-up area gaining would be higher
given the current situation of dense industrial development in this watershed.
Changing the driver variables or modifying other modeling parameters may result
differently and could produce more reasonable output for built-up area.
During the modeled period, wetland to forest conversion will remain active
with probabilities of 0.22 for coniferous, 0.14 for deciduous and 0.30 for mixed
forest. There is only a 31% chance is that wetland extent will be unaffected, yet there
is notable likelihood of alteration of other classes to wetland, such as: chance of
change from cropland to wetland is 0.13; planted or re-growth forest to wetlands is
0.13 and for cut block and pasture, chance is 0.25 and 0.15 respectively.
Summarizing all the transition probabilities explained above, the MC model
was used to create another matrix of expected transition of pixels for enhanced
understanding (Table 9). This matrix exhibits how many pixels of a LULC type are
expected to transition to other LULC types in 2020. In this table, the highlighted
pixels will remain unchanged for each LULC class. Based on this latest matrix, LCM
produced a hard prediction using multi-objective land allocation algorithm. Also, a
118
soft prediction is crafted using the principle that a location is more vulnerable to
change if it is desired by several transitions at the same time than if it is desired by a
single transition.
Table 9: Expected transition of pixels
Pixels in 11 CL CF DF MF P/RF CB PS WT WL BACL 20913 0 0 0 0 0 0 0 0 0CF 0 1192282 4501 10791 5488 7338 0 0 9558 3453DF 0 0 831158 28544 43795 0 0 0 3960 3141MF 0 38905 29027 319499 6464 0 0 0 12602 0
P/RF 0 0 29036 0 145759 0 0 0 0 0CB 0 0 0 0 0 29312 0 0 0 0PS 0 0 0 0 0 0 66995 0 0 0
WT 0 0 0 0 0 0 0 15051 0 0WL 0 22838 14210 30223 0 0 2068 0 134392 0BA 0 0 0 0 0 0 0 0 0 52499
A hard prediction result was two land use-land cover maps for 2015
(intermediate stage) and 2020 (final stage) with all the classes that exist in 2010
LULC map while forest fire affected area remained the same as it was excluded from
the transition modeling process. Figure 63 and Figure 64 exhibit the maps generated
from hard prediction.
119
2015
Built-up area
Coniferous forest
Cropland
Cut block
Deciduous forest
Forest fire
Mixed forest Pasture
Planted or regrowth forest
Water
Wetland 25 kmN
Figure 63: Interm ediate stage LULC map from hard prediction
120
2020
Built-up area
Coniferous forest
Cropland
Cut block
Deciduous forest
Forest fire
M xed forest
Pasture
Planted or regrowth forest Water
Wetland
kr j j
25 km
Figure 64: Final stage LULC map from hard prediction
121
Table 10 lists the area calculated for both predicted LULC maps. Though
minor changes are projected for few classes, no significant change is observed. This
is further realized in Figure 65 which demonstrates the difference in area for each
LULC type between the predicted LULC maps and existing (2010) LULC map.
Forest cover will remain almost the same with minor shifting from one to another; a
45 km2 increase in total forest cover is predicted. The depleting trend of wetland area
appears to be continuing; depletion of another 67.89 km2 of wetland was estimated
by the model. It also forecasted an increase of 11.57 km2 of built-up area and 7.74
km2 of cut block area from which it may be inferred that the ongoing industrial
activity will keep escalating during the prediction period. The depletion of wetlands
needs further examination to identify the probable cause of change. Some possible
reasons may be climate change, enhanced anthropogenic activities including natural
gas development, logging, agriculture etc. as explained in the preceding section.
122
Table 10: Area calculated for predicted LULC maps
LULC 2010 Intermediate Stage (2015) Final Stage (2020)type km2 % km2 % km2 %CL 18.82 0.66 18.82 0.66 18.82 0.66CF 1107.84 39.06 1126.51 39.72 1164.88 41.08DF 815.34 28.75 817.06 28.81 800.01 28.21MF 365.88 12.90 348.19 12.28 324.51 11.44
P/RF 157.23 5.55 179.37 6.32 201.88 7.12CB 26.38 0.93 32.59 1.15 34.13 1.21PS 60.30 2.13 62.16 2.19 63.57 2.24WT 20.48 0.72 20.67 0.73 20.79 0.73WL 183.30 6.46 144.26 5.09 115.41 4.07BA 47.24 1.67 53.18 1.88 58.81 2.07FF 33.19 1.17 33.19 1.17 33.19 1.17
1200.00
1000 00
800 00
600.00x,
400 00
200.00
0 00
2010
Intermediate Stage (201S)
Final Stage (2020)
CL CF DF MF PRF CB PS WT WL BA FF
Figure 65: Difference in area for each LULC type between existing and predicted maps
123
The soft prediction output consists of maps representing the probability of
change for a given set of transitions. Soft prediction was also performed in two
stages. Figure 66 exhibits the soft prediction output from LCM for 2020. The soft
output was a continuous mapping of vulnerability to change for selected set of
transitions. This prediction identified the extent to which the land area has the
propensity and right criteria to be altered. While the hard prediction created only a
single realization of the future LULC status, the soft prediction was a comprehensive
assessment of change potential. This is why the soft output detected the areas with
varying degree of vulnerability instead of identifying w hat and how much of LULC
area would be changed. From the modeled output, it is evident that most of the
southern portion of the watershed is highly vulnerable to transition under the
current set of driver variables and identified individual transitions from one type to
another. This is reasonable as this part of the watershed has large area of wetlands
which has exhibited the most significant depletion during the study period.
Considerable vulnerability is observed in the northern portion as well where
Halfmoon-Oetata, Brassey and Mainstem sub-basins are located. Reasons for this
vulnerability may be attributed to the recent intensified gas development activity in
this area along with land use change derived by agricultural and farming activities.
The middle portion of the watershed is characterized by mixed probability of change
while most of the open water body showed almost no vulnerability to change.
124
n '
2020Vulnerability to change
0
0 .0 1 - 0.20
0.20-0.30 0.30 0 40 0.40-0.50 0.50 - 0.60 O.SO - 0.70 0.70-0.80 0.80-0.90 0.90 100
20 kmt
Figure 66: Soft prediction output for 2020
125
CHAPTER 5: CONCLUSION AND RECOMMENDATIONS
5.1 Summary
Remote sensing, GIS and modeling techniques were employed in this
research to study the land use-land cover dynamics in Kiskatinaw River Watershed
in north-eastern BC. The study provided an avenue for comprehensive LULC
analysis in KRW. LULC maps were created to understand the characteristics of
LULC change which were later used for transitional potential modeling and
projecting future LULC pattern.
Landsat satellite images were successfully exploited in the present research
for producing LULC information for the study area. Three satellite images of the
years 1984, 1999 and 2010 were studied. An object oriented image classification
technique was executed for analyzing the satellite images with a high degree of
accuracy (90.45%). The analysis of the images generated three LULC maps for the
study area. The produced maps identified that the major share of the LULC within
the study area is occupied by different types of forest which was estimated as 80% in
1984, 85.60% in 1999 and 86.28% in 2010. Forest harvesting has been compensated by
planting or allowing re-growth of the forests according to forestry practice in this
watershed.
126
Land use practices, such as agriculture, farming etc. are mostly limited to the
northern part of the watershed. Other than these, the most striking land use activity
during the past few years is natural gas development. The study area falls within the
Montney shale gas play which is one of the large underground shale gas reservoirs
of the world. The extraction of this natural gas is heavily impacting the study
watershed's land use dynamics. Until now, the signature of this intensified gas
development activity is narrow in the LULC maps; however, the impact could be
significant if not properly planned and managed. East KRW sub-basin is facing the
most extensive natural gas development. Activities are also prominent in Halfmoon-
Oetata, Brassey as well as in Mainstem sub-basins.
Another distinguishing feature of this LULC change analysis is the extensive
depletion of the wetland area. A total of 270.78 km2 of wetlands have disappeared
during the study period from 1984 to 2010 while most of the depletion occurred
between 1984 and 1999 estimated as 233.22 km2. This wetland depletion demands
further investigation to satisfy the queries of 'w hat are the reasons of this depletion'
and 'w hat is its impact on the watershed system'. The northern location of the
watershed made the wetland depletion more serious due to the implications of
climate change.
127
LULC changes were also realized at the sub-watershed scale. The dynamics
of change vary from sub-basin to sub-basin since the land use activity is different in
each area. The cropland and pasture land use type exist mainly in the Mainstem and
Brassey sub-basins and thus, these two sub-basins experienced associated alteration
in the extent of cropland and pasture. Mainstem was identified as more dynamic
than Brassey for agriculture and farming. The change in forest cover remains active
for all the sub-watershed during the study period. However, shrinking of wetland
area varied for different sub-basins w ith higher amount of depletion observed in the
southernmost areas. The increase in built-up area occurred across the watershed,
although east KRW sub-basin featured the most rigorous gas development activity.
Based on the change analysis performed for KRW, a modeling approach was
applied in this study to forecast the future LULC dynamics of the watershed. Multi
layer perception (MLP) neural network was utilized to model the transition
potential of each identified LULC transition during the study period while Markov
Chain (MC) model was used to generate future LULC scenario. Both models were
simulated up to 2020. Several driving variables were identified for the modeling
analysis. These variables are expected to alter LULC in KRW and were tested for
their effectiveness before final model was run. The transition probability matrix and
expected transition area matrix were produced from the first phase of modeling.
These matrices exhibited the likelihood of change for each LULC class. Different
128
forest types and open water body showed the highest probability that they will keep
their current extent unaltered after the modeled period in 2020. Wetlands will
continue to deplete during this period though at a slower rate. Dynamics for other
LULC types will remain more or less the same. Based on the transition probability
estimation, two types of output were produced from MC modeling: hard prediction
and soft prediction. Soft prediction generated maps show overall vulnerability of
change for the whole watershed, whereas hard prediction generated LULC maps for
2015 and 2020.
In summary, the study combined contemporary satellite image classification
methods with a robust modeling environment and proved effective for the land-use
and land cover analysis in Kiskatinaw River Watershed. The research produced a
LULC inventory for the study area which will benefit the land use planner and
stakeholders of the watershed to formulate and implement an efficient water
resources management. Being the first comprehensive LULC inventory of the
watershed, its leverage could be multi-faceted for this fast economically growing
watershed. The LULC information could be realized in every sector related to
natural resources extraction and development within Kiskatinaw River Watershed.
129
5.2 Limitations and recommendations for future work
In this study, the application of RS, GIS and modeling has proven as an
effective means for land use and land cover change analysis and produced some
meaningful information about the overall land use system of Kiskatinaw watershed.
But it involves some inherent limitations, such as:
1. Challenges in acquiring cloud free useable satellite imagery;
2. With the spatial resolution (30 m) of the Landsat satellite data, it was hard
to differentiate the signature of some LULC features, e.g. cropland and pasture etc.;
3. Inaccessibility of the study watershed for which some ground truth survey
locations could not be sampled;
4. The time frame and scope of the study did not allow furthering the
investigation of the wetland depletion findings.
The experiences obtained from this study put forward some
recommendations for future research. Aerial photograph or higher resolution
satellite imagery of dry and wet season for each year may produce more robust
results, particularly to justify the change of the wetlands whose remarkable
depletion poses a serious urgency. Also, incorporation of a soil moisture study with
130
LULC change analysis would be effective as soil moisture varies w ith land-use types
and have the potential to make the detection process more robust.
For more comprehensive study, RS and GIS analysis can also be aligned with
hydrological modeling to investigate if there is any connection between the land-use
changes, alteration of hydrologic regime and anthropogenic activity which may also
be able to tell us about the causes of the wetlands decrease.
131
Bibliography
Almeida, C. (2008). Using neural networks and cellular automata for modelling
intra-urban land use dynamics. International Journal o f Geographic Information Science,
22 (9), 943-963.
Aplin, P., Atkinson, P. M., & Curran, P. J. (1997). Fine spatial resolution satellite
sensors for the next decade. International Journal o f Remote Sensing, 18 (18), 3873-3881.
Araya, Y. H. (2009). Urban land use change analysis and modeling: a case study ofSetubal
and Sesimbra, Portugal. Unpublished thesis. Munster, Germany: Institute for
Geoinformatics, University of Munster.
Aspinall, R. (2004). Modeling lsnd use change with generalized linear models - a
multi-model analysis of change between 1860 and 2000 in Gallatin Valley, Montana.
Journal o f Environmental Management, 72, 91-103.
Ayele, H. (2011). Land use/ land cover change and impact ofjatropha on soil fertility : the
case ofM ieso and Bati districts, Ethiopia. Unpublished thesis. Alemaya,Ethiopia:
Haramaya University.
Baker, W. L. (1989). A review of models of landscape change. Landscape Ecology, 2 (2),
111-133.
Batty, M. (2005). Cities and Complexity: Understanding Cities with Cellular Automata,
Agent-Based Models, and Fractals. Cambridge, Massachusetts: The MIT Press.
Beck, P. S., Homing, N., Goetz, S. J., Loranty, M. M., & Tape, K. D. (2011). Shrub
Cover on the North Slope of Alaska: a circa 2000 Baseline Map. Arctic, Antarctic, and
Alpine Research, 43 (3), 355-363.
132
Benito, P., Cuevas, J., Parra, R., Prieto, F., Barrio, J., & Zavala, M. (2010). Land use
change in a Mediterranean metropolitan region and its periphery: assessment of
conservative policies through CORINE Land Cover data and Markov models. Forest
Systems, 19 (3), 315-328.
Benz, U. C., Hofmann, P., Willhauck, G., Lingenfelder, I., & Heynen, M. (2004).
Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready
information. ISPRS Journal o f Photogrammetry and Remote Sensing, 58, 239-258.
Blaschke, T. (2010). Object based image analysis for remote sening. ISPRS Journal of
Photogrammetry and Remote Sensing, 65, 2-16.
Brink, A. B., & Eva, H. D. (2009). Monitoring 25 years of land cover change dynamics
in Africa: A sample based remote sening approach. Applied Geography, 29 (4), 501-
512.
Bruce, D. (2008). Object orinted classification : case studies using different image
types different spatial resolutions. The International Archives o f the Photogrammetry,
Remote Sensing and Spatial Information Sciences (pp. 515-520). Beijing: International
Society of Photogrammetry and Remote Sensing (ISPRS).
Campbell, J. B. (2002). Introduction to remote sensing. New York: the Guilford Press.
Casals-Carrasco, P., Kubo, S., & Madhavan, B. B. (2000). Application of spectral
mixture analysis for terrain evaluation studies. International Journal o f Remote Sensing,
21 (16), 3039-3055.
Chan, J., Chan, K., & Yeh, A. G. (2001). Detecting the nature of change in an urban
environment: A comparison of machine learning algorithms. Photogrammetric
Engineering and Remote Sensing, 67 (2), 213-225.
133
Chen, J., Mao, Z., Philpot, B., Li, J., & Pan, D. (2013). Detecting changes in high-
resolution satellite coastal imagery using an image object detection approach.
International Journal o f Remote Sensing, 34 (7), 2454-2469.
Civco, D. L. (1993). Artificial neural networks for land cover classification and
mapping. International Journal o f Geographic Information Systems, 7 (2), 173-186.
Congalton, R. G. (1991). A review of assessing the accuracy of classifications of
remotely sensed data. Remote Sensing o f Environment, 37, 35-46.
Congalton, R., & Green, K. (2009). Assessing the accuracy of remotely sensed data:
principles and practices. Boca Raton, Florida, USA: CRC/Taylor & Francis.
De Jong, S. M., Homstra, T., & Maas, H. (2001). An integrated spatial and spectral
approach to the classification of Mediterranean land cover types: the SSC method.
International Journal o f Applied Earth Observation and Geoinformation, 3 (2), 176-183.
De Kok, R., Schneider, T., & Ammer, U. (1999). Object based classification and
application in the Alpine forest environment. Joint ISPRS/EARSeL: Fusion o f sensor
data, knowledge sources and algorithms. Valladolid, Spain: ISPRS.
Dean, A. M., & Smith, G. M. (2003). An evaluation of per-parcel land cover mapping
using maximum likelihood class probabilities. International Journal o f Remote Sensing,
24, 2905-2920.
Desclee, B., Bogeart, P., & Defourny, P. (2006). Forest change detection by statistical
object-based menthod. Remote Sensing of Environment, 102,1-11.
Dietzel, C., & Clarke, K. C. (2007). Towards optimal calibration of the SLEUTH land
use change model. Transactions in GIS, 11, 29-45.
134
Dobson Engineering and Urban Systems Ltd. (2003). Kiskatinaw river watershed
management plan. Dawson Creek, BC, Canada: City of Dawson Creek.
Dobson Engineering Ltd. (2007). Kiskatinaw River watershed source protection plan.
Dawson Creek, BC: City of Dawson Creek.
Dwivedi, R. S., Kandrika, S., & Ramana, K. V. (2004). Comparison of classifiers of
remote-sensing data for land-use/land-cover mapping. Current Science, 86 (2), 328-
335.
Dymond, J. R., Shepherd, J. D., Newsome, P. F., Gapare, N., Burgess, D. W., & Watt,
P. (2012). Remote sensing of land-use change for Kyoto protocol reporting: the New
Zealand case. Environmental Science & Policy, 16,1-8.
Eastman, J. R. (2012). "Run Transition Sub-Model - Land Change Modeler", IDRISI
help system. Accessed in IDRISI selva. Worcester, MA, USA: Clark University.
Echeverria, C., Coomes, D. A., Hall, M., & Newton, A. C. (2008). spatially explicit
models to analyze forest loss and fragmentation betweenl976 and 2020 in southern
Chile. Ecological Modelling, 212, 439-449.
Egberth, M., & Nilsson, M. (2010). KNN-Sweden-Current map data on Swedish
forests. ForestSat 2010: Operational tools in forestry using remote sensing techniques (pp.
265-267). Lugo, Spain: ForestSat.
El-Kawy, O. R., Rod, J. K., Ismail, H. A., & Suliman, A. S. (2011). Land use and land
cover change detection in the western Nile delta of Egypt using remote sensing data.
Applied Geography, 31 (2), 483-494.
135
Enderle, D., & Weih, R. C. (2005). Integrating supervised and unsupervised
classification methods to develop a more accurate land cover classification. Journal of
the Arkansas Academy of Science, 59, 65-73.
Foley, J.A., DeFries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin,
F.S., Coe, M.T., Daily, G.C., Gibbs, H.K., Helkowski, J.H., Holloway, T., Howard,
E.A., Kucharik, C.J., Monfreda, C., Patz, J.A., Prentice, I.C., Ramankuty, N., Snyder,
P.K. (2005). Global consequences of land use. Science, 309, 570-574.
Foley, Jonathan; Ramankutty, Navin; Brauman, Kate; Cassidy, Emily; Gerber, James;
Johnston, Matt; Mueller, Nathaniel; O'Connell, Christine; Ray, Deepak; West, Paul;
Balzer, Christian; Bennett, Elena; Carpenter, Stephen; Hill, Jason; Monfreda, Chad;
Polasky, Stephen; Rockstrom, Johan; Sheehan, John; Siebert, Stefan; Tilman, David;
Zaks, David (2011). Solutions for a cultivated planet. Nature, 478, 337-342.
Forest Practices Board. (2011). Cumulative effects assessment: A case study for the
Kiskatinaw River Watershed. Victoria, BC: Forest Practices Board .
Freier, K. P., Schneider, U. A., & Finckh, M. (2011). Dynamic interations between
vegetation and land use in semi-arid Morocco: using a Markov process for modeling
rangelands under climate change. Agriculture, Ecosystems and Environment, 140, 462-
472.
Frohn, R., Autrey, B., Lane, C., & Reif, M. (2011). Segmentation and object-oriented
classification of wetlands in a karst Florida landscape using multi-season Landsat-7
ETM+ imagery. International Journal o f Remote Sensing, 32 (5), 1471-1489.
Gao, Y., & Mas, J. (2008). A comparision of the performance of pixel-based and
object-based classifications over images with various spatial resolution. GEOBIA
2008-GEOgraphic object based image analysis for the 21st century. Calgary, AB: ISPRS.
136
Geneletti, D., & Gorte, B. (2003). A method for object-oriented land cover
classification combining Landsat TM data and aerial photographs. International
Journal o f Remote Sensing, 24 (6), 1273-1286.
Goldewijk, K., Beusen, A., Drecht, G., & Vos, M. (2011). The HYDE 3.1 spatially
explicit database of human-induced global land-change over the past 12,000 years.
Global Ecology and Biogeography, 20 (1), 73-86.
Gregersen, H. M., Ffolliott, P., & Brooks, K. (2007). Integrated Watershed
M anagementConnecting people to their land and water. Oxfordshire: CAB International.
Hasbani, J. (2008). Semi-automated calibration of a cellular automata model to simulate
land-use chnages in the Calgary region. Unpublished thesis. Calgary, AB: University of
Calgary.
Hay, G. J., & Gastilla, G. (2006). Object-based image analysis: strength, weakness,
opportunities, and threats (SWOT). 1st International Conference on Object-based Image
Analysis (OBIA 2006), (pp. 4-5). Salzburg, Austria.
Hay, G., & Castilla, G. (2008). Geographic object-based image analysis (GEOBIA): A
new name for a new discipline. In Object based image analysis (pp. 93-112).
Heidelberg, Berlin: Springer.
He, C., Okada, N., Zhang, Q., Shi, P., & Li, J. (2008). Modeling dynamic urban
expansion process incorporating a potential model with cellular automata. Landscape
and Urban Planning, 86, 79-91.
Herold, M., Goldstein, N., & Clarke, C. (2003). The spatiotemporal form of urban
growth: measurement, analysis and modeling. Remote Sensing of Environment, 86 (3),
286-302.
137
Islam, M. S., & Ahmed, R. (2011). Land use change prediction in Dhaka city using
GIS aided Markov chain modeling. Journal o f Life and Earth Science, 6, 81-89.
Jacklin, J., French, T. D., & Carmichael, B. (2003). Assessment of the City o f Dawson
Creek's drinking water supply (Kiskatinaw River): source water characteristics. Prince
George, BC: BC Ministry of Water, Land and Air Protection.
Jha, M. K., Schilling, K. E., Gassman, P. W., & Wolter, C. F. (2010). Targeting land-
use change for nitrate-nitrogen load reductions in an agricultural watershed. Journal
o f Soil and Water Conservation, 65 (6), 342-352.
Kamusoko, C., Aniya, M., Adi, B., & Manjoro, M. (2009). Rural sustainability under
threat in Zimbabwe-Simulation of future land use/cover changes in the Bindura
district based on the Markov-cellular automata model. Applied Geography, 29 (3), 435-
447.
Kamusoko, C., Wada, Y., Furuya, T., Tomimura, S., Nasu, M., & Homsysavath, K.
(2013). Simulating future forest cover change in Pakxeng District, Lao People's
Democratic Republic (PDR): Implications for sustainable forest management. Land,
2,1-19.
City of Dawson Creek (1991). Kiskatinaw River IW M P. Dawson Creek, BC.
City of Dawson Creek (2009). Kiskatinaw River Watershed Brochure. Dawson Creek,
BC.
Kline, J., Moses, A., Lettman, G. J., & Azuma, D. L. (2007). Modeling forest and range
land development in rural locations, with example from eastern Oregon. Urban
Plannin , 80, 320-332.
138
Lambin, E. F., Geist, H. J., & lepers, E. (2003). Dynamics of land-use and land-cover
change in tropical regions. Annual Review o f Environmental Resources, 28, 205-241.
Lee, P., & Hanneman, M. (2012). Atlas o f land cover, industrial land uses and industrial-
caused land change in the Peace Region of British Columbia. Edmonton, AB: Global
Forest Watch Canada report #4 International Year of Sustainable Energy for All.
Li, X., & Yeh, A. G.-O. (2002). Neural-network-based cellular automata for
simulating multiple land use changes using GIS. International Journal o f Geographic
Information Science, 16 (4), 323-343.
Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2008). Remote sensing and image
interpretation (6th ed.). Hoboken, NJ: John Wiley & Sons, Inc.
Lin, Y., Chu, H., Wu, C., & Verburg, P. (2011). Predictive ability of logistic
regression, auto-logistic regression and neural network models in empirical land-use
change modeling - a case study. International Journal of Geographic Information Science,
25 (1), 65-87.
Ly, K. (2010). Assessment o f the effects o f land use change on Water quality and quantity in
the Nam N gum river basin using better assessment science integrating point and nonpoint
sources 4.0 (BASINS 4.0) model. Unpublished thesis. Christchurch, New Zealand:
Lincoln University.
Lyons, M., Phinn, S., & Roelfsema. (2012). Long term land cover and seagrass
mapping using Landsat and object-based image analysis from 1972 to 2010 in the
coastal environment of South East Queensland, Australia. ISPRS Journal o f
Photogrammetry and Remote Sensing, 71, 34-46.
139
MacLean, M., Campbell, M., Maynard, D., Ducey, M., & Congalton, R. (2013).
Requirements for labelling forest polygons in an object-based image analysis
classification. International Journal o f Remote Sensing, 34 (7), 2531-2547.
Martinuzzi, S., Gould, W. A., & Gonz'alez, O. M. (2007). Land development, land use
and urban sprawl in Puerto Rico integrating remote sensing and population census
data. Landscape and Urban Planning, 79, 288-297.
Mas, J. F. (1999). Monitoring land-cover changes: a comparison of change detection
techniques. International Journal o f Remote Sensing, 20 (1), 139-152.
Mas, J. F., Puig, H., Palacio, J. L., & Sosa-Lopez, A. (2004). Modeling deforestation
using GIS and artificial neural networks. Environmental Modelling & Software, 19, 461-
471.
Matinfar, H. R., Sarmadian, F., Panah, S. K., & Heck, R. J. (2007). Comparisions of
obejct-oriented and pixel-based classification of land use/land cover types based on
Landsat7 ETM+ spectral bands (case study: Arid region of Iran). American-Eurasian
Journal o f Agriculture and Environmental Science, 2, 448-456.
Mubea, K., Ngigi, T., & Mundia, C. (2010). Assessing application of Markov chain
analysis in predicting land cover change: a case study of Nakuru Municipality.
Journal o f Agriculture, Science and Technology, 12 (2), 126-143.
Musaoglu, N., Coskun, M., & Kocabas, V. (2005). Land use change analysis of
Beykoz-Istanbul by means of satellite images and GIS. Water Science & Technology, 51
(11), 245-251.
Myint, S., & Wang, L. (2006). Multicriteria decision approach for land use land cover change using Markov chain analysis and a cellular automata approach. Canadian
Journal o f Remote Sensing, 32 (6), 390-404.
140
NASA. (2011). Landsat science data users handbook. Washington, D.C.: National
Aeronautics and Space Administration.
NASA. (2011, March 21). National Aeronautics and Space Administration. Retrieved
June 24, 2013. http://Landsat.gsfc.nasa.gov/education/compositor/diff_color.html
NASA. (2011, March 21). National Aeronautics and Space Administration: Goddard Space
Flight Center. Retrieved June 21, 2013.
http://Landsat.gsfc.nasa.gov/education/compositor/invisible_light.html
Oumer, H. A. (2009). Land use and land cover change, drivers and its impact: A
comparative study from Kuhar Michael and Lenche Dima of Blue Nile abd Awash Basins of
Ethiopia. Unpublished thesis. Ithaca, NY: Cornell University.
Overmars, K. P., Verburg, P. H., & Veldkamp, T. (2007). Comparison of a deductive
and an inductive approach to specify land suitability in a spatiality explicit land use
model. Land Use Policy, 24, 584-599.
Paiboonvorachat, C. (2008). Using remote sensing and GIS techniques to assess land
use/land cover changes in the Nan Watershed, Thailand. Unpublished thesis. Carbondale:
Dept, of Geography and Environmental Resources, Southern Illinois University.
Paul, S. S., Hasan, K., Akhter, S. H., & Li, J. (2012). Application of remote sensing
and GIS for investigating urbanization induced surface water body change in and
around Bangladesh. 12th International Environmental Specialty Conference. Edmonton:
Canadian society for Civil Engineering (CSCE).
Peace Forest District. (2010). A udit o f forestry, oil and gas and range activities in the
Kiskatinaw river watershed. Victoria, BC: Forest Practices Board.
141
Pijanowski, B. C., Brown, D. G., Shellito, B. A., & Manik, G. A. (2002). Using neural
networks and GIS to forecast land use changes: a Land Transformation Model.
Computer, Environment and Urban Systems, 26 (6), 553-575.
Pijanowski, B., Pithadia, S., & Shellito, B. (2005). Calibrating a neural network-based
urban change model for two metropolitan areas of the Upper Midwest of the United
States. International Journal o f Geographic Information, 19 (2), 197-215.
Redoux, J., & Defourny, P. (2007). A quantitative asessment of boundaries in
automated forest stand delineation using high resolution imagery. Remote Sensing of
Environment, 110 (4), 468-475.
Richards, J. A., & Jia, X. (1999). Remote Sensing Digital Image Analysis: A n Introduction
(3rd ed.). New York: Springer.
Ridd, K. M., & Liu, J. (1998). A comparision of four algorithms for change detection
in urban environment. Remote Sensing of Environment, 63, 95-100.
Robertson, L., & King, D. (2011). Comparison of pixel- and object-based classification
in land cover change mapping. International Journal o f Remote Sensing, 3 2 ,1505-1529.
Sabins, F. F. (2007). Remote Sensing: Principles and Interpretation. Long Grove, IL:
Waveland Press Inc.
Saha, G., Paul, S., Li, J., Hirshfield, F., & Sui, J. (2013). Investigation ofland-use change
and groundwater-surfacewater interaction in the Kiskatinaw River Watershed, British
Columbia. Vancouver, Canada: Geoscience BC.
Sangermano, F., Eastman, J., & Zhu, H. (2010). Similarity weighted instance-based
learning for the generation of transition potentials in land use change modeling.
Transactions in GIS, 14 (5), 569-580.
142
Schulp, C., Nabuurs, G., & Verburg, P. (2008). Future carbon sequestration in
Europe-effects of land use change. Agriculture, Ecosystems & Environment, 127, 251-
264.
Serneels, S., & Lambin, E. F. (2001). Proximate causes of land-use change in Narok
District, Kenya: a spatial statistical model. Agriculture, Ecosystems & Environment, 85
(1-3), 65-81.
Singh, A. K. (2003). Modeling land use land cover changes using cellular automata in a
geo-spatial environment. Unpublished thesis. Enschede, The Netherlands: ITC.
Smith, A. P., Western, A. W., & Hannah, M. C. (2013). Linking water quality trends
with land use intensification in dairy farming catchments. Journal o f Hydrology, 476,
1- 12.
Sorensen, R., Zinko, U., & Seibert, J. (2005). On the calculation of the topographic
wetness index: evaluation of different methods based on field observations.
Hydrology and Earth System Sciences Discussions, 2 ,1807-1834.
Stabile, M. (2012). Deconstructing the complexity o f land use and cover classification and
land change modeling. Unpublished thesis. New South Wales, Australia: Faculty of
Agriculture, Food and Natural Resources, The University of Sydney.
Sun, H., Forsythe, W., & Waters, N. (2007). Modeling urban land use change and
urban sprawl: Calgary, Alberta, Canada. Networks and Spatial Economics, 7, 353-376.
Tong, S., & Chen, W. (2002). Modeling the relationship between land use and surface
water quality. Journal o f Environmental Management, 66 (4), 377-393.
Torrens, P. M., & O'Sullivan, D. (2001). Editorial: Cellular automata and urban
simulation: where do we go from here? Environment and Planning, 28,163-168.
143
Tu, J. (2011). Spatially varying relationships between land use and water quality
across an urbanization gradient explored by geographically weighted regression.
Applied Geography, 31, 376-392.
Turner, B., Lambin, E., & Reenberg, A. (2007). The emergence of land change science
for global environmental change and sustainability. Proceedings o f the National
Academy of Sciences, 104, 20666-20671. PNAS.
United Nations. (1986). Principles relating to remote sensing o f the Earth from space.
United Nations.
USGS. (2007, February 13). Global Geographic Information Systems. Retrieved May 29,
2013. http://webgis.wr.usgs.gov/globalgis/tutorials/what_is_gis.htm
Veldkamp, A., & Lambin, E. F. (2001). Predicting land-use change. Agriculture,
Ecosystems & Environments, 85 (1-3), 1-6.
Verburg, P. H., Eickhout, B., & Meijl, FI. (2008). A multi-scale, multi-model approach
for analyzing the future dynamics of European land use. Annals o f Regional Science,
42 (1), 57-77.
Verburg, P. H., Schot, P. P., Dijst, M. J., & Veldkamp, A. (2004). Land use modelling:
current practice and research priorities. Geojournal, 61, 309-324.
Walsh, S., Messina, J., Mena, C., Malanson, G., & Page, P. (2008). Complexity theory,
spatial simulation models and land use dynamics in the Northern Ecuadorian
Amazon. Geoforum, 39 (2), 867-878.
Weinzettel, J., FFertwich, E., Peters, G., Steen-Olsen, K., & Galli, A. (2013). Affluence
drives the global displacement of land use. Global Environmental Change, 23 (2), 433-
438.
144
Weng, Q. (2002). Land use change analysis in the Zhujiang Delta of China using
satellite remote sensing, GIS and stochastic modelling. Journal o f Environmental
Management, 64, 273-284.
Wilson, C. O., & Weng, Q. (2011). Simulating the impacts of future land use and
climate changes on surface w ater quality in the Des Plaines River watershed,
Chicago Metropolitan Statistical Area, Illinois. Science o f the Total Environment, 409,
4387-4405.
Zhang, Y., Lu, D., Yang, B., Sun, C., & Sun, M. (2011). Coastal wetland vegetation
classification with a Landsat Thematic M apper image. International Journal o f Remote
Sensing, 32 (2), 545-561.
Zubair, A. O. (2006). Change detection in land use and land cover using remote sensing
data and GIS (a case study ofllorin and its environs in Kwara State). Unpublished thesis.
Ibadan, Nigeria: University of Ibadan.
145