A Comparative Analysis of Remote Sensing Image

Classification Techniques
Pushpendra Singh Sisodia Vivekanand Tiwari Anil Kumar
Computer Science and Technology Electronics and Communication Computer Science and Technology
Manipal University Jaipur, Jaipur, India Manipal University Jaipur, Manipal University Jaipur,
[email protected] Jaipur, India Jaipur, India

Abstract— In this paper, we have compared the accuracy of four Unsupervised classification is not required any prior
supervised classification as Mahalanobis, Maximum Likelihood knowledge of investigator to identify classes for classification.
Classification (MLC), Minimum distance and Parallelepiped
classification with remote sensing Landsat images of different time period Instead of prior knowledge, unsupervised classification used
and sensors. We have used Landsat Multispectral Scanner (MSS), characteristic of classes directly from input data for
Thematic Mapper (TM) and Enhanced Thematic Mapper+ (ETM+) classification. Unsupervised classification calculated the
images of 1972, 1998 and 2013 respectively of Jaipur district, Rajasthan, minimum distance between object mean and every possible
India. Accuracy has been calculated using Producer accuracy, User
accuracy, Overall accuracy and Kappa statistics. We have found that class mean and this process runs until mean remain same and
remote sensing images with a different time period and sensors, when assign that object to that corresponding class of that object
classified with supervised algorithms produced different results. [4].
Minimum distance classification produced better accuracy with Landsat
MSS image than other three classifications while Maximum Likelihood Supervised classification required training samples in the
Classification produced better accuracy with Landsat TM and ETM+ form of pixel to teach classifier to determine decision
images than other three classifications.
boundary in feature space. The decision boundary accuracy is
Keywords— Supervised classification, Remote sensing, Landsat MSS, dependent on the number of training sample and type of
TM, ETM+, Jaipur classes acquired from information classes [4].
Unsupervised classification seems more elegant and
I. INTRODUCTION automated than supervised classification but in remote sensing
when an image have same reflectance for more than one
Remotely sensed image classification can be represented
information classes, classification accuracy of unsupervised
as a combination of pattern recognition, image analysis and
classification is decreased due to mixed pixel problem. In this
classification technique [1]. Classification of remote sensing
case supervised classification promised to give better
image is an important part of pattern recognition. The
Accuracy [4].
principle application of classified images is to display
meaningful pattern in the form of classes [2]. Classified Accuracy of supervised classification also dependent on
images are widely used in various applications like urban quality of satellite images because different satellite sensor
planning, agriculture, forestry, geology, hydrology, ocean & produced different quality images. So, there is need of study to
coastal monitoring and land cover/land use, etc. The output of analyze which supervised classification suitable to which
classified remotely sensed images comes in the form of satellite images for better accuracy. Supervised classification
thematic maps that represent land use, vegetation, water can be further differentiating as Mahalanobis, Maximum
bodies, etc. Environmental and socioeconomic study is based Likelihood Classification (MLC), Minimum distance and
on results of classified images, which is the reason why many Parallelepiped classification respectively [4, 7-9].
researchers have keen interest in classification of remote
In this paper, an attempt has been made to compare the all
sensing images [3].
supervised classifications with different Landsat satellite
Classification of remote sensing image and image images of different time period of 1972 (MSS), 1998 (TM)
processing can be viewed as a same phenomenon. Generally, and 2013 (ETM+).
the classification of remotely sensed images can be known as
a process of grouping pixel in finite set of individual classes II. STUDY AREA AND DATA USED
based on their data values. The pixel is assigned to a particular
The capital of the state Rajasthan, Jaipur has been chosen
class if it satisfies the certain set of rules to fit in a particular
for our study. The study area extends from 26 92’ N to 75 82’
class. The classes can be known or unknown, if investigator is
E and having the total municipal area of 467 sq.km. The
able to separate the classes from giving study area, it is known
graphical representation of the study area has shown in Fig. 1.
classes otherwise classes are unknown but number of class
categories can be determined [4]. Using this prior knowledge
classification can be broadly categorized in two way first is
supervised classification and unsupervised classification.

978-1-4799-3080-7/14/$31.00 2014
c IEEE 1418
 A. Geometric corrections
Geometric correction of the satellite image is very
necessary because MSS, TM and ETM+ are obtained in meter
format and to check the post classification accuracy it is
required geometric correction. Geometric correction has been
done by using Geographic (lat/lon) projection, Everest datum
and nearest neighbour techniques. At the time of geometric
correction, Root Mean Square Error (RMSE) maintained to be
very low.
B. Image enhancement
Fig.1 Study area The main objective of image enhancement is to achieve
good visual interpretability of the satellite image. Image
Satellite Landsat MSS, TM and ETM+ images has been enhancement helped to increase the apparent distinction
used in this study and acquired from public domain of U.S.
between the different features. The contrast stretching
Geological Survey [6]. We have taken data with minimum
technique has been used for better interpretation.
cloud cover for better accuracy. For image accuracy
assessment Ground control points collected from field survey, C. Extracting the signatures
Survey of India toposheets of 1:50,000, 1:250,000 scale and
After image enhancement, we have obtained different
public domain of U.S. Geological Survey [6] to register and
signatures from satellite images of MSS, TM and ETM+
geo-correct remote sensing data.
respectively. Signatures are group of pixel that represents the
Table I. Data used in this study corresponding class of that signature. We obtained different
signatures form images to train the classifiers.
S. Type of Sensor/ Year Scale and Path Row
No Data Toposheet Resolution in D. Image classification
meter Our aim is to use supervised classification and comparison
1 Landsat ETM+ 2013 30 147 041
of all four supervised classification to determine the best
suitable classification technique for different satellite image.
We classified the three Landsat satellite images as MSS, TM
2 Landsat TM 1998 30 147 041 and ETM+ of the years 1972, 1998 and 2013 respectively,
with four different supervised classifications as Mahalanobis,
3 Landsat MSS 1972 80 Maximum Likelihood Classification (MLC), Minimum
147 041
G43D12, distance and Parallelepiped classification. Accuracy
G43D16 assessment has been carried out by taking a 300 random
G43J9, 1:50000,
4 Toposheets
G43J13
2013
1:2,50,000
- - sample pixel for each satellite image and check accuracy of
G43J10, image by field data and toposheets. Classification results have
G43J14 shown in Fig.2 and Table: 2, 3 and 4 respectively.

III. METHODOLOGY IV. RESULTS AND ACCURACY ACCESSMENT

In our study, we have used Landsat MSS, TM, and ETM+ Accuracy assessment has been performed by using overall
images of year 1972, 1998 and 2013 respectively for accuracy, producer accuracy, user accuracy and overall kappa
generating Landuse classification image by using statistics [5]. Results has been calculated for MSS, TM and
Mahalanobis, Maximum Likelihood Classification (MLC), ETM+ images for year 1972, 1998 and 2013 respectively, and
Minimum distance and Parallelepiped supervised evaluate supervised classification by using Mahalanobis,
classification. Before image classification, geometric Maximum Likelihood Classification (MLC), Minimum
correction and image enhancement technique has been applied distance and Parallelepiped classification.
for better training signatures that were used to train classifiers.
Erdas Imagine 9.1 software has been used for supervised
classification.

2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI) 1419

 Fig.2 Satellite image classification of MSS, TM and ETM+ images for the years of 1972, 1998 and 2013

used Maximum Likelihood Classification (MLC) classifier

A. Results of 1972 MSS image
and got producer accuracy at 94.20%, user accuracy as
We have classified the 1972 MSS image (a) used 93.50%, overall accuracy as more than 93% and overall
Mahalanobis classifier and got producer accuracy at 85.46%, Kappa statistics as 89.60%.
user accuracy as 82.72%, overall accuracy as more than 82% Table II. Accuracy Results of 1972 MSS image
and overall Kappa statistics as 76.37%. Classified Image (b)
Classification Accuracy
used Maximum Likelihood Classification (MLC) classifier
and got producer accuracy at 83.21%, user accuracy as Overall Classification Accuracy Overall Kappa
1972 MSS images
79.36%, overall accuracy as more than 79% and overall Statistics
Kappa statistics as 75.32%. Classified Image (c) used Producer User
Accuracy (%) Accuracy (%) (%)
Minimum distance classifier and got producer accuracy at
86.52%, user accuracy as 83.23% overall accuracy as more (a) Mahalanobis 85.46 82.72 76.37
than 83% and overall Kappa statistics as 78.01%. Classified
Image (d) used Parallelepiped classifier and got producer (b) Maximum
83.21 79.36 75.32
likelihood
accuracy at 83.11%, user accuracy as 80.43%, overall
accuracy as more than 80% and overall Kappa statistics as (c) Minimum
86.52 83.23 78.01
75.89%. Results in tabular form are displayed in Table: 2. distance

B. Results of 1998 TM image (d) Parallelepiped 83.11 80.43 75.89

We have classified the 1998 TM image (e) used
Mahalanobis classifier and got producer accuracy as 93.78%, Classified Image (g) used Minimum distance classifier and
user accuracy as 92.05%, overall accuracy as more than 92% got producer accuracy at 91.32%, user accuracy as 90.54%,
and over all Kappa statistics as 88.78%. Classified Image (f) overall accuracy as more than 90% and overall Kappa
statistics as 89.22%. Classified Image (h) used Parallelepiped

1420 2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI)

classifier and got producer accuracy at 92.55%, user accuracy Parallelepiped classification with a different time period of
as 91.23%, overall accuracy as more than 91% and overall satellite images as MSS, TM and ETM+. We have found that
Kappa statistics as 89.32%. for year 1972 MSS image Minimum distance classification
Table III. Accuracy Results of 1998 TM image has given better accuracy than other three classifications, but
Classification Accuracy for year 1998 TM and 2013 ETM+ Minimum distance
classifier has given comparatively low accuracies. However,
Overall Classification Accuracy Overall Kappa when images of year 1998 TM and 2013 ETM+ classified,
1998 TM image
Statistics Maximum Likelihood Classification (MLC) has given better
Producer User results than other three classifications. Overall Maximum
Accuracy (%) Accuracy (%) (%) Likelihood Classification technique has been given most
accurate results for remotely sensed images according to our
(e) Mahalanobis 93.78 92.05 88.78 study.
(f) Maximum
likelihood
94.20 93.50 89.60 VI. FUTURE WORK
In the future, we will use Maximum Likelihood
(g) Minimum Classification as a classifier to classify remotely sensed
91.32 90.54 89.22
distance
images for Land Use and Land cover analysis. These Land
(h) Parallelepiped 92.55 91.23 89.32
Use and Land cover images can be used for measuring and
monitoring of urban growth of the city, forest, agricultural
C. Results of 2013 MSS image land monitoring etc.

We have classified the 2013 ETM+ image (e) used ACKNOWLEDGEMENT

Mahalanobis classifier and got producer accuracy at 93.58 %, We are highly thankful to Manipal University Jaipur for
user accuracy as 95.45%, overall accuracy as more than 95% providing partial financial support under the project of
and overall Kappa statistics as 90.02%. Classified Image (f) “Urbanization of State of Rajasthan” with the office order
used Maximum Likelihood Classification (MLC) classifier number of MUJ/REG/OO/2013-116.
and got producer accuracy at 94.80 %, user accuracy as
96.09%, overall accuracy as more than 96% and overall REFERENCES
Kappa statistics 90.52%. Classified Image (g) used Minimum
distance classifier and got producer accuracy at 91.78%, user [1] Perumal, K. and R Bhaskaran. "Supervised classification performance of
accuracy 93.65%, overall accuracy as more than 93% and multispectral images." arXiv preprint arXiv: 1002.4046 (2010).
overall Kappa statistics 89.12%. Classified Image (h) used [2] Jasinski, Michael F. "Estimation of sub-pixel vegetation density of
Parallelepiped classifier and got producer accuracy 92.78%, natural regions using satellite multispectral imagery." Geoscience and
user accuracy 94.67%, overall accuracy as more than 94% and Remote Sensing, IEEE Transactions on 34.3 (1996): 804-813.
[3] D. Lu, Q. Weng, "A survey of image classification methods and
overall Kappa statistics 89.38%.
techniques for improving classification performance", International
Table IV. Accuracy Results of 2013 ETM+ images Journal of Remote Sensing, Vol. 28, No. 5, pp. 823-870, January 2007.
[4] Mather, Paul, and Brandt Tso. Classification methods for remotely sensed
Classification Accuracy data. CRC press, 2010.
[5] Congalton, Russell G. "A review of assessing the accuracy of
2013 ETM+ Overall Classification Accuracy Overall Kappa classifications of remotely sensed data." Remote sensing of
images Statistics Environment 37.1 (1991): 35-46.
Producer User
Accuracy (%) Accuracy (%) (%) [6] http://landsatlook.usgs.gov
[7] Strahler, Alan H. "The use of prior probabilities in maximum likelihood
(i) Mahalanobis 93.58 95.45 90.02 classification of remotely sensed data." Remote Sensing of
Environment 10.2 (1980): 135-163.
(j) Maximum [8] Gong, Peng, and PHILIPJ HOWARTH. "The use of structural
94.80 96.09 90.52
likelihood information for improving land-cover classification accuracies at the
rural-urban fringe.” Photogrammetric engineering and remote
(k) Minimum
91.78 93.65 89.12 sensing 56.1 (1990): 67-73.
distance
[9] Wacker, A. G., and D. A. Landgrebe. "Minimum distance classification
(l) Parallelepiped 92.78 94.67 89.38 in remote sensing." LARS Technical Reports (1972): 25.

V. CONCLUSION
In this paper, we have compared four supervised
classification techniques as Mahalanobis, Maximum
Likelihood Classification (MLC), Minimum distance and

2014 International Conference on Advances in Computing,Communications and Informatics (ICACCI) 1421