Deforestation Study of Phnom Aural Cambodia Using Change Detection Techniques

GEOG 4440/ENVS 4521
Term Paper
Deforestation Study of Phnom Aural, Cambodia,
Using Change Detection Techniques
Submitted by: Katrina Gordon
Student # 210380053
Submitted to: Gang Hong
Lab session: 02
Due Date: April 8, 2013

Introduction
The rate of deforestation in Cambodia is one
of the highest in the world. In 1969, Cambodia's
primary rainforest was estimated to encompass 70%
of its land mass. As of 2007, it dropped to 31%
(Cambodia 2013). Phnom Aural is the tallest peak of
the Cardamom Mountains located in central
Cambodia west of the Mekong River and is a region
that has experienced a rapid rate of deforestation
despite being a protected area (Mansfield and Parnell
2009). The specific study region consists of Phnom
Aural the mountain region north east of the study
region, the major city of Phnom Penh and extending
to Prey Veaeng the city east of the Mekong River as
seen in Figure 1. Cambodia consists of three main
types of forest: relatively open deciduous forest;
closed semi-deciduous forest; and closed coniferous
forest (Piazza 2007). The Phnom Aural Region of
Cambodia consists of a mix of deciduous and
coniferous forests. The dominant trees, which can
reach a height of 20 m, include oaks such as Litho-
carpus spp, Quercus cambodiensis and Castanopsis
cambodiana (Cambodia 2013).
Deforestation in Cambodia can be attributed
to a variety of sources and time periods. Historically,
the Vietnam War (1965-1975) was a major
contributor to deforestation as the U.S army sprayed
“Agent Orange” to remove foliage and reveal the
opposition (Phat et al. 1998). This was particularly
devastating to areas around the Mekong river where
approximately 2 million hectares of forest were
destroyed (Phat et al. 1998). Additionally, the rapid
increase of population has continuously led to the
increase of demands for biofuels such as firewood
and agricultural lands. Logging is also a major
industry in Cambodia. In 1999, the Cambodian
government issued a declaration on Measures to
Management of Forests and the Elimination of Illegal
Forest Activities (Energy, Environment and
Resources Chatham House 2013). This declaration
only allows tree harvesting by licensed companies
and individuals and prohibits harvesting on national
parks and protected land such as Phnom Aural.
Abstract
This study takes place in the Phnom Aural Region of Cambodia in South Asia at the coordinates of 12.0333°
N, 104.1667°E for the years 1990, 2003 and 2010. The Phnom Aural Region has been characterized by high rates of
deforestation and Cambodia in general has been recognized as having one of the highest deforestation rates globally.
The causal factors of deforestation in this region have been attributed to the Vietnam War with the use of Agent
Orange, agricultural purposes, population growth and illegal logging (Phat, Ouk & Ueki, 1998). The objective of this
study is to monitor deforestation concentrated in and around the Phnom Aural Region of Cambodia north-west of
Phnom Penh in addition to draw conclusions on the effectivity of forestry laws in Cambodia through the applications
of unsupervised classification remote sensing techniques in addition to, the execution of a Matrix analysis (MAT) and
a statistical Local Area Reports to detect change in the region. The study ultimately found unsupervised classification
methods were effective in studying vegetation changes however, the use of canopy structure could improve the level
of detail of the study and would be more indicative of the type of forestry that exists.

Figure 1. Study Area: Phnom Aural, Cambodia

However, the implementation of policies and
regulatory frameworks is hampered by a lack of
institutional capacity and illegal logging continues to
be an issue which questions the effectiveness of the
legislation (Energy, Environment and Resources
Chatham House 2013).
There are a variety of remote sensing
applications used to study forestry. Both
unsupervised and supervised classifications are
evident in a variety of forestry studies and are often
studied over a period of time. The alterations over a
specified time period are then understood better
through change detection techniques such as local
area reports which quantify land classification
statistics. Furthermore, a more sophisticated analysis
can be determined using a Matrix Analysis which
determines where exactly changes in the landscape
are occurring and by a statistical percentage of land.
There are numerous examples of studies using
classification techniques to map land classes
specifically in the use of forestry. For example,
Matuyama and Yamaguchi analyzed deforestation in
Mato Grosso, Brazil in 2010 caused by soya bean
plantations and relied on an unsupervised
classification with an NDVI classification to view
changes to the region over the span of 1991 to 2009.
Kumar et al in 2010 applied a supervised
classification technique to monitor deforestation in
Ranchi, Jharkland, India in addition to relying on a
statistical area report to describe how much of the
land was changing over that period. Classification
using ISODATA was used in a study conducted by
Nijomo (2008) who used ISODATA which calculates
class means evenly distributed in the data space then
iteratively clusters the remaining pixels using
minimum distance techniques (Geo Hul 2008). The
study performed an unsupervised classification of the
Congo Basin using ISODATA and then mapped
areas of change. These methods proved effective and
less time consuming in forestry studies that were not
able to get in situ measurments. Going beyond
specific image processing techniques, Defourny et al
(1993), identifies other key factors to ensure accuracy
of conclusions especially in tropically sensitive
regions subjected to wet and dry seasons. Defourny et
al studied land cover change in Cambodia and
discusses seasonal changes in Cambodia that affect
data collection. For example heavy cloud cover is
often seen through May to October as a result of their
Monsoon season making images from November to
the following March optimal images in addition to
avoiding seasonal additions of vegetation in images
skewing results.
The purpose of this study is to assess the
effectivity of Cambodia’s deforestation laws through
spatial analysis of deforestation in Phnom Aural

region of Cambodia. The use of unsupervised
ISODATA classifications will be relied on to define
areas of vegetation and where they exists throughout
the years of 1990, 2003 and 2010 to track patterns of
deforestation. Local Area reports in addition to a
Matrix Analysis (MAT) will be relied on as a form of
change detection to better understand these patterns.
This study is important as forestry studies of
Cambodia are rather limited despite the high rate of
deforestation.
Methods
Data was acquired using Global Land
Survey images from United States Geological Survey
(USGS). After researching the worst areas of
deforestation in Cambodia Phnom Aural was selected
and entered in the USGS search engine. Phnom Aural
is located at 12.0333° N, 104.1667°E.While the years
1975, 1990, 2000, 2003 and 2010 were retrieved,
only 1990, 2003 and 2010 could be accepted as the fit
within the same Path and Row (WRS Path 126, Row
52) and were images taken in Cambodia’s dry season
(November to March) which was important to
maintain an accurate and fair representation of
vegetation conditions. Meta data on Phnom Aural
images used can be found in the appendix i. Climatic
conditions are important to consider due to the wet
monsoons in this tropical region and would affect
results based on seasonal vegetational changes.
Figure 2, depicts the image processing techniques
which began with image retrieval of the Phnom Aural
region from 1990 (Figure 3), 2003 (Figure 4) and
2010 (Figure 5) were overlaid and clipped so that the
extent of the region that is being studied is consistent
throughout the years. Atmospheric correction was
applied using ATCOR in PCI Geomatica to help
reduce atmospheric noise and enhance the images.
Band combinations were set to 4,5,2 to contrast
differences in vegetation so that the degree of
photsynthetically active areas of vegetation display in
shades of deep red (Quinn 2001). The varying
shades of vegetation allowed for a better
classification of the vegetation. Simply knowing
where vegetation exists is trivial in a tropical region.
Being able to differentiate between healthy
vegetation and sparse regrowth allowed for a better
analysis. The band combination of 4,5,2 was also
used as a reference in carrying out the unsupervised
classification. The unsupervised classification was
applied using ISODATA to classify with a maximum
of 20 classes and 20 iterations. Classification
involved five different land classes: Dense
Vegetation, Sparse Vegetation, Bare Soil/Urban,
Water, and Clouds. Originally, Soil and Urban

Figure 2. Image processing model.

Figure 3. 1990 True colour image of Phnom Aural,
Cambodia. This image depicts Phnom Aural as
region characterized by a green and highly vegetated
environment.
True Colour Images of Phnom Aural,
Cambodia
Cambodia. This image depicts Phnom Aural with a
much more bare environment.
Cambodia. This image depicts Phnom Aural as being
similar to the 2003 conditions just slightly
accelerated.

had been classified separately, however, due to the
amount of dirt roads, many urban areas were
classified as bare soil that actually made up urban
areas Phnom Penh, Cambodia’s Capital city. For the
purpose of this study, in that deforestation is the
target variable, both bare soil and urban were
classified under one class.
To assess the accuracy, of classification, two
methods were relied on. First, a pan-sharpened image
was used as a reference image for the 2003 image of
Phnom Aural, Cambodia which had increased clarity
through a higher pixel resolution of 15 meters in
comparison to the 30 meter pixel resolution in the
classified images. Pan-sharpening of remote sensing
multispectral imagery has been found to directly
influences the accuracy of interpretation of
classification (Makaru et al. 2012). Pan-sharpened
ETM+ images such as the 2003 image of Phnom
Aural, have been found to enable better
discrimination of finer change detail than the original
ETM+ multispectral images (Makaru et al. 2012).
This is because a pan-sharpen image incorporates the
spatial detail features present in the panchromatic
image and missing in the initial multispectral one
(Makaru et al. 2012). The second method used true
colour images as the reference for the 1990 and 2010
images as a visual comparison this is because the
1990 and 2010 images were not supplied with a
panchromatic band. Pan-sharpening is a more widely
accepted version of an accuracy assessment however,
in the absence of the panchromatic band, a secondary
accuracy assessment method was needed. Neither
methods define a true accuracy assessment as they
are not referencing data that is truly determined such
as in situ data but in the absence of in situ data, they
provide a secondary method to assist in better
classifying the images. In both methods, a set of 100
random points were reassessed to determine of
classification was accurate represented.
In terms of change detection, a Local Area
report was relied on which quantifiably produced
area statistics in terms of percentage of land
classified in addition to square kilometers. To further
understand deforestation in the Phnom Aural region,
a Matrix Analysis was used to determine where
exactly change was occurring in addition to how
much vegetative land was being logged or turn into
urban areas.
Results
Figures 6, 7, 8 depict the unsupervised
classifications of Phnom Aural, Cambodia. From
1990 to 2003 it is clear that there is a progressive
amount of deforestation experienced. From 2003 to
22010 the deforestation shows much less progression
but an increase none the less. Table 1, quantifies

Figure 8. 2010 Unsupervised Classification of
Phnom Aural, Cambodia. This image depicts Phnom
Aural as having a high reduction in vegetation
however, it is evident that the reduction has slowed.
Aural with a dramatically reduced appearance of
vegetation.
Aural as region with copious amounts of vegetation.
Unsupervised Classifications of Phnom
Aural, Cambodia

Area % Total Area (km2) Area % Total Area (km2) Area % Total Area (km2)
CombinedVegetation 78.8 17615.63 CombinedVegetation 62.73 14023.28 CombinedVegetation 50.43 11274.5
Soil/Urban 17.2 3845.61 Soil/Urban 33.9 7578.01 Soil/Urban 45.71 10218.42
Water 3.5 782.25 Water 3.38 754.68 Water 3.85 860.92
Clouds 0.5 112.48 Clouds 0 0 Clouds 0.01 2.13
Total 100 22355.97 Total 100 22355.97 Total 100 22355.97
1990 20102003
these changes into the percentage taken up by the
land class and the square kilometers. The report was
generated by the Local Area Report in PCI.
MAT was also used to identify where the
changes of vegetation are occurring in the Phnom,
Aural Region of Cambodia. Figures 9, 10 and 11 map
out where strictly vegetation that has changed into a
clear cut area or an urban feature shown in grey.
Discussion
The results would indicate that deforestation
in the Phnom Aural region is progressing with the
most rapid rate occurring in the earlier period of
1990-2003. From 2003 to 2010 there is a slower
progression of deforestation. This also corresponds
with the enactment of the forestry laws from 1999
and onwards making it plausible that although the
legislation did not comprehensively halt deforestation
it did prove to be effective enough to slow the
progress. This can be confirmed in the Local Area
Report statistics. In 1990, Phnom Aural and the
surrounding region consisted of 78.8% vegetation
and dropped almost 16.07% to 62.73%. From 2003 to
2010 vegetation dropped another 12.3%. Overall
from the period of 1990 to 2010 vegetation was
reduced by 28.37%.
The MAT analysis provided some very
interesting information of where change was
occurring. It seems that the protected parks while not
free from deforestation and therefore illegal logging,
have stayed the most intact such a Phnom Aural, the
upper left area of the study region. It seems like in
addition to logging, it is urban sprawl from the main
city Phnom Penh and even east into Prey Veaeng that
experience the most amount of vegetation change.
From 1990 to 2003 the application of MAT
discovered that 65% of the combined vegetation
classes was changed into the Soil/Urban class. From
2003 to 2010 43% of the vegetation from 2003 was
Table 1. Local Area Report Statistics for the Phnom Aural, Cambodia images from 1990, 2003, 2010 respectively.
Dense and Sparse vegetation were combined to make an overall statistic due to the degree of subjectivity involved
in the visual analysis in determining dense vegetation from sparse vegetation which could have led to areas of
overlap and therefore misinterpretation.

Deforestation Study of Phnom Aural Cambodia Using Change Detection Techniques

transformed into the Soil/Urban class. This is
generally seen in the outskirts of the major cities
Phnom Penh and Prey Veaeng in the study site.
When looked at from an overall vegetation change
from 1990 to 2010, 88% of the original vegetation
from 1990 was changed into the Soil/Urban class.
This is highly indicative of deforestation and
confirms that Cambodia has been experience a very
high rate of deforestation. The legislation has
certainly reduced the rate at which Cambodia’s
forests are being lost, however, it is limited to the
protected areas.
In terms of the Accuracy of each classified
image, each image demonstrated a high degree of
accuracy. The 1990 image was found to have an
overall accuracy of 87% and a Kappa Statistic of
0.79 which would signify a substantial agreement
between the producers and users accuracy. Despite
the pan-sharpened image having a higher pixel
resolution, it was found to be the least accurate with
an overall accuracy of 86% however, the Kappa
Statistic was also 0.79. This could be due to the
overall colour change that resulted after the image
was pan-sharpened making it more difficult to
visually assess what land class was truly present. The
2010 Image was found to have the highest degree of
the overall accuracy being 91% and a Kappa Statistic
of 0.86 signifying an almost perfect agreement. The
higher accuracy is most probably due to the severe
degree of deforestation drawn by strict boundaries of
forestry parks such as Phnom Aural. Again, it is
important to remember that the accuracy assessment
involved in all three images cannot be defined as a
true accuracy assessment so these statistics would
likely be subject to change if in situ data were
presented. In all cases, the dense and sparse
vegetation were more easily confused as seen in the
error matrix appendix iii, confirming the importance
of combining the vegetation groups to discuss an
overall deforestation. Complete results of accuracy
statistics can be seen in appendix iv.
Conclusion
It is evident that the Phnom Aural and
surrounding region of Cambodia has experienced
very rapid rates of deforestation especially around the
major cities. Since the implementation of forestry
laws in Cambodia in the late 90’s, deforestation has
slowed making a case that such implementation of
legislation has assisted to reduce the deforestation
experienced however, there is still a considerable
amount of deforestation occurring and further action
to enforce forestry legislation should be pursued.
This study provided information on overall levels of
vegetation changes in the Phnom Aural, Region of
Cambodia but could be improved by implementing a

canopy analysis of Cambodia’s forestry. Cambodia
naturally is a tropical environment so even in the
presence of logging, vegetation to a lesser degree
could still exist. Therefore, studying tree canopies of
vegetated areas would allow for a better indication of
the forestry structure and whether old growth or
primary forest are being clear cut or altered into
urban areas. Overall, using an unsupervised
classification determined by ISODATA, proved to
be an effective and less labour intensive strategy to
study deforestation where in situ data is absent
however, final results should be taken in a more
general context to understand the overall changes that
are occurring rather than as an exact measurement.

Data Set Attribute Attribute Value
Entity ID P126R052_4X19891115
Acquisition Date 15-Nov-89
Satellite Number Landsat4
Sensor ID TM
Datum WGS84
Zone Number 48
Orientation NUP
Resampling Technique CC
Product Size 329.1
File Size 143156301
Sun Azimuth 137.63
Sun Elevation 50.05
WRS Path 126
WRS Row 52
Center Latitude 11°33'10.31"N
Center Longitude 104°45'16.80"E
NW Corner Lat 12°25'39.12"N
NW Corner Long 104°05'16.27"E
NECorner Lat 12°11'01.97"N
NECorner Long 105°45'02.62"E
SW Corner Lat 10°55'05.15"N
SW Corner Long 103°45'46.19"E
SECorner Lat 10°40'36.99"N
SECorner Long 105°25'03.37"E
Center Latitude dec 11.5528643
Center Longitude dec 104.7546671
NW Corner Lat dec 12.4275333
NW Corner Long dec 104.0878538
NECorner Lat dec 12.1838819
NECorner Long dec 105.7507289
SW Corner Lat dec 10.9180979
SW Corner Long dec 103.7628318
SECorner Lat dec 10.6769423
SECorner Long dec 105.417604
1990 Data Set Attribute Attribute Value
Landsat Scene Identifier LE71260522003038SGS00
Sensor Mode N/A
Station Identifier SGS
Day/Night DAY
WRS Path 126
WRS Row 52
Date Acquired 07/02/2003
Start Time 2003:038:03:08:26.0288125
StopTime 2003:038:03:08:53.1304374
Image Quality VCID 1 9
Image Quality VCID 2 9
CloudCover 0
CloudCover QuadUpper Left 0
CloudCover QuadUpper Right 0
CloudCover QuadLower Left 0
CloudCover QuadLower Right 0
Sun Elevation 48.6229095
Sun Azimuth 130.0988312
NECorner Lat 105°45'21.24"E
2003
Full Aperture Calibration N
Gain Band 1 H
Gain Band 2 H
Gain Band 3 H
Gain Band 4 L
Gain Band 5 H
Gain Band 6 VCID1 L
Gain Band 6 VCID 2 H
Gain Band 7 H
Gain Band 8 L
Browse Exists Y
Data Category NOMINAL
Data Type Level 1 ETM+ L1T
Elevation Source GLS2000
Output Format GEOTIFF
Ephemeris Type DEFINITIVE
Corner UL Latitude Product 12.52881 (12°31'43.72"N)
Corner UL Longitude Product 103.65901 (103°39'32.44"E)
Corner UR Latitude Product 12.53092 (12°31'51.31"N)
Corner UR Longitude Product 105.81550 (105°48'55.80"E)
Corner LL Latitude Product 10.62746 (10°37'38.86"N)
Corner LL Longitude Product 103.66804 (103°40'04.94"E)
Corner LR Latitude Product 10.62924 (10°37'45.26"N)
Corner LR Longitude Product 105.81001 (105°48'36.04"E)
Panchromatic Lines 14021
Panchromatic Samples 15621
Reflective Lines 7011
Refelective Samples 7811
Thermal Lines 7011
Thermal Samples 7811
Ground Control Points Model 59
Geometric RMSEModel 5.692
Geometric RMSEModel X 2.573
Geometric RMSEModel Y 5.077
Gain Change Band 1 HH
Gain Change Band 4 LL
Gain Change Band 6 VCID 1 LL
Gain Change Band 6 VCID 2 HH
Gain Change Band 8 LL
Map Projection L1 UTM
Datum WGS84
Ellipsoid WGS84
UTM Zone 48
Grid Cell Size Panchromatic 15
Grid Cell Size Reflective 30
Grid Cell Size Thermal 30
Orientation NORTH_UP
Resampling Option CUBIC_CONVOLUTION
Appendix i

Data Set Attribute Attribute Value
Entity ID LT51260522009014BKT00
Acquisition Date 14/01/2009
WRS Path 126
WRS Row 52
Satellite Landsat5
Zone Number 48
Datum WGS84
Resampling Technique CC
Orientation NUP
Scene Size 179827506
Product Type L1T
Sun Azimuth 136.8860938
Sun Elevation 45.0801299
GapFill Percent
GapFill Acquisition Date
Registration Acquisition Date
NECorner Lat 12°13'51.13"N
NECorner Long 105°48'41.94"E
2010
Class Name Dense Vegetation Sparse Vegetation Soil/Urban Water Clouds Total
Dense Vegetation 31 2 0 0 0 33
Sparse Vegetation 2 8 0 0 0 10
Soil/Urban 2 3 45 0 0 50
Water 0 0 0 7 0 7
Clouds 0 0 0 0 0 0
Total 35 13 45 7 0 100
2010 Error Matrix
Dense Vegetation 17 2 o 0 0 19
Soil/Urban 0 2 44 0 0 46
Water 3 0 0 2 0 5
Clouds 0 0 0 0 0 0
Total 25 27 46 2 0 100
2003 Error Matrix
Dense Vegetation 37 6 0 0 0 43
Soil/Urban 0 0 14 0 14
Water 0 0 0 2 0 2
Clouds 0 0 0 0 0 0
Total 44 40 14 2 0 100
1990 Error Matrix
Appendix ii Appendix iii

Class Name Producers Accuracy User's Accuracy
Dense Vegetation 84.09 86.05
Sparse Vegetation 85 82.93
Soil/Urban 100 100
Water 100 100
Clouds 0 0
Overall Accuracy 87%
Overall Kappa Statistic 0.79%
1990 Accuracy Statistics
Dense Vegetation 68 89.47
Sparse Vegetation 85.19 76.67
Soil/Urban 95.65 95.65
Water 100 40
Clouds 0 0
Dense Vegetation 88.57 93.94
Sparse Vegetation 61.54 80
Soil/Urban 100 90
Water 100 100
Clouds 0 0
Appendix iv

References
Cambodia. 2013. In Encyclopædia Britannica. Retrieved from
http://www.britannica.com/ EBchecked/topic/
90520/Cambodia
Cambodia. 2013. About Cambodia: Geography. Retrieved
from http://www.cambodiaun.org /index.php?option=
com_content&view=category&layout=blog&id=36&
Itemid=23
Defourny, P., U, Pradhan, and S. Shrestha. 1993. A land
cover monitoring system for Cambodia: Preliminary
results. Proceedings of IGARSS '93 - IEEE
International Geoscience and Remote Sensing
Symposium, (6)4: 1685-7.
Energy, Environment and Resources Chatham House. 2013.
Illegal Logging, Cambodia. Retrieved from
http://www.illegallogging.info/approach.php?a_id=8
Geol Hu. 2005. Vegetation indices. Retrieved from
http://geol.hu/data/online_help/Vegetation_Indic
es.html
Makarau, A., G. Palubinskas, and P. Reinartz. 2012. Analysis
and selection of pan-sharpening assessment
measures. Journal of Applied Remote Sensing
(6)4:380-401.
Mansfield, N and T. Parnell. 2009. Helping communities
protect Cambodia’s biodiversity. The East-West
Management Institute, NY: New York.
Maruyama, M., and Y. Yamaguchi. 2010. Analysis of
deforestation in Mato Grosso using multi-temporal
landsat TM imageries. 2010 IEEE International
Geoscience and Remote Sensing Symposium
(IGARSS 2010):367-70.
Panday, P. K., and B. Ghimire. 2012. Time-series analysis of
NDVI from AVHRR data over the Hindu Kush-
Himalayan region for the period 1982-2006.
International Journal of Remote Sensing, 33(21):
6710-6721.
Phat, K., S. Ouk, and T. Ueki. 1998. An outline of the causes
of deforestation in Cambodia. Journal for Nature
Conservation 18(7), 232-254.
Piazza, M. 2007. Brief on national forest inventory: Cambodia.
Forestry Department of Food and Agriculture
Organization of the United Nations. Retrieved from
http://www.fao.org /docrep/016/ap183e/ap183e.pdf
Quinn, J. 2001. Band combinations. Retrieved from
http://web.pdx.edu/~emch/ip1/bandcombinations.html
Zhang, L., M. Liao, L. Yang. and H. Lin. 2007. Remote
sensing change detection based on canonical
correlation analysis and contextual bayes decision.
Photogrammetric Engineering and Remote Sensing,
73(3):311-318.

Deforestation Study of Phnom Aural Cambodia Using Change Detection Techniques

More Related Content

What's hot (19)

Viewers also liked (19)

Similar to Deforestation Study of Phnom Aural Cambodia Using Change Detection Techniques (20)

Deforestation Study of Phnom Aural Cambodia Using Change Detection Techniques