You are on page 1of 5

2015 IEEE 11th International Colloquium on Signal Processing & its Applications (CSPA2015), 6 -8 Mac.

2015, Kuala Lumpur, Malaysia

Differences of Image Classification Techniques for


Land Use and Land Cover Classification
Nur Anis Mahmon1, Norsuzila Ya’acob 1,* Azita Laily Yusof 2,*

1
Faculty of Electrical Engineering,Universiti Teknologi MARA,
404500 Shah Alam Selangor, Malaysia
*
Wireless Communication Technology (WiCoT),
Faculty of Electrical Engineering, Universiti Teknologi MARA,
404500 Shah Alam Selangor, Malaysia
nuranismahmon@yahoo.com, norsuzilayaacob@yahoo.com, azita968@salam.uitm.edu.my

Abstract—Land use and land cover classification of remotely different image classification procedures used for different
sensed data is an important research and commonly used in purposes by various researchers [4].
remote sensing application. In this study, the different types of The objectives of this study is to test and evaluate the
classification techniques were used by using satellite image of classification techniques to classifying land use and land cover
some part of Selangor, Malaysia. For this objective, the land of some area in Selangor, Malaysia using three types of
use and land cover was classified with Landsat 8 satellite image classification techniques: Maximum Likelihood classifier,
and ERDAS Imagine software as the image processing Mahalanobis Distance classifier and Minimum Distance
packages. From the classification output, the accuracy
classifier.
assessment and kappa statistic were evaluated to get the most
accurate classifier. The optimal performance would be II. BACKGROUND
identified by validating the classification results with ground
truth data. Of classified image, the Maximum Likelihood A. Land Use and Land Cover
technique (overall accuracy 82.5%) is the highest and more
applicable for satellite image classification compared with Land cover refers to the characteristic or features of the
Mahalanobis Distance and Minimum Distance. The accurate Earth's surfaces and also as subsurface, water surface,
classification can produce the correct Land Use and Land containing soil, topographic, and well as human structures.
Cover map that can be used for many varieties purposes. Examples of land cover include forest, grassland, cropland,
wetland, urban structure. While the land use is refer to
Index Terms—Land use and land cover, classification human use of land which involved the management and
techniques, accuracy assessment.
modification of the natural environment such as urban.
Therefore, the land use and land cover is illustrated for
I. INTRODUCTION human made coverings and both natural of the earth's
Remote sensing application is the activities of the surface. The relationship between land use and land cover is
surveying and recording the objects on Earth surfaces. The not always direct and apparent. A single category of land
satellite images became the main sources of remotely sensed cover may support many use, while a single land use could
data to get the information for different application such as involve the maintenance of several distinct land covers [5].
monitoring land use changes, forestry and [1]. In land use
B. Image Classification
and land cover planning and management the remotely
sensed is very helpful tool for identification and Image processing is the one technique that uses a
classification the features of the land surfaces from any computer to collect for manipulation of digital images.
satellite images [2]. Image classification normally included in four steps: Firstly
Analyzing the image and recognition the pattern is the is pre-processing image, which is to finding the band ratio,
most important part of remote sensing application [3]. In reduce haze and atmospheric correction. Secondly, is
some cases, the classification itself can be the object of training sample, it is the process to do the selection of
analysis. For instance, the remotely sensed data can produce particular criteria feature for describes the pattern. Thirdly, is
a map like the image for land use classification which is as the decision of selecting the suitable technique for
the final product of analysis. Image classification it becomes comparing the target with the image pattern. Lastly, is
an important tool for digital image classification. The classifier assessing the accuracy of the image classification [6]. After
loosely refers to computer programs that perform specific the satellite image classification was done, the land use and
procedures for classification of images. At this time, there are

978-1-4799-8249-3/15/$31.00 ©2015 IEEE 90


2015 IEEE 11th International Colloquium on Signal Processing & its Applications (CSPA2015), 6 -8 Mac. 2015, Kuala Lumpur, Malaysia

land cover mapping will be produced with the accurate


classification.
III. STUDY AREA AND DATA SET

A. Study Area
The study area was referring coverage area of this study
that will be used. Fig.1 shows the study area selected for
this study is some district of Selangor which are Klang,
Petaling, Gombak and Hulu Langat. These areas were
selected to cover the classes of land use and land cover
classification. Remote sensing techniques will be used to
classify the land use and land cover. Selection of study area
can assist in the selection of satellite image correctly.
Fig. 2. Satellite image of study area

IV. METHODS

A. Research Structure
Based on Fig. 3, the methodology structure in this study
is divided into three stages. Firstly, is processing the satellite
image including image pre-processing process and image
classification. Secondly, the analysis part which is analyzed
the data with ground truth data and then get the accuracy
assessment for each classifier. While the third stage is
producing the land and land cover map as the final output for
image classification.

Landsat 8 Imagery 2014

Topographic map
Image pre-processing

Image classification
Fig. 1. Study area

B. Satellite Imagery Maximum likelihood classifier


Mahalanobis distance classifier
In this case area of study, the Landsat 8 satellite sensor Minimum distance classifier
will be used to analyze the land use and land cover
classification. The resolution of Landsat 8 is arrange 28.50
Ground validation
meter until 90.00 meter. This resolution can show the Earth
surfaces to land cover mapping, support resource
assessment and to see the changes of the land use and land Accuracy Assessment
cover area. The spectral information was collected which
have the ability to assess changes in Erath surfaces.
Land use and land cover
map

Fig. 3. Structure of research methodology

B. Pre-processing Image
This part is an important stage before the actual image
classification and image analysis. Software that used in this
part is ERDAS Imagine software. Pre-processing involves
three processes which are Geo-reference, radiometric
correction and subset image. Geo-referenced image need for

91
2015 IEEE 11th International Colloquium on Signal Processing & its Applications (CSPA2015), 6 -8 Mac. 2015, Kuala Lumpur, Malaysia

specific geodetic coordinate system which is RSO classifiers was calculated by analyzed with the confusion
coordinate system to locate the coordinate in the correct matrix and also called error matrix. Other from that, there is
position according to the ground coordinate. Radiometric some indicator that used to show the classification results
correction processed to make the image clearer without the such as overall accuracy, producer accuracy, user accuracy
haze and lines. Pre-processing ensure the quality of the and Kappa coefficient value [11].
welds. r r
N ¦ x ii − ¦ xi + .x + i
C. Image Classification k = i =1
r
i =1

In this study, supervised classification classifiers have been V. N 2


− ¦ xi + .x + i (1)
used to classify the image of the study area for land use and i =1

land cover classification. The Maximum Likelihood, Where:


Mahalanobis classifier, and Minimum Distance classifiers are Xii = Sum of diagonal input of error matrix
used for this paper. Xi+ = Sum of row I of error matrix
For land cover classification, the algorithm that widely and X+I = Sum of column I of error matrix
most commonly used is maximum likelihood algorithm for N = No. of elements in error
supervised classification technique [7]. Maximum Likelihood
Classification is a statistical decision criterion to assist in the Based on [12], producer accuracy has been calculated by
classification of overlapping signatures; pixels are assigned to dividing the number of objects that are right for a particular
the class of highest probability. The Maximum Likelihood class of objects with the actual amount of reference data for
classifier are considered to provide more accurate results that class. Although the accuracy of the user who has been
compare the minimum distance classification, but it is slower
determined by dividing the number of objects that are right
due to the extra calculation. It was found that the maximum
for a particular class with a given number of objects in that
likelihood method gives more accurate result and both
Minimum distance and Mahalanobis distance method trivia class. To carry out precision parts manufacturer labeled
agricultural land and urban areas [8]. objects in the reference data properly informed. Accuracy of
Mahalanobis distance classifier is equal to the Minimum users, however, that quantifies the proportion of objects that
Distance, except that the covariance matrix used in the are assigned to a certain class who agreed with the objects
equation. This algorithm assumes that the histogram of the in the reference data. Accuracy user indicates the
band has a normal distribution [9]. Covariance is reflected in probability that the object belongs to also specifically label
the group so very different to lead the class also vary. For a particular class in reality. It can show the errors of
example, when classifying urban areas- usually classes vary commission.
widely-right pixel is usually not a very different class it is
slower to calculate the minimum distance then [10]. VI. RESULTS AND ANALYSIS
Mahalanobis distance is a parametric, meaning that it is largely
dependent on the normal distribution of data in each input A. Analyzing Kappa Coefficient
band. In classification stages, three methods of supervised were
The minimum distance classifier is based on data from the selected to classify the image. Three accuracies of
training site. This classifier treating each class with an average classification output were tested in this study namely kappa
position of each band. The minimum distance classifier is coefficient, error matrix and overall accuracy. The kappa
highly recommended in all applications for image coefficient of each classifies is shown in Table 1. Based on
classification. The classification is done by placing pixels in a Fig. 4, Maximum Likelihood (0.8216) method produces the
nearby classroom min. Minimum distance algorithm is also highest value of the kappa statistic compares with other
more attractive because it is a technique that is faster than the techniques.
maximum likelihood classification.
D. Field data processing TABLE I. KAPPA COEFFICIENT FOR DIFFERENT TYPE OF CLASSIFIER
Field observation which is ground truthing was used to Classifier Kappa Coefficient
identify the true type of classes in the study area. In this
Maximum Likelihood 0.8216
work Global Positioning System (GPS) has been used to get
the observation of the study area are observed in 90 points Mahalanobis Distance 0.6982
for ground categories according to classes in this study to Minimum Distance 0.7893
cover land use and land cover of study area.
For validation stages the coordinate data of ground data
used in the accuracy assessment stages to be compared with
the reference data in the confusion matrix.
E. Accuracy Assessment
The results of the classification are depending on the
accuracy assessment and Kappa coefficient value. The
percentage of accuracy of the classification result for all

92
2015 IEEE 11th International Colloquium on Signal Processing & its Applications (CSPA2015), 6 -8 Mac. 2015, Kuala Lumpur, Malaysia

Urban 17 17 16 85.71% 85.71%


Open land 18 36 17 87.50% 26.92%
Totals 90 90 67
Overall Classification Accuracy = 74.44%

TABLE IV. CONFUSION MATRIX FOR MINIMUM DISTANCE


CLASSIFICATION

Class Ref. Classified Producers Users


Number
Name Total Total Accuracy Accuracy
Correct
Water
18 14 14 50.00% 100.00%
bodies
Fig. 4. Comparison of Kappa Statistic for different type of classifier. Forest 18 15 14 50.00% 80.00%
Agriculture 19 18 15 55.56% 62.50%
B. Analyzing Accuracy Classification Urban 17 28 16 85.71% 33.33%
The accuracy assessment was calculated by comparing Open land 18 12 12 25.00% 100.00%
the references data (validation point) with the output of the Totals 90 90 71
classification using Erdas Imagine Software. From the
accuracy assessment report, the confusion matrix was Overall Classification Accuracy = 78.88%
obtained to shows the producer accuracy, user accuracy and
overall accuracy. The confusion matrix results of three TABLE V. LAND USE AND LAND COVER CLASSIFICATION EFFICIENCY
OF DIFFERENT METHODS
techniques classifiers are shown in the Table II, Table III
and Table IV. From this table the overall accuracy is Maximum Mahalanobis Minimum
expressed as a percentage of the test pixel successfully Likelihood Distance Distance
Techniques/ PA UA PA UA PA UA
assigned to the correct classes that want to be classified. Class Name (%) (%) (%) (%) (%) (%)
Based on Table V, the Maximum Likelihood classifier
Water bodies 50.00 100.00 25.00 100.00 50.00 100.00
produced the greater accuracy with overall accuracy of
88.88% for land use and land cover classification. While Forest 25.00 66.67 25.00 100.00 50.00 80.00
Mahalanobis Distance classifier gives the lowest of overall Agriculture 11.11 100.00 25.00 100.00 55.56 62.50
accuracy.
Urban 85.71 85.71 85.71 85.71 85.71 33.33
TABLE II. CONFUSION MATRIX FOR MAXIMUM LIKELIHOOD Open land 87.50 31.82 87.50 26.92 25.00 100.00
CLASSIFICATION Overall
Classification 88.88% 74.44% 78.88%
Class Ref. Classified Number Producers Users
Accuracy
Name Total Total Correct Accuracy Accuracy
Water
18 14 14 50.00% 100.00% According to Fig. 5, the comparison of classification
bodies
Forest 18 13 12 25.00% 66.67% techniques shows the Maximum Likelihood techniques
always be the top of the percentage value for overall
Agriculture 19 11 11 11.11% 100.00%
accuracy.
Urban 17 17 16 85.71% 85.71%
Open land 18 35 27 87.50% 31.82%
Totals 90 90 80
Overall Classification Accuracy = 88.88%

TABLE III. CONFUSION MATRIX FOR MAHALANOBIS DISTANCE


CLASSIFICATION

Class Ref. Classified Producers Users


Number
Name Total Total Accuracy Accuracy
Correct
Water
18 12 12 25.00% 100.00%
bodies
Forest 18 12 12 25.00% 100.00% Fig. 5. Comparison of overall accuracy for different type of classifier.
Agriculture 19 10 10 25.00% 100.00%

93
2015 IEEE 11th International Colloquium on Signal Processing & its Applications (CSPA2015), 6 -8 Mac. 2015, Kuala Lumpur, Malaysia

C. Land Use and Land Cover Map International Earth Science Information Network of Columbia
University. 2002, pp. 67.
Land use and land cover maps can play an important role
in the environment management. The five classes of image [3] Dean, A.M and Smith, G.M. “ An evaluation of per – parcel
land cover mapping using maximum likelihood class
classification from the land use and land cover
probabilities” in International Journal of Remote Sensing, 24
classifications can describe the environment of the study (14), 2905-2920, 2003.
area. So, this project obtained classified of Land Use and
[4] Bruzzone, L., and S.B. Serpico “A technique for feature
Land Cover classification with analysis of Landsat 8 selection in multiclass problems” in International Journal of
satellite image. Based on Fig. 6 the classified map produced Remote Sensing, 21:549-563, 2000.
according to the best result of image classification. [5] M.Seetha, I.V.Muralikrishna “Artificial neural network and
other methods of image classification” in Journal of
theoretical and Applied Information Technology.
[6] Geospatial Innovation Facilities in Center for Biodiversity
and Conservation.
[7] Jia, X. and Richards, J.A. “Remote sensing digital image
analysis, an introduction”. Taylor and Industrial Arts, 2006.
[8] Al-Ahmadi, F. S. and Hames, A. S. “Comparison of Four
Classification Methods to Extract Land Use and Land Cover
from Raw Satellite Images for Some Remote Arid Areas”,
Kingdom of Saudi Arabia JKAU; Earth Sci., 20(1), 167-191,
2009.
[9] Perumal, K. and Bhaskaran, R. “Supervised classification
performance of multispectral images” in Journal of
computing, 2(2), arXiv:1002.4046v1, 2010.
Fig. 6. Land Use and Land Cover map of study area.
[10] Swain, P. H. and Davis, S. M. “Remote Sensing: The
Quantitative Approach”, US, McGraw-Hill, Inc,1978.
VII. CONCLUSION [11] Sultan AlSultan “Innovative Satellite Image Map Of
R.Alkhabrta Area, Saudi Arabia Using High Resolution
From the investigation in this study, the results of image Image”, School of Physics.
classification show the efficiency of the three methods that
[12] Tso, B. and P.M. Mather. “Classification Methods for
have been used to classify the Land Use and Land Cover map
Remotely Sensed Data”. 1st Edn., Taylor and Francis,
of Selangor district. From different three classifiers, Maximum
London. 2001.
Likelihood classification techniques produced the highest
overall accuracy with 88.88% of accuracy and also the kappa
coefficient which is 0.8216. Compare with the Mahalanobis
and Minimum classifier give the overall accuracy and kappa
value lowest which are 74.44% and 78.88%. For classes to
cover the land use and land cover for this study are, five
features from the image have been classified. The features are
Forest, Agriculture, Water bodies, Urban and Open Land. The
high resolution images gave more detail information of the
classified map. The classified images could be used for Natural
Resource management planning and development purposes in
the future.

ACKNOWLEDGMENT
This research is sponsored by Universiti Teknologi MARA
(UiTM). The authors would like to thank Agency Remote
Sensing Malaysia (ARSM) for providing the satellite
imagery. The author also wishes to thank UiTM Library-
Tun Abdul Razak Library for supplying topographic map of
the study area.

REFERENCES
[1] S. E. Franklin “Remote Sensing for Sustainable Forest
Management”. New York .Lewis Publishers, 2001.
[2] A. De Sherbinin “Land use and land cover change” in 4 A
CLESIN Thematic Guides, Palisades, YN. Centre for

94

You might also like