Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Detection and Analyses of Land-cover Change: A Case of Two Mindanao Provinces
with History of Forest Resource Utilization
Thesis by
Meriam M. Makinano
B.S. Geodetic Engineering
Submitted to the Graduate Division
College of Engineering
University of the Philippines Diliman
In Partial Fulfillment of the Requirements
For the Degree of Master of Science in
Remote Sensing
College of Engineering
University of the Philippines Diliman
Quezon City
May 2010
ii
This thesis, entitled DETECTION AND ANALYSES OF LAND-COVER
CHANGE: A CASE OF TWO MINDANAO PROVINCES WITH HISTORY OF
FOREST RESOURCE UTILIZATION, prepared and submitted by MERIAM M.
MAKINANO, in partial fulfillment of the requirements for the degree of MASTER OF
SCIENCE IN REMOTE SENSING is hereby accepted.
ENRICO C. PARINGIT, Dr. Eng.
Thesis Adviser
Accepted as partial fulfillment of the requirements for the degree of MASTER
OF SCIENCE IN REMOTE SENSING.
ROWENA CRISTINA L. GUEVARA, Ph.D.
Dean
iii
Acknowledgment
This thesis would not have been possible without the help of so many people. I
am heartily thankful to all of those who supported me in any respect during the
completion of this thesis.
I am sincerely thankful to my adviser, Dr. Enrico C. Paringit, whose
encouragement, guidance and support enabled me to develop a good understanding of my
thesis topic.
I am indebted to the Department of Science and Technology-Philippine Council
for Advanced Science and Research Development (DOST-PCASTRD) through its
Human Resources and Institution Development Division, for the financial support
extended during my graduate studies here in UP Diliman, and for financing this thesis.
I am also very thankful to Dr. Tolentino B. Moya and Prof. Florence A. Galeon
for accepting my invitation to be my thesis defense chairman and member, respectively.
Their insightful comments and suggestions helped me better understand my thesis and
these were very helpful in the improvement of the manuscript.
Acknowledgements are also extended to DENR-Caraga Region XIII and to the
Forest Management Bureau Main Office for providing the datasets needed in the
analysis.
To Dr. Edgar W. Ignacio (former President of the Northern Mindanao State
Institute of Science and Technology, now Caraga State University), thank you sir for
your encouragement for me to pursue graduate studies here in UP Diliman.
My deep appreciation is also extended to the Caraga State University (through our
President, Dr. Joanna B. Cuenca) and also to my CEIT family especially to Engr.
Jonathan M. Tiongson and Engr. Alexander T. Demetillo for all the support given to me.
Thank you also to Engr. Lorie Cris S. Asube for taking care of my responsibilities in the
CEIT while I’m away.
To Engr. Michelle V. Japitana, my ever loving and helpful friend, thank you Kay
for the companionship and for always being there for me.
To all my Research Groupmates at the Applied Geodesy and Space Technology
Laboratory namely: Cecil, Ate Beth, Nimol, Alex O., Rose, Ate Merlie, Mitch and Jene.
Thank you for all your comments and suggestions during my progress reports.
My special thanks to Engr. Alexander S. Caparas and Engr. Jessi Lin P. Ablao for
the friendship and for being so accommodating especially during my initial stays in UP
Diliman.
iv
Thank you also to Ma’am Lynn Serrano and all the staff of the Engineering
Graduate Office for all your assistance, especially during the preparations for my thesis
defense.
Thank you to Tatay, Gigie, Dodong, and most especially to Nanay for
understanding my absence during her time of illness. Thank you for your support and
believing that I can pursue graduate studies in UP Diliman. Kining tanan na akong
pagpaningkamot ay para sa inyo.
To my best friend and boyfriend, Engr. Jojene R. Santillan, thank you gá for
helping me in all aspects of my graduate studies, especially during the conduct of this
thesis. Thank you for being my second adviser, my proof reader, critique, and personal
assistant. Salamat kaayo gá sa tanan nimong pagpalangga sa ako.
And most of all, to our Almighty God for making all these things possible.
v
For My Dearest Mother,
Virgincita M. Makinano.
vi
Abstract
This study presents an integrated approach involving Remote Sensing (RS), Geographic
Information System (GIS) and statistical analysis to detect and analyze 25-year land-
use/land-cover change (LULCC) in the provinces of Agusan del Norte and Agusan del
Sur in Northeastern Mindanao, Philippines with history of forest resource utilization in
the context of limited land-cover information due to cloud contamination of RS images.
Using cloud and shadow masking algorithm and state-of-the-art RS image analysis
techniques provided by the Support Vector Machine classifier, highly accurate land-cover
maps were obtained from Landsat Multi-Spectral Scanner (MSS) and Enhanced Thematic
Mapper + (ETM+) images and used to detect land-cover transitions in the study area
from 1976-2001. The differences in deforestation and other land-cover change types in
the two provinces were then characterized and compared using GIS-based spatial analysis
techniques. The significance and magnitude of the relationship between the detected
deforestation and various georeferenced socio-economic and bio-physical factors were
determined through logistic regression analysis. Major results showed that the detected
changes in land-cover were found to be different in the Agusan provinces. Forest to
rangeland is the major land-cover change in Agusan del Norte from 1976 to 2001; in
Agusan del Sur, the two most prominent land-cover change types are the conversions of
rangeland to forest and of forest to palm trees. The results of GIS-based characterization
of deforestation and logistic regression analysis based on combined bio-physical and
socio-economic factors provided significant results as to what factors were associated
with deforestation in the Agusan provinces. For Agusan del Norte, the bio-physical
factors DISTRIV (distance to rivers) and ELEV (elevation) were found to be the most
positively and negatively related to deforestation, respectively. For Agusan del Sur,
DISTNEWBUILT (distance to new built-up areas) and ELEV are found to be the most
positively and negatively related to deforestation, respectively. With the identification of
the factors associated with deforestation, this study has provided a first step in controlling
forest loss which is very useful in comprehensive forest management planning and in
formulation of appropriate forest policy. This study is a significant contribution to
LULCC research by providing a series of techniques to understand deforestation and
relate it to bio-physical and socio-economic factors using an un-ideal dataset. An
important finding of this study is that it is possible to analyze deforestation using cloud
contaminated RS images. Local agencies in the Agusan provinces may use the land-cover
maps and statistics obtained in this study to further evaluate the process of deforestation
in these provinces in order to create and evaluate strategies that attempt to mitigate its
negative effects.
vii
Table of Contents
Acknowledgment ............................................................................................................... iii
Abstract .............................................................................................................................. vi
Table of Contents .............................................................................................................. vii
List of Figures .................................................................................................................... ix
List of Tables .................................................................................................................... xii
List of Abbreviations ....................................................................................................... xiv
Chapter 1. Introduction ........................................................................................................1
1.1 Background of the study ........................................................................................... 1
1.2 Objectives of the study.............................................................................................. 5
1.3 Research significance ................................................................................................ 5
Chapter 2. Review of Related Literature .............................................................................7
2.1 Drivers of land-use/land-cover change ..................................................................... 7
2.2 Deforestation and land-cover change in the Philippines......................................... 12
2.3 Land-cover change detection .................................................................................. 17
2.3.1 Review of RS change detection techniques ..................................................... 17
2.3.2 Post-classification change detection: review of classification methods .......... 19
2.3.3 Classification by Support Vector Machine ...................................................... 21
2.4 GIS in LULCC studies ............................................................................................ 23
2.5 Review of statistical methods in LULCC studies ................................................... 26
Chapter 3. The Study Area.................................................................................................34
3.1 Background ............................................................................................................. 34
3.2 The Province of Agusan del Norte.......................................................................... 34
3.3 The Province of Agusan del Sur ............................................................................. 36
3.4 Status of Forest Resources in the Agusan Provinces .............................................. 38
3.5 Forest License Agreements Issued in the Agusan Provinces.................................. 40
Chapter 4. Methodology ....................................................................................................45
4.1 Overview ................................................................................................................. 45
4.2 Remote sensing image analysis .............................................................................. 47
4.2.1 Landsat images................................................................................................. 47
4.2.2 Image geometric accuracy assessment............................................................. 50
4.2.3 Image pre-processing ....................................................................................... 54
4.2.4 Cloud and shadow masking ............................................................................. 58
4.2.5 Image classification and accuracy assessment ................................................. 59
4.2.6 Post-classification change detection ................................................................ 66
4.3 GIS spatial change analysis .................................................................................... 67
4.4 Statistical analysis of land-cover change ................................................................ 70
Chapter 5. Results and Discussions ...................................................................................73
5.1 Land-cover maps ..................................................................................................... 73
5.1.1 The 1976 land-cover map ................................................................................ 73
5.1.2 Accuracy of the 1976 land-cover map ............................................................. 79
viii
5.1.3 The 2001 land-cover map and accuracy .......................................................... 80
5.2 Land-cover change in the Agusan Provinces .......................................................... 85
5.3 Deforestation in the Agusan Provinces ................................................................... 92
5.4 Characterizing 25-year deforestation in the Agusan Provinces .............................. 95
5.5 Logistic regression analysis results ....................................................................... 104
5.5.1 Logistic regression based on bio-physical factors only ................................. 105
5.5.2 Logistic regression based on socio-economic factors only............................ 108
5.5.3 Logistic regression using combined socio-economic and bio-physical factors
................................................................................................................................. 112
5.5.4 Logistic regression analysis using new set of 5% sample ............................. 114
5.6 Characterization of “No Data” pixels ................................................................... 119
5.7 Summary of findings............................................................................................. 120
Chapter 6. Conclusions and Recommendations ...............................................................125
6.1 Conclusions ........................................................................................................... 125
6.2 Recommendations ................................................................................................. 127
References ........................................................................................................................128
Appendices .......................................................................................................................135
Appendix 1. Maps showing the location of retained and deforested areas. ................ 136
Appendix 2. Factor Maps ............................................................................................ 137
Appendix 3. Maps showing the location of retained forest and deforested areas with
CBFMA, CBRM, TLA, and IFMA. ........................................................................... 141
Appendix 4. Maps showing the distance to new roads of retained forest and deforested
areas. ........................................................................................................................... 142
Appendix 5. Maps showing the distance to river of retained forest and deforested areas.
..................................................................................................................................... 143
ix
List of Figures
Figure 1. Map showing the provinces of Agusan del Norte and Agusan del Sur. ...............3
Figure 2. Drivers of tropical deforestation [22],[10]. ..........................................................8
Figure 3. Deforestation trend in the Philippines from 1903-2001 [19]. ............................15
Figure 4. Map showing the municipalities and cities in the Agusan provinces .................35
Figure 5. Log Production of Agusan del Norte and Agusan del Sur from 1984-2001 ......38
Figure 6. Graph showing the timber processing plants and sawmills in the Agusan
Provinces .............................................................................................................39
Figure 7. Map showing the location of TLAs and IFMAs issued in the Agusan provinces.43
Figure 8. Map showing the location of CBFMAs and CBRMs issued in Agusan
Provinces. ............................................................................................................44
Figure 9. The three phases of the study’s methodology. ...................................................46
Figure 10. Process flow diagram of remotely-sensed image analysis ...............................47
Figure 11. The two Landsat images of the study area that were subjected to image
analysis to derive land-covers maps for the years 1976 and 2001. .....................49
Figure 12. Location of points used to determine the geometric accuracy of the 2001
Landsat image and the resulting RMSE vectors of the comparisons with
NAMRIA maps. The numerical values and the lines indicate the magnitude and
direction of the differences in coordinates (local RMSE), with the arrows
pointing to the “actual” (i.e., NAMRIA map) coordinates. .................................52
Figure 13. Location of points used to determine the geometric accuracy of the 1976
Landsat image and its co-registration with the 2001 Landsat image. Also shown
are the resulting RMSE vectors of the comparisons. The numerical values and
the lines indicate the magnitude and direction of the differences in coordinates,
with the arrows pointing to the “actual” (i.e., 2001 Landsat) coordinates. .........53
Figure 14. Flowchart of the simple cloud and shadow detection and masking technique
developed and applied in this study. ....................................................................60
Figure 15. The 1976 land-cover map of Agusan del Norte and Agusan del Sur resulting
from the classification of the April 17, 1976 Landsat MSS image using SVM.
All white areas within the provincial boundaries classified as “No Data” are
clouds and shadow pixels in the image. ..............................................................74
x
Figure 16. Bar charts showing the Producer’s (a) and User’s (b) Accuracies of land-cover
types in three SVM-classified land-cover maps for 1976. ..................................78
Figure 17. The 2001 land-cover map of Agusan del Norte and Agusan del Sur resulting
from the classification of the May 22, 2001 Landsat ETM+ image using SVM.
All white areas within the provincial boundaries classified as “No Data” are
clouds and shadow pixels in the image. ..............................................................81
Figure 18. Bar charts showing the Producer’s (a) and User’s (b) Accuracies of land-cover
types in the three land-cover maps. .....................................................................84
Figure 19. The 1976-2001 land-cover maps of Agusan del Norte province. Areas with
data comprise 66.98% (or 2044.67sq.km.) of the total land area of Agusan del
Norte. ...................................................................................................................86
Figure 20. The 1976-2001 land-cover maps of Agusan del Sur province. Areas with data
comprise 51.10% (or 4,133.82 sq. km.) of the total land area of Agusan del Sur.87
Figure 21. Land-cover change in Agusan del Norte province from 1976-2001 for cloud
free areas only. Upper and lower error bars represent errors of omission and
commission, respectively, of the land-cover classifications. ..............................88
Figure 22. Top 10 land-cover change types in Agusan del Norte province from 1976-
2001 for cloud-free areas only. Upper and lower error bars represent errors of
omission and commission, respectively, of the land-cover classifications .........89
Figure 23. Land-cover change in Agusan del Sur province from 1976-2001 for cloud-free
areas only. Upper and lower error bars represent errors of omission and
commission, respectively, of the land-cover classifications. ..............................90
Figure 24. Top 10 land-cover change types in Agusan del Sur province from 1976-2001
for cloud-free areas only. Upper and lower error bars represent errors of
omission and commission, respectively, of the land-cover classifications. ........91
Figure 25. Comparison of magnitude of forest cover area reduction by types of change. 93
Figure 26. Mean elevation of location of forest cover occurrences in Agusan del Norte
and Agusan del Sur. Error bars indicate 95% confidence interval of the mean. .96
Figure 27. Mean SLOPE of location of forest cover occurrences in Agusan del Norte and
Agusan del Sur. Error bars indicate 95% confidence interval of the mean. ........97
Figure 28. Mean DISTRIV of location of forest cover occurrences in Agusan del Norte
and Agusan del Sur. Error bars indicate 95% confidence interval of the mean. .98
Figure 29. Mean DISTNEWBUILT of location of forest cover occurrences in Agusan del
Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the
mean. ...................................................................................................................99
Figure 30. Mean DISTNEWRD of location of forest cover occurrences in Agusan del
Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the
mean. .................................................................................................................100
xi
Figure 31. Mean DIST_TLA-IFMA of location of forest cover occurrences in Agusan del
Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the
mean. .................................................................................................................101
Figure 32. Mean DIST_CBFMA-CBRM of location of forest cover occurrences in
Agusan del Norte and Agusan del Sur. Error bars indicate 95% confidence
interval of the mean. ..........................................................................................103
Figure 33. Mean POPDENCHANGE of location of forest cover occurrences in Agusan
del Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the
mean. .................................................................................................................104
Figure 34. Diagram for interpreting the logistic regression coefficients. ........................105
Figure 35. Graph showing β values indicating the magnitude of association of bio-
physical factors with deforestation. Error bar indicate standard error. .............106
Figure 36. Graph showing the magnitude of association of socio-economic factors with
deforestation. Error bars indicate +/- standard error. ........................................109
Figure 37. Graph showing the magnitude of association of the combined bio-physical and
socio-economic factors with deforestation. .......................................................113
Figure 38. Graph showing the comparison between the original and the new 5% samples
in Agusan del Norte. Error bars indicate +/- standard error. .............................116
Figure 39. Graph showing the comparison between the original and the new 5% samples
in Agusan del Sur. Error bars indicate +/- standard error. .................................118
Figure 40. Graph showing the mean factor values of no data and with data pixels for
Agusan del Norte and Agusan del Sur. Error bars indicate +/- standard error. .120
xii
List of Tables
Table 1. Drivers of tropical deforestation presented by Geist and Lambin. ........................8
Table 2. List of Timber License Agreements (TLAs) issued in Agusan del Norte and
Agusan del Sur with date of TLA issuance and expiry, and area covered.
(Source: Yearly Forestry Statistics, DENR-FMB). .............................................42
Table 3. Characteristics of the Landsat images used in the study. ....................................48
Table 4. Values used for the calibration of the Landsat MSS image to radiance. .............55
Table 5. Values used for the calibration of the Landsat ETM+ image to radiance. ..........55
Table 6. Landsat MSS mean solar exoatmospheric spectral irradiances [87]. ..................56
Table 7. Landsat ETM+ mean solar exoatmospheric spectral irradiances [86]. ................57
Table 8. Values used for the computation of the surface reflectance. ...............................57
Table 9. Definitions of land-cover types used in this study. ..............................................61
Table 10. Image keys used in visual interpretations of the 1976 Landsat MSS image. ....62
Table 11. Image keys used in visual interpretations of the 2001 Landsat ETM+ image. ..63
Table 12. Number of pixels collected for image classifications and accuracy assessments.64
Table 13. Various combinations of input bands used in image classification ...................65
Table 14. Definitions of georeferenced bio-physical and socio-economic factors. ...........67
Table 15. The 5% samples used in logistic regression analysis. .......................................71
Table 16. Matrix of percent overall classification accuracies of 32 classified images (from
various band combinations of the1976 Landsat MSS image and image by-
products (Ground truth pixels = 2, 276) ..............................................................75
Table 17. Error matrix of the SVM-classified Landsat MSS reflectance bands with NDVI
and DEM (the source of the 1976 land-cover map of the study area). ................76
Table 18. Error matrix of the SVM-classified Landsat MSS reflectance bands with DEM.77
Table 19. Error matrix of the SVM-classified Landsat MSS reflectance bands with
simulated Red and Green bands and DEM. .........................................................77
Table 20. Summary of Producer’s and User’s Accuracies of 1976 land-cover types in
three SVM-classified land-cover maps. ..............................................................79
Table 21. Matrix of percent overall classification accuracies of 8 classified images from
various band combinations of the 2001 Landsat ETM+ image and DEM.
(Ground truth pixels= 6,581). ..............................................................................82
xiii
Table 22. Error matrix of the SVM-classified Landsat ETM+ reflectance bands with
normalized temperature and DEM (the source of the 2001 land-cover map of the
study area). ..........................................................................................................83
Table 23. Error matrix of the SVM-classified Landsat ETM+ reflectance bands with
temperature band (normalized from 0 to 1). ........................................................83
Table 24. Error matrix of the Maximum likelihood-classified Landsat ETM+ reflectance
bands with temperature band (normalized form 0 to 1) and DEM (also
normalized from 0 to 1) .......................................................................................84
Table 25. Summary of the Producer’s and User’s Accuracies of land-cover types in three
derived land-cover maps. .....................................................................................85
Table 26. Forest cover change statistic (1976-2001) in the Agusan Provinces. ................93
Table 27. Binary logistic regression of FCOVER versus bio-physical factors for Agusan
del Norte and Agusan del Sur ............................................................................106
Table 28. Binary logistic regression of FCOVER versus socio-economic factors for
Agusan del Norte and Agusan del Sur ..............................................................109
Table 29. Binary logistic regression of FCOVER versus the combined bio-physical and
socio-economic factors for Agusan del Norte and Agusan del Sur ...................112
Table 30. Comparison between the β values for Agusan del Norte ................................115
Table 31. t-Test results for Agusan del Norte ..................................................................116
Table 32. Comparison between the β values for Agusan del Sur ...................................117
Table 33. t-Test results for Agusan del Sur .....................................................................118
xiv
List of Abbreviations
ADN Agusan del Norte
ADS Agusan del Sur
ANN Artificial Neural Network
CBFMA Community-Based Forest Management Agreement
CBRM Community-Based Resource Management
DENR Department of Environment and Natural Resources
DT Decision Tree
ETM+ Enhanced Thematic Mapper Plus
FMB Forest Management Bureau
GIS Geographic Information System
IFMA Integrated Forest Management Agreement
ITP Industrial Tree Plantation
LULCC Land-use/Land-cover Change
MLC Maximum Likelihood Classifier
MSS Multi-spectral Scanner
RBF Radial Basis Function
RS Remote Sensing
SVM Support Vector Machine
TLA Timber License Agreement
1
Chapter 1
Introduction
1.1 Background of the study
Understanding the drivers of land-use/land-cover change (LULCC) is a complex
issue and presently remains to be a very active area of research. LULCCs are the result of
the interplay between socio-economic, institutional and environmental factors [1]. The
causes attributed to LULCC are considered multivariate in nature, interrelated and differ
at local, regional as well as national scale and can be summed up as complex socio-
economic processes such that it is impossible to isolate a single cause [2]. It is because of
these complexities that questions of LULCC have constantly attracted interests among a
wide variety of researchers concerned with understanding the causes and consequences of
these changes [3]. Studying the dynamics of land-cover change is essential because it
could generate primary data on the location, type, and rate of land development and, in
turn, provide a basis for analyzing the impacts of these dynamics not only on socio-
economic processes but also on such environmental processes as energy flux, runoff,
erosion, air and water quality, and biodiversity.
In northeastern Mindanao, Philippines, the provinces of Agusan del Norte and
Agusan del Sur (Figure 1) have been widely known for its rich forest resource; hence,
2
making them the major timber producers in the whole country since the 1950s up to the
present. In fact, the two provinces belong to the so-called “Eastern Mindanao Corridor”
where 75 percent of the country’s timber extraction comes from [4]. The two provinces
have utilized their forest resources extensively resulting from the establishment of
logging and timber industries way back in the 1950s [5] that continue to operate until
this time by way of forest license agreements issued by the Philippine government to
private corporations and non-government organizations. These industries have
contributed greatly to the economy of both provinces and to the Philippines as a whole
[6]; however, they are often blamed for decades of rampant upland forest destruction and
significant changes in land-cover whose ecological aftermath continues to unfold in the
valleys below. In fact, logging in the primary watersheds of the two provinces between
the 1950s and 1970s has resulted in massive upland erosion and lowland siltation,
combined with rapid runoff and flooding [7]. In 1981, for example, heavy rains spilling
into the Agusan River were blocked by huge silt deposits near the mouth of the river,
causing a series of floods which killed hundreds and left thousands homeless [7],[8].
Recently, the same environmental impacts of deforestation and land-cover change are
still a common problem that environmentalists, watershed planners, and policy makers
face today in the Agusan Provinces [9].
3
Figure 1. Map showing the provinces of Agusan del Norte and Agusan del Sur.
While the logging industries may have direct connection to deforestation and
other types of land-cover changes in the Agusan provinces, the contributions of other
equally relevant factors associated with deforestation such as agricultural expansion,
wood extraction, expansion of infrastructure, population growth, economic and
4
technological factors, policy/institutional factor, land characteristics, bio-physical
environment, and government policy failures, among others [10] maybe overlooked.
Hence, there arises a necessity to ascertain what were the factors associated with
deforestation in these two provinces.
The roles of Remote Sensing (RS) and Geographic Information Systems (GIS)
have become significant recently in LULCC researches (e.g., [11-17]). Imageries from
satellite RS platforms provides valuable sources of land-cover and other information
related to topography, and surface conditions especially in areas which are difficult to
monitor and could be very expensive when using conventional techniques [11]. Despite
the high regard accorded to RS and GIS in LULCC studies, the studies of land-cover
change are hampered by lack of good RS images due to the presence of clouds and cloud
shadows, especially in tropical countries like the Philippines. Hence, the use of medium
resolution optical RS images (e.g., those provided by the Landsat satellite) for land-cover
change detection are often limited because of the presence of clouds and shadows that
prevents the derivation of land-cover characteristics from the images. The utilization of
RS and GIS technologies to understand the process of land-cover change especially
deforestation at a finer scale are limited in these areas. Furthermore, none of numerous
studies attempted to consider and take into account the case when RS images used for
deriving land-cover change are contaminated with clouds and cloud shadows. The use of
radar RS images that overcome weather obstacles may be a solution but the availability
of such images and their long-term and multi-temporal capabilities are often inadequate
for studies that requires immediate images.
5
1.2 Objectives of the study
This study is an attempt to detect and analyze deforestation and ascertain what
were the factors associated to it in an area with a history of forest resource utilization (the
Agusan Provinces) in the context of limited land-cover information due to cloud
contamination of RS images. Using an integrated approach involving Remote Sensing
(RS), Geographic Information System (GIS) and statistical analysis, 25-year land-cover
change in the Agusan Provinces was detected and analyzed. Specifically, this involved:
1. Detecting deforestation and other types of LULCC in the two provinces
through analysis of Landsat MSS and ETM+ images;
2. Characterizing and comparing the differences in deforestation in the two
provinces using GIS-based spatial analysis techniques; and
3. Determining, through logistic regression analysis, the significance and
magnitude of the relationship between the detected deforestation and
georeferenced socio-economic and bio-physical factors such as presence of
logging and timber industries, population growth, road infrastructures,
elevation, slope, soil quality and proximity to water resources.
1.3 Research significance
This study provides an integrated RS-GIS-Statistical Analysis approach in
understanding as to which factors were associated with deforestation in the Agusan
provinces. From a socioeconomic perspective, studying deforestation and other types of
6
land-cover change in the Agusan Provinces is important because it provides data that may
be used to explore relationships with potential causal mechanisms, thereby increasing our
understanding of the development process. Conversely, analyzing LULCC and
identifying its major drivers are important from a planning perspective because they
provide a means to create and evaluate strategies that attempt to mitigate its negative
effects [18].
Deforestation, a widely recognized problem in the study area [7],[9], is the major
reason behind flooding, acute water shortages, rapid soil erosion, siltation, and mudslides
that have proved to be costly not only to the environment and properties but also in
human lives [19]. In this context, identification of factors contributing to deforestation,
among other LULCC in the Agusan Provinces, is a first step in controlling forest loss
[20] and is necessary in comprehensive forest management planning and formulation of
appropriate forest policy [12]. Furthermore, results of this study can be utilized as a
temporal LULCC model for the provinces of Agusan del Norte and Agusan del Sur that
can help in quantifying the extent and nature of change and aid planning agencies in
developing sound and sustainable land-use practices.
7
Chapter 2
Review of Related Literature
This chapter presents a review of literatures relevant to the nature and scope of the
study. The review aims to provide a clearer understanding on the different processes and
drivers with LULCC. Studies on the LULCC in the Philippines are also presented. The
state of the art of the detection and analysis of the drivers of land-use/land-cover change
through RS, GIS and statistical analysis are discusses as well.
2.1 Drivers of land-use/land-cover change
Understanding the drivers of LULCC is a complex issue and presently remains to
be a very active area of research. Lesschen et al. [1] reported that LULCC are the result
of the interplay between socio-economic, institutional and environmental factors. The
most common form of LULCC is deforestation. This is probably due to the already
established knowledge that the process of deforestation is a first step in LULCC [21].
Geist & Lambin [22] presented a grouping of the drivers of tropical deforestation
(Table 1 and Figure 2). These are a complex set of actions and factors involved in
deforestation.
8
Table 1. Drivers of tropical deforestation presented by Geist and Lambin.
Cluster Major examples
Proximate causes Agricultural expansion
Wood extraction
Expansion of infrastructure
Underlying causes Demographic (population growth)
factors
Economic factor
Technological factor
Policy/institutional factor
Cultural or socio-political factors
Other factors (land characteristics, bio-
physical drivers and social trigger events)
Land characteristics
Bio-physical environment
Health and economic crisis
Government policy failures
Figure 2. Drivers of tropical deforestation [22],[10].
Proximate causes of deforestation are human activities at the local level, that
originate from intended land-use and that have direct impact on forest cover [22].
Examples of such causes are agricultural expansion, wood extraction and infrastructure
expansion. Underlying driving factors are fundamental social processes associated with
9
deforestation, such as human population dynamics or agricultural policies that underpin
the proximate causes, and which either operate at the local level or have indirect impacts
that are felt at the local level (e.g. national or global policies). These factors include: (1)
demographic, (2) economic, (3) technological, (4) policy and institutional and (5)
cultural.
The Other factors are defined as those factors that can also play an important role
in driving deforestation; these factors include pre-disposing environmental factors (e.g.,
land characteristics, including soil quality and topography), bio-physical drivers or
triggers (fires, droughts, floods and pest outbreaks) and social trigger events (e.g.
revolution, social disorder and economic shocks) [22].
The conceptual framework developed by Geist & Lambin [22] as presented in
Figure 2 is based on the analysis of 152 case studies of tropical forest cover loss in Asia.
According to Verbist et al. [21], this framework is probably the most comprehensive in
identifying which factors drives tropical forest decline but he asserted that it needs to be
mentioned that in most of these studies, deforestation has been regarded as a unilinear
process, whereby little or no attention has been given either to the land-cover types that
were replacing the forests, or to the factors driving that replacement.
Lambin et al. [23] highlighted the complexity of land-use/cover by stating that
land-cover changes do not always occur in a progressive and gradual way, but they may
show periods of rapid and abrupt change followed either by a quick recovery of
ecosystems or by a non-equilibrium trajectory. Such short-term changes are often caused
by the interaction of climatic and land-use factors (for example, periodic El Niño-driven
droughts lead to an increase in the forest’s susceptibility to fires).
10
Lambin et al. [23]’s study further indicates that slow and localized land-cover
conversion takes place against a background of high temporal frequency regional-scale
fluctuations in land-cover conditions caused by climatic variability, and it is often linked
through positive feedback with land-cover modifications. These multiple spatial and
temporal scales of change, with interactions between climate-driven and anthropogenic
changes, are a significant source of complexity in the assessment of land-cover changes.
Lambin et al. assessed that it is not surprising that the land-cover changes for which the
best data exist—deforestation, changes in the extent of cultivated lands, and
urbanization—are processes of conversion that are not strongly affected by inter-annual
climatic variability. By contrast, few quantitative data exist at the global scale for
processes of land-cover modification that are heavily influenced by inter-annual climatic
fluctuations, e.g., desertification, forest degradation and rangeland modifications.
The roles that the proximate and underlying factors play in the complex dynamics
of LULCC are described by Lambin et al. [23] as follows. In general, proximate causes
operate at the local level (e.g., individual farms, households, or communities). By
contrast, underlying causes may originate from the regional (districts, provinces, or
country) or even global levels, with complex interplays between levels of organization.
Underlying causes are often exogenous to the local communities managing land and are
thus uncontrollable by these communities. Only some local-scale factors are endogenous
to decision makers. An important system property associated with changes in land-use is
feedback that can either accentuate or amplify the speed, intensity, or mode of land
change, or constitute human mitigating forces, for example via institutional actions that
dampen, impede, or counteract factors or their impacts. Examples are the direct
11
regulation of access to land resources, market adjustments, or informal social regulations
(e.g., shared norms and values that give rise to shared land management practices).
According to Lambin et al. [23], land-use change is always caused by multiple
interacting factors originating from different levels of organization of the coupled human-
environment systems. Changes are generally driven by a combination of factors that work
gradually and factors that happen intermittently [24]. The mix of driving forces of land-
use change varies in time and space, according to specific human-environment
conditions. Driving forces can be slow variables, with long turnover times, which
determine the boundaries of sustainability and collectively govern the land use trajectory
(such as the spread of salinity in irrigation schemes or declining infant mortality), or fast
variables, with short turnover times (such as food aid or climatic variability associated
with El Niño oscillation) [23]. Summarizing a large number of case studies, Lambin et al.
[23] concluded that land-use change is driven by a combination of the following
fundamental high-level causes:
1. resource scarcity leading to an increase in the pressure of production on
resources,
2. changing opportunities created by markets,
3. outside policy intervention,
4. loss of adaptive capacity and increased vulnerability, and
5. changes in social organization, in resource access, and in attitudes.
Lambin et al. [23] explained that some of these fundamental causes are
experienced as constraints. They force local land managers into degradation, innovation,
or displacement pathways. The other causes are associated with the seizure of new
12
opportunities by land managers who seek to realize their diverse aspirations. Each of
these high-level causes can apply as slow evolutionary processes that change
incrementally at the timescale of decades or more, or as fast changes that are abrupt and
occur as perturbations that affect human-environment systems suddenly. Only a
combination of several causes, with synergetic interactions, is likely to drive a region into
a critical trajectory. Lambin et al. explained further that some of the fundamental causes
leading to land-use change are mostly endogenous, such as resource scarcity, increased
vulnerability and changes in social organization, even though they may be influenced by
exogenous factors as well.
2.2 Deforestation and land-cover change in the Philippines
In the Philippines, deforestation and forest degradation are the most important
land-use change processes [25],[26]. These processes are an important threat to the highly
rated biodiversity of the country. Only a small fraction of the natural forest that once
covered the country remains.
It has been reported in the vast literature of Philippine LULCC that the country
was 90% forested when the Spaniards conquered the islands in the middle of the
sixteenth century, decreasing to 70% by 1900 and approximately 23% by 1987 [25],[27-
29]. The establishment of plantations of export crops led to deforestation in the
nineteenth century while unrestricted forest harvesting caused enormous losses in the
post-war years. Of almost 15 million hectares of natural dipterocarp forest in 1950 only 4
million remained in 1992. A large part of these 4 million hectares is heavily logged-over
13
forest of varying quality [29],[25]. In the late 1970s the emphasis began to shift from
timber harvesting and utilization to the protection, rehabilitation, and development of
forestlands. Log production steadily decreased through prescribed annual allowable cuts
for each logging concession. From 1992 onward, logging became officially prohibited in
virgin forests, in areas over 1000 meters in elevation and in areas with slopes of 50% and
above. The conservation impact of this order was limited because by the time it was
issued, only a small portion of the Philippines’ remaining natural forest had not yet been
logged over. The log ban in virgin forests did not mean the end of corporate logging: it
simply led companies to transfer their attention to the secondary forests. In the mid 1990s
a political discussion was held concerning the implementation of a total log ban. This
total log ban was never implemented, but policies reducing logging in fragile areas have
become stronger by the years [27],[29],[25]. In spite of different policies that aim to
reduce logging recent commercial deforestation, illegal logging and agricultural
expansion pose an important threat to the remaining forest areas in the Philippines [26].
For the past four decades, a number of studies have been conducted to detect and
analyze LULCC in the Philippines. In particular, Kummer [30] presented a model of
deforestation in which logging and agriculture (both shifting and permanent) have been
identified as the two main agents of forest destruction for the post-war (late 1940’s to
1986) Philippines. He postulated that logging is primary responsible for converting the
primary forest to secondary forest and that agriculture activities then convert the
secondary forest to farmland. His postulation attributed conversion of tropical moist
forest to logger and forest farmer’s interaction, i.e. loggers log the forest and then leave,
while the farmers follow logging roads to new accessible forest areas for cultivation.
14
Kummer tested his model of deforestation using multiple regression cross-sectional
analysis, panel analysis, and path analysis. The results of these statistical analyses
indicate that absolute forest cover is negatively related to road and population density but
there is a positive relation between the actual deforestation from 1970 to 1980 and the
forest area in 1970, distance from Manila, change in agricultural area and logging quotas
in 1970. An important conclusion of Kummer's research is that studies of deforestation
which uses percentage forest cover as the dependent variable are of limited importance in
depicting the process of deforestation. Kummer further stated that deforestation in the
post-war Philippines is the result of the failure of the Philippine economy to provide jobs
and elite control of government which has concentrated the financial returns from logging
in the hands of concessionaires and their allies which means that deforestation in the
Philippines is amenable to policy intervention.
The findings of Kummer [30] was supported by the study conducted by Verburg
et al [26] in which he reported that land-cover change in the Philippines between 1970
and the early 1990s are generally caused by large-scale logging of the forest areas
followed by agriculture. This process was accompanied by road construction for logging
and non-logging purposes and by both internal population growth and migration. Logging
opened up the forests both by constructing roads into the forests and, at the same time, by
removing large amount of timber, facilitating the clearing of the remaining degraded
forests by subsistence migrant farmers.
Moya & Malayang III [19] presented a very good background on the rate and
extent of deforestation in the Philippine which was believed to have been covered
partially, if not wholly covered with forest vegetation at the start of 20th
century
15
amounting to 21 million hectares of forest or about 70% of the national land area of the
Philippines in 1903. But in 2001, only 5.1 million ha of forest cover remain intact which
is only 17% of the national total land area. Figure 3 shows the deforestation trend of the
Philippine forest between 1903 and 2001 according to Moya & Malayang III.
Figure 3. Deforestation trend in the Philippines from 1903-2001 [19].
Moya & Malayang III [19] further stated that the conversion of forest into
croplands has been the leading cause of deforestation in the tropics, especially in the
Philippines. Population growth, inequitable land distribution, and the expansion of export
agriculture have reduced cropland available for subsistence farming, forcing many
farmers to clear virgin forest to grow food.
The interference of the Philippine governments’ policy into deforestation cited by
Kummer [30] has been supported by the result of the study conducted by Moya &
Malayang III [19]. They stated that the main culprits for continued deforestation in the
Philippines are the unchecked illegal logging and government’s negligence to combat the
same resulting to devastation of the forests.
16
In consonance with Moya & Malayang III [19]’s study, Verburg & Veldkamp
[25] reported that agricultural expansion is the main cause of further degradation of the
Philippine forests. Citing earlier reports [31-34], Verburg & Veldkamp [25] stated that
the Philippines is an example of unchecked agricultural expansion in uplands, within a
policy setting that encourages it. The area devoted to upland agriculture in the Philippines
increased six-fold between 1960 and 1987, and coincided with a rapid decline in forest
cover. According to the authors, the main reasons for this enormous expansion in upland
agriculture are population growth, inadequate labor absorption and agricultural price
policies. The high rates of forest clearing in the uplands are driven, in part, by the efforts
of low-income farmers to secure subsistence [34]. The policy bias (through price and
technology policies) in favor of crops, such as corn and temperate vegetables, whose
cultivation is most strongly associated with upland agricultural lands, is another cause of
forest frontier expansion [25],[31].
Apan & Peterson [12] probed tropical deforestation in two municipalities (Abra
de Ilog and Mamburao) of Mindoro Occidental, a province south of Manila. Licensed
logging in the area began in the late 1960s and ended in 1983. In 1978, there were about
40 pasture lease agreement holders covering some 23,825 ha. The authors aimed to (1)
determine the significance and magnitude of the relationship between forest cover and
some georeferenced environmental factors (such as population, land-use, land-ownership,
geology, soil depth, soil fertility, distance from water resources, distance from road,
aspect, elevation and slope), (2) characterize and analyze the deforested lands using GIS-
based spatial analysis techniques, and (3) gain insights as to the causes of this
deforestation. The results of their statistical analysis using Pearson chi-square test
17
indicated that all the georeferenced environmental factors are significantly related to
forest cover. However, additional testing using Cramer’s V revealed that magnitude of
relationship for all variable ranges from weak to very weak. Those variables with very
low magnitude of association (almost no relation between factors) with forest cover
include population, distance to water and distance to road. One of the appealing results
that Apan and Peterson found was that accessibility factors (i.e., distance from road and
distance from water) were very weakly associated with forest cover. Deforested lands are
significantly present in both accessible and inaccessible areas (arbitrarily, > 4 km for
roads and > 1 km for rivers/creeks).
2.3 Land-cover change detection
2.3.1 Review of RS change detection techniques
Imageries from satellite RS platforms provides valuable sources of land-cover and
other information related to topography, and surface conditions especially in areas which
are difficult to monitor and could be very expensive when using conventional techniques
(e.g., ground-based mapping and aerial photography) [11]. In land-cover change
detection and analysis, one of the most interesting applications of RS concerns the
analysis of multi-temporal images for detecting land-cover changes [35]. This process
involves the comparison of two co-registered images acquired in the same geographical
area at two different times. In the vast literature on digital change detection, two main
approaches to the change-detection problem have been adopted for RS images: the pre-
18
classification approach and the post-classification comparison approach [36]. The former
is based on existing classification methods, which require the availability of a multi-
temporal ground-truth. The latter performs change detection by making a direct
comparison of the two multispectral images considered, without relying on any additional
information [35].
In general, pre-classification change detection techniques apply various
algorithms to multiple dates of satellite imagery to generate “change” vs. “no-change”
maps [36]. They are sets of image enhancement procedures where mathematical
combinations of satellite imagery from different dates are involved such as univariate
image differencing, image ratioing, image regression or principal components
transformation [37]. Thresholds are applied to the enhanced image to isolate the pixels
that have changed. These techniques locate changes but do not provide information on
the nature of change [38],[39],[37].
Post-classification comparison methods use separate classifications of images
acquired at different times to produce difference maps from which ‘‘from–to’’ change
information can be generated [40]. The objective of post classification change detection
is to achieve the best possible independent classification for each data set and then assess
any change as accurately as the data allow [13]. Post-classification approach exhibits
some important advantages over the pre-classification approach because of its capability
of explicitly recognizing the kind of land-cover transitions which occurred in the
investigated area and its ability to process multisensor/multisource images [41]. The post-
classification comparison approach also compensates for variation in atmospheric
conditions and vegetation phenology between dates since each classification is
19
independently produced and mapped [14],[37],[42],[41]. Factors that limit the application
of post-classification change detection techniques can include cost, consistency, and error
propagation [39]. One of the major requirements of using this approach is the availability
of ground truth information for the individual classification of the images taken at
different times [35]. As the land-cover transitions are usually detected by comparing the
thematic maps obtained by classifying independently the two considered images, the
accuracy yielded strongly depends on the errors present in the classification maps. For
this reason, in the context of the detection of land-cover transitions, it is of great
importance to develop effective classification approaches capable of achieving
classification accuracies as high as possible.
2.3.2 Post-classification change detection: review of classification
methods
The usual flow of analysis in detecting land-cover change using the post-
classification approach involves applying traditional supervised classification algorithms
such as the Maximum Likelihood Classifier [43] to each image (e.g., date 1 and date 2
images) in order to categorize each pixel in the image to a particular land-cover type. The
two-independently classified images are then compared pixel-by-pixel to determine the
type of change [14]. While the use of traditional classifiers, especially Maximum
Likelihood, has been effective in a number of post-classification comparison change
detection studies e.g., [44],[15],[45],[46], a major problem with it is the errors attributed
to misclassification caused by similarities in spectral responses of certain land-cover
classes [42]. Another limitation of maximum likelihood is its assumption of normal
20
distribution of class signatures. In some cases, the number of training samples to obtain
class signatures is actually limited and may not have normal distributions [47], which
make the Maximum Likelihood classifier can not get ideal result [48]. Some studies
addressed these problems by using a hybrid supervised–unsupervised training approach
with post-classification refinements [42], or reclassifying inaccurately classified or
“mixed” pixels using several filter algorithms [13] to improve the classification accuracy.
Others refrained from using the Maximum Likelihood classifier and instead resorted to
other means of classification such as decision tree rules [49], artificial neural networks
[16],[50],[51], and support vector machines [49],[52-54]. Decision trees in particular
offer advantages not provided by other approaches [49]. They are computationally fast
and make no statistical assumptions regarding the distribution of data. However, the
challenge to using decision trees lies in the determination of the “best” tree structure and
the decision boundaries [49]. Artificial neural networks (ANN), on the other hand, are
non-linear mapping structures based on the function of the human brain [16]. Advantages
of the ANN approach include ability to handle non-linear functions, to perform model-
free function estimation, to learn from data relationships that are not otherwise known
and, to generalize to unseen situations. In land-cover classification, ANNs can produce
the most accurate maps and could be resistant to training data deficiencies [50]. ANNs
avoid some of the problems of the Maximum Likelihood Classifier (i.e., the normal
distribution assumption) by adopting a non-parametric approach [54]. Paola &
Schowengerdt [55] compared Maximum Likelihood classifier with ANN and showed that
ANN is more robust to training site heterogeneity and the use of class labels for land use
that are mixtures of land cover spectral signatures. The differences between the two
21
algorithms may be viewed, in part, as the differences between nonparametric (neural
network) and parametric (maximum-likelihood) classifiers. Computationally, the back
propagation neural network is at a serious disadvantage to maximum-likelihood, taking
nearly an order of magnitude more computing time when implemented on a serial
workstation.
2.3.3 Classification by Support Vector Machine
The support vector machine (SVM) is a classification system derived from
statistical learning theory [56],[57]. It represents a group of theoretically superior
machine learning algorithms, and employs optimization algorithms to locate the optimal
boundaries between classes [54]. It separates the classes with a decision surface that
maximizes the margin between the classes. The surface is often called the optimal
hyperplane, and the data points closest to the hyperplane are called support vectors. The
support vectors are the critical elements of the training set. In practice, the SVM has been
applied to optical character recognition, handwritten digit recognition and text
categorization [56],[58]. These experiments found the SVM to be competitive with the
best available classification methods, including neural networks and decision tree
classifiers [54]. Recently, support vector machines are becoming popular for
classification of multispectral RS images [54],[52],[59],[53],[60]. SVM achieves a higher
level of classification accuracy than either the Maximum Likelihood Classifier or the
ANN classifier, and that the SVM can be used with small training datasets and high-
dimensional data [60]. The superior performance of the SVM was also demonstrated in
classifying hyperspectral images acquired from the Airborne Visible/Infrared Imaging
22
Spectrometer (AVIRIS) [61]. Nemmour & Chibani [53] introduced the use of SVM for
land cover change detection with an application for mapping urban extensions and
showed that SVMs have higher recognition rates compared to neural networks, hence
confirming their efficiency for land cover change detection.
A detailed assessment has been conducted by Huang et al. [54] with regards to the
relative performance of Maximum Likelihood classifier (MLC), Decision Tree Classifier
(DT), Neural Network Classifier (NNC) and SVM in land-cover classification from a
Landsat TM image. Of the four algorithms evaluated, the MLC had lower accuracies than
the three non-parametric algorithms. The SVM was more accurate than DT in 22 out of
24 training cases. It also gave higher accuracies than NNC when seven TM bands were
used in the classification. The higher accuracies of the SVM should be attributed to its
ability to locate an optimal separating hyperplane. Statistically, the optimal separating
hyperplane found by the SVM algorithm should be generalized to unseen samples with
fewer errors than any other separating hyperplane that might be found by other
classifiers. Generally, the absolute differences of classification accuracy were small
among the four classifiers. However, many of the differences were statistically
significant. In terms of algorithm stability, the SVM gave more stable overall accuracies
than the other three algorithms except when trained using 6% pixels with three variables.
Of the other three algorithms, DT gave slightly more stable overall accuracies than NNC
or the MLC, both of which gave overall accuracies in wide ranges. In terms of training
speed, the MLC and DTC were much faster than the SVM and NNC. While the training
speed of NNC depended on network structure, momentum rate, learning rate and
converging criteria, that of the SVM was affected by training data size, kernel parameter
23
setting and class separability. All four classifiers were affected by the selection of
training samples. It was not possible to determine the minimum number of samples for
sufficiently training an algorithm according to results from this experiment. However, the
initial trends of improved classification accuracies for all four classifiers as training data
size increased emphasize the necessity of having adequate training samples in land cover
classification. Feature selection is another factor affecting classification accuracy.
Substantial increases in accuracy were achieved when all six TM spectral bands and the
NDVI were used instead of only the red, NIR and the NDVI. The additional four TM
bands improved the discrimination between land classes. Improvements due to the
inclusion of the four TM bands exceeded those due to the use of better classification
algorithms or increased training data size, underlining the need to use as much
information as possible in deriving land cover classification from satellite images [54].
2.4 GIS in LULCC studies
Much LULCC researches have been devoted to the analysis of relations between
LULCC and socio economic and biophysical variables that act as the ‘driving factors’ of
change [24],[1]. The roles of GIS have become significant recently in this LULCC
researches.
GIS has been used extensively by Hietel et al. [62] to develop spatial-temporal
database of land cover and of environmental variables (such as elevation, slope, aspect,
available water capacity and soil texture, and structural variables such as patch size,
shape and distance) that are needed in investigating land-cover trajectory types, land-
cover transitions at individual time intervals and their relationships to the environmental
24
variables in Hesse, Germany. GIS aided in preparing the complex set of variables
required in their conduct of the statistical analyses such as the creation of “trajectories of
change” maps at eight time intervals, and to introduce these datasets to multivariate
statistical analysis through binary encoding into a presence/absence map layers.
Similarly, Apan & Peterson [12] used GIS capabilities to probe tropical
deforestation in Mindoro, Philippines. Specifically, they used GIS to store and analyze
forest-cover data from Landsat TM and thematic maps of georeferenced physical
variables (elevation, geology, slope, distance to road, distance to water, etc) and to further
process these datasets using Pearson’s chi square test, Cramer’s V calculations and
logistic regression analysis. They were able to show the utility and effectiveness of the
GIS environment, in tandem with statistical packages, to handle large datasets, to obtain
samples of almost unlimited number for a study area, and to analyze the relationship
between variables expressed as data layers. Data formatting for transfer to statistical
software was likewise unconstrained. However, they recognized that all the thematic
layers should be digitized and georeferenced with high accuracy to maximize the
effectiveness of GIS-coupled statistical analyses.
Helmer [63] used GIS to satisfy his LULCC study’s requirement of comparing
multitemporal land-cover maps of Puerto Rico derived from 1977-1978 aerial
photographs and 1991-1992 Landsat TM images in order to analyze patterns of land
development. He used GIS to (1) rasterize the polygon-level maps from aerial
photographs into 30-m cell size, (2) co-register it with the map from Landsat TM image,
(3) edit both maps to a comparable set of classes through overlays and class
generalizations, and (4) cross-tabulate the number of pixels of each class in 1977-78 that
25
changed or did not change to each 1991-92 class. In addition to these, he also used GIS to
create the large amount of datasets necessary for binomial logistic regression analysis of
land cover change, which included among others, distances to road, distance to nearest
urban area, distance to nearest road, elevation, slope, geology, and forest and urban patch
size.
Indeed, GIS plays a crucial role in LULCC research. From the three articles
reviewed above, it can be observed that GIS is being used as pre-processor of land-cover
information and related variables prior to statistical analysis. Some studies have coupled
GIS with RS to visualize and quantify land-cover changes [64] and some even used GIS
native functions to derive spatial patterns and statistics of land-cover change vis-à-vis sets
of environmental variables [65] and to supplement statistical change detection analysis
[66].
In Jung et. al [65]’s study, GIS was used extensively to investigate the
relationship between deforested area and spatial data (e.g., topography, road, and existing
protection area. The basic topographic information (altitude, slope, distance from access
roads, and distance from protected zones) were computationally derived from geographic
data of 30-m resolution. By performing an overlay analysis technique with the
topographic and spatial information, the relationship between these basic spatial factors
and deforestation distribution patterns on each polygon was analyzed. Consequently, they
were able to produce various graphs that depict the relationship between the spatial
factors and deforestation such as the correspondence of distance from the nearest access
roads and protected area borders with quantity of deforestation and with the frequency of
occurrence of deforestation.
26
Porter-Bolland et al. [66] used GIS buffering techniques to supplement their
change detection analysis of three satellite images to assess forest clearing for agricultural
use from two time periods (1988-2000 and 2000-2005) in La Montaña, Campeche,
Mexico. Buffers from specific variables (distance to roads, distance to settlements and
proximity to lowland flooded forest) were used to assess deforestation patterns and
identify potential variables that determine land use change. Through this buffer analysis,
they were able to establish the trend that (1) deforestation appears to be prominent near
lowland flooded forests on soil types identified by local people as suitable for agriculture,
particularly for pasture establishment., and (2) recent occurrence of deforestation in the
area is strongly associated with soil/vegetation characteristics, infrastructure development
and settlement locations while the proximate causes are related to agricultural expansion.
2.5 Review of statistical methods in LULCC studies
It has been stated in the earlier sections of this chapter that LULCC’s are the
result of the interplay between socio-economic, institutional and environmental factors.
Lesschen et al. [1] provide a comprehensive discussion on statistical methods for
analyzing the spatial dimension of changes in LULCC in relation to these factors. In their
report, the analysis of relations between land use and the socio-economic and biophysical
variables that act as the ‘driving forces’ of land use change is given great emphasis, i.e.,
to understand LULCC by identifying its proximate and underlying causes through
empirical data analysis. Two categories of empirical analysis techniques are presented
27
based on objective and data structure: (1) exploratory spatial data analysis, and (2)
regression analysis.
The main uses of exploratory data analysis techniques are related to data
reduction and structure detection [1]. These methods aim (i) to reduce the number of
variables; (ii) to describe the underlying structure between variables in the data; and (iii)
to classify variables into groups. Examples of these are factor analysis and principal
component analysis (PCA), which are applied as data reduction or structure detection
methods, and cluster analysis for classification. This is useful in LUCC analysis because
land use change is often assumed to be influenced by a large set of driving and
conditioning factors. PCA and factor analysis are suited to exploration of the structure of
interrelationships between these different driving factors. Furthermore, the methods can
also be used to characterize land use systems based on a number of indicators.
The study conducted by Veldkamp & Fresco [67] is an example of exploratory
data analysis wherein factor analysis was used to investigate land use and land cover in
Costa Rica at six different scales. Spatial distributions of potential biophysical and LUCC
drivers were statistically related to the distribution of pastures, arable lands, permanent
crops, and natural and secondary vegetation. The factor analysis demonstrated that factor
contributions and compositions change with scale, confirming spatial scale dependence in
the structure of the spatial data. The total variance in the data set could be described by
four significant factors for all scales, describing between 68% and 81% of the total
variance.
Lesschen et al. [1] provide caution on the use of factor analysis. All variables
should be quantitative at the interval or ratio level. Categorical data, e.g. ethnicity or soil
28
type, are not suitable for factor analysis. The data should have a bivariate normal
distribution for each pair of variables, and observations should be independent. Those
variables that do not show variability can be discarded. It is established on an a priori
basis that the variables with a coefficient of variation of less than 50% are normally not
considered [68]. Second, some variables may not be relevant to the typification required
for the purposes of a particular study and can therefore be discarded, even though the
typology obtained initially is consistent with observations. Thus one has to assess if the
information imparted by a variable is consistent with the research objectives. Third,
highly correlated variables can be eliminated, as an uncritical use of such variables.
Canonical correlation analysis is another exploratory data analysis technique. It is
a multivariate technique that has the same computational basis as factor analysis, but in
its concept and objectives it is closely related to multiple regression [1]. Multiple
regression is concerned with the relationship between a single dependent variable Y and a
set of predictor variables X1, X2, …, Xm. An extension of this concern is the
relationship(s) between a set of Y variables and a second set of X variables measured on
the same objects. These relationships may be investigated by finding linear combinations
of the X and Y variables that give the highest correlation between the two sets. Such
correlations are called canonical correlations and the linear combinations are called
canonical variables. In effect, the set of X variables is converted into a single new
variable and the set of Y variables into another single new variable. Then the correlation
between these new variables is determined [69]. This statistical method is particularly
appropriate when the dependent variables themselves are correlated with each other. In
29
such cases, canonical correlation analysis can uncover complex relationships that reflect
the structure between predictor and dependent variables.
Canonical correlation analysis was used by Hietel et al. [70] to identify key socio-
economic indicators of land-cover changes in Hesse, Germany. Canonical analysis was
used as an explorative process to reduce large set of socioeconomic variables and to
define a plausible land cover model. Correlation coefficients computed were used to
identify key socio-economic indicators of land-cover changes. The results showed that a
relatively high percentage of variance in land-cover data can be explained by socio-
economic factors. The types of land-cover changes can be characterized by combinations
of key socio-economic indicators.
Regression analysis falls in the second category of empirical analysis techniques
[1]. Regression analysis is used to investigate the association of a dependent variable with
one or more independent variables. In linear regression a straight line is used to represent
the association of the explanatory variables with the dependent variable. More complex
methods of regression exist, intended for different types of dependent variables and data
structures.
Linear regression is a method that estimates the coefficients of a linear equation,
involving one or more independent variables that best predict the value of the dependent
variable. Linear regression is a frequently used technique; however, in LUCC modelling,
this regression is less popular because linear regression can only be applied for
continuous dependent variables [1]. Instead logistic or multinomial regression is used,
because land use is normally expressed as a discrete variable. In linear regression
analysis, it is possible to test whether two variables (or transformed variables to allow for
30
non-linearity) are linearly related and to calculate the strength of the linear relationship if
the relationship between the variables can be described by an equation of the form Y = α
+ βX. Y is the variable being predicted (the dependent, criterion, outcome or endogenous
variable), X is a variable whose values are being used to predict Y (the independent,
exogenous or predictor variable), and α and β are population parameters to be estimated
[71].
The study of Weiss et al. [72] is an example study where linear regression was
used to assess the condition of rangelands in Saudi Arabia and evaluate the effects of
grazing. The coefficient of variation (COV) of the monthly normalized difference
vegetation index (NDVI) was used as a measure of vegetative biomass change. A higher
NDVI COV for a given pixel represented a greater change in vegetation biomass for that
area. The trend in COV values was assessed with linear regression over a 12-year period.
The COV regression line for each pixel reflects the overall long-term trend in the data. A
t-test of the value of the slope was performed to test whether the data used to compute the
regression line were statistically significant at a certain confidence level. Another
example of linear regression is the study of López et al. [73], who used linear regression
between urban growth and population growth for the prediction of urban expansion in
Morelia, Mexico.
Lesschen et al. [1] noted that linear regression for the analysis of multiple land
use types is only used when the land use data are represented as continuous values
instead of dichotomous. Such a representation is used in the case of a coarse spatial
resolution at which the data land use situation cannot adequately be presented by
31
dichotomous data. This is exemplified in the works of Verburg & Chen [74] and Wood &
Skole [75].
Logistic regression is useful for situations where the dependent variable has a
binary output, e.g. the presence or absence of a characteristic or outcome [1]. The method
is very appropriate in predicting the probability that a case will be classified into one as
opposed to the other of the two categories of the dependent variable. The odds that Y = 1,
written odds(Y=1), is the ratio of the probability that Y = 1 to the probability that Y ≠ 1.
The odds that Y = 1 is equal to P(Y=1) / [1– P(Y=1)]. Unlike P(Y=1), the odds has no
fixed maximum value, but like the probability, it has a minimum value of 0 [71].
Logistic regression is a very popular and widely used method in LULCC studies
[1]. Apan & Peterson [12] used logistic regression to probe tropical deforestation in
Mindoro, Philippines by determining the statistical significance of the relationships of
various georeferenced environmental variables to forest cover. Their results indicated
that, except from distance to road, all variables (elevation, slope, distance to water,
geology, land form, soil fertility, etc.) are statistically related to forest cover.
Serneels & Lambin [76] used logistic regression to identify how much
understanding of the driving forces of land use changes can be gained through a spatial
statistical analysis for the Mara ecosystem in Kenya. All explanatory variables suggested
by the conceptual model for the study area were introduced in the statistical mode and,
based on the full model information, they analyzed which variables contribute
significantly to the explanation of land use changes. Schneider & Pontius [77] used
logistic regression for modelling deforestation in the Ipswich watershed of
Massachusetts. Geoghegan et al. [78] used logistic regression to model tropical
32
deforestation and land use intensification in the southern Yucatán peninsular region, in
combination with household survey data on agricultural practices.
Verburg et al. [79] used logistic regression to analyze the factors determining land
use patterns in the Netherlands. The method was based on an extensive database,
including land use, biophysical, socio-economic, neighborhood and policy characteristics.
All data were aggregated to 500×500-meter grids covering the Netherlands. Historic and
recent land use changes were studied. The long-term effects of land use changes were
studied by analyzing current land use patterns. Many factors that are commonly used to
explain land use change patterns are endogenous at a long timescale, e.g. measures
indicating current accessibility. Therefore the assumption was made that long-term land
use change was mainly determined by biophysical factors. A binomial logit model was
compiled for each land use type:
Logit P = α + β1Xsoil + β2Xaltitude + β3Xdist-hist-town
The exp(β) values (odds ratio) for the logit models describing the land use pattern for the
main land use types in 1989 indicated that a very clear association exists between the pH
and the location of forest, which is mainly found on poor sandy soils. Model fit for forest
is good, while the independent variables for residential and industrial areas only explain a
small fraction of the spatial variability. The logit models indicated which factors were
important determinants of land use patterns in the Netherlands.
Lesschen et al. [1] noted that multicollinearity of dependent variables needs to be
accounted for in exploratory and regression analysis of LULCC data. Multicollinearity,
or the dependency between the explanatory variables, is an important issue to account for
in all multivariate methods. Collinearity arises when independent variables are correlated
33
with one another. It is suggested that data should be checked on multicollinearity before
any regression analysis [1]. Perfect collinearity means that an independent variable is a
perfect linear combination of the other independent variables. If each independent
variable in turn is treated as the dependent variable in a model with all of the other
independent variables as predictors, perfect collinearity would result in R2 = 1 for each of
the independent variables. When perfect collinearity exists, it is impossible to obtain a
unique estimate of the regression coefficients; any of an infinite number of possible
combinations of linear or logistic regression coefficients will work equally well [71].
Collinearity is easy to detect, but there are only few acceptable remedies for it [80].
Deleting a variable involved in collinearity runs the risk of omitted variable bias.
Methods to prevent multicollinearity include factor analysis, a priori correlation analysis
and stepwise regression [1].
34
Chapter 3
The Study Area
3.1 Background
The Provinces of Agusan del Norte and Agusan del Sur (Figure 4) are located in
the Caraga Region XIII in the northeastern part of Mindanao, Philippines. The two
provinces were previously part of one province, the Agusan province, not until the
issuance of Republic Act 4979 on June 17, 1967 separating it into two independent
provinces as they are now.
3.2 The Province of Agusan del Norte
The province of Agusan del Norte (ADN) is located in the northeastern part of
Mindanao. It is bounded on the north by Butuan Bay and Surigao del Norte, east by
Surigao del Sur; west by Misamis Oriental, and south by Agusan del Sur. The capital of
ADN is Butuan City until August 16, 2000 where the seat of provincial government was
transferred to Cabadbaran by virtue of RA 8811 and Butuan City became an
administratively independent city. The province has 10 municipalities namely:
Buenavista, Carmen, Jabonga, Kitcharao, La Nievas, Magallanes, Nasipit, Santiago,
35
Tubay, and Remedios T. Romualdez and 1 component city, City of Cabadbaran, which is
also the capital of the province. (Butuan City was integrated in Agusan del Norte in all
the analyses conducted in this study.)
Figure 4. Map showing the municipalities and cities in the Agusan provinces
ADN has a total land area of 259,052 hectares or 2,590.52 sq. km. where 25.72%
of it is classified as alienable and disposable while 74.28% are forestlands [81]. Land use
36
in Agusan del Norte is primarily for agriculture making it one of the country’s leading
rice producer. Other major products are coconut, corn, mango, bananas, palm oil,
vegetables, and prawns. The province also has 23 lumber and plywood production plants
operating in Butuan City, thus, the province continues to be a major timber producer.
Based on the May 1, 1975 census, Agusan del Norte (including Butuan City) had
a population of 300,735 which gives a population density of 116 persons per sq. km. As
of May 1, 2000 census, the province population already reach 551,503 with an annual
growth rate of 1.89% hence making it the 47th
most populous province in the Philippines.
Population density of the province in the year 2000 is 213 persons per sq. km. With a
span of 25-years, difference in population is 250,768 and the difference in population
density is 97 persons per sq. km. Climate in the province is moderate, having no definite
dry season. ADN is strategically located outside the typhoon belt area in the Philippines.
Rainfall is evenly distributed throughout the whole year. The terrain of ADN is generally
composed of lowlands plains.
3.3 The Province of Agusan del Sur
Agusan del Sur (ADS) is a landlocked province located south of Agusan del
Norte. It is bounded on the east by Surigao del Sur; south by Davao Oriental, Compostela
Valley and Davao, and west by Bukidnon. The province has 14 municipalities namely:
Bayugan, Bunawan, Esperanza, La Paz, Loreto, Prosperidad, Rosario, San Francisco, San
Luis, Santa Josefa, Sibagat, Talacogon, Trento and Veruela. The provincial capital is
Prosperidad.
37
ADS has a total land area of 896,550 hectares or 8,965.50 sq. km. making it the
seventh largest province in the country where 75.28% of it is forestland and the
remaining 24.72% is alienable and disposable. ADS is an elongated basin formation with
mountain ranges in the eastern and western sides forming a valley, and half of the
southern province is an area filled with swamps and lakes. Agusan del Sur falls under the
Type II climate classification system: no dry season with very pronounced wet season of
heavy precipitation. Maximum rainfall generally occurs from the months of December to
January. The average temperature is 27 °C.
Based on the May 1, 1975 census, Agusan del Sur had a population of 174,682
which gives a population density of 19 persons per sq. km. As of May 1, 2000 census, the
province population already reach 559,294 with an annual growth rate of 1.79% giving
the place the most populous province in the Caraga region and ranked 43rd
in the most
populous provinces in the Philippines. Population density of the province in 2000 is 62
persons per sq. km. With a span of 25-years, difference in population is 384,612 and
difference in population density of 43 persons per sq. km. Residents in ADS are mostly
engaged in forestry and agriculture with rice, corn, and fruits as the major agricultural
crops. Palm oil plantation, crude oil processing and coconut trees products are also
produced in ADS. The terrain of ADS is generally rugged with slope that extends up to
more than 50 percent.
38
3.4 Status of Forest Resources in the Agusan Provinces
Records of the Department of Environment and Natural Resources-Forest
Management Bureau (DENR-FMB) show that in the year 1984 to 1992, the log
production of the two Agusan provinces is relatively high (Figure 5). The log production
of Agusan del Norte is very high until year 1992. After year 1992, the log production of
Agusan del Norte was observed to decline rapidly. This may be due to the
implementation of the total log ban in year 1992. The province of Agusan del Sur also
experienced decline in log production after the implementation of the total log ban but the
province was slowly recovering and was observed that its log production increased
starting 1996. Unlike Agusan del Sur, Agusan del Norte was not able to recover from its
decline in log production after the implementation of the total log ban.
LOG PRODUCTION (1984-2001)
0
50000
100000
150000
200000
250000
300000
350000
400000
450000
500000
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
Year
Vo
lum
e o
f P
ro
du
cti
on
(cu
.m.)
Agusan del Norte
Agusan del Sur
Figure 5. Log Production of Agusan del Norte and Agusan del Sur from 1984-2001
39
0
5
10
15
20
25
30
A DN ADS A DN ADS A DN ADS
1995 1998 2000
Year
Nu
mb
er
of
Pro
du
cti
on
Pla
nts Active Regular Sawmills
Existing Mini-Sawmills
Active Plywood Plants
Active Veneer Plants
Existing other Wood-Based
Plants
Figure 6. Graph showing the timber processing plants and sawmills in the Agusan
Provinces
Figure 6 shows the number of sawmills, plywood plants, veneer plants and other
wood-based plants existing in the two provinces. It can be noticed that there was an
abrupt increase in the establishment of mini-sawmills in Agusan del Norte from the year
1995 to 1998. From 7 sawmills in 1995, it increased to 25 in the year 1998 but slightly
goes down to 18 in the year 2000. No sawmill is present in Agusan del Sur from 1995 to
1998. Only 2 mini-sawmills was established in Agusan del Sur in 2000. Although
decreasing in number from year 1995 to year 2000, regular sawmills and plywood plants
are also found in Agusan del Norte. On the other hand, only 1 plywood plant was
established in Agusan del Sur 1995. Establishment of veneer plants however in Agusan
del Norte increased from 2 in year 1995 to 7 in year 1998. Establishment of other wood-
based plants started in 1998.
40
3.5 Forest License Agreements Issued in the Agusan Provinces
Due to the vast forest resource of the Agusan Provinces, several privately owned
companies and community-based organizations applied and where awarded with different
forest licenses in the Agusan Provinces. Forest licenses issued in the Agusan Provinces
include Timber License Agreement (TLA), Integrated Forest Management Agreement
(IFMA), Community-Based Forest Management Agreement (CBFMA), and Community-
Based Resource Management (CBRM).
Timber License Agreement (TLA) is an agreement between the government and
privately owned companies to explore and exploit the forest resource in the area. Since
the implementation of TLA in 1958, it consequently covered major part of the country’s
forest land resulting to denudation of the forest resources. In order to answer the
denudation problem caused by TLA, the establishment of Industrial Tree Plantation (ITP)
was implemented in September 9, 1981 under Executive Order No. 725. The areas
available were the open, denuded and inadequacy stocked residual natural forest areas
within the concession. It was renamed as Industrial Forest Management (IFM) since its
coverage expanded to allow planting of non-timber products. Likewise, the activities
under the program were expanded to include not just the industrial plantation
development and related activities but also the management and protection of the natural
forest. Industrial Forest Management Agreement was again renamed as Integrated Forest
Management Agreement (IFMA). An IFMA is a production sharing contract entered into
by and between the DENR and a qualified applicant wherein the DENR grants to the
latter the exclusive right to develop, manage, protect and utilize a specified area of
forestland and forest resources therein for a period of 25 years and may be renewed for
41
another 25-year period. Aside from the TLA which was subsequently replaced by IFMA,
another timber license introduced last July 19, 1995 was the Community-Based Forest
Management Agreement (CBFMA). Anchoring on the concept of "people first and
sustainable forestry will follow", CBFMA is a production sharing agreement between the
DENR and the participating people’s organization (POs) for a period of 25 years
renewable for another 25 years and shall provide tenurial security and incentives to
develop, utilize and manage specific portions of forest lands. Along with this, another
program introduced by the government is the Community-Based Resource Management.
CBRM is a $US50M project financed by the World Bank and the Philippine government,
designed to address the twin objectives of ameliorating rural poverty and resource
degradation through support for locally generated and implemented natural resource
management projects. The project aimed to strengthen the capacity of local communities
in forest, upland and near-shore areas, and that of Local Government Units (LGUs) to
plan and implement investments for community-initiated development projects to reduce
poverty and environmental degradation [82].
Table 2 shows the timber license agreements issued in the provinces of Agusan
del Norte and Sur. It can be observed that as of the year 1959 to 1983, there were a total
of 200,144 hectares issued with TLA in Agusan del Norte while during the same time
period, there were a total of 323,931 hectares issued with TLA in Agusan del Sur.
Looking at the percent area of the total TLA coverage in each province with respect to its
total land area, 77% of the total land area of Agusan del Norte is covered with TLA while
in Agusan del Sur only 36% of its total land area is covered with TLA. Based on these
statistics, it is evident that more TLAs were issued in Agusan del Norte than in Agusan
42
del Sur during the time period of 1959-1983. Figure 7 shows the location of the TLAs
and IFMAs issued in the Agusan provinces.
Table 2. List of Timber License Agreements (TLAs) issued in Agusan del Norte and
Agusan del Sur with date of TLA issuance and expiry, and area covered. (Source: Yearly
Forestry Statistics, DENR-FMB).
Name Date Issued Expiry Date Area Covered (ha.)
Agusan del Norte
Nasipit Lumber Corp. 4-Dec-1959 30-Jun-2007 98,310
Butuan Lumber Manufacturing Co. 14-Sep-1961 30-Jun-1986 12,109
Sibagat Timber Corp. 22-Feb-1973 30-Jun-1997 19,050
Adgawan Timber Inc. 24-Sep-1975 30-Jun-1985 9,175
Mainit Lumber and Dev. 5-Dec-1975 30-Jun-1997 27,870
Ventura Timber Corp. 11-May-1983 31-Mar-2008 33,630
Agusan del Sur
Bueno Industrial and Dev't. Corp 14-Sep-1961 3-Jun-1982 23,402
Grecan Co., Inc. 14-Aug-1970 30-Jun-1982 14,975
JCA Lumber and Plywood Ind. 8-Sep-1970 30-Jun-1990 12,940
CVC Lumber Ind. 8-Sep-1970 30-Jun-1990 30,295
Republic Timber Corp. 2-Aug-1972 30-Jun-1997 19,270
JJ Tirador Lumber Ind. 4-Feb-1974 3-Jun-1997 47,980
SPV Timber and Construction Inc.
19-Apr-1974 30-Jun-1974 37,160
Del Rosario and Sons Logging Ent. 25-Apr-1974 30-Jun-1984 14,470
Agusan Wood Ind. Inc. 16-Jul-1974 30-Jun-1998 60,390
Southern Agusan Timber Co. 27-Feb-1976 30-Jun-1982 5,560
Woodland Domain Corp. 11-Nov-1982 30-Jun-2007 57,489
Prudent Logging Dev't. Corp. 9-Jul-1985 30-Jun-2000 37,160
El Salvador Lumber Co. 16-Sep-1985 3-Sep-1995 49,115
43
Figure 7. Map showing the location of TLAs and IFMAs issued in the Agusan provinces.
On the other hand, issuance of CBFMA in Agusan del Norte and Agusan del Sur
started in 1997. Total area of CBFMA in Agusan del Norte is 37,806.61 hectares or
14.59% of its total area is covered with CBFMA while in Agusan del Sur the total area
44
covered by CBFMA is 68654.37 hectares or 7.66% of its total land area. Figure 8 shows
the location of issued CBFMAs and CBRMs issued in the Agusan provinces.
Figure 8. Map showing the location of CBFMAs and CBRMs issued in Agusan
Provinces.
45
Chapter 4
Methodology
4.1 Overview
The overall flow of methodology of the study is presented in Figure 9. The
methodology is subdivided into three phases: (i.) RS image analysis to derive multi-
temporal land-cover and change maps, (ii.) GIS analysis of detected forest cover change
or deforestation, and (iii.) statistical analysis to determine the degree of association of
bio-physical and socio-economical factors with deforestation.
In Phase 1, RS images of the study area acquired by the Landsat Multi-spectral
Scanner (MSS) and Enhanced Thematic Mapper Plus (ETM+) sensors for 1976 and
2001, respectively, were analyzed in order to derive land-cover maps and to determine
the changes in land-cover during this time period. The detected changes, in the form of a
land-cover change map, are directed to Phase 2 wherein it is visualized and analyzed in a
GIS. Much of the GIS analysis focused on the characterization and visualization of the
detected deforestation vis-à-vis sets of georeferenced bio-physical and socio-economical
factors which are hypothesized to be associated with deforestation [12] such as elevation,
slope, distance to road, distance to water resources, distance to forest resource industries
(TLA, IFMA, CBFMA, and CBRM) and population density, among others. The analysis
is further expanded in Phase 3 where exploratory statistical data analysis and logistics
46
regression techniques are used to determine the significance and magnitude of the
relationship between the detected deforestation and the georeferenced factors.
Figure 9. The three phases of the study’s methodology.
Phase 1
Remotely-sensed image analysis
Processing of 1976 Landsat MSS and 2001 Landsat ETM+
images of Agusan del Norte and Agusan del Sur to derive land-
cover maps
Detection of land-cover change and derivation of a land-cover
change map
Phase 2
GIS Spatial Data Analysis
Visualization of land-cover change
Characterization of deforestation vis-à-vis sets of georeferenced
environmental factors such as elevation, slope, distance to road,
distance to water resources, distance to forest licenses and
population density, etc.
Phase 3
Statistical Data Analysis
Exploratory statistical analysis to determine the magnitude and
relationship of deforestation vis-à-vis georeferenced bio-physical
and socio-economical factors
Logistic regression to model the significance and magnitude of
influence of the factors to deforestation
47
4.2 Remote sensing image analysis
Figure 10 shows the process flow for the analysis of remotely-sensed images of
the study area to derive the land-cover maps for the years 1976 and 2001, as well as the
land-cover change map between these years. Each step is discussed in the following
subsections.
Figure 10. Process flow diagram of remotely-sensed image analysis
4.2.1 Landsat images
Landsat 2 MSS image acquired on April 17, 1976 (path 120/row 54 of World
Reference System 1 – WRS1 tiling) covering the study area was downloaded free-of-
charge from the University of Maryland - Global Land Cover Facility (GLCF) website
Geometric
accuracy
assessment,
image
processing,
cloud and
shadow masking
DENR 2003 Land-
cover map, ground
truth data,
2005Quickbird image
Post-classification
change detection
Classification
accuracy
assessments
1976 – 2001
Land-cover Change
Map
1976 Land-cover
Map
2001 Land-cover
Map
Supervised image
classifications
1976
Landsat
MSS
2001
Landsat
ETM+
SRTM
DEM
48
(http://glcf.umiacs.umd.edu), while Landsat 7 ETM+ image acquired on May 22, 2001
(path 112/row 54 WRS2 tiling) was obtained free-of-charge from the U.S. Geological
Survey through the web-based application of Earth Explorer
(http://edcsns17.cr.usgs.gov/EarthExplorer/). These 8-bit images (Figure 11) were
already orthorectified upon download and are georeferenced in Universal Transverse
Mercator Zone 51 (UTM51) projection on the World Geodetic System (WGS) 1984
datum. Characteristics of these images are listed in Table 3.
Table 3. Characteristics of the Landsat images used in the study.
a. April 17, 1976 Landsat MSS image.
Band
No.*
Spectral Range (μm) Band Name Spatial Resolution (m)
4 0.5 – 0.6 Green 57
5 0.6 – 0.7 Red 57
6 0.7 – 0.8 Near infra red 1 57
7 0.8 – 1.1 Near infra red 2 57
*Bands 1 to 3 were assigned to three Return Beam Vidicon (RBV) cameras on board the
Landsat 2 satellite. The MSS bands were numbered to follow on in this sequence [43].
b. May 22, 2001 Landsat ETM+ image.
Band
No. Spectral Range (μm) Band Name
Spatial Resolution
(m)
1 0.45 – 0.52 Blue 30
2 0.52 – 0.60 Green 30
3 0.63 – 0.69 Red 30
4 0.76 – 0.90 Near infra red 30
5 1.55 – 1.75 Middle infra red 1 30
6 10.4 – 12.5 Thermal 60
7 2.08 – 2.35 Middle infra red 2 30
49
126°30'0"E
126°30'0"E
126°0'0"E
126°0'0"E
125°30'0"E
125°30'0"E
125°0'0"E
125°0'0"E
9°3
0'0
"N
9°3
0'0
"N
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®
20KM
April 17, 1976
Landsat 2 MSS
RGB = 6-5-4
Agusan
del Norte
Agusan
del Sur
126°30'0"E
126°30'0"E
126°0'0"E
126°0'0"E
125°30'0"E
125°30'0"E
125°0'0"E
125°0'0"E
9°3
0'0
"N
9°3
0'0
"N
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®
20KM
May 22, 2001
Landsat 7 ETM+
RGB = 4-2-1
Agusan
del Norte
Agusan
del Sur
Figure 11. The two Landsat images of the study area that were subjected to image analysis to derive land-covers maps for the
years 1976 and 2001.
50
4.2.2 Image geometric accuracy assessment
Prior to any image pre-processing, the geometric accuracy of the Landsat images
were first assessed, i.e. to determine the degree of accuracy of the UTM 51 WGS 1984
coordinates (Easting, Northing) of the pixels in the Landsat images as well as their co-
registration. This task is important as it could minimize the error in change detection.
The assessment was conducted by first checking the geometric accuracy of the
2001 Landsat image by comparing it to 1:50,000 NAMRIA topographic maps of the
image coverage (in paper form). The rule-of-thumb of a Global Root Mean Square Error
(GRMSE) of less than or equal to half a pixel (i.e., 15-m) was adapted to test whether the
geometric accuracy of the 2001 Landsat image is acceptable or not. In case the GRMSE
is more than 15-m, re-georeferencing the image must be done. However, if the GRMSE
is ≤ 15-m, the 2001 image’s geometric accuracy is deemed acceptable and re-
georeferencing is not needed. If this is the case, the next step is to test the co-registration
of the 1976 Landsat image to the 2001 Landsat image, with the 2001 Landsat image as
the reference or base image.
In testing the 2001 Landsat image’s geometric accuracy, a total of thirty eight (38)
points identifiable on both the image and the NAMRIA maps were used. These points
(Figure 12) are mostly road intersections, bridges, and river bends and their intersections.
The UTM 51 WGS 1984 coordinates of each point were determined both on the image
and on the maps. Linear interpolation by scaling was used to determine the coordinates of
the points in the NAMRIA maps while a direct coordinate readout using ITT Visual
Information Solutions’ Environment for Visualizing Images (ENVI) 4.4 software [83]
was used to determine the coordinates of the same points in the 2001 Landsat image.
51
Prior to comparison, a datum transformation of the UTM 51 coordinates of the points
determined from the NAMRIA maps was done because NAMRIA maps have the Clarke
1866 spheroid as the datum. Datum transformation from Clarke 1866 to WGS 1984 was
done using the Environmental Systems Research Institute’s Arcview GIS 3.2 Projection
Utility software [84].
Comparison of UTM51 WGS 1984 coordinates of 38 points in the 2001 Landsat
image with those in the NAMRIA maps showed that the geometric accuracy of the image
is acceptable because the GRMSE is 10.25 meters, and is less than 15-m (Figure 12).
Also, the local RMSE of the points are all less than 15-m with a mean of 10.05 m.,
further indicating the good accuracy of the UTM 51 WGS 1984 coordinates of the pixels.
The co-registration of the 1976 Landsat image to the 2001 Landsat image was
next performed. In this case, the 2001 Landsat image is the reference image where the
UTM coordinates of points on the 1976 Landsat image will be compared. Considering
that the pixel size of the 1976 image is 57 m., the target value of the GRMSE must be ≤
28.5 m in order for the 1976 image to be geometrically acceptable. Based on 22 points
common on both images (Figure 13), the GRMSE was computed as 16.62 m, with
average local RMSE equal to 15.20 m. This indicates that the geometric accuracy of the
1976 Landsat image is acceptable and its co-registration with the 2001 Landsat image is
good. Furthermore, as the GRMSE and average local RMSE is less than 28.5 m., the 30-
m. resolution land-cover map derived from the 2001 Landsat image after undergoing
resampling to 57-m resolution, will perfectly align with the land-cover map derived from
the 1976 Landsat image. This minimizes the error due to image mis-registration in the
change detection and analysis.
52
!
!
!
!
! !
!
!
!
!
!
!
!
!
!
!
!
!
!
! !
!
!
!
! !
!
!
!
!
!
!
!
!
!
!
!
!
8.3
9.877.39
11.49.17
7.31
8.47
5.79
9.288.37
9.65
9.82
9.17
9.97
9.76
8.95
6.68 8.43
5.61
12.24
11.41
11.69
12.52
11.25
11.41
10.12
10.03
11.05
10.46
11.36
11.98
13.23
13.66
12.34
13.54
10.57
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°3
0'0
"N9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®
25
Kilometers
Agusan del
Norte
Agusan del
Sur
Local RMS Error
Magnitude (m.)
and Direction
! Test Point Location
Geometric Accuracy of the
2001 Landsat ETM+ Image
Average Local
RMS Error: 10.05 m.
Global RMS Error: 10.25 m.
Figure 12. Location of points used to determine the geometric accuracy of the 2001
Landsat image and the resulting RMSE vectors of the comparisons with NAMRIA maps.
The numerical values and the lines indicate the magnitude and direction of the
differences in coordinates (local RMSE), with the arrows pointing to the “actual” (i.e.,
NAMRIA map) coordinates.
53
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
11.4
3.27
9.93
5.66
5.56
8.51
14.9
20.81
26.84
14.66
27.73
17.03
22.83
24.32
22.31
12.97
13.86
19.83
10.63
12.22
14.66
14.51
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°3
0'0
"N
9°3
0'0
"N
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®
25
Kilometers
Agusan del
Norte
Agusan del
Sur
Geometric Accuracy
of the1976 Landsat MSS
Image
Average Local
RMS Error: 16.62 m.
Global RMS Error: 15.20 m.
Local RMS Error
Magnitude (m.)
and Direction
! Test Point Location
Figure 13. Location of points used to determine the geometric accuracy of the 1976
Landsat image and its co-registration with the 2001 Landsat image. Also shown are the
resulting RMSE vectors of the comparisons. The numerical values and the lines indicate
the magnitude and direction of the differences in coordinates, with the arrows pointing to
the “actual” (i.e., 2001 Landsat) coordinates.
54
4.2.3 Image pre-processing
After testing for the geometric accuracy, the Landsat MSS and ETM+ images
were subjected to radiometric calibration and atmospheric correction using the steps
provided by Schowengerdt [85]. All image processing was done using ENVI 4.4.
First, the pixel values of each band of the image which are in digital numbers
(DN), with grey-scale level from 0-255 were converted to at-sensor or “top-of-
atmosphere” radiance. For the Landsat ETM+ image, the spatial resolution of Band 6
(which is 60 meters) was resampled using nearest neighbor method to 30 meters first
prior to conversion. This is to make it compatible with the other 6 bands.
The conversion of DN values to top-of-atmosphere radiance was done using the
standard formula available from the Landsat 7 Science Data Users Handbook [86] and
from Markham & Barker [87]:
( )( )
( )
cal calmin
calmax calmin
LMAX LMIN Q QL LMIN
Q Q (1)
where Lλ = spectral radiance at the sensor's aperture in W/(m2
· sr · μm);
Qcal = quantized calibrated pixel value in DN's;
Qcal min = minimum quantized calibrated pixel value corresponding to LMINλ;
Qcal max = maximum quantized calibrated pixel value corresponding to LMAXλ;
LMINλ = spectral radiance that is scaled to Qcal min in W/(m2
· sr · μm);
LMAXλ = spectral radiance that is scaled to Qcal max in W/(m2
· sr · μm).
All the variables needed by Eqn. 1 are available from the metadata file of the
Landsat images, and are presented in Table 4 (for Landsat MSS) and in Table 5 (for
Landsat ETM+).
55
Table 4. Values used for the calibration of the Landsat MSS image to radiance.
MSS Band No. LMAXλ LMINλ Qcal max Qcal min
4 26.300 0.800 255.0 1.0
5 17.600 0.600 255.0 1.0
6 15.200 0.600 255.0 1.0
7 13.00 0.400 255.0 1.0
Table 5. Values used for the calibration of the Landsat ETM+ image to radiance.
ETM+ Band
No. LMAXλ LMINλ Qcal max Qcal min
1 191.600 -6.200 255.0 1.0
2 196.500 -6.400 255.0 1.0
3 152.900 -5.000 255.0 1.0
4 241.100 -5.100 255.0 1.0
5 31.060 -1.000 255.0 1.0
6 17.040 0.000 255.0 1.0
7 10.800 -0.350 255.0 1.0
Second, a fast atmospheric correction by means of the dark-object subtraction
method using band minimum [85] was applied to the at-sensor radiance image. A further
nominal calibration using a standard atmospheric model was not done because the
necessary input data and software facilities for such model were not available during the
conduct of image calibrations. With the dark-object subtraction method, the atmospheric
absorption was disregarded and atmospheric scattering was assumed to be an additive
component that has the effect of adding a constant value to each pixel in a spectral band
of the images. The pixel in each band of the images with the minimum value was
considered as the “dark object”. The method further assumes that the dark object has
uniformly zero radiance for all bands, and that any non-zero measured radiance must be
due to atmospheric scattering into the object’s pixels [85]. This correction was applied
uniformly to each band of the Landsat MSS and ETM+ images, thus assuming a constant
56
atmosphere across the images. The results were surface radiance images for the years
1976 and 2001, respectively.
After conversion to surface radiance, the 1976 and 2001 radiance images (except
for Band 6 of the 2001 radiance image) was further calibrated to surface reflectance using
the formula [86],[87]:
2
p
s
L d
ESUN cos
(2)
where ρP is the surface reflectance, Lλ is the surface radiance, d is the earth-sun distance in
astronomical units, ESUNλ is the mean solar exoatmospheric irradiances, and θs is the
solar zenith angle in degrees. ESUNλ values are listed in Table 6 (for Landsat MSS) and
Table 7 (for Landsat ETM+).
The earth-sun distance was approximated as [88]:
2 ( 93.5)
1 0.0167sin365
Dd
(3)
where D is the Julian day number of the day of acquisition. Other needed values used for
the computation of Eqn. 2 and 3 are shown in Table 8.
Table 6. Landsat MSS mean solar exoatmospheric spectral irradiances [87].
Band ESUNλ (Units: W/m2∙µm)
4 185.600
5 155.900
6 126.900
7 90.600
57
Table 7. Landsat ETM+ mean solar exoatmospheric spectral irradiances [86].
Band ESUNλ (Units: W/m2∙µm)
1 1969.000
2 1840.000
3 1551.000
4 1044.000
5 225.700
7 82.070
Table 8. Values used for the computation of the surface reflectance.
Variable Landsat MSS Landsat ETM+
Sun Elevation Angle 54.58 61.07
Solar Zenith Angle, θs 34.42 28.93
D 107 142
d
1.000067 1.00024
In the case of the Landsat ETM+ image, the band 6 radiance image was converted
to surface temperature under an assumption of unity emissivity using the formula [86]:
2
1ln 1
KT
K
L
(4)
where T is the surface temperature, Lλ is the band 6 radiance image, and K1 and K2 are
Landsat ETM+ pre-calibration coefficients where K1 = 666.09 W/(m2
· sr · μm) and K2 =
1282.71 Kelvin. After conversion to surface temperature, a linear normalization was done
to re-scale the temperature values from 0-1 so that it will be compatible with the 6
reflectance bands. This normalization is necessary prior to image classification.
From the reflectance images, Normalized Difference Vegetation Index (NDVI)
and synthetic reflectance bands were also created to supplement the limited number of
bands of the Landsat 2 MSS. This was found necessary in order to increase the number of
R-G-B band combinations that could be made such that the image could be interpreted
properly during the course of image classification. Using the Landsat 2 MSS reflectance
58
bands as inputs, three additional bands were created that when combined in an R-G-B
mode, the band combination simulates a “true color” image. The following equations
were implemented in ENVI 4.4 in order to create the additional bands:
RED = MSS Band 5
2 1GREEN = (MSS Band 4) MSS Band 6
3 3
2 1BLUE = (MSS Band 4) MSS Band 6
3 3
Simulated
Simulated
Simulated
(5) (a to c)
The resulting RGB image using the simulated Red, Green and Blue bands were
subjected to photographic stretching to produce another RGB image that correspond well
to the response of the human eye [83], an image that is “more realistic” and very useful
for identifying various land-cover classes present in it.
In preparation for the succeeding image analysis procedures, the radiometrically
calibrated and atmospherically corrected Landsat images (as well as other by-products)
were sub-setted to the portions bounded by the study area.
4.2.4 Cloud and shadow masking
The presence of clouds and shadows in the images especially in highly elevated
areas and mountain ranges was a great obstacle in the extraction of accurate land-cover
information. To minimize the error and confusion that cloud cover and shadows may
introduce to the extraction of land-cover information during the image classification
process, a simple cloud and shadow detection and masking technique was developed and
used to mask them in the images. The technique (Figure 14) is generally composed of
manual segmentation of cloud and shadow contaminated regions of the image, and
application of Maximum Likelihood supervised classification to label pixels
59
contaminated and not contaminated with clouds and shadows. The technique was applied
individually to the 1976 and 2001 subset Landsat images. In the manual segmentation of
cloud and shadow contaminated regions, portions of the images with clouds and shadows
were delineated through on-screen digitizing in ENVI 4.4. The delineated regions may
contain areas not contaminated by clouds and shadows. The purpose of this step is to
limit the classification to regions were cloud and shadow contaminations area present.
This is more appropriate than doing a cloud and shadow detection by subjecting the
whole image to a supervised or unsupervised classification, which by experience, is more
prone to misclassification especially of urban areas which are usually mislabeled as
“clouds”. In Step 2, the segmented regions were subjected to Maximum Likelihood
classification. Groups of pixels representing clouds, shadows and others (non-cloud and
no-shadow) were collected from the segmented regions and used to train the classifier.
Then, the classifier was run to label each pixel in the segmented regions, thereby
separating clouds and shadows pixels from others. For both the two Landsat images, all
bands were used as input for classification. Pixels labeled as clouds and shadows were
then merged to create a cloud-shadow mask that was applied to their corresponding
Landsat image.
4.2.5 Image classification and accuracy assessment
The cloud free, radiometrically calibrated and atmospherically corrected Landsat
images (as well as other by-products such as NDVI and synthetic bands) which were sub-
setted to the portions bounded by the study area were subjected to supervised
classification to derive the 1976 and 2001 land-cover maps.
60
Figure 14. Flowchart of the simple cloud and shadow detection and masking technique
developed and applied in this study.
Eight (8) land-cover classes namely, Forest, Rangeland, Built-up, Palm Trees,
Cropland, Bare Soil, Exposed Rocks and Water, were identified from the images through
visual interpretation using existing land-cover maps of the DENR, topographic maps and
Clouds
Shadow
Non-cloud and non-shadow
a. Input Image b. Segmentation
d. Segmented Image c. Maximum Likelihood
Classification
e. Classified Image f. Cloud and Shadow-free
Image
61
Google Earth images as references. Ground truth dataset collected from fieldwork
conducted in October-December 2006 and April-May 2007 were also used. Definitions of
these land-cover types are presented in Table 9. Table 10 and Table 11 show various
image keys used in the visual interpretations.
Table 9. Definitions of land-cover types used in this study.
Land-cover Type
Bare Soil Areas with exposed soil and in which less than one half of an area
unit has vegetation or other cover.
Built-up Areas
Comprised of areas of intensive use with much of the land covered
by structures. Includes settlement areas, buildings, farmsteads, and
surrounding lots.
Cropland Comprised of areas planted with crops (e.g. rice)
Exposed Rocks
Exposed rocks, sands, stones, cobbles, and boulders along rivers,
streams, and shorelines that are not covered by water during the
time of image acquisition.
Forest
Parcels of lands having a tree-crown areal density (crown closure
percentage) of 10 percent or more and are stocked with trees
capable of producing timber or other wood products. Includes
deciduous and evergreen forestlands.
Palm Trees Areas covered with palm oil, nipa and coconut plantation.
Rangeland
Areas where the potential natural vegetation is predominantly
grasses, grasslike plants, or shrubs and less permanently used for
that purpose.
Water Area covered with water. Includes sea water, rivers, and streams.
Representative samples of each class (training set) were collected from the images
for supervised image classification; another independent set of samples (validation set)
were likewise collected for accuracy assessment (Table 12). A minimum number of 30
pixels were chosen randomly for each class, following the guidelines of Van Genderen et
al. [89] to obtain a reliable estimate of classification accuracy of at least 90%. The
classification algorithms included traditional classifiers such as Minimum Distance,
Mahalanobis Distance and Maximum Likelihood and the recently developed SVM
classifier. SVM was implemented as a non-linear classifier using the Radial Basis
Functions (RBF) kernel available in the ENVI 4.4 image analysis software.
62
Table 10. Image keys used in visual interpretations of the 1976 Landsat MSS image.
Land Cover
Type
Band Combinations (R-G-B)
SimR-
SimG-
SimB
SimR-NDVI-
Band7
NDVI-Band
7-Band 6
BDVI-Band
7-Band 6
Band 6-
Band 5-
Band 4
NDVI-
SimG-SimB
NDVI-
SimR-
Band 7
Band 5-
Band 4-
SimB
Bare Soil
Built-up Areas
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
63
Table 11. Image keys used in visual interpretations of the 2001 Landsat ETM+ image.
Land Cover
Type
Band Combinations (R-G-B)
3-2-1 4-2-1 4-5-1 5-3-1 5-4-7 7-4-2 7-5-3 7-5-4
Bare Soil
Built-up Areas
Cropland
Exposed
Rocks
Forest
Palm Trees
Rangeland
Water
64
Table 12. Number of pixels collected for image classifications and accuracy assessments.
Land-cover Class
1976 Land-cover Classification 2001 Land-cover
Classification
Training Accuracy
Assessment Training
Accuracy
Assessment
Bare Soil 592 122 1028 298
Built-up 553 50 2356 640
Cropland 1465 354 3928 1287
Exposed Rocks 265 36 714 198
Forest 2059 602 2680 1148
Palm Trees 539 160 929 422
Rangeland 1073 565 928 374
Water 4874 387 7670 2214
Total 11420 2276 20233 6581
For both the 1976 and 2001 image dataset, each classifier was implemented using
various combinations of input bands (Table 13). The use of 4 classifiers and various
combinations of image bands and by-products was done to generate several classified
images and selecting from these outputs the best classified image. A 90-m spatial
resolution digital elevation model (DEM) from the Shuttle Radar Topography Mission
(SRTM) was also included as additional band during image classification as it has been
found that DEMs could significantly increase the classification accuracy [90-93]. This
DEM was first re-sampled (using bilinear interpolation) to 57-m to be compatible with
the Landsat MSS dataset, and to 30-m to be compatible with the Landsat ETM+ datasets.
Then the elevation values were normalized from 0 to 1 to be compatible with the data
range of the multispectral bands and by-products.
65
Table 13. Various combinations of input bands used in image classification
1976 image
classification
4 Reflectance Bands (Bands 4, 5, 6 and 7)
4 Reflectance Bands with NDVI
4 Reflectance Bands with DEM
4 Reflectance Bands with NDVI and DEM
4 Reflectance Bands with Simulated Red and Green Bands
4 Reflectance Bands with Simulated Red and Green Bands
and NDVI
4 Reflectance Bands with Simulated Red and Green Bands
and DEM
4 Reflectance Bands with Simulated Red and Green Bands,
NDVI and DEM
2001 image
classification
Reflectance Bands, Temperature (normalized, 0 -1)
Reflectance Bands, Temperature (normalized, 0 -1) and
DEM (normalized 0 – 1)
The computation of accuracy is done by comparing the validation set pixels with
the classification results. From these checks the percentage of pixels from each class in
the image labeled correctly by the classifier is estimated, along with the proportions of
pixels from each class erroneously labeled into every other class. The results are
presented in a confusion or error matrix that lists the number of validation set pixels, in
each case, correctly and incorrectly labeled by the classifier. Two measures of accuracy
were employed to test the classified images, namely the Overall Classification Accuracy
(in percent), Producer’s Accuracy and User’s Accuracy. The overall classification
accuracy is the percentage of correct classifications of the ground truth pixels. It is
computed by dividing the sum of the diagonals of the error matrix (which pertains to the
number of correctly classified pixels for each class) with the total number of validation
set pixels. The Producer’s Accuracy, which is computed for each land-cover class using
the column values of the error matrix (no. of correctly classified x 100% divided by
column total), is the probability that the classifier has labeled the image pixel exactly as
66
its actual ground truth land-cover type [94]. The User’s Accuracy, on the other hand, is
the probability that the pixels belonging to actual land-cover class in the classified image
have been labeled correctly [94] (i.e., do all the pixels labeled as “forest” in the land-
cover map are actually “forest” on the ground?). This measure of accuracy is computed
using the row values of the error matrix ((no. of correctly classified over row total x
100%).
For 1976 and 2001 image classifications, the selection of the “best” classified
image for each year which will be the source of the land-cover maps is based on the
criteria that the classified image must have the highest overall classification accuracy (at
least 90%) among all the classifications and that the Producer’s and User’s Accuracy of
land-cover types relevant to this study which include forest, built-up, rangeland, palm
trees, cropland and bare soil are at least 85% each [95] and must also be highest among
all classification results.
4.2.6 Post-classification change detection
The two land-cover maps derived were then subjected to post-classification
comparison change detection analysis [14] to examine the location, extent and
distribution of land-cover change in the study area. The 2001 land-cover map was first re-
sampled to 57-m resolution using nearest neighbor method prior to change detection.
Because of cloud and shadows present in the images used (“No Data” in the LC maps),
only portions of the LC maps that both have data in 1976 and 2001 were subjected to
change detection analysis. Land-cover change statistics were also computed. Overall
accuracy of the land-cover change detection was computed by multiplying the 1976 LC
67
Map Overall Classification Accuracy and the 2001 LC Map Overall Classification
Accuracy times 100 [42].
4.3 GIS spatial change analysis
The detected changes in forest cover, in the form of a change-no change in forest
cover map, were visualized and analyzed in a GIS. The GIS analysis involved
characterization and visualization of the detected changes vis-à-vis sets of georeferenced
bio-physical and socio-economic factors hypothesized to be associated with deforestation
[12] such as presence of forest licenses, population density change, road infrastructures,
increase in built-up areas, elevation, slope, soil quality and proximity to water resources.
These factors are described in Table 14.
Table 14. Definitions of georeferenced bio-physical and socio-economic factors. Factor Description
Bio-physical
ELEV Elevation of the Agusan provinces
SLOPE Slope of the Agusan provinces
DISTRIV Distance to major river networks in the Agusan
provinces
SOILQUAL Soil quality of Agusan provinces, coded as 0 for non-
suitable for agriculture and 1 for suitable.
Socio-
economical
DISTNEWRD Distance to the new roads since 1976 to 2001
DISTNEWBUILT Distance to the new built-up since 1976 to 2001
POPDENCHANGE Change in population density of the Agusan
provinces from 1976-2001.
DIST_TLA-IFMA Distance to the combined land parcels subjected to
Timber License Agreements and Integrated
Forest Management Agreement DIST_CBFMA-CBRM Distance to the combined land parcels subjected to
Community Based Forest Management
Agreement and Community-Based Resource
Management
68
Point shapefiles of “changed” and “no changed” in forest cover (hereafter referred
to as “FCOVER”) were made from the forest-cover change map. Overlay analysis was
then performed to populate the attribute of the FCOVER pixels with their corresponding
socio-economic and bio-physical factor values. The resulting tabular data was exported to
a spreadsheet file and further analyzed. For each factor, the mean values of all ‘change’
and ‘no change’ samples were computed and were displayed graphically for both
qualitative and quantitative analyses. The analyses of the mean factor values was made in
order to gain insights on the possible similarities or differences in trends between the two
provinces’ forest cover in relation to the identified bio-physical factors and socio-
economical.
The geo-referenced socio-economic and bio-physical factors were prepared as
follows. The presence of logging and timber industries were represented as maps of forest
license agreements issued by the Philippine government between 1976 and 2001. This
spatial data, in Arcview polygon shapefile, was obtained from the DENR-Caraga
Regional Office. There were two kinds of license agreements: (1) those issued to private
corporations that include TLAs and IFMAs; and (2) those issued to non-government
organizations that CBFMAs and CBRMAs. Proximity grids were then computed from
these two factors to determine the Euclidean distance (in meters) of pixels within the
study area from the polygons of these license agreements. A value of 0 indicates that the
pixel is within a particular type of license agreement. These two factors were aptly
labeled as “DIST_TLA-IFMA” and “DIST_CBFMA-CRBMA”, respectively. These two
separate proximity grids were created so as to determine how the pattern of deforestation
and forest retention would vary as the type of licensee differs.
69
Population density change, instead of population count change, was used in this
study as an indicator of forest cover change based on the acceptable assumption that it is
the change and increase in number of persons per unit area that the retention or the
change in forest cover could be expected. Municipal-level population data for the years
1976 (estimated from the 1975 census) and 2001 (estimated from the 2000 census) were
obtained from the National Statistics Offices in Agusan del Norte and Agusan del Sur.
Population density change was computed by subtracting the 1976 population to 2001
population of each municipality, and dividing this by the GIS-computed area of the
particular municipality. This resulting factor map was labeled as “POPDENCHANGE”.
Road infrastructure and increase in built-up areas as determinants of forest cover
change were examined in this study by taking only those new roads and new built-up
areas since 1976. This “new” road data was obtained by overlaying road network
(digitized from the 1954 NAMRIA topographic map and from the 1976 Landsat MSS
image) to the 2001 Landsat ETM+ image. “New” roads, which are those not intersecting
the 1976 roads were then digitized. A proximity grid was then computed and labeled as
“DISTNEWRD”. The rationale behind the use of “new roads” as indicator of forest cover
change is based on the hypothesis that it is the construction of new roads (that may have
resulted from economic development or due to the proliferation of logging industries)
that forest cover in the two provinces became more accessible to change. Similarly, the
relative contribution of the increase in built-up areas since 1976 (e.g., difference in 2001
and 1976 built-up areas) was taken account in this study by calculating a distance to new
built areas grid (DISTNEWBUILT).
70
The four biophysical factors namely, elevation, slope, soil quality and proximity
to water resources were also prepared in the same manner as those of the socio-economic
factors. Elevation (ELEV) and percentage slope (SLOPE) grids were computed using the
90-m SRTM DEM. This DEM was first calibrated with spot heights derived from
NAMRIA topographic maps to linearly transform the elevation values to mean sea level.
Soil quality (SOILQUAL) was obtained from the digital 1:250,000 soil taxonomy map
published by the Bureau of Soils and Water Management. Soil quality was coded as “1”
if the soil has low fertility, 2 if moderately fertile and 3 if highly fertile. The proximity to
water resources was considered in this study by calculating a distance to river grid
(DISTRIV). The river network data was digitized from NAMRIA topographic maps and
from the Landsat images. ArcView GIS 3.2 software was used in the analysis.
4.4 Statistical analysis of land-cover change
Logistic regression analysis was employed to ascertain the degree of association
of bio-physical and socio-economical variables with FCOVER. The multivariate logistic
regression equation used in the analysis is of the form [96]:
1 1 2 2
1 1 2 2( )
1
i i
i i
x x x
x x x
ex
e
(6)
where π(x) is the probability that the dependent variable y equals 1, is the equation
constant, and βi is the coefficient of predictor variable xi (i.e. the socio-economic and bio-
physical factors). Each of the regression coefficients describes the size of the contribution
of that factor. A positive regression coefficient implies that as the value of the factor
increases, the probability of deforestation increases. A negative regression coefficient
71
means that as factor values increases, the probability of deforestation decreases. A near-
zero regression coefficient means that that factor has little influence on the probability of
deforestation.
Logistic regression analysis was employed because of its advantage of analyzing
variables that maybe either continuous or discrete or any combination of both types and
they do not necessarily have normal distributions [97]. Three separate logistic regression
analysis were conducted for each province. These include testing for the (i) bio-physical
factors only, (ii) socio-economic factors only, and (iii) combined bio-physical and socio-
economic factors. It should be noted that logistic regression was mainly used as a way to
explain forest cover change in ADN and ADS vis-à-vis bio-physical and socio-economic
factors using the regression coefficients as indicators, and not as a predictor of FCOVER
change.
Because of the large number of FCOVER pixels in each province, representative
samples (about 5% each of ‘changed’ and ‘no-change’ collected in a stratified random
manner) were subjected to logistic regression analysis. Table 15 shows the 5% samples
subjected to logistic regression analysis.
Table 15. The 5% samples used in logistic regression analysis.
Agusan del Norte Agusan del Sur
No Change (0) 6,581 23,051
Change (1) 5,719 8,196
Total 12,300 31,247
For each province, ‘no change’ pixels were coded as “0” and ‘change’ pixels as
“1”. The 5% sample size was chosen as it could be representative of the existing forest-
cover change in the study area as long as the number of samples is large [12] (e.g.,
72
thousands). To allow less-biased comparison of the forest cover change characteristics of
the two provinces, it was necessary to normalize the factor values (transform the values
so that the minimum is 0 and the maximum is 1) to the global minimum and maximum of
each factor; i.e. maximum and minimum of combined ADN and ADS factor values.
Logistic regression analysis requires absence of collinearity (or multicollinearity)
among the independent variables. Collinearity (or multicollinearity) is the undesirable
situation when one independent variable is a linear function of other independent
variables [1]. Tests for collinearity of variables using Tolerance Statistics and Variance
Index Factor (VIF) were made in order to determine the variables that are correlated with
each other before conducting logistic regression analysis. All linear combinations of
biophysical and socioeconomic variables were tried through linear regression to
determine collinearity of the variables (e.g., is slope a linear combination of all the other
variables?). Computed values for each linear combination of the variables indicate
absence of multi-collinearity if the Tolerance Statistic >0.5 and close to 1; and VIF <2.
Aside from the Tolerance Statistic and Variance Index Factor, computation of a Pearson
Pairwise Correlation Matrix (Pearson’s R) was also done to easily determine which
variables are correlated. If the Pearson’s R > 0.5, collinearity exists and one of the
variables can be dropped off from the logistic regression model (Millington et al., 2007).
The correlation matrix was computed using merged ADN and ADS FCOVER datasets.
All statistical analyses were done in SPSS Version 16.
73
Chapter 5
Results and Discussion
5.1 Land-cover maps
5.1.1 The 1976 land-cover map
Figure 15 shows the results of the classification done on the April 17, 1976
Landsat MSS image of the study area using the Support Vector Machine (SVM)
algorithm implemented with the Radial Basis Function (RBF) as the mathematical
surface for land-cover class separation. This land-cover map, already masked out with
clouds and cloud shadows, has an overall classification accuracy of 94.99%, the highest
among 32 classifications that utilized 6 various combinations of inputs bands subjected to
four classifications algorithms (Table 16). The input bands used in the classification to
derive this final land-cover map were the Landsat MSS surface reflectance bands (Bands
4, 5, 6 and 7), NDVI and DEM (normalized from 0 to 1). The total number of ground
truth pixels used for accuracy assessment is 2,276.
The results of SVM-classification of 4 reflectance bands with NDVI and DEM as
the source of the 1976 land-cover map is based on the criteria that the classified image
must have the highest overall classification accuracy in all the classifications and that the
Producer’s and User’s Accuracy of land-cover types relevant to this study which include
74
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
April 17, 1976
Land-Cover Map
®
25
Kilometers
Agusan del
Norte
Agusan del
Sur
Land-cover Types:
Bare Soil
Built-up
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
No Data
forest, built-up, rangeland, palm trees, cropland and bare soil are at least 85% each and
must also be highest in all classification results.
Figure 15. The 1976 land-cover map of Agusan del Norte and Agusan del Sur resulting
from the classification of the April 17, 1976 Landsat MSS image using SVM. All white
areas within the provincial boundaries classified as “No Data” are clouds and shadow
pixels in the image.
75
Table 16. Matrix of percent overall classification accuracies of 32 classified images (from
various band combinations of the1976 Landsat MSS image and image by-products
(Ground truth pixels = 2, 276)
Input Band
Combinations
Classification Algorithm
Minimum
Distance
Mahalanobis
Distance
Maximum
Likelihood
Support
Vector
Machine
4 Reflectance Bands
(Bands 4, 5, 6 and 7) 68.96 66.40 85.43 85.63
4 Reflectance Bands
with NDVI 74.04 70.07 83.49 85.41
4 Reflectance Bands
with DEM 65.17 65.47 89.80 94.07
4 Reflectance Bands
with NDVI and DEM 73.51 68.43 89.45 94.99
4 Reflectance Bands
with Simulated Red and
Green Bands
69.14 73.42 82.91 85.90
4 Reflectance Bands
with Simulated Red and
Green Bands and NDVI
74.08 78.68 83.80 85.59
4 Reflectance Bands
with Simulated Red and
Green Bands and DEM
65.34 70.86 90.55 93.76
4 Reflectance Bands
with Simulated Red and
Green Bands, NDVI
and DEM
73.60 75.23 90.60 93.63
Table 17 shows the confusion (or error) matrix of the SVM-classified Landsat
MSS reflectance bands with NDVI and DEM. The error matrices of the classification
results with the 2nd
and 3rd
highest overall classification accuracies: SVM-classified 4
reflectance bands with DEM and SVM-classified 4 reflectance bands with simulated Red
and Green bands and DEM, respectively, are shown in Table 18 and Table 19. These
matrices were used in computing the overall classification accuracy, Producer’s Accuracy
76
and User’s Accuracy, that were then used in evaluating the criteria for the selection of
the 1976 land-cover map of the study area.
Table 17. Error matrix of the SVM-classified Landsat MSS reflectance bands with NDVI
and DEM (the source of the 1976 land-cover map of the study area).
Cla
ssif
ied P
ixel
s
Land-cover
Type
Validation Pixels
Forest Range-
land
Built-
up
Palm
Trees Cropland
Bare
Soil
Exposed
Rocks Water Total
Forest 590 4 0 6 3 0 0 0 603
Rangeland 0 557 0 0 0 0 0 0 557
Built-up 0 0 45 0 7 0 9 2 63
Palm Trees 4 4 0 154 1 0 0 0 163
Cropland 2 0 3 0 311 1 0 9 326
Bare Soil 0 0 0 0 1 120 0 2 123
Exposed
Rocks 0 0 1 0 0 0 27 3 31
Water 6 0 1 0 31 1 0 371 410
Total 602 565 50 160 354 122 36 387 2276
In Figure 16, bar charts of Producer’s and User’s Accuracy of each land-cover
type in the three classified images are shown. It is very clear from these charts that the
SVM-classified reflectance bands with NDVI and DEM gained the highest values of the
three measures of accuracies (see Table 20 for values) and modestly satisfied the final
land-cover map selection criteria with >85% Producer’s and User’s accuracies for forest,
rangeland, palm trees, cropland and bare soil. In the case of pixels classified as “built-
up”, the producer’s accuracy in the classified image is 90% and passed the selection
criteria. However, the User’s Accuracy is quite low at 71.43%. Nevertheless, this value is
still acceptable as it is still higher compared to the User’s Accuracy of the two other
classifications results (68.85% and 68.33%, respectively).
77
Table 18. Error matrix of the SVM-classified Landsat MSS reflectance bands with DEM. C
lass
ifie
d P
ixel
s
Land-cover
Type
Validation Pixels
Forest Range-
land
Built-
up
Palm
Trees Cropland
Bare
Soil
Exposed
Rocks Water Total
Forest 582 4 0 7 1 1 0 0 595
Rangeland 0 557 0 0 0 0 0 0 557
Built-up 0 0 42 0 7 0 8 4 61
Palm Trees 4 4 0 153 2 0 0 0 163
Cropland 5 0 3 0 307 3 0 9 327
Bare Soil 0 0 0 0 1 115 0 2 118
Exposed
Rocks 0 0 2 0 0 0 28 4 34
Water 11 0 3 0 36 3 0 368 421
Total 602 565 50 160 354 122 36 387 2276
Table 19. Error matrix of the SVM-classified Landsat MSS reflectance bands with
simulated Red and Green bands and DEM.
Cla
ssif
ied P
ixel
s
Land-cover
Type
Validation Pixels
Forest Range-
land
Built-
up
Palm
Trees Cropland
Bare
Soil
Exposed
Rocks Water Total
Forest 589 5 0 11 6 1 0 0 612
Rangeland 0 556 0 0 0 0 0 0 556
Built-up 0 0 41 0 6 0 9 4 60
Palm Trees 4 4 0 149 1 0 0 0 158
Cropland 1 0 4 0 308 3 0 10 326
Bare Soil 0 0 0 0 6 117 0 10 133
Exposed
Rocks 0 0 1 0 0 0 26 3 30
Water 8 0 4 0 27 1 1 360 401
Total 602 565 50 160 354 122 36 387 2276
78
Figure 16. Bar charts showing the Producer’s (a) and User’s (b) Accuracies of land-cover
types in three SVM-classified land-cover maps for 1976.
50
55
60
65
70
75
80
85
90
95
100
Fores
t
Ran
gela
nd
Bui
lt-up
Palm
Tre
es
Cro
plan
d
Bar
e Soi
lExp
osed
Roc
ks
Wat
er
Land-cover Type
Pro
duce
r's
Acc
ura
cy (
%)
Reflectance Bands with NDVI
and DEM
Reflectance Bands with DEM
Reflectance Bands with
Simulated Red and Green Bands,
and DEM
50
55
60
65
70
75
80
85
90
95
100
Fores
t
Ran
gela
nd
Bui
lt-up
Palm
Tre
es
Cro
plan
d
Bar
e Soi
lExp
osed
Roc
ks
Wat
er
Land-cover Type
Use
r's
Acc
ura
cy (
%) Reflectance Bands with NDVI
and DEM
Reflectance Bands with DEM
Reflectance Bands with
Simulated Red and Green Bands,
and DEM
`
(a.)
(b.)
79
Table 20. Summary of Producer’s and User’s Accuracies of 1976 land-cover types in
three SVM-classified land-cover maps.
Land-
cover
Type
Classified Image
4 Reflectance Bands
with NDVI and DEM
4 Reflectance Bands
with DEM
4 Reflectance Bands
with Simulated Red
and Green Bands, and
DEM
Producer’s
Accuracy
User’s
Accuracy
Producer’s
Accuracy
User’s
Accuracy
Producer’s
Accuracy
User’s
Accuracy
Forest 98.01 97.84 96.68 97.82 97.84 96.24
Rangeland 98.58 100.00 98.58 100.00 98.41 100.00
Built-up 90.00 71.43 84.00 68.85 82.00 68.33
Palm
Trees 96.25 94.48 95.63 93.87 93.13 94.30
Cropland 87.85 95.40 86.72 93.88 87.01 94.48
Bare Soil 98.36 97.56 94.26 97.46 95.90 87.97
Exposed
Rocks 75.00 87.10 77.78 82.35 72.22 86.67
Water 95.87 90.49 95.09 87.41 93.02 89.78
5.1.2 Accuracy of the 1976 land-cover map
A short discussion of the Producer’s and User’s Accuracy of the 1976 land-cover
map may be necessary to better understand the accuracy of the land-cover type labeling
that was done by the SVM classifier. In the 1976 land-cover map, it could be explained
that for land-cover type “forest”, only 590 of the 602 ground truth pixels were correctly
labeled by the classifier (Table 17). Hence, the producer’s accuracy for this type is
590/602 = 0.9801 or 98.01% (as listed in Table 20). Considering the representativeness
of all ground truth pixels used in the accuracy assessment, it could be stated that of all the
(actual) forest in the study area, only 98.01% were labeled in the 1976 land-cover map.
The remaining 1.99% of forest were not labeled by the classifier as “forest” but labeled
them as otherwise, thereby omitting them in the forest class of the land-cover map (hence
the term “errors of omission”). Looking back to the error matrix, these 1.99% pixels were
80
actually labeled by the classifier as “palm trees” and “cropland”. The same logic applies
for the other land-cover types. It could be stated that there is above 90% probability that
all the forest, rangeland, built-up, palm trees, bare soil and water present in the study area
on April 17, 1976 were accurately labeled in the land-cover map. Only a few percentage
(<10%) of these classes were omitted and labeled incorrectly. In the case of “cropland”
and “exposed rock” classes, there is below 90% probability that all these land-cover types
were correctly labeled. The errors of omissions of these classes are more than 10%.
On the land-cover map’s User’s Accuracy, it could be explained that the user’s
accuracy for “forest” class is 590/603 = 0.9784 (or 97.84%). This means that only
97.84% of the pixels labeled as “forest” in the land-cover map are correct. The remaining
2.16% were supposed to belong to another land-cover class (in this case, “rangeland”,
“palm trees” and “cropland”) but erroneously labeled as “forest”. These are known as the
“errors of commission”. In the final 1976 land-cover map, it is notable that, in the
exception of the “built-up” and “exposed rocks” classes, all the land-cover types have
above 90% probability that they were actually the same land-cover types when checked
on the ground on April 17, 1976.
5.1.3 The 2001 land-cover map and accuracy
Figure 17 shows the results of the classification done on the May 22, 2001
Landsat ETM+ image of the study area using the SVM algorithm implemented with RBF
as the mathematical surface for land-cover class separation. This land-cover map, already
masked out with clouds and cloud shadows, has an overall classification accuracy of
98.25%, the highest among 8 classifications that utilized 2 combinations of input bands
subjected to four classification algorithms (Table 21). The input bands used in the
81
classification to derive this final land-cover map are the Landsat ETM+ surface
reflectance bands (Bands 1,2,3,4,5 and 7), temperature band 6 (normalized from 0 to 1),
and DEM (normalized from 0 to 1). The total number of ground truth pixels used for
accuracy assessment is 6,581.
Figure 17. The 2001 land-cover map of Agusan del Norte and Agusan del Sur resulting
from the classification of the May 22, 2001 Landsat ETM+ image using SVM. All white
areas within the provincial boundaries classified as “No Data” are clouds and shadow
pixels in the image.
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
May 22, 2001
Land-Cover Map
®
25
Kilometers
Agusan del
Norte
Agusan del
Sur
Land-cover Types:
Bare Soil
Built-up
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
No Data
82
Table 21. Matrix of percent overall classification accuracies of 8 classified images from
various band combinations of the 2001 Landsat ETM+ image and DEM. (Ground truth
pixels= 6,581).
Input Band
Combinations
Classification Algorithm
Minimum
Distance
Mahalanobis
Distance
Maximum
Likelihood
Support
Vector
Machine
Reflectance Bands,
Temperature
(normalized, 0 -1)
82.48 84.93 94.99 95.87
Reflectance Bands,
Temperature
(normalized, 0 -1) and
DEM (normalized 0 – 1)
83.03 74.78 95.76 98.25
Table 22 shows the confusion (or error) matrix of the SVM-classified Landsat
ETM+ reflectance bands with normalized temperature and DEM. The error matrices of
the classification results with the 2nd and 3rd highest overall classification accuracies:
SVM-classified 6 reflectance bands with temperature band (normalized from 0 to 1) and
SVM-classified 6 reflectance bands with temperature band (normalized form 0 to 1) and
DEM (also normalized from 0 to 1), respectively, are shown in Table 23 and Table 24 .
These matrices were used in computing the overall classification accuracy, Producer’s
Accuracy and User’s Accuracy that were then used in evaluating the criteria for the
selection of the 2001 land-cover map of the study area.
In Figure 18, bar charts of Producer’s and User’s Accuracy of each land-cover
type in the three classified images are shown. The SVM-classified reflectance bands with
normalized temperature and DEM gained the highest values of the three measures of
accuracies (see Table 22 for values) and satisfied the final land-cover map selection
criteria with >85% Producer’s and User’s accuracies for forest, rangeland, palm trees,
cropland, and bare soil.
83
Table 22. Error matrix of the SVM-classified Landsat ETM+ reflectance bands with
normalized temperature and DEM (the source of the 2001 land-cover map of the study
area). C
lass
ifie
d P
ixel
s
Land-class
Types
Ground Truth Pixels
Forest
Range-
land
Built-
up
Palm
Trees Cropland
Bare
Soil
Exposed
Rocks Water Total
Forest 1131 0 0 2 0 0 0 0 1133
Rangeland 3 374 0 1 1 2 0 0 381
Built-up 0 0 633 1 2 3 28 3 670
Palm Trees 14 0 0 414 0 2 0 0 430
Cropland 0 0 4 0 1276 3 0 8 1291
Bare Soil 0 0 0 0 0 285 0 0 285
Exposed
Rocks
0 0 3 1 1 1 168 18 192
Water 0 0 0 3 7 2 2 2185 2199
Total 1148 374 640 422 1287 298 198 2214 6581
Table 23. Error matrix of the SVM-classified Landsat ETM+ reflectance bands with
temperature band (normalized from 0 to 1).
Cla
ssif
ied P
ixel
s
Land-class
Types
Ground Truth Pixels
Forest
Range-
land
Built-
up
Palm
Trees Cropland
Bare
Soil
Exposed
Rocks Water Total
Forest 1113 0 0 24 0 0 0 0 1137
Rangeland 3 373 0 10 2 1 0 0 389
Built-up 0 0 625 2 2 3 71 11 714
Palm Trees 32 0 0 382 0 3 0 0 417
Cropland 0 1 3 1 1243 22 0 8 1278
Bare Soil 0 0 0 0 4 266 0 0 270
Exposed
Rocks 0 0 11 0 1 1 122 10 145
Water 0 0 1 3 35 2 5 2185 2231
Total 1148 374 640 422 1287 298 198 2214 6581
84
Table 24. Error matrix of the Maximum likelihood-classified Landsat ETM+ reflectance
bands with temperature band (normalized form 0 to 1) and DEM (also normalized from 0
to 1) C
lass
ifie
d P
ixel
s Land-class
Types
Ground Truth Pixels
Forest
Range-
land
Built-
up
Palm
Trees Cropland
Bare
Soil
Exposed
Rocks Water Total
Forest 1125 0 0 2 0 0 0 0 1127
Rangeland 4 369 0 5 1 1 0 0 380
Built-up 0 5 626 9 110 7 24 2 783
Palm Trees 19 0 0 399 0 0 0 0 418
Cropland 0 0 0 2 1153 2 0 22 1179
Bare Soil 0 0 0 0 4 287 0 0 291
Exposed
Rocks 0 0 14 2 1 0 174 21 212
Water 0 0 0 3 18 1 0 2169 2191
Total 1148 374 640 422 1287 298 198 2214 6581
Figure 18. Bar charts showing the Producer’s (a) and User’s (b) Accuracies of land-cover
types in the three land-cover maps.
50
55
60
65
70
75
80
85
90
95
100
Fores
t
Range
land
Built-
upPal
m
Cropl
and
Bare
Soil
Expos
ed R
ocks
Wat
er
Land-cover Type
Use
r's A
ccur
acy
(%)
SVM: Reflectance Bands with
Temperature and DEM
SVM: Reflectance Bands with
Temperature
MaxLike: Reflectance Bands with
Temperature and DEM
(b.)
50
55
60
65
70
75
80
85
90
95
100
Fores
t
Range
land
Built-
upPal
m
Cropl
and
Bare
Soil
Expos
ed R
ocks
Wat
er
Land-cover Type
Prod
ucer
's A
ccur
acy
(%)
SVM: Reflectance Bands with
Temperature and DEM
SVM: Reflectance Bands with
Temperature
MaxLike: Reflectance Bands with
Temperature and DEM
(a.)
(b.)
85
Table 25. Summary of the Producer’s and User’s Accuracies of land-cover types in three
derived land-cover maps.
5.2 Land-cover change in the Agusan Provinces
The 57-m resolution land-cover maps of Agusan del Norte and Agusan del Sur for
1976 and 2001, showing areas with data common to both year (i.e., cloud covered areas
in 1976 and 2001 were excluded), are depicted in Figure 19 and Figure 20.
Land-
cover
Type
Classified Image
SVM-classified 6
Reflectance Bands
with Normalized
Temperature Band and
DEM
SVM-classified 6
Reflectance Bands with
Normalized Temperature
Band
Maximum Likelihood-
classified 6 Reflectance
Bands with Normalized
Temperature Band and
DEM
Producer's
Accuracy
User's
Accuracy
Producer's
Accuracy
User's
Accuracy
Producer's
Accuracy
User's
Accuracy
Forest 98.52 99.82 96.95 97.89 98.00 99.82
Rangeland 100.00 98.16 99.73 95.89 98.66 97.11
Built-up 98.91 94.48 97.66 87.54 97.81 79.95
Palm
Trees 98.10 96.28 90.52 91.61 94.55 95.45
Cropland 99.15 98.84 96.58 97.26 89.59 97.79
Bare Soil 95.64 100.00 89.26 98.52 96.31 98.63
Exposed
Rocks 84.85 87.50 61.62 84.14 87.88 82.08
Water 98.69 99.36 98.69 97.94 97.97 99.00
86
Figure 19. The 1976-2001 land-cover maps of Agusan del Norte province. Areas with data comprise 66.98% (or
2044.67sq.km.) of the total land area of Agusan del Norte.
125°45'0"E
125°45'0"E
125°30'0"E
125°30'0"E
125°15'0"E
125°15'0"E
9°1
5'0
"N
9°1
5'0
"N
9°0
'0"N
9°0
'0"N
8°4
5'0
"N
8°4
5'0
"N
2001 Land-Cover Map
AGUSAN DEL NORTE
®20
Kilometers
Land-cover Types:
Bare Soil
Built-up
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
No Data
Butuan City
125°45'0"E
125°45'0"E
125°30'0"E
125°30'0"E
125°15'0"E
125°15'0"E
9°1
5'0
"N
9°1
5'0
"N
9°0
'0"N
9°0
'0"N
8°4
5'0
"N
8°4
5'0
"N
1976 Land-Cover Map
AGUSAN DEL NORTE
®20
Kilometers
Land-cover Types:
Bare Soil
Built-up
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
No Data
Butuan City
87
Figure 20. The 1976-2001 land-cover maps of Agusan del Sur province. Areas with data comprise 51.10% (or 4,133.82 sq.
km.) of the total land area of Agusan del Sur.
126°0'0"E
126°0'0"E
125°30'0"E
125°30'0"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
1976 Land-Cover Map
AGUSAN DEL SUR
®
25
Kilometers
Land-cover Types:
Bare Soil
Built-up
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
No Data
126°0'0"E
126°0'0"E
125°30'0"E
125°30'0"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
2001 Land-Cover Map
AGUSAN DEL SUR
®
25
Kilometers
Land-cover Types:
Bare Soil
Built-up
Cropland
Exposed Rocks
Forest
Palm Trees
Rangeland
Water
No Data
88
0
100
200
300
400
500
600
700
800
900
Bar
e Soil
Bui
lt-up
Cro
plan
d
Expos
ed R
ocks
Forest
Palm
Tre
es
Ran
gela
nd
Wat
er
Land-cover Type
Are
a, in
sq
. k
m.
1976 Land-cover area
2001 Land-cover area
Figure 21. Land-cover change in Agusan del Norte province from 1976-2001 for
cloud free areas only. Upper and lower error bars represent errors of omission and
commission, respectively, of the land-cover classifications.
Changes in land-cover of Agusan del Norte from 1976-2001 is very evident in the
areas surrounding the Butuan City. This is where drastic increases in built-up, cropland
and water (e.g., expansion of fishpond) areas can be found. Yet, the most pronounced
change in land-cover is that of forest and rangeland. Quantitative assessments through
change detection using the land cover change map (93.33% accurate) show significant
decrease in forest cover by 32% (or about 255.30 sq. km.) while rangeland areas
increased by 92% (about 327.86 sq. km.) during the 25-year period. Forest to rangeland is
the major land-cover change in Agusan del Norte from 1976 to 2001 (Figure 22).
Although deforestation due to increase in rangeland is significantly evident, “re-
forestation” of rangeland areas from 1976 to 2001 was also present. It can be observed
that large tract of lands planted with palm trees in 1976 have been converted into
croplands in 2001. Perhaps, this is due to the fact that croplands and palm trees in the
89
study area are usually located near each other than any other land-cover type. Hence,
expansion of one type (in this case the cropland) will result to reduction in another type
(in this case the palm tree lands).
0
50
100
150
200
250
300
Forest
to R
angel
and
Palm
Tre
es to
Cro
pland
Range
land
to F
orest
Palm
Tre
es to
Ran
gela
nd
Bare
soil
to R
angel
and
Forest
to P
alm
Tre
es
Cropl
and to
Pal
m T
rees
Cropl
and to
Ran
gela
nd
Range
land
to P
alm
Tre
es
Bare
soil
to P
alm
Tre
es
Land-cover change type
Are
a o
f ch
ang
e, i
n s
q.
km
.
Figure 22. Top 10 land-cover change types in Agusan del Norte province from 1976-
2001 for cloud-free areas only. Upper and lower error bars represent errors of omission
and commission, respectively, of the land-cover classifications
In the case of Agusan del Sur, increase in cropland and decrease in forest cover is
the most significant land-cover change in terms of change in land area. Quantitatively,
these translate to 156% increase in cropland (or about 198.47 sq. km.) and about 6%
decrease in forest cover (or 113.42 sq. km.). In terms of specificity, the two most
prominent land-cover change types from 1976 to 2001 in this province is the conversion
of rangeland to forest and forest to palm trees. Considering errors in classifications, these
two land-cover change types are almost identical in magnitude. A third major type of
change is that of conversion of forest to rangeland. It is very apparent that the changes in
90
land-cover between the years 1976-2001 are somehow different for each province based
on the top 10 land-cover change types (). In Agusan del Norte, the major land-cover
change type is “forest to rangeland”. A decrease in forest area in this province was found
to be due to the conversion of 269 sq. km. of forest area in 1976 to rangeland areas in
2001. This converted tract of land is about 33% of intact forest cover of Agusan del Norte
in 1976. In the case of Agusan del Sur, an opposite type of major change was found
which is “rangeland to forest”. Here, about 300 sq. km. of rangeland has been converted
to forest lands. This area is about 46% of intact rangeland of Agusan del Sur in 1976.
0
200
400
600
800
1000
1200
1400
1600
1800
2000
2200
Bar
e Soil
Bui
lt-up
Cro
plan
d
Expos
ed R
ocks
Forest
Palm
Tre
es
Ran
gela
nd
Wat
er
Land-cover Type
Are
a, in
sq
. k
m.
1976 Land-cover area
2001 Land-cover area
Figure 23. Land-cover change in Agusan del Sur province from 1976-2001 for cloud-free
areas only. Upper and lower error bars represent errors of omission and commission,
respectively, of the land-cover classifications.
91
0
50
100
150
200
250
300
350
Ran
gela
nd to
For
est
Fores
t to
Palm
Tre
es
Fores
t to
Ran
gela
nd
Palm
Tre
es to
Cro
plan
d
Palm
Tre
es to
Ran
geland
Ran
gela
nd to
Palm
Tre
es
Palm
Tre
es to
For
est
Cro
plan
d to
Palm
Tre
es
Bar
e so
il to
Palm
Tre
es
Bar
e so
il to
For
est
Land-cover change type
Are
a o
f ch
ang
e, in
sq
. km
.
Figure 24. Top 10 land-cover change types in Agusan del Sur province from 1976-2001
for cloud-free areas only. Upper and lower error bars represent errors of omission and
commission, respectively, of the land-cover classifications.
It can be deduced from the computed land-cover change statistics that forest cover
in Agusan del Norte have been reduced drastically by conversion to rangeland. This may
have been due to unsustainable logging activities, where, after the trees have been
harvested, the logged-over areas were left behind without replanting that made it suitable
for grasses to grow. In Agusan del Sur, a much slighter decrease in forest cover was
detected compared to that of Agusan del Norte.
The major types of land-cover change in Agusan del Norte is almost similar to
that of Agusan del Sur. It is only the magnitude of change (i.e., the area converted) that
differs. The most interesting, as discussed earlier, is the “forest-to-rangeland” and
“rangeland-to-forest” types of changes. In Agusan del Norte, “forest-to-rangeland” is the
number one type of change but this is number three in Agusan del Sur. The type “palm
trees to cropland” is next to “forest to rangeland” in Agusan del Sur. The same pattern
92
can be observed in Agusan del Sur. It is only that “palm trees to cropland” is number two
in Agusan del Norte but number 4 in Agusan del Sur. While the ranking maybe different,
the conversion of palm tree lands to cropland is the most pronounced type of agricultural
conversions. There were no other types of land-cover except for “palm trees” that have
been put into agricultural use. This type of change is purely an effect brought by the vast
existence of “palm tree” lands in both provinces due to soil suitability and climatic
conditions. As the demand for cropland products such as rice and corn intensifies,
farmers have to clear tracts of land occupied by palm trees for agricultural use. On the
other hand, while there was a reduction in palm tree lands for cropland purposes, majority
of the type of changes in both provinces are conversions to “palm trees”. This is highly
indicative of the proliferation of coconut and palm oil plantations in these provinces. This
indicates some sort of balance between usage of lands for “palm trees” and for crop
production.
5.3 Deforestation in the Agusan Provinces
It has been stated earlier that forest cover change in Agusan del Norte and Agusan
del Sur was found to be mainly factors of (i.) reduction of forest cover by conversion to
rangeland and (ii.) increase in forest cover by conversion of rangeland, respectively. The
contribution by other types of land-cover change in forest cover retention and reduction is
summarized in Table 26. The relative contribution of each of this change relative to the
original forest cover area in 1976 is shown in Figure 25.
93
Table 26. Forest cover change statistic (1976-2001) in the Agusan Provinces.
a. Agusan del Norte
Conversion: from forest to Area converted
(sq. km.)
% of original (1976) forest
cover area
Retained Forest 427.58 53.26
Rangeland 269.26 33.54
Palm Trees 67.73 8.44
Bare Soil 18.83 2.35
Cropland 15.18 1.89
Water 3.32 0.41
Built-up 0.6 0.07
Exposed Rocks 0.32 0.04
Total area of forest cover in 1976 802.82 100.00
b. Agusan del Sur
Conversion: from forest to Area converted
(sq. km.)
% of original (1976) forest
cover area
Retained Forest 1,497.86 73.55
Palm Trees 286.61 14.07
Rangeland 208.27 10.23
Bare Soil 19.53 0.96
Cropland 18.02 0.88
Water 6.15 0.30
Built-up 0.11 0.01
Exposed Rocks 0.09 0.00
Total area of forest cover in 1976 2,036.64 100.00
0
10
20
30
40
50
60
70
80
90
100
Retained
Forest
Rangeland Palm Trees Bare Soil Cropland Water Built-up Exposed
Rocks
Type of forest cover change (from forest to - )
Per
cen
t o
f co
nve
rted
19
76
fo
rest
co
ver
Agusan del Norte
Agusan del Sur
Figure 25. Comparison of magnitude of forest cover area reduction by types of change.
94
It can be observed that deforestation in Agusan del Norte is significant with only
about 53% of its initial forest cover in 1976 remaining. On the other hand, deforestation
in Agusan del Sur is less significant compared to Agusan del Norte because its forest
cover in 1976 was reduced by about 37%. Yet these estimates of deforestation provide
little comfort as these were computed only based on cloud-free portions of the Landsat
images. Hence, these values may not be true as the total area of the Agusan provinces
were not considered in the computation.
Considering only those portions with land-cover data common to both provinces,
the total areas of forest cover in Agusan del Norte and Agusan del Sur in the year 1976
are 802.82 sq. km. and 2,036.64 sq. km., respectively. Clearly, forest cover in Agusan del
Sur is much larger than that of Agusan del Norte due to the fact that Agusan del Sur is
bigger than Agusan del Norte in terms of land area. The forest cover change statistics
showed that conversion to rangeland, palm trees, bare soil and cropland are among the
four major contributors to forest cover reduction in both provinces. In terms of magnitude
of change, conversion of forest to rangeland is very pronounced in Agusan del Norte than
in Agusan del Sur. Based on Table 26, about 33% of forest cover in 1976 have been
converted to rangeland in 2001 for the province of Agusan del Norte. On the other hand,
only 10% of forest cover of Agusan del Sur was converted to rangeland in 2001. While it
can be stated that the greatest contributor in deforestation of Agusan del Norte is
conversion to rangeland, the same can not be confirmed in the Agusan del Sur. For the
latter, it is a mix of conversion to palm trees and rangeland (14% and 10%, respectively)
that explains forest cover reduction. The contributions of bare soil and cropland are also
greater in magnitude in Agusan del Norte than in Agusan del Sur. Each account for about
95
2% of forest cover reduction in Agusan del Norte. In Agusan del Sur, their contributions
is minimal at approximately 1%. Other changes in forest cover due to conversion to built-
up as well as natural deforestations (e.g., conversion to exposed rocks and water) are
minimal in both provinces.
Although the statistics varies, it can be speculated that deforestation in the Agusan
provinces is largely due to conversion of forest cover to rangeland and palm trees, with
minimal contributions from conversion to cropland and bare soil. The question on why
there were such kinds of conversions that drove deforestations may be answered by
taking into account the interplay between various bio-physical and socio-economical
factors in both provinces.
5.4 Characterizing 25-year deforestation in the Agusan Provinces
Remote sensing image analysis was able to provide data on the location and
magnitude of deforestation and other types of land-cover change in the Agusan
provinces. A further analysis as to the factors associated with deforestation was made
through GIS overlay analysis. In this section, the location of retained forest and the
location of occurrences of all types of deforestation with respect to the mean values of
bio-physical and socio-economical factors (e.g. ELEV, SLOPE, DISTRIV,
DISTNEWBUILT, DISTNEWRD, DIST_TLA-IFMA, DIST_CBFMA-CBRM, and
POPDENCHNGE) of the two Agusan provinces are described.
Figure 26 shows the mean elevation of location of forest cover change
occurrences. In both Agusan del Norte and Agusan del Sur, conversion from forest to
bare soil and rangeland occurred in areas with higher elevation (elevation from 200 m. to
96
230 m.). In areas with elevation less than 55 meters, the evident forest conversion that
occurred is forest to built-up, forest to cropland, and forest to palm trees. Combining all
types of deforestation (forest to bare soil, built-up, cropland, palm trees, and rangeland),
it was observed that it occurred in areas with mean elevation of 230 meters in Agusan del
Norte and 140 meters in Agusan del Sur. In terms of the retained forest, it is mostly
located in areas with an elevation of 380 meters in Agusan del Norte and 250 meters in
Agusan del Sur. It can be stated that retained forest and deforestation in Agusan del Norte
are located in areas with higher elevation compared to Agusan del Sur. This strongly
implies that remaining forest in ADN are located in higher areas because forest in
relatively lower areas have been deforested, since 1976.
Mean ELEV of Location of Forest Cover Change Occurences
382
224 223
2642
78
271
249
141
209
4049
65
247
0
50
100
150
200
250
300
350
400
450
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Ele
vati
on
, m
.
Agusan del Norte
Agusan del Sur
Figure 26. Mean elevation of location of forest cover occurrences in Agusan del Norte
and Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
97
In terms of SLOPE (Figure 27), retained forest in Agusan del Norte are mostly
located in areas with mean slope of 30%, while in Agusan del Sur, retained forest are
located in areas with mean slope of 20%. However, deforestation in Agusan del Norte is
also located areas with higher mean slope (22%) as compared to Agusan del Sur where
deforestation occurred in areas with 18% mean slope. With respect to the specific forest
cover conversion, the changes forest to built-up and forest to cropland occurred in areas
with mean slope of 5-10%. The conversion from forest to palm trees and forest to
rangeland in Agusan del Norte occurred in areas with mean slope of 16% and 18%,
respectively. In Agusan del Sur, conversion from forest to palm trees and forest to
rangeland occurred in areas with milder slope (9% and 19%, respectively) as compared to
Agusan del Norte.
Mean SLOPE of Location of Forest Cover Change Occurences
30
22 22
5
10
16
24
20
13
16
6
9
19
6
0
5
10
15
20
25
30
35
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mea
n S
lop
e, %
Agusan del Norte
Agusan del Sur
Figure 27. Mean SLOPE of location of forest cover occurrences in Agusan del Norte and
Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
98
Mean DISTRIV of Location of Forest Cover Change Occurences
1770
2863
5382
4030
3126
1922
2906
691 759 745
440
741 766 752
0
1000
2000
3000
4000
5000
6000
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Froest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Dis
tan
ce t
o R
iver,
m.
Agusan del Norte
Agusan del Sur
Figure 28. Mean DISTRIV of location of forest cover occurrences in Agusan del Norte
and Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
With respect to DISTRIV (Figure 28), all type of forest cover change in Agusan
del Norte occurred in areas 2000 meters up to 5,300 meters away from the river. In
Agusan del Sur, these types occurred in areas 500-800 meters from the rivers. In terms of
retained forest in Agusan del Norte, it is mostly located in areas 1900 meters away from
the river while in Agusan del Sur, it is mostly located in areas nearer the river or about
800 meters from the river. Looking at the general occurrences of the combined types of
deforestation, in Agusan del Norte, deforestation occurred in areas 2,900 meters away
from the river while in Agusan del Sur, deforestation occurred in areas much nearer in the
river, about 900 meters, compared to Agusan del Norte.
99
Mean DISTNEWBUILT of Location of Forest Cover Change Occurences
8381
4302 4169
0
811
1806
5145
10357
5992
9364
0
2535
3451
9474
0
2000
4000
6000
8000
10000
12000
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Dis
tan
ce t
o N
ew
Bu
ilt-
up
s, m
.
Agusan del Norte
Agusan del Sur
Figure 29. Mean DISTNEWBUILT of location of forest cover occurrences in Agusan del
Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
With respect to DISTNEWBUILT (Figure 29), the following could be observed:
in both Agusan del Norte and Agusan del Sur, retained forest can be found in areas
farther from the newly constructed built-up. On the other hand, deforestation in Agusan
del Norte occurred in areas 4000 meters away from the new built-up while in Agusan del
Sur, deforestation occurred in areas 6000 meters away from the new built-up. In terms of
the specific forest cover change, change in forest to bare soil in Agusan del Norte
occurred in areas 4000 meters away from the new built-up while in Agusan del Sur it
occurred in areas much farther from the new built-up (i.e. 9000 meters away from the
new built-up). The changes: forest to cropland and forest to palm trees in Agusan del
Norte occurred in areas 2000 meters or less from the new built-up areas. In Agusan del
100
Sur, these changes occurred 2500 meters to 3800 meters away from new built-up areas.
The conversion from forest to rangeland in Agusan del Norte occurred at areas 5000
meters away from the new built-ups while in Agusan del Sur it occurred in areas 9,500
meters away from the new built-ups.
Mean DISTNEWRD of Location of Forest Cover Change Occurences
7745
3821 3621
7511091
1624
4549
11647
7129
9374
47014275
4901
10232
0
2000
4000
6000
8000
10000
12000
14000
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Dis
tan
ce t
o N
ew
Ro
ad
s, m
.
Agusan del Norte
Agusan del Sur
Figure 30. Mean DISTNEWRD of location of forest cover occurrences in Agusan del
Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
In terms of occurrences of deforestation with respect to the new roads in the
Agusan provinces (Figure 30), it can be observed that in Agusan del Norte deforestation
occurred in areas 3900 meters away from the new roads while in Agusan del Sur,
deforestation occurred in areas 7000 meters away from new roads. This implies that
accessibility does not hinder deforestation in Agusan del Sur. Retained forest, however,
are found in areas much farther from the new roads in Agusan del Sur than in Agusan del
101
Norte. In terms of specific forest cover change, in Agusan del Norte, forest to bare soil
and forest to rangeland occurred in areas 3,900 meters and 4,200 meters away from the
new road, respectively. In Agusan del Sur, forest to bare soil and forest to rangeland
occurred in areas 4,200 meters and 10,100 meters away from the new road, respectively.
The change from forest to built-up in Agusan del Norte occurred in areas near the new
roads, 900 meters away, while in Agusan del Sur, the same forest cover change occurred
in areas much farther from the new road, 5000 meters away. Forest to cropland and forest
to palm trees in Agusan del Norte, respectively, occurred in areas 1000 meters and 1900
meters away from the new road, while in Agusan del Sur forest to cropland occurred
4,100 meters away from new road and forest to palm trees occurred 5,000 meters away
from the new road.
Mean DIST_TLA-IFMA of Location of Forest Cover Change Occurences
49065454
5252
11303
9520
6440
4978
2548
5024 5099
62026706
3342
7186
0
2000
4000
6000
8000
10000
12000
14000
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Dis
tan
ce t
o T
LA
's a
nd
IF
MA
's, m
.
Agusan del Norte
Agusan del Sur
Figure 31. Mean DIST_TLA-IFMA of location of forest cover occurrences in Agusan del
Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
102
With respect to DIST_TLA-IFMA (Figure 31), the change: forest to bare soil
occurred in almost the same distance from the TLA’s and IFMA’s both in Agusan del
Norte and Agusan del Sur but retained forest in Agusan del Sur are observed to be
located nearer in TLA’s and IFMA’s compared to Agusan del Norte (e.g., retained forest
Agusan del Norte are observed more in areas farther in TLA’s and IFMA’s as compared
to Agusan del Sur). The change forest to built-up in Agusan del Sur occurred in areas
11,000 meters away from TLA’s and IFMA’s while in Agusan del Norte the same forest
cover change occurred in areas about 6,000 meters away from TLA’s and IFMA’s. The
change forest to cropland and forest to palm trees in Agusan del Norte occurred farther
from TLA’s and IFMA’s than in Agusan del Sur but the change forest to rangeland in
Agusan del Norte occurred in areas nearer TLA’s and IFMA’s compared to Agusan del
Sur. In general, deforested areas in both ADN and ADS are located at approximately the
same distance from TLAs/IFMAs. In ADN, unchanged and changed forest areas are
located farther from TLAs/IFMAs; in ADS unchanged forest areas are nearer to
TLAs/IFMAs.
With regards to the occurrences of forest cover change with respect to distance to
CBFMA’s and CBRM’s (Figure 32), all types of forest cover change in Agusan del Sur
occurred in areas farther from CBFMA’s and CBRM’s compared to Agusan del Norte.
Retained forest in Agusan del Norte are located in municipalities with mean
population density change of 50 person per sq. km. while in Agusan del Sur, retained
forest are located in municipalities with population density change of 22 persons per sq.
km (Figure 33) . On the other hand, deforestation in Agusan del Norte are located in areas
with population density change of 75 persons per sq. km. while in Agusan del Sur,
103
deforestation occurred in municipalities with population density change of 35 persons per
sq. km. This implies that more deforestation happens in areas with higher population
density change. All types of forest cover change showed that in Agusan del Norte
deforestation occurred in areas with higher population density change than in Agusan del
Sur.
Mean DIST_CBFMA-CBRM of Location of Forest Cover Change Occurences
3623
29692834
3754
3199
2719
3026
4937
43374500 4570
5092
4500
4033
0
1000
2000
3000
4000
5000
6000
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Dis
tan
ce t
o C
BF
MA
's a
nd
CB
RM
's, m
. Agusan del Norte
Agusan del Sur
Figure 32. Mean DIST_CBFMA-CBRM of location of forest cover occurrences in
Agusan del Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the
mean.
104
Mean POPDENCHANGE of Location of Forest Cover Change Occurences
51
74
64
119
101
92
69
23
33 34
41 39
26
42
0
20
40
60
80
100
120
140
Retained
Forest
Deforested Forest to
Bare Soil
Forest to
Built-up
Forest to
Cropland
Forest to
Palm Trees
Forest to
Rangeland
Forest-Cover Change Type
Mean
Po
pu
lati
on
Den
sity
Ch
an
ge (
19
76
-20
01
),
no
./sq
. k
m.
Agusan del Norte
Agusan del Sur
Figure 33. Mean POPDENCHANGE of location of forest cover occurrences in Agusan
del Norte and Agusan del Sur. Error bars indicate 95% confidence interval of the mean.
5.5 Logistic regression analysis results
GIS-based characterization of FCOVER change vis-à-vis bio-physical and socio-
economic factors can only provide a generalization on the location of FCOVER with
respect to these factors. In order to determine which variables are strongly or least
associated with changes and no change in FCOVER, logistic regression analysis is
necessary. Prior to the analysis, test for multicollinearity of variables was performed.
Computed values of Tolerance Statistic and Variance Index Factor for all linear
combinations of the variables indicated absence of multi-collinearity (Tolerance Statistic
>0.5 and VIF <2). The computed Pearson Pairwise Correlation Matrix (Pearson’s R) also
indicates absence of collinearity, with all values less than 0.5. Hence, all variables were
105
included in the logistic regression analysis. In order to have an easier understanding of
the logistic regression coefficient, a simple interpretation diagram was prepared (Figure
34). The results of the logistic regression analysis are presented in the next sub-sections.
Figure 34. Diagram for interpreting the logistic regression coefficients.
5.5.1 Logistic regression based on bio-physical factors only
The results of binary logistic regression of FCOVER with bio-physical variables
for the Agusan provinces are presented in Table 27 and Figure 35.
Large Factor Values
Negative β
Small Factor Values
Positive β
106
Table 27. Binary logistic regression of FCOVER versus bio-physical factors for Agusan
del Norte and Agusan del Sur
Biophysical
Factors
β
Agusan del
Norte
Agusan del
Sur
ELEV -6.247 -5.768
SLOPE -3.501 -1.243
DISTRIV 4.2 1.801
SOILQUAL -0.386 -0.988
Magnitude of Association of Bio-physical Factors with Deforestation
-6.247
-3.501
4.2
-0.386
2.067
-5.768
-1.243
1.801
-0.988
1.543
-7
-5
-3
-1
1
3
5
ELEV SLOPE DISTRIV SOILQUAL Constant
Bio-physical Factors
Logis
tic
Reg
ress
ion
Coef
fici
ent
(B)
Agusan del Norte
Agusan del Sur
Figure 35. Graph showing β values indicating the magnitude of association of bio-
physical factors with deforestation. Error bar indicate standard error.
The results for Agusan del Norte indicates that among the 4 bio-physical
variables, the distance to rivers (DISTRIV) has the largest and positive regression
coefficient of 4.2. As the value is positive, this means that as distance to river increases,
the probability of forest cover change increases (similarly, an area nearer to river would
have lesser probability of being deforested). This result is quite contrary to what can be
107
expected in forested areas nearer to rivers. In the study area, rivers provide a means of
transportation for loggers to access forested lands as well as for transporting logs.
Following this logic, the nearer a forestland to rivers, the higher the probability that it
will be deforested. Apparently, the result of the logistic regression deviates from this
assumption. Interestingly, however, this result is consistent with the GIS characterization
of FCOVER versus DISTRIV for Agusan del Norte, where it was shown that changes in
FCOVER are found in areas farther from rivers (Figure 28). A plausible explanation for
this maybe the historical deforestation that have occurred during the 1950s to the 1970s
in the study area, most especially in the upland watersheds [7],[8]. During this time
period, forestlands near rivers have been drastically harvested through logging and very
little were left to regenerate. As a consequence, forestlands from 1970s onwards can no
longer be found near rivers but rather in areas far from it. This is heavily supported by the
results of GIS-based characterization (Figure 28) where deforested areas are very far
from rivers (average of 2,863 m). Another explanation would be the implementation of
National Integrated Protected Areas System (NIPAS) Act of 1992 which aimed to protect
areas comprising of large natural parks, landscapes and small watersheds from forestry
activities [98]. Because a network of interconnecting rivers comprises protected
watersheds in the study area, their protection under the NIPAS Act may have prevented
loggers to use the rivers as access to forest. The loggers may have used alternative routes
to conduct logging activities far from these rivers.
With regards to elevation and slope, the computed regression coefficients are
negative which indicate that as its value increases, the probability of deforestation in
Agusan del Norte decreases. This is acceptable because higher elevation and slope values
108
would naturally be indicative of inaccessibility. Again, these results confirm the initial
findings of GIS characterization using the factors values of ELEV and SLOPE (see
Figure 26 and Figure 27). The computed regression coefficient for soil quality
(SOILQUAL) is also negative (-0.386) which indicate that soils with good quality
decreases the probability of deforestation. This result contradicts the expected, that
supposedly good soil quality would increase the probability of a certain area to be
deforested. However, the coefficient is so small that it may not strongly influence the
probability of deforestation.
The logistic regression results for Agusan del Sur show similar patterns on
contributions of ELEV, SLOPE, DISTRIV and SOILQUAL as those of Agusan del
Norte. However, regression coefficient of ELEV and SLOPE in Agusan del Sur is larger
than those of Agusan del Norte. This implies that the influence of elevation and slope in
decreasing the probability of deforestation is more pronounced in Agusan del Norte than
those in Agusan del Sur. However, the contribution of soil quality in decreasing the
probability of deforestation is larger compared to that of Agusan del Norte.
5.5.2 Logistic regression based on socio-economic factors only
The results of binary logistic regression of FCOVER with socio-economic
variables for the Agusan provinces are presented in Table 28 and Figure 36.
109
Table 28. Binary logistic regression of FCOVER versus socio-economic factors for
Agusan del Norte and Agusan del Sur
Socio-economic
Factors
β
Agusan del
Norte
Agusan del
Sur
CHNGROAD -3.19 -3.775
BUILTCHNGE -3.302 3.117
POPDENCHNG 1.074 0.615
DIST_TLAIFMA -1.594 0.684
DIST_CBFMACBRM 2.379 -3.534 *significant at 95% level (p<0.05).
Magnitude of Association of Socio-economic Factors with Deforestation
-3.19 -3.302
1.074
-1.594
2.379
0.792
-3.775
3.117
0.615 0.684
-3.534
-0.397
-5
-4
-3
-2
-1
0
1
2
3
4
5
DISTNEWRD DISTNEWBUILT POPDENCHNG DIST_TLAIFMA DIST_CBFMACBRM Constant
Socio-economic Factors
Logis
tic
Reg
ress
ion
Coef
fici
ent
(B)
Agusan del Norte
Agusan del Sur
Figure 36. Graph showing the magnitude of association of socio-economic factors with
deforestation. Error bars indicate +/- standard error.
In Agusan del Norte, DIST_CBFMA-CBRM and POPDENCHANGE have strong
influence in increasing the probability of deforestation. The positive regression
coefficient for the DIST_CBFMA-CBRM implies that the farther a forested land is to a
110
CBFMA/CBRM, the higher the probability that it will be deforested. The result of the
logistic regression is meaningful in the sense that it provides a confirmation that CBFMA
and CBRMs are restrictive in nature, perhaps in Agusan del Norte; that it follows
regulations in proper planting, harvesting and re-planting of forest lands. Because of this
restrictive nature, the probability of deforestation decreases as the distance of forestlands
to CBFMA and CBRMs decreases, and vice-versa. On the other hand, the regression
coefficient value of DIST_TLA-IFMA showed a negative value. This value indicates that
as DIST_TLA-IFMA increases, the probability of deforestation in Agusan del Norte
decreases; this actually means that forestlands nearer to TLAs and IFMAs have higher
probabilities of being deforested. These results of the logistic regression analysis for
ADN are not surprising albeit some of it is contrary to expectations. Referring back to the
results of the GIS-based characterization of FCOVER vis-à-vis DIST_TLA-IFMA and
DIST_CBFMA-CBRM (Figure 31 and Figure 32), the averaged distance to TLA and
IFMAs of deforested pixels is lesser compared to the averaged distance of unchanged
(retained) forests. This implies that deforestation occurs in areas nearer to TLAs and
IFMAs while unchanged forest areas are far from them. This confirms the result of the
logistic regression that decreasing distance to TLAs and IFMAs increases deforestation.
The same logic applies in the case of the association of FCOVER with DIST_CBFMA-
CRBMs. Distances to CBFMAs and CBRMs in ADN of deforested pixels are, on the
average, greater that of retained forest. The logistic regression results is confirmed by this
where increasing distance to CBFMAs and CBRMs increases the probability of
deforestation. With these confirmations by the GIS-based characterizations, it can be
111
stated that the logistic regression analysis, more or less, describes the deforestation
process in the ADN as far as the location of occurrences is concerned.
On the other hand, the regression coefficient of POPDENCHANGE is indicative
of strong influence of population density change to deforestation in Agusan del Norte.
This result is true because of highly dynamic economy of Agusan del Norte brought
about by the timber industries as well as urbanization that accommodated in-migration,
some of which maybe work forces in logging and timber industries. In the case of
DISTNEWRD and DISTNEWBUILT, regression coefficients for these factors are
indicative of decreasing the probability of deforestation as these factor values increases.
The influences by these two factors are almost similar, signifying that a forestland,
whether near to a new road or to new built-up areas, will likely to be deforested.
In Agusan del Sur, the influence of POPDENCHANGE in increasing the
probability of deforestation is still evident but it is not as strong as to that of Agusan del
Norte. This maybe due to the fact that population growth in this province is low
compared to Agusan del Norte. While the influences of DIST_CBFMA-CBRM and
DIST_TLA-IFMA is opposite in Agusan del Norte (i.e., probability of deforestation is
higher in areas nearer to TLA/IFMAs than those nearer to CBFMA/CBRMs), the same
can also be said in the case of Agusan del Sur. The only difference is that the probability
of deforestation is higher in areas nearer to CBFMA/CBRMs than those nearer to
TLA/IFMAs. This is quite expected especially that deforestation occurrences in ADS
between 1976-2001 are farther from TLAs and IFMA than those of retained forest
(Figure 31). Similarly, deforested pixels in ADS are mostly found in areas nearer to
CBFMAs and CBRMs (Figure 32) compared to unchanged forest that are found farther.
112
Again, the results of the GIS-based characterization complements the results of the
logistic regression analysis. On the other hand, the influence of DISTNEWRD in
decreasing the probability of deforestation is stronger compared to its influence in
Agusan del Norte. The influence of DISTNEWBUILT is opposite to the effect of
DISTNEWRD. DISTNEWBUILT in Agusan del Sur has positive regression coefficient
indicating that the farther an area from new built up, the higher the probability that gets
deforested. This result is true because of the very large land area of Agusan del Sur,
hence the density of built-up is less compared to that of Agusan del Norte.
5.5.3 Logistic regression using combined socio-economic and bio-
physical factors
The results of binary logistic regression of FCOVER with combined bio-physical
and socio-economic variables for the Agusan provinces are presented in Table 29 and
Figure 37.
Table 29. Binary logistic regression of FCOVER versus the combined bio-physical and
socio-economic factors for Agusan del Norte and Agusan del Sur
Combine Bio-physical
and
Socio-economic Factors
Β
Agusan del
Norte
Agusan del
Sur
ELEV -3.155 -8.124
SLOPE -3.247 -1.013
DISTRIV 4.785 1.61
SOILQUAL -0.105 -1.111
DISTNEWRD -3.752 0.513
DISTNEWBUILT -0.836 3.247
POPDENCHNG 1.618 1.38
DIST_TLAIFMA -1.412 2.263
DIST_CBFMACBRM 1.252 -2.986
113
Magnitude of Association of Factors with Deforestation
-3.155 -3.248
4.785
-0.105
-3.752
-0.836
1.618
-1.412
1.252
-8.124
-1.013
1.61
-1.111
0.513
3.247
1.38
2.263
-2.986
1.2011.425
-9
-7
-5
-3
-1
1
3
5
ELEV
SLOPE
DIS
TRIV
SOIL
QU
AL
DIS
TNEW
RD
DIS
TNEW
BU
ILT
POPD
ENCHN
G
DIS
T_T
LAIF
MA
DIS
T_C
BFM
ACBRM
Con
stan
t
Combined Bio-physical and Socio-economic Factors
Lo
gis
tic
Reg
ress
ion
Co
effi
cien
t (B
)
Agusan del Norte
Agusan del Sur
Figure 37. Graph showing the magnitude of association of the combined bio-physical and
socio-economic factors with deforestation.
In Agusan del Norte, DISTRIV has the largest positive regression coefficient,
indicating that its influence in increasing the probability of deforestation increases as its
value increases. The result is similar to the regression analysis based on bio-physical
factors alone. The explanations provided earlier with regards to historical deforestation in
the 1950s to 1970s and the implementation of NIPAS Act of 1992 may be helpful in
qualifying this result. Next to DISTRIV, two socio-economic factors were found to have
influenced deforestation in Agusan del Norte. These are POPDENCHANGE and
DIST_CBFMA-CBRM. It can be observed that the values of regression coefficients for
these two factors are very near to those computed earlier using the socio-economic
factors only-based logistic regression analysis. The results are also consistent with
114
regards to factors that decrease the probability of deforestation: ELEV, SLOPE,
SOILQUAL, DISTNEWRD, DISTNEWBUILT, and DIST_TLA-IFMA. This means that
when the values of these factors increase, the probability of deforestation decreases.
In Agusan del Sur, the major factor with the strongest influence in increasing the
probability of deforestation, as its value increases, is a combination of DISTNEWBUILT
and DIST_TLA-IFMA. One thing noticeable is the change of sign of DISTNEWRD from
negative when computed using the socio-economic factors only that became positive
when computed using the combined bio-physical and socio-economic factors. The
change of its value from positive to negative may have been affected by the roles of
socio-economic factors incorporated during the computation. This result exemplified the
fact that land-cover change is interplay between socio-economic and bio-physical factors.
As for the factors that decreases the probability of deforestation, the results of the logistic
regression is consistent to earlier findings (ELEV, SLOPE, SOILQUAL, and
DIST_CBFMA-CBRM). This means that when the value of these factors increases, the
probability of deforestation decreases and vice versa.
5.5.4 Logistic regression analysis using new set of 5% sample
In order to determine the representativeness of the samples used in the logistic
regression analysis, another set of 5% sample was analyzed. The expected result should
have insignificant difference with the original set of sample analyzed using logistic
regression in order to confirm that the samples used are indeed representative of the study
area. A simple t-Test was employed to determine whether the results of the analysis using
the two sets of samples have significant or insignificant difference.
115
Table 30 and Figure 38 show the results of the comparison done between the
logistic regression coefficients computed using the original and the new 5% samples. t-
Test results for Agusan del Norte (Table 31) showed that there was an insignificant
difference between the regression coefficient, β, of the original samples and the new 5%
samples collected (i.e., t(9)=-0.361, p>0.05, both for one-tailed and two-tailed tests). The
Pearson correlation coefficient between the two sets of samples is 0.987.
Table 30. Comparison between the β values for Agusan del Norte
Agusan del Norte Combined Factors
Factors Original 5% New 5%
ELEV -3.155 -3.204
SLOPE -3.248 -2.609
DISTRIV 4.785 4.741
SOILQUAL -0.105 -0.038
DISTNEWRD -3.752 -3.223
DISTNEWBUILT -0.836 -1.797
POPDENCHNG 1.618 1.614
DIST_TLAIFMA -1.412 -1.199
DIST_CBFMACBRM 1.252 1.354
116
Comparison between original and new 5% samples
Agusan del Norte
-6
-4
-2
0
2
4
6
ELEV
SLOPE
DIS
TRIV
SOIL
QUAL
DIS
TNEW
RD
DIS
TNEW
BUIL
T
POPD
ENCHNG
DIS
T_TLA
IFM
A
DIS
T_CBFM
ACBR
M
Fcators
Re
gre
ss
ion
co
eff
icie
nt
Original
New
Figure 38. Graph showing the comparison between the original and the new 5% samples
in Agusan del Norte. Error bars indicate +/- standard error.
Table 31. t-Test results for Agusan del Norte
t-Test for Agusan del Norte Original Sample New Sample
Mean -0.539 -0.484
Variance 7.718 7.116
Observations 9 9
Pearson Correlation 0.987
Hypothesized Mean
Difference 0
df 8
t Stat -0.361
p value (T<=t) one-tailed 0.363*
t Critical one-tail 1.859
p value (T<=t) two-tailed 0.727**
t Critical two-tail 2.306
*not significant; p>0.05 (one-tailed)
**not significant; p>0.05 (two-tailed)
117
Table 32 and Figure 39 show the comparison of the logistic regression coefficient
values for Agusan del Sur. Although the sign of the regression coefficients of some
factors has changed, the analysis using t-test (Figure 39) revealed that the difference
between the sets of samples is insignificant. The computed Pearson correlation
coefficient computed was 0.826.
Table 32. Comparison between the β values for Agusan del Sur
Agusan del Sur Combined
Factors Original 5% New 5%
ELEV -8.124 -8.674
SLOPE -1.013 -0.705
DISTRIV 1.61 -0.337
SOILQUAL -1.111 -2.338
DISTNEWRD 0.513 0.861
DISTNEWBUILT 3.247 -0.5
POPDENCHNG 1.38 1.779
DIST_TLAIFMA 2.263 2.199
DIST_CBFMACBRM -2.986 0.562
118
Comparison between original and new 5% samples
Agusan del Sur
-10
-8
-6
-4
-2
0
2
4
6
ELEV
SLOPE
DIS
TRIV
SOIL
QUA
L
DIS
TNEWRD
DIS
TNEWBU
ILT
POPD
ENCH
NG
DIS
T_TLAIF
MA
DIS
T_CB
FMAC
BRM
Factors
Reg
ress
ion
co
effi
cien
tOriginal
New
Figure 39. Graph showing the comparison between the original and the new 5% samples
in Agusan del Sur. Error bars indicate +/- standard error.
Table 33. t-Test results for Agusan del Sur
t-Test Results for Agusan del Sur Original Samples New Samples
Mean -0.469 -0.795
Variance 11.952 10.617
Observations 9 9
Pearson Correlation 0.826
Hypothesized Mean Difference 0
df 8
t Stat 0.491
p value (T<=t) one-tailed 0.318*
t Critical one-tail 1.859
p value (T<=t) two-tailed 0.636**
t Critical two-tail 2.306
*not significant; p>0.05 (one-tailed)
**not significant; p>0.05 (two-tailed)
119
5.6 Characterization of “No Data” pixels
This section is a discussion on the limitation on the analysis brought by the “no
data” pixels due to the presence of clouds and cloud shadow in the image of the study
area.
In order to determine as to what factors where biases were made during the
analysis due to the absence of complete land-cover information, the mean factor values of
the “no data” and with data pixels were compared. The comparison revealed that pixels
with “no data” are found in areas with much higher elevation and farther from new roads
and new built-ups (Figure 40). This result was expected because clouds and cloud
shadows are often found in elevated areas in the study area. Consequently, there are no
roads and built-ups that can be found in highly elevated areas. The implication of this is
that: should complete dataset will be available, the factors that might have a change in its
influence or degree of association to deforestation in the study area is ELEV. With
regards to the other factors, “no data” and “with data” pixels are both found in almost the
same mean factor values.
120
Mean factor values of No data and with Data Pixels
(Agusan del Norte and Agusan del Sur)
-5000
0
5000
10000
15000
20000
25000
30000
35000
40000
45000
Elev*
100
slope
*100
DIS
TRIV
SOIL
QUA
L
DIS
T_NEW
RD
DIS
T_NEW
BUIL
T
POPD
ENCH
AN
GE
DIS
T_TLA-I
FMA
DIS
T_CB
FMA-C
BRM
Factors
Mea
n f
acto
r v
alu
es
No data
With data
Figure 40. Graph showing the mean factor values of no data and with data pixels for
Agusan del Norte and Agusan del Sur. Error bars indicate +/- standard error.
5.7 Summary of findings
This research was able to detect deforestation and other types of land-cover
change in the provinces of Agusan del Norte and Agusan del Sur for the 1976-2001
periods. Using state of the art RS image analysis techniques provided by Support Vector
Machine classification algorithm, highly accurate land-cover maps of 1976 and 2001 with
overall accuracy 94.99% and 98.25% respectively, were obtained and used to detect land-
cover transitions in the study area. The land cover change map desired is 93.33%
accurate. The limitations brought by cloud cover contamination in the images were
addressed by a simple cloud masking algorithm developed in this study that was
comprised of image segmentation and maximum likelihood classification.
The detected changes in land-cover were found to be different in the Agusan
provinces. Forest to rangeland is the major land-cover change in Agusan del Norte from
121
1976 to 2001. Although deforestation due to increase in rangeland is significantly
evident, “re-forestation” of rangeland areas from 1976 to 2001 was also present. It was
also observed that large tract of lands planted with palm trees in 1976 have been
converted into croplands in 2001. In the case of Agusan del Sur, increase in cropland and
decrease in forest cover is the most significant land-cover change in terms of change in
land area. Quantitatively, these translate to 156% increase in cropland (or about 198.47
sq. km.) and about 6% decrease in forest cover (or 113.42 sq. km.). In terms of
specificity, the two most prominent land-cover change types from 1976 to 2001 in this
province is the conversion of rangeland to forest and forest to palm trees. Considering
errors in classifications, these two land-cover change types are almost identical in
magnitude. A third major type of change is that of conversion of forest to rangeland.
The forest cover change statistics showed that conversion to rangeland, palm
trees, bare soil and cropland are among the four major contributors to forest cover
reduction in both provinces.
Using GIS, deforestation in the Agusan provinces were characterized with respect
to bio-physical and socio-economic factors that were hypothesized to be the main drivers
of deforestation. GIS-based characterization provided an overview on the relative
relationship of these factors to forest cover retention and reduction. GIS-based
characterization indicates that deforestation in Agusan del Norte are located in areas with
higher elevation and steeper slope compared to Agusan del Sur. Deforestation occurrence
in Agusan del Norte is located much farther from the river (approx. 3km.) compared to
Agusan del Sur where deforestation can be found in areas less than 1km away from the
river. However, with regards to the occurrence of deforestation with respect to the new
122
built-up and new road, it was observed that deforestation in Agusan del Norte are located
in areas nearer to new built-up and new road compared to Agusan del Sur. Deforestation
in Agusan del Norte also occurred in areas with higher population density compared to
Agusan de Sur. This may be due to the fact that population density, built-up and roads in
Agusan del Norte is more dense than in Agusan del Sur. With regards to the occurrence
of deforestation in Agusan del Norte with respect to the TLA-IFMA and CBFMA-CBRM
parcels, it is located in areas farther from TLA-IFMA but nearer to CBFMA-CBRM
parcels compared to Agusan del Sur. This scenario was attributed to the fact that the
largest and the earliest TLA-IFMA concession in the Agusan province were awarded to
Nasipit Lumber Company which is located in Agusan del Norte (see Table 2).
The logistic regression analysis of bio-physical variables as sole factors
influencing deforestation in the Agusan provinces showed promising results. The results
are highly indicative of the inverse relationship of deforestation with elevation and slope
values. Only the factor DISTRIV was found to have positive regression coefficient that
influence deforestation.
The logistic regression analysis of socio-economic variables as sole factors
influencing deforestation in the Agusan provinces also showed promising results. The
results are highly indicative of the inverse relationship of deforestation with distance to
new roads which confirms the fact that accessibility is a major factor in deforestation.
Population density change was found to be a major indicator of forest cover change in
both provinces, with its influence stronger in Agusan del Norte than in Agusan del Sur.
The results of the influence of new road and population density to deforestation is similar
to the results of the studies conducted by Kummer [30], Vagen [45], and Geist &Lambin
123
[22]. With regards to the influence of TLA/IFMAs and CBFMA/CRBMs to deforestation
in Agusan del Norte, increasing distance from TLAs/IFMAs indicates decrease in the
probability of deforestation; while increasing distance to CBFMA/CBRM indicates
increase in the probability of deforestation. The exact opposites were found in the case of
Agusan del Sur.
The results of logistic regression based on combined bio-physical and socio-
economic factors provided significant results as to what has influenced deforestation in
the Agusan provinces. For Agusan del Norte, the bio-physical factor DISTRIV was found
to be positively related to deforestation, followed by socio-economic factors
POPDENCHANGE and DIST_CBFMA-CBRM. Compared to DISTRIV, the
contributions of these two socio-economic factors are minimal. The bio-physical factors
ELEV and SLOPE, and the socio-economic factors DISTNEWRD and DIST_TLA-
IFMA, were all found to be negatively related to deforestation, thus, the probability of
deforestation decreases as the values of these factors increases. Although the socio-
economic factor DISTNEWBUILT was found to be a contributor to deforestation, its
effect is minimal. For Agusan del Sur, DISTNEWBUILT, DIST_TLA-IFMA,
POPDENCHANGE, and DISTRIV are found to be positively related to deforestation (i.e.
when there value increases, the probability of deforestation increases). Although the
socio-economic factor DISTNEWRD was found to be positively related to deforestation,
its effect is minimal. ELEV, SLOPE, SOILQUAL, and DIST_CBFMA-CBRM were the
factors that decrease the probability of deforestation (i.e. negatively related to
deforestation).
124
The results of logistic regression using the combined bio-physical and socio-
economic factors emphasized the fact that deforestation is the result of the interplay
between socio-economic, institutional and environmental factors [1]. Running the
logistic regression using only the bio-physical or the socio-economic factors may not
provide adequate description of their influence to deforestation in the Agusan Provinces.
The combined approach re-affirmed initial findings based on the logistic regression using
bio-physical or socio-economic factors alone that the most prominent factor that is positively
related to deforestation in Agusan del Norte and Sur are DISTRIV and DISTNEWBUILT,
respectively. On the other hand, the prominent factors that are negatively related to
deforestation in Agusan del Norte are DISTNEWRD, ELEV and SLOPE. In Agusan del Sur,
the factor negatively related to deforestation are ELEV and DIST_CBFMA-CBRM.
The result of the logistic regression analysis using another set of 5% sample revealed
that the original set of 5% sample used in logistic regression analysis indeed represents the
characteristic of the study area. Should complete dataset will be available, the factors that
might have a change in its influence to the deforestation in the study area is ELEV.
125
Chapter 6
Conclusions and Recommendations
6.1 Conclusions
This study has provided a comprehensive analysis of land-cover change in the
Agusan provinces. A series of methodology was developed to understand the history of
forest utilization using RS and GIS. This study also provided a series of techniques to
understand deforestation as well as to relate to bio-physical and socio-economic factors
using an un-ideal dataset. This study was able to establish a technique that has bridged
the gap in RS problems. The results of this study made sense although datasets used to
derive land-cover information were contaminated with clouds. Thus, it is possible to
analyze deforestation using cloud contaminated RS images.
Results of this study exemplified the fact that although the two Agusan provinces
have history of forest resource industry, the presence of these industries is not the most
prominent factor that influenced deforestation. The most prominent factors that
influenced deforestation in Agusan del Norte is the socio-economic factor DISTNEWRD.
Its influence was found to be negatively correlated with deforestation: the nearer an area
to a newly-constructed road, the higher the probability it will be deforested. Conversely,
the influence of the biophysical factor DISTRIV was found to be positively correlated
126
with deforestation in ADN: the farther the area to a river, the higher the probability of
deforestation.
In Agusan del Sur, the biophysical factor ELEV has the greatest influence with
deforestation: low-lying areas are more prone to deforestation. This factor is seconded by
the socio-economic factor DIST_CBFMA-CBRM: forested areas nearer to CBFMAs and
CBRMs are more prone to be deforested. The factors DISTNEWBUILT, DIST_TLA-
IFMA and DISTRIV were found to be positively correlated with deforestation: increasing
distance from these factors increases the probability of deforestation.
The influences of timber industries to deforestation in the Agusan provinces were
found to be different. In ADN, DIST_TLA-IFMA is negatively correlated (increasing
distance decreases deforestation) while DIST_CBFMA-CRBM is positively correlated
(increasing distance increasing deforestation). The exact opposite was found in ADS.
As the factors associated with deforestation vary in ADN and ADS, it is
concluded that the factors influencing deforestation in one area may not be the same
factors that can influence deforestation in another area. Deforestation is indeed a
combination and the interplay between several bio-physical and socio-economic factors.
This study demonstrated the usefulness of RS and GIS not only in obtaining
accurate information on the location, extent and type of land-cover change (including
deforestation) but also in characterizing the relationships of the detected changes with
bio-physical and socio-economic factors. With statistical analysis, the information and
the characterizations can be expounded further leading to a more comprehensive analysis
of the deforestation process.
127
6.2 Recommendations
This study was sorely limited by the incomplete land-cover information from the
Landsat images due to cloud cover contamination. Hence, results of the analysis should
be used carefully.
Although a simple cloud masking algorithm was developed, it did not avoid the
problem of missing land-cover information due to the masking process. Moreover, it is
admitted that this study may not provide a complete analysis of the deforestation process
in the Agusan provinces. In this regard, continuation of this study is hereby
recommended. The use of the developed cloud masking algorithm is also recommended
because of its ease-of-use and accuracy. The problem of cloud cover contamination may
be addressed by producing a cloud-free mosaic from multi-temporal remotely-sensed
images obtained at closer time interval.
The use of Support Vector Machine algorithm is also recommended for land
cover classification as it could provide very accurate land cover maps provided that it is
implemented with greater number of bands (at least 7) and sufficient number of training
pixels.
Results of this study already gave fair description on the interplay of the bio-
physical and socio-economic factors that are associated with deforestation in the Agusan
provinces. Local agencies in Agusan del Norte and Agusan del Sur may use the land-
cover maps and statistics generated in this study to further evaluate the process of
deforestation in these provinces in order to create and evaluate strategies that attempt to
mitigate its negative effects.
128
References
[1] J. Lesschen, P. Verburg, and S. Staal, Statistical methods for analysing the spatial
dimension of changes in land use and farming systems. LUCC Report Series No. 7,
Co-published by the International Livestock Research Institute, Nairobi, Kenya and
LUCC Focus 3 Office, Wageningen University, The Netherlands.: 2005.
[2] APN, Land-use and Land-cover Change in Southeast Asia: A Synthesis Report,
Asia Pacific Network for Global Change Research, Kobe, Japan, Earth Observation
Center, Universiti Kebangsaan Malaysia: 2001.
[3] E. Irwin and J. Geoghegan, “Theory, data, methods: developing spatially explicit
economic models of land-use change,” Agriculture, Ecosystems and Environment,
vol. 85, 2001, pp. 7-23.
[4] Critical Ecosystems Partnership Fund, “An Overview of CEPF’s Portfolio in the
Philippines Hotspot: Creating Corridors and Alliances,” Medan, Indonesia: Asia
Forum, 2005.
[5] D. Paler, T. Shinohara, R. del Castillo, and I. Nomura, “Development of industrial
tree plantation in the Administrative Region No. 10 in Mindanao Island,
Philippines: with two case studies.,” Journal of Forest Research, vol. 3, 1998, pp.
193-197.
[6] R. Rapera, V. Fernandez, Villaflor, A.A., E. Corpuz, M. Villanueva, F. Bacani, P.
Caraan, and S. Garnace, Rationalizing the Wood Industry, Los Banos, Laguna:
Forest Development Center, College of Forestry, University of the Philippines,
1987.
[7] Southeast Asia Sustainable Forest Management Network, Upland Philippine
Communities: Guardians of the Final Forest Frontiers, California, USA: Center for
Southeast Asia Studies, International and Area Studies, University of California,
1993.
[8] E. Tandem, “Conflict over Land-Based Natural Resources in the ASEAN
Countries,” Conflict Over Natural Resources in South-East Asia and the Pacific,
Lim Tech Ghee and M. Valencia, Eds., Tokyo, Japan: United Nations University
Press, 1990.
[9] Asian Development Bank, Technical Assistance to the Republic of the Philippines
for a Master Plan for the Agusan River Basin, Mandaluyong City, Philippines:
Asian Development Bank, 2004.
[10] H.J. Geist and E.F. Lambin, “Proximate causes and underlying driving forces of
tropical deforestation,” BioScience, vol. 52, 2002, pp. 143–150.
[11] D.O. Fuller and R.R. Chowdhury, “Monitoring and modelling tropical deforestation:
Introduction to the Special Issue,” Singapore Journal of Tropical Geography, vol.
27, 2006, pp. 1-3.
[12] A. Apan and J. Peterson, “Probing tropical deforestation: the use of GIS and
statistical analysis of georeferenced data,” Applied Geography, vol. 18, 1996, pp.
137-152.
129
[13] M. Emch, J.W. Quinn, M. Peterson, and M. Alexander, “Forest Cover Change in the
Toledo District, Belize from 1975 to 1999: A Remote Sensing Approach,” The
Professional Geographer, vol. 57, 2005, pp. 256–267.
[14] P. Coppin, I. Jonckheere, K. Nackaerts, and B. Muys, “Digital change detection
methods in ecosystem monitoring: a review,” International Journal of Remote
Sensing, vol. 25, 2004, pp. 1565–1596.
[15] W. Muttitanon and N.K. Tripathi, “Land use/land cover changes in the coastal zone
of Ban Don Bay, Thailand using Landsat 5 TM data,” International Journal of
Remote Sensing, vol. 26, 2005, pp. 2311–2323.
[16] J.F. Mas, H. Puig, J.L. Palacio, and A. Sosa-López, “Modelling deforestation using
GIS and artificial neural networks,” Environmental Modelling & Software, vol. 19,
May. 2004, pp. 461-471.
[17] T. Fung, “An assessment of TM imagery for land cover change detection,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 28, 1990, pp. 681-684.
[18] J. Wilson and G. Lindsey, “Socioeconomic correlates and environmental impacts of
urban development in a Central Indiana landscape,” Journal of Urban Planning and
Development, vol. 3, 2005, pp. 159-169.
[19] T.B. Moya and B.S. Malayang III, “Climate variability and deforestation–
reforestation dynamics in the Philippines,” Environment, Development and
Sustainability, vol. 6, 2004, pp. 261-277.
[20] A. Grainger, Controlling Tropical Deforestation, Earthscan Publications, 1993.
[21] B. Verbist, D. Putra, A. Eka, and S. Budidarsono, “Factors driving land use change:
Effects on watershed functions in a coffee agroforestry system in Lampung,
Sumatra,” Agricultural Systems, vol. 85, 2005, pp. 254–270.
[22] H.J. Geist and E.F. Lambin, What Drives Tropical Deforestation? A Meta-Analysis
of Proximate and Underlying Causes of Deforestation Based on Subnational Case
Study Evidence, Louvain-la-Neuve, Belgium: LUCC International Project Office,
2001.
[23] E.F. Lambin, H.J. Geist, and E. Lepers, “Dynamics of land-use and land-cover
change in tropical regions,” Annual Reviews of Environment and Resources, vol.
28, 2003, pp. 205-241.
[24] E.F. Lambin, B. Turner II, H.J. Geist, S. Agbola, and A. Angelsen, “The causes of
land-use and land-cover change: moving beyond the myths,” Global Environmental
Change, vol. 11, 2001, pp. 261-269.
[25] P.H. Verburg and A. Veldkamp, “Projecting land use transitions at forest fringes in
the Philippines at two spatial scales,” Landscape Ecology, vol. 19, 2004, pp. 77–98.
[26] P.H. Verburg, K.P. Overmars, M.G. Huigen, W.T. de Groot, and A. Veldkamp,
“Analysis of the effects of land use change on protected areas in the Philippines,”
Applied Geography, vol. 26, 2006, pp. 153–173.
[27] D. Garrity, D. Kummer, and E. Guiang, “The Philippines,” Sustainable Agriculture
and the Environment in the Humid Tropics, National Research Council Committee
on Sustainable Agriculture and the Environment in the Humid Tropics, Ed.,
Washington D.C., USA: National Academy Press, 1993, pp. 549-624.
[28] D. Liu, L. Iverson, and S. Brown, “Rates and patterns of deforestation in the
Philippines: application of geographic information system analysis,” Forest Ecology
and Management, vol. 57, 1993, pp. 1-16.
130
[29] G.M. Van den Top, The Social Dynamics of Deforestation in the Sierra Madre,
Philippines, Amsterdam, The Netherlands: NIASS Press, 1998.
[30] D. Kummer, Deforestation in the Post-War Philippines, Ann Arbor, Michigan,
USA: University Microfilms International, 1990.
[31] I. Coxhead, “Consequences of a food security strategy for economic welfare,
income distribution and land degradation: The Philippine Case,” World
Development, vol. 28, 2000, pp. 111-128.
[32] I. Coxhead, A. Rola, and K. Kim, “How do national markets and price policies
affect land use at the forest margin? Evidence from the Philippines,” Land
Economics, vol. 77, 2001, pp. 250-267.
[33] I. Coxhead, G. Shively, and X. Shuai, “Development policies, resource constraints,
and agricultural expansion on the Philippine land frontier,” Environment and
Development Economics, vol. 7, 2002, pp. 341-363.
[34] G.E. Shively, “Agricultural change, rural labor markets, and forest clearing: an
illustrative case from the Philippines,” Land Economics, vol. 77, 2001, p. 268.
[35] L. Bruzzone, R. Cossu, and G. Vernazza, “Detection of land-cover transitions by
combining multidate classifiers,” Pattern Recognition Letters, vol. 25, 2004, pp.
1491–1500.
[36] D. Yuan, C.D. Elvidge, and R.S. Lunetta, “Survey of multispectral methods for land
cover change analysis,” Remote sensing change detection: Environmental
Monitoring Methods and Applications, Michigan, USA: Ann Arbor Press, 1998, pp.
21-39.
[37] J.F. Mas, “Monitoring land-cover changes: a comparison of change detection
techniques,” International Journal of Remote Sensing, vol. 20, 1999, pp. 139–152.
[38] M.K. Ridd and J. Liu, “A comparison of four algorithms for change detection in an
urban environment,” Remote Sensing of Environment, vol. 63, 1998, pp. 95–100.
[39] A. Singh, “Digital change detection techniques using remotely-sensed data,”
International Journal of Remote Sensing, vol. 10, 1989, pp. 989–1004.
[40] J.R. Jensen, “Digital change detection,” Introductory Digital Image Processing: A
Remote Sensing Perspective, New Jersey: Prentice-Hall, 2004, pp. 467-494.
[41] L. Bruzzone and S.B. Serpico, “An iterative technique for the detection of land-
cover transitions in multitemporal remote-sensing images,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 35, 1997, pp. 858–867.
[42] F. Yuan, K.E. Sawaya, B.C. Loeffelholz, and M.E. Bauer, “Land cover
classification and change analysis of the Twin Cities (Minnesota) Metropolitan Area
by multitemporal Landsat remote sensing,” Remote Sensing of Environment, vol.
98, 2005, pp. 317–328.
[43] J. Richards and X. Jia, Remote Sensing Digital Image Analysis: An Introduction,
Berlin, Germany: Springer, 1999.
[44] J. Helmschrot and W. Flügel, “Land use characterisation and change detection
analysis for hydrological model parameterisation of large scale afforested areas
using remote sensing,” Physics and Chemistry of the Earth, vol. 27, 2002, pp. 711-
718.
[45] T.G. Vagen, “Remote sensing of complex land use change trajectories—a case
study from the highlands of Madagascar,” Agriculture, Ecosystems and
Environment, vol. 115, 2006, pp. 219–228.
131
[46] R. Chowdhury, “Landscape change in the Calakmul Biosphere Reserve, Mexico:
Modeling the driving forces of smallholder deforestation in land parcels,” Applied
Geography, vol. 26, 2006, pp. 129–152.
[47] P.H. Swain and S.M. Davis, Remote sensing: the quantitative approach, McGraw-
Hill New York, 1978.
[48] J. Benediktsson, P.H. Swain, and O.K. Ersoy, “Neural network approaches versus
statistical methods in classification of multisource remote sensing data,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 28, 1990, pp. 540–552.
[49] J. Otukei and T. Blaschke, “Land cover change assessment using decision trees,
support vector machines and maximum likelihood classification algorithms,”
International Journal of Applied Earth Observation and Geoinformation, vol. 12,
Feb. 2010, pp. S27-S31.
[50] J. Rogan, J. Franklin, D. Stow, J. Miller, C. Woodcock, and D. Roberts, “Mapping
land-cover modifications over large areas: A comparison of machine learning
algorithms,” Remote Sensing of Environment, vol. 112, May. 2008, pp. 2272-2283.
[51] D.L. Civco, “Artificial neural networks for land-cover classification and mapping,”
International Journal of Geographical Information Systems, vol. 7, 1993, p. 173.
[52] T. Kavzoglu and I. Colkesen, “A kernel functions analysis for support vector
machines for land cover classification,” International Journal of Applied Earth
Observation and Geoinformation, vol. 11, Oct. 2009, pp. 352-359.
[53] H. Nemmour and Y. Chibani, “Multiple support vector machines for land cover
change detection: An application for mapping urban extensions,” ISPRS Journal of
Photogrammetry and Remote Sensing, vol. 61, 2006, pp. 125–133.
[54] C. Huang, L.S. Davis, and J.R.G. Townshend, “An assessment of support vector
machines for land cover classification,” International Journal of Remote Sensing,
vol. 23, 2002, pp. 725–749.
[55] J. Paola and R. Schowengerdt, “A detailed comparison of backpropagation neural
network and maximum-likelihood classifiers for urban land use classification,”
IEEE Transactions on Geoscience and Remote Sensing, vol. 33, 1995, pp. 981-996.
[56] V. Vapnik, The Nature of Statistical Learning Theory, New York, USA: Springer-
Verlag, 1995.
[57] V. Vapnik, Statistical Learning Theory, New York, USA: Wiley, 1998.
[58] T. Joachims, C. Nedellec, and C. Rouveirol, “Machine Learning: ECML-98 10th
European Conference on Machine Learning, Chemnitz, Germany,” 1998, pp. 137–
142.
[59] Jin Shuying, Li Deren, and Wang Jingwen, “A comparison of support vector
machine with maximum likelihood classification algorithms on texture features,”
Proceedings of the IEEE International Geoscience and Remote Sensing Symposium,
vol. 5, 2005, pp. 3717-3720.
[60] M. Pal and P.M. Mather, “Support vector machines for classification in remote
sensing,” International Journal of Remote Sensing, vol. 26, 2005, pp. 1007–1011.
[61] J.A. Gualtieri and R.F. Cromp, “Proceedings of SPIE- The International Society for
Optical Engineering,” 1999, pp. 221–232.
[62] E. Hietel, R. Waldhardt, and A. Otte, “Analyzing land-cover changes in relation to
environmental variables in Hesse, Germany,” Landscape Ecology, vol. 19, 2004,
pp. 473-489.
132
[63] E. Helmer, “Forest conservation and land development in Puerto Rico,” Landscape
Ecology, vol. 19, 2004, pp. 29-40.
[64] A. Shalaby and R. Tateishi, “Remote sensing and GIS for mapping and monitoring
land cover and land use changes in the northwestern coastal zone of Egypt,” Applied
Geography, vol. 27, 2007, pp. 28-41.
[65] H. Jung, D. Lee, S. Jeon, and W. Song, “Analysis of deforestation patterns in
Baekdudaegan preservation area using land cover classification and change
detection techniques: the feasibility of restoration,” Landscape Ecology and
Engineering, vol. 1, 2005, pp. 177-190.
[66] L. Porter-Bolland, E.A. Ellis, and H.L. Gholz, “Land use dynamics and landscape
history in La Montaña, Campeche, Mexico,” Landscape and Urban Planning, vol.
82, 2007, pp. 198–207.
[67] A. Veldkamp and L.O. Fresco, “CLUE-CR: an integrated multi-scale model to
simulate land use change scenarios in Costa Rica,” Ecological Modelling, vol. 91,
1996, pp. 231–248.
[68] C. Köbrich, T. Rehman, and M. Khan, “Typification of farming systems for
constructing representative farm models: two illustrations of the application of
multi-variate analysesin Chile and Pakistan,” Agricultural Systems, vol. 76, 2003,
pp. 141-157.
[69] J. Davis, Statistics and Data Analysis in Geology, New York, USA: John Wiley &
Sons, 1986.
[70] E. Hietel, R. Waldhardt, and A. Otte, “Statistical modeling of land-cover changes
based on key socio-economic indicators,” Ecological Economics, vol. 62, 2007, pp.
496-507.
[71] S.W. Menard, Applied Logistic Regression Analysis, Thousand Oaks, California,
USA: Sage Publications, 2002.
[72] E. Weiss, S.E. Marsh, and E.S. Pfirman, “Application of NOAA-AVHRR NDVI
time-series data to assess changes in Saudi Arabia's rangelands,” International
Journal of Remote Sensing, vol. 22, 2001, pp. 1005–1027.
[73] E. López, G. Bocco, M. Mendoza, and E. Duhau, “Predicting land-cover and land-
use change in the urban fringe: A case in Morelia City, Mexico,” Landscape and
Urban Planning, vol. 55, 2001, pp. 271-285.
[74] P.H. Verburg and Y.Q. Chen, “Multiscale characterization of land-use patterns in
China,” Ecosystems, vol. 3, 2000, pp. 369–385.
[75] C. Wood and D. Skole, “Linking satellite, census, and survey data to study
deforestation in the Brazilian Amazon,” People and Pixels: Linking Remote Sensing
and Social science, D. Liverman, E. Moran, R. Rindfuss, and P. Stern, Eds.,
Washington D.C., USA: National Academy Press, 1998.
[76] S. Serneels and E.F. Lambin, “Proximate causes of land-use change in Narok
District, Kenya: a spatial statistical model,” Agriculture, Ecosystems &
Environment, vol. 85, 2001, pp. 65–81.
[77] L.C. Schneider and G. Pontius, “Modeling land-use change in the Ipswich
watershed, Massachusetts, USA,” Agriculture, Ecosystems & Environment, vol. 85,
2001, pp. 83–94.
[78] J. Geoghegan, S. Villar, P. Klepeis, M. Mendoza, Y. Ogneva-Himmelberger, R.R.
Chowdhury, B. Turner II, and C. Vance, “Modeling tropical deforestation in the
133
southern Yucatán peninsular region: Comparing survey and satellite data,”
Agriculture, Ecosystems and Environment, vol. 85, 2001, pp. 25-46.
[79] P.H. Verburg, J.R. van Eck, T.C. de Nijs, M.J. Dijst, and P. Schot, “Determinants of
land-use change patterns in the Netherlands,” Environment and Planning B, vol. 31,
2004, pp. 125–150.
[80] I. Dohoo, C. Ducrot, C. Fourichon, A. Donald, and D. Hurnik, “An overview of
techniques for dealing with large numbers of independent variables in
epidemiologic studies,” Preventive Veterinary Medicine, vol. 29, 1997, pp. 221-
239.
[81] FMB, Philippine Forestry Statistics, Manila, Philippines: Forest Management
Bureau, Department of Environment and Natural Resources, 2000.
[82] W. Osita, “The Philippines' community-based resource management program,”
Annals of Tropical Research, vol. 25, 2003, pp. 19-24.
[83] ITT VIS, ENVI 4.4 User's Guide, Boulder, Colorado, USA: ITT Visual Information
Solutions, 2007.
[84] ESRI, Arcview GIS 3.2, Redlands, California, USA: Environmental Systems
Research Institute, Inc., 1999.
[85] R. Schowengerdt, Remote Sensing: Models and Methods for Image Processing, San
Diego, California: Academic Press, 1997.
[86] R.R. Irish, “Landsat 7 Science Data Users Handbook,” National Aeronautics and
Space Administration, 2000.
[87] B. Markham and J. Barker, Landsat MSS and TM post-calibration dynamic ranges,
exoatmospheric reflectances and at-satellite temperatures, EOSAT Landsat
Technical Notes No. 1, Lanham, Maryland, USA: 1986.
[88] F. van der Meer, “Spectral mixture modelling and spectral stratigraphy in carbonate
lithofacies mapping,” ISPRS Journal of Photogrammetry and Remote Sensing, vol.
51, 1996, pp. 150–162.
[89] J. Van Genderen, B. Lock, and P. Vass, “Remote sensing: statistical testing of
thematic map accuracy,” Remote Sensing of Environment, vol. 7, 1978, pp. 3-14.
[90] S. Franklin, “Terrain analysis from digital patterns in geomorphometry and Landsat
MSS spectral response,” Photogrammetric Engineering & Remote Sensing, vol. 53,
1987, pp. 59-65.
[91] A. Elumnoh and R. Shrestha, “Application of DEM data to Landsat image
classification: evaluation in a tropical wet-dry landscape of Thailand,”
Photogrammetric Engineering & Remote Sensing, vol. 66, 2000, pp. 297-304.
[92] H. Su, E. Kanemasu, M. Ransom, and S. Yang, “Separability of soils in a tallgrass
prairie using SPOT and DEM data,” Remote Sensing of Environment, vol. 33, 1990,
pp. 157-163.
[93] A. Fahsi, T. Tsegaye, W. Tadesse, and T. Coleman, “Incorporation of digital
elevation models with Landsat-TM data to improve land cover classification
accuracy,” Forest Ecology and Management, vol. 128, 2000, pp. 57-64.
[94] R. Congalton, “A review of assessing the accuracy of classifications of remotely
sensed data,” Remote Sensing of Environment, vol. 37, 1991, pp. 35-36.
[95] J. Anderson, E. Hardy, J. Roach, and W. Witmer, “A Land Use and Land Cover
Classification System For Use With Remote Sensing Data,” USGS Professional
Paper 964, 1976, pp. 138-145.
134
[96] J. Millington, G. Perry, and R. Romero-Calcerrada, “Regression techniques for
examining land use/cover change: a case study of a Mediterranean landscape,”
Ecosystems, vol. 10, 2007, pp. 562-578.
[97] B. Pradhan and A. Youssef, “Manifestation of remote sensing data and GIS on
landslide hazard analysis using spatial-based statistical models,” Saudi Society of
Geosciences, 2009.
[98] R. Senga, “Establishing protected areas in the Philippines,” The George Wright
FORUM, vol. 18, 2001, pp. 56-65.
135
Appendices
136
Appendix 1. Maps showing the location of retained and deforested areas.
Figure A1. 1 Map showing the retained and
deforested areas in Agusan del Sur
Figure A1. 2 Map showing the retained and
deforested areas in Agusan del Norte
137
Appendix 2. Factor Maps
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
SRTM ELEVATION
Factor Map
Elevation,above mean sea level
1 - 26 meters
26 - 42
42 - 67
67 - 108
108 - 174
174 - 280
280 - 452
452 - 731
731 - 1,180
1,180 - 1,907
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
SLOPE (%)
Factor Map
Slope, in %
0 - 3
3 - 6
6 - 8
8 -18
18 - 32
32 - 50
>50
Figure A1. 4 The SLOPE Factor Map Figure A1. 3 The ELEV Factor Map
138
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
DISTANCE TO
RIVERS
Factor Map
Rivers
Distance to Rivers
0 - 1 Kilometers
2 - 2
3 - 3
4 - 5
6 - 7
8 - 10
11 - 14
15 - 20
21 - 28
29 - 38
39 - 52
53 - 70
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
DISTANCE TO NEW
ROADS SINCE 1976
Factor Map
1976 Roads
New Roads (since 1976)
Distance to Road
0 - 1 Kilometer
1 - 3
3 - 5
5 - 8
8 - 12
12 - 18
18 - 26
26 - 36
36 - 50
50 - 70
Figure A1. 5 The DISTNEWRD Factor Map Figure A1. 6 The DISTRIV Factor Map
139
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
DISTANCE TO NEW
BUILT-UP AREAS
SINCE 1976
Factor Map
1976 Built-up Areas
New Built-up since 1976
Distance to Built-up Areas
0 - 1 Kilometer
1 - 3
3 - 5
5 - 8
8 - 12
12 - 18
18 - 26
26 - 36
36 - 50
50 - 70
LA PAZ
LORETO
SAN LUIS
BAYUGAN
ESPERANZA
BUNAWAN
VERUELA
BUTUAN CITY
TRENTO
ROSARIO
LAS NIEVES
PROSPERIDAD
CABADBARAN
BUENAVISTA
SAN FRANCISCO
JABONGA
SANTIAGO
KITCHARAO
CARMEN
NASIPIT
TUBAY
STA. JOSEFA
TALACOGON
MAGALLANES
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
POPULATION DENSITY
CHANGE
Factor Map*
:.mk .qs rep snosrep fo .oN
21 - 6
71 - 31
32 - 81
23 - 42
54 - 33
36 - 64
78 - 46
221 - 88
171 - 321
042 - 271
* as per May 1975 and May 2000
National Census of the
Philippine Government.
Figure A1. 8 The DISTNEWBUILT Factor Map Figure A1. 7 The POPDENCHANGE Factor Map
140
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
DISTANCE TO
TLAs & IFMAs
Factor Map
TLAs & IFMAs
Distance
0 - 1 Kilometer
2 - 2
3 - 4
5 - 7
8 - 11
12 - 16
17 - 24
25 - 34
35 - 49
50 - 70
126°1'0"E
126°1'0"E
125°30'30"E
125°30'30"E
9°0
'0"N
9°0
'0"N
8°3
0'0
"N
8°3
0'0
"N
8°0
'0"N
8°0
'0"N
®25
Kilometers
Agusan del
Norte
Agusan del
Sur
DISTANCE TO
CBFMAs & CBRMs
Factor Map
CBFMAs & CBRMAs
Distance
0 - 1
2 - 2
3 - 4
5 - 7
8 - 11
12 - 16
17 - 24
25 - 35
36 - 50
51 - 71
Figure A1. 9 The DIST_TLA-IFMA Factor Map Figure A1. 10 The DIST_CBFMA-CBRM Factor Map
141
Appendix 3. Maps showing the location of retained forest and deforested areas with CBFMA, CBRM,
TLA, and IFMA.
Figure A1. 11 TLA and IFMA locations with retained
forest and deforested areas in the Agusan provinces
Figure A1. 12 CBFMA and CBRM locations with retained
forest and deforested areas in the Agusan provinces
142
Appendix 4. Maps showing the distance to new roads of retained forest and deforested areas.
Figure A1. 13 Distance to new roads of 5% retained
forest samples
Figure A1. 14 Distance to new roads of 5% deforested
samples
143
Appendix 5. Maps showing the distance to river of retained forest and deforested areas.
Figure A1. 15 Distance to rivers of 5% deforested areas
samples
Figure A1. 16 Distance to rivers of 5% retained
forest samples