Change Detection Analysis on Support Vector

Machine Classifier of Land use and Land cover

Changes (Case Study on Yangon)
Khin Mar Yee1, Mu Mu Than2, Kyi Lint3, Aye Aye Oo4, Chan Mya Hmway5, Khin Zar Chi Winn6

al. (1994) indicates there are variations in the observed cause to

cover relationships at different levels.

Abstract—The dynamic changes of Land Use and Land Cover

(LULC) changes in Yangon have generally resulted the improvement
of human welfare and economic development since the last twenty
years. Making map of LULC is crucially important for the sustainable
development of the environment. However, the exactly data on how
environmental factors influence the LULC situation at the various as a result of different socioeconomic characteristics, politics,
scales because the nature of the natural environment is naturally levels of affluence, and technological development as well as
composed of non-homogeneous surface features, so the features in the culture in different parts of the world. The rapid urban
satellite data also have the mixed pixels. The main objective of this expansions have caused LULC changes, which affect the
study is to the calculation of accuracy based on change detection of ecological environmental process at local and regional levels,
LULC changes by Support Vector Machines (SVMs). For this research especially the urban heat island [3] [4]. Urbanization is believed
work, the main data was satellite images of 1996, 2006 and 2015.
one the most prevalent anthropogenic causes of the losing
Computing change detection statistics use change detection statistics
to compile a detailed tabulation of changes between two classification
arable land, devastating habitats, and the decline in natural
images and Support Vector Machines (SVMs) process was applied vegetation cover [5]. As a result, the simplistic assumption that
with a soft approach at allocation as well as at a testing stage and to LULC is driven by only a few forces has moved to a
higher accuracy. The results of this paper showed that vegetation and complicated understanding that involves interactions among a
cultivated area were decreased (average total 29 % from 1996 to 2015) large number of factors at different spatio-temporal scales [6].
because of conversion to the replacing over double of the built up area LULC information is essential for managing natural resources
(average total 30 % from 1996 to 2015). The error matrix and and monitoring of environmental changes [7]. Human activities
confidence limits led to the validation of the result for LULC mapping. are one the most widespread causes of the decrease forest land
Keywords— Land use and Land cover change, change detection,
and habitat destruction and decline the natural environment.
image processing, Support Vector Machines.
GIS (Geographic Information Systems) and RS (remote
sensing) are great helpful and cost-effective tools for assessing
the distribution of spatial and temporal dynamics of LULC [8].
I. INTRODUCTION
Research on LULC change is increasingly becoming important
Remote sensing data provide valuable multi- change due to the fact that it is a primary factor for global change,
detection temporal data on the processes and patterns of LULC reason being its interactions with biodiversity, biogeochemical
change, and GIS is useful for mapping and analyzing these cycles, climate, human activities and ecosystem processes [9].
patterns [1]. Image processing and making map of LULC is a This research work was considered the change detection of
key application of remote sensing data. Very updated land use support vector machine classifier on land use and land cover
and land cover information is needed at local, regional and dynamic changes of Yangon area.
national administrative levels for land use planning and
managements. Global environmental change and management I. STUDY AREA
practices in the urban environment is closely related to the
distribution of LULC characteristics. LULC change is most Yangon City was the former capital of Myanmar and lying
important in understanding environmental change; between 16˚ 41' 00'' N to 17˚ 5' 30'' N latitudes and 95 ˚ 59' 30''
necessitating the investigation of cause to cover relationship. E and 96˚ 27' 30'' E longitudes (Fig. 1). In this paper, our
LULC change and its relation to cause is very important as it research interest focused on the Yangon City area covering an
has the greatest implications on the environment [2]. Turner et area of approximately 633 square kilometer including water
body.

1 5
Associate Professor and Head, Geospatial Technology Center, Department of Assistant Lecturer, Department of Electronics, Technical University
Geography, Dagon University (Thanlyin)
2 6
Associate Professor, Geospatial Technology Center, Department of Geography, Deputy Staff Officer, Information Technology Directorate, Internal Revenue
Dagon University Department, Ministry of Finance
3
Professor and Head, Department of Geography, Dagon University
4
Professor, Department of Archeology, Dagon University
 accuracy assessment based on the post field work data
collection and ground truth.

B. Georeferencing and Taking Training Sample

Landsat images have been georeferenced by the using Survey

of Myanmar toposheet, 1696-01. Survey of Myanmar
topographic maps was much helpful to prepare the base maps,
and geocoding the satellite imageries in the case of digital
image interpretation. Training sites were the areas defined for
each LULC types of the image. The chosen colour compostie
was used for selecting pixcels around each training site for the
similar land cover. After that, a unique identifier defined an
assigned to each known land cover type [10] (Eastman, 2009).
Taking training sample for this paper, about 100 training
samples were selected for each LULC class. These training
samples were as pure as possible and their locations over the
images. More than one training site for each type has been
Fig. 1 Location of Study Area identified for making the land cover images. Therefore, at the
locations were ground truth data or training data was selected
II. DATA AND METHODOLOGY
To prepare the base maps for analysis purposed with Areas of Interest (AOI) were created.
applying the different methods to achieve the research
objectives, Landsat satellite images (1996, 2006 and 2015) C. Image Band Combination
were obtained from the official website of U.S. Geological
Image enhancement is a kind of image modification that
Survey (USGS) and used in order to reach the research
enables the capabilities of human vision to identify and select
objectives.
regions of interests [11]. Landsat 5 TM records 7 spectral bands
and Landsat 8 records 11 bands. For visual purpose any 3 bands
were combined that were acting a False Colour Composite
(FCC). The FCC of RGB was combined with bands 4, 3 and 2
for 1996 and 2006; bands 5, 4 and 3 for 2015.

D. Defining LULC Classes

A modern nation, as a modern business, must have adequate

information on many complex interrelated aspects of its
Fig. 2 Satellite Images L5 TM (1996, 2006) and L8 OLI (2015) activities in order to make decisions.
TABLE I
DESCRIPTION OF LULC CLASSIFICATION
The used images satellite for this study, the follow criteria
were considered for the applying satellite images should be
Land cover Type Description
enough time to compare detecting the LULC and LST changes
and less than 10 percent cloudless condition. For the processing Built up area All infrastructure-residential, commercial mixed use
of Landsat Thematic Mapper (Landsat 5 TM) dated March 25, and industrial areas, villages, settlements, road
1996 (0 % cloud cover) and March 5, 2006 (8% cloud cover) network, pavements, and man-made structures
and Landsat 8 Operational Land Imager (OLI) dated March 14,
Cultivated land Crops fields, pastures, orchards, vineyards and
2015 (3.56 % cloud cover) were used very good quality. The
nurseries
study area is located in the Landsat path 132 and row 48 for
Vegetation Trees, natural vegetation, mixed forest, gardens,
Landsat 5 and Landsat 8 OLI. The Universal Transverse gardens, parks and playgrounds, grassland, vegetated
Mercator (UTM) projection (within Zone 47 North) and World lands
Geodetic System (WGS)-1984 datum were applied to the image
Barren land Scatter Vegetation and mixed vegetation which has a
(Fig. 2). scattered distribution, mostly shrubs and rangeland
A. Generating of Land use and Land cover Classification
The procedures of this topic consisted of three phases. The
River and stream Water body of river and stream area
first phase was a pre field work including collection of training
samples. The second phase was image processing of LULC and Lake and pond Water body of permanent open water, lakes, ponds,
the third phase was the data analysis of the change detection and canals, seasonal wetland, low lying areas, mushy land,
and swamps
 LULC has become increasingly important as the national [12].
plans to overcome the problems of haphazard, uncontrolled SVM is basically a linear learning machine based on the
development, deteriorating environmental quality etc,. One of principle of optimal separation of classes. The aim is to find a
the prime prerequistes for the better use of land is information linear separating hyperplane that separates classes of interest.
on existing LULC patterns and changes in land use through The hyperplane is a plane in a multidimensional space and is
time. The six classes of land cover classification considered for also called a decision surface or an optimal separating
the study area are built up, cultivated land, vegetation, barren hyperplane or an optimal margin hyperplane. The linear
land, river and stream and lake and pond. The detail description separating hyperplane is placed between classes in such a way
of LULC classification can be seen in Table 1 in detail. that it satisfies two conditions. First, all the data vectors that
belong to the same class are placed on the same side of the
E. Supervised Classification of LULC hyperplane. Second, the distance or margin between the closest
data vectors in both the classes is maximized [13] [14].
Supervised classification is approached that the user Statistically, the optimal boundaries should be generalized to
develops statistical description for various known land use and unseen samples with least errors among all possible boundaries
land cover categories that is called signature development. separating the classes, therefore minimizing the confusion
Then, the procedure is used to identify the similarlar pixels for between classes. SVMs have often been found to provide higher
various land use and land cover types for the image of study clssification accuracies than other widely used pattern
area. The procedure that are followed for this supervised recogniton accuracies than other widely used patten recognition
classification are as follow: techniques, such as tha maximum likelihood and the multilaer
Image classification refered to grouping image pixels into perceptron neural network classifiers. Therefore, SVMs have
categories or classes to produce a thematic representative. For often been found to provide higher clssification accuracies than
this paper, about 100 training samples were selected for each other widely used pattern recogniton accuracies than other
image. These training samples were as pure as possible and widely used patten recognition techniques, such as tha
their locations was maintained over the three images. Image maximum likelihood and the multilaer perceptron neural
classification was made over band 2, 3 and 3 for 1996 and 2006, network classifiers. For the image processing of LULC
and 3, 4 and 5 for 2015 because they were found to be the ones classification, SVM classifier assisted to good LULC classes to
that best discriminated considered classes. For the making map calculate the change detection algorithm.
of LULC classification, the LULC spatial distribution pattern
was mapped by supervised classification with the Support F. Change Detection Technique of LULC
Vector Machine (SVM).
SVMs, first introduced by Boser et al., 1992 and discussed in LULC change detection is very essential for better
more detail by Vapnik, 1995, 1998, have their roots in statistical understanding of landscape dynamic during a known period of
learning theory whose goal is to create a mathematical time having sustainable management. LULC changes is wide
framework for learning from input training samples with known spread and accelerating process, mainly driven by natural
identify and predict the outcome of data points with unknown phenomena and anthropogenic activities, which in turn drive
identity. changes that would impact natural ecosystem [15].
For this paper, LULC classes processed SVM of supervised Understanding landscape patterns, changes and interactions
classification. Because, SVM approach has some advantages between human activities and natural phenomenon are essential
compared to others classifiers. SVM is a classification system for proper land management and decision improvement. Today,
derived from statistical learning theory. It separated the classes earth resource satellites data are very applicable and useful for
with a decision surface that maximizes the margin between the LULC change detection studies [16] [17]. For this research
classes. The surface is often called the optimal hyperplane, and work, the results of change detection algorithm were good
the data point closes to the hyperplane are called support driver to continue for the change matrix and accuracy
vectors. The support vectors are the critical elements of the assessment.
training set. While SVM is a binary classifier in its simplest
form. It can function as a multiclass classifier by combining G. Accuracy Assessment of LULC
several binary SVM classifiers. ENVI classic’s implementation
of SVM uses the pairwise classification strategy for multiclass For image classification purpose, a fisher supervised
classification. SVM classification output is the decision values classification method has been used. Then after achieving
of each pixel for each class, which are used for probability satisfactory accuracy results, the base maps have been finalized.
estimates. The probability values, stored in ENVI Class as the On this basis an error matrix was produced for each result
rule images, represent “true” probability in the sense that each presenting the overall accuracy, the users and producer’s
probability falls in the range of 0 to 1, and the sum of these accuracy as well as the kappa coefficient. For all classes, the
values for each pixel equal 1. ENVI Classic performs accuracy assessment was done by generating stratified random
classification by selecting the highest probability. The points for the classified images, using the 50 points per each
probability values, SVM technique also shows greater ability to LULC classes were taken.
generalize and greater likelihood of generating good classifiers
 Most of the expansion of built up area extended to all the
direction especially to the northern side along the Yangon-
III. RESULTS AND DISCUSSION Mandalay and Yangon-Magway highway roads and
transportation roads.
All of the results were the post image processing stage of
LULC classification. LULC was and still continue to make one The area of water body was contained river and stream, and
of the environmental issues mostly tightly linked to other lake and pond which were a nearly stable area and changed a
processing of physical, social and ecomomical even climate little. Barren land was fluctuated because of the fallow land of
changes in a complex manner. The following views were spatial cultivated area and construction and project sites of the Yangon
distribution of LULC changes in 1996, 2006 and 2015 and the environment. Table 2 showed the area coverage and represented
description of change detection metrix and gain and loss percentage of the different type of LULC classes. In 1996, built
between various LULC classes. up area was 118.02 sq km (18.64 %) and increased to 217.48 sq
km (34.36 %) in 2006, and elevated again to 307.77 sq km
A. Analysis on Spatial Distribution of LULC Pattern (48.62 %) in 2015.
The cultivated land of 1996 was 345.21 sq km (54.54 %)
Fig. 3 was monitored the spatial distribution patterns of decreased to 272.01 sq km (42.97 %) in 2006 and went down to
various LULC categories between 1996, 2006 and 2015. The 215.16 sq km (33.99 %) in 2015. Although the vegetation area
visual view can be found clearly the built up area was increased, was 91.31 sq km (14.42%) in 1996, 79.86 sq km (12.62 %) in
and vegetation and cultivated area were decreased from 1996 to 2006 and only left 42.63 sq km (6.73%) in 2015. Barren land
2015. The results show that the built up area was wider and was 19.82 sq km (3.13%) in 1996 to 0.66 sq km (0.1%) in 2006
wider from 1996 to 2015. At the same time, the vegetation area and 5.19 sq km (0.82%) in 2015. The area of water body of river
was lower to between 1996 and 2015. Cultivated land was and stream was fluctuated from 1996 to 2015. The area of river
reduced and replaced to the built up area especially residential and stream demonstrated 52.65 sq km (8.32 %), 55.04 sq km
areas because of horizontal urban expansion as the new satellite (8.7 %) and 47.81 sq km (7.55%) for three periods. The area of
townships. lake and pond was very small but wider and wider from 1996
to 2015.

Fig. 3 Spatial Distribution of LULC in Yangon (a) 1996, (b) 2006 and (c) 2015

TABLE 2
AREA OF LULC CLASSES OF YANGON (1996, 2006 AND 2015) (unit : sq km and %)

1996 2006 2015

Area % Area % Area %
LULC
Built up area
118.02 18.64 217.48 34.36 307.77 48.62
Cultivated land
345.21 54.54 272.01 42.97 215.16 33.99
Vegetation
91.31 14.42 79.86 12.62 42.63 6.73
Barren land
19.82 3.13 0.66 0.10 5.19 0.82
River and stream
52.65 8.32 55.04 8.70 47.81 7.55
Lake and pond
5.99 0.95 7.94 1.25 14.43 2.28
B. Analysis on the Spatial Temporal Changes of LULC overall accuracy, kappa coefficient, and the overall kappa value
was calculated.
The calculation of the change area based on the change For this study, selected pixels to ensure that the correct clss
detection algorithm of the LULC classes are distinguished by was identified for each pixel. The classified images were then
Table 3. According to table, the built area, cultivated land and assessed for accuracy based on a random selection of 900
vegetation were rapidly and clearly change from 1996 to 2015. reference pixels for each time period.
Built up area was raised up 99.46 sq km in 1996-2006 and 90.29 Overall accuracy was computed by dividing the total
sq km in 2006-2016. The total change of the area from 1996 to corrected pixels by the total number of pixels in the error
2015 was 189.75 sq km. The cultivated land loosed about 73.2 matrix. This statistic indicates the probability of a reference
sq km in 1996-2006 and 56.85 sq km in 2006-2015 and total pixel was being correctly classified and is the measurement of
changed area was about 130.05 sq km of 1996-2015. The omission error. In addition, the training areas were verified by
change area of the vegetation was 11.45 sq km for 1996-2006 references data. After the post classification comparison change
and 37.23 sq km in 2006-2015 and total change was 48.68 sq detection algorithm was selected to detect changes in LULC
km for 1996-2015 respectively. The other LULC class of barren from 1996, 2006 and 2015 in study area. For the accuracy
land, river and stream, and lake and pond area were very little assessment, the 50 samples for each land use class were selected
change within three periods. using random method to check ground survey and present
different LULC classes of the study area.
Fig. 4 presented the situation of changes area of the various The overall accuracies for 1996, 2006 and 2015 were found
LULC pattern of 1996, 2006 and 2015. Among the LULC 93.33 percent, 94.56 percent and 97.11 percent respectively.
classes, build up area was increased much amount and the other The accuracy assessment shows clearly that all of the results
LULC classes were decreased in all three periods. The graph fulfil an overall accuracy higher than 85 percent. The total
showed clearly that built up area was clearly changed among overall accuracy of the classified images for 1996, 2006 and
the other LULC types. 2015 was 95 percent. The omission error was done error that
consists of not doing something or not including something
such as an amount sample point.
TABLE 3 Error of commission represented a mistake that consists of
LULC CHANGE AREA OF YANGON FOR THREE PERIOD (unit : sq km)
doing something wrong, such as a wrong amount, or including
LULC 1996-2006 2006-2015 1996-2015 an amount in the wrong place. Kappa coefficient is a measure
Built up area 99.46 90.29 189.75 of the proportional (or percentage) improvement by the
Cultivated land -73.20 -56.85 -130.05 classifier over a purely random assignment to classes. The
Vegetation measurement of the Kappa statistics resulted 0.928, 0.95 and
-11.45 -37.23 -48.68
0.96.
Barren land -19.16 4.53 -14.63
River and stream 2.39 -7.23 -4.84 V CONCLUSION
Lake and pond 1.95 6.49 8.44 Nowadays, Yangon will be undergoing major alterations in
its economic, social and infrastructures accompanied with
country’s changes. According to Yangon vision 2040 with the
current growth rate, Yangon city will become 10 million
population and turn into one of the world’s megacities [18].
This study was investigated change detection analysis on the
SVM classification of LULC by using by multi-temporal
remotely sensed data and statistical analysis, this paper
presented better results for the spatiotemporal patterns of LULC
changes with the fewer remotely sensed data images focusing
on the relatively the Yangon area.
In this work, it has been proven, that the SVM of supervised
classification of multi-temporal satellite image is an effect tool
to quantify current LULC as well as detect changes of three
periods (1996, 2006 and 2015) with high accuracy LULC
Fig. 4 changes area of LULC (1996, 2006 and 2015)
within the test areas could examined.
The analysis carried out found that a notable decreasing of
D. Statistical Analysis of LULC Validation
vegetation and cultivated areas (178.73 sq km from 1996 to
2015) because of conversion to the replacing over double up of
This section presented the evaluation of classification
the built up area (189.75 sq km from 1996 to 2015). The overall
performance for the classified images of 1996, 2006 and 2015.
accuracy was almost perfect agreement.
For each class the procedure’s accuracy, user’s accuracy,
Yangon City, the old capital and megacity of Myanmar, is
also home to more than 7 million people and the growth rate,
urbanization rate, industrialization and motivation primarily by [7] Setturu B., K.S. Rajan, and T.V. Ramachandra, 2013. Land Surface
Temperature Responses to Land Use Land Cover Dynamics, Geoinformatic
rural to urban migration, is highest in Myanmar. The challenges
and Geostatistics: An Overview, 1(4).
of the urban environmental problems widely experienced, due [8] Hathout, S. 2002. “The Use of GIS for Monitoring and Predicting Urban
to the luck of planning and mismanagement of limited land Growth in East and West St Paul, Winnipeg, Manitoba, Canada.” J.
resources. Although Yangon is believed to be highly vulnerable Environmental Management 66 (3): 229-38.
[9] Xiao, J., Shen, Y., Ge, J., Tateishi, R., Tang, C., Liang, Y., and Huang, Z.
to climate change, no previous study has investigated the
2006. “Evaluating Urban Expansion and Land Use Change in Shijiazhuang,
impacts of land cover changes on the spatial and temporal China, by Using GIS and Remote Sensing.” Landscape and Urban Planning 75
distribution of temperatures. A study of the relationships (1-2): 69-80.
between LULC and thermal properties in Yangon will be of [10] Eastman, J.R. 2009. IDRISI Taiga guide to GIS and image processing.
Clark Labs Clark University, Worcester, MA.
great useful to appropriate land use planning. The results of this
[11] Billah, M. and G.A. Rahman, 2004. Land cover mapping of Khulna city
paper LULC changes and the issue of increasing LST catches applying remote sensing technique. Geoinformatics, 707-714.
more attention during the urban development. Yangon will still [12]Thome, A.C.G. 2012. Svm classifiers-concepts and applications to
perform almost certain to become a key engine in the nation’s character recognition. INTECH Open Access Publisher.
[13] Vapnik, V. and A. Chervonenkis, 1974. Theory of Pattern Recognition.
economic growth as Myanmar’s largest city, commercial
Moscow Nauka.
capital, most important port and tourist destination, and most [14] Vapnik, V.N. 1982. Estimation of dependences based on empirical data
logical site for export-oriented manufacturing. But how well (Vol. 40). New York: Springer-Verlag.
Yangon fulfils these roles depend on how well good [15] Berlanga-Robles, C.A. and A. Ruiz-Luna, 2011. Integrating remote
sensing techniques, geographical information systems (GIS), and stochastic
management of the city plan. All these are really a matter of
models for monitoring land use and land cover (LULC) changes in the northern
special consideration of impacts, challenges and opportunities. coastal region of Nayarit, Mexico. GIScience and Remote Sensing, 48(2): 245-
263.
REFERENCES [16] Diallo, Y., G. Hu, and X. Wen, 2009. Applications of remote sensing in
[1] Zhang, Q., Wang, J., Peng, X., and Shi, P. 2002. “Urban Built-Up Change land use/land cover change detection in Puer and Simao Counties, Yunnan
Detection with Road Density and Spectral Information from Multi-temporal Province. Journal of American Science, 5(4): 157-166.
Landsat TM Data.” Int. J. Remote Sensing 23 (15): 3057-78. [17] Lu, D. P. Mausel, E. Brondizio, and E. Moran, 2004. Change detection
[2] Turner, B.L. 1994. Local faces, global flowers: the role of land use and land techniques. International journal of remote sensing, 25(12): 2365-2401.
use changes in Datong basin, China, Environmental Geology, 5(2): 1825-1837. [18] AsiaNews, 2013. Yangon, by 2014 a “megacity” of 10 million inhabitants.
[3] Gallo, K.P., A.L. McNab, T.R. Karl, J.F. Brown, J.J. Hood, and J.D. AsiaNews.it, 05/10/2013.
Tarpley, 1993. The use of NOAA AVHRR data for assessment of the urban
heat island effect. Journal of Applied Meteorology, 32(5): 899-908.
[4] Weng Q., D. Lu and J. Schubring, 2004. Relations between NDVI and tree
productivity in the central Great Plains. International Journal of Remote
Sensing of Environment, 25(16): 3127-3138.
[5]Ahmed. B, 2011. Urban Land Cover Change Detection Analysis and
Modelling Spatio-Temporal Growth Dynamics Using Remote Sensing and GIS
Techniques, Dissertation of Master of Science in Geospatial Technologies
[6] Lambin, E.E., H.J. Geist, and E. Lepers, 2003. Dynamics of Land Use and
Cover Change in Tropical Regions. Annual review of environment and
resources, 28(1): 205-41.