International Journal of Computer Applications (0975 - 8887)
Volume 134 - No.16, January 2016
Supervised Techniques and Approaches for Satellite
Image Classification
Minu Nair S.
PG scholar
Department of Computer Science
College of Engineering Perumon, Kerala, India
Bindhu J.S.
Assistant Professor in CSE
Department of Computer Science and Engineering
College Of Engineering Perumon, Kerala, India
ABSTRACT
Remote Sensing is a multi-disciplinary technique for image acqui-
sition and measurement of information. Remote sensing analysis
paved way for satellite image classification which facilitates the
image interpretation of large amount of data. Satellite Images cov-
ers large geographical span and results in the exploitation of huge
information which includes classifying into different sectors. Dif-
ferent classification algorithms exist for image classification, but
with the wide range of applications an algorithm with improved
performance in terms of accuracy is required. Here in this paper
we analyze different methods of supervised classification, different
post classification techniques, spectral contextual classification and
provide a comparative study on their efficiency.
Keywords
Classification, Supervised classifiers, Contextual classification,
Cellular Automata
1. INTRODUCTION
Remote Sensing [1] is a technique introduced in early 1960’s
for data analysis and interpretation. Remote sensing collects huge
amount of satellite data. Satellite remote sensing imagery covers
large geographical area with high temporal frequency as compared
to other imagery. Interpretation of these satellite images helps in
a variety of applications as environmental conservation and man-
agement, water resource research, soil quality studies, environmen-
tal study after natural disasters, meteorology simulations, deriving
land use and land cover information, preventing natural disasters,
studying climatic change evolution.
Different techniques are used for data extraction from remote sens-
ing images. Classification technique is the most useful technique
for image interpretation and information extraction. Satellite image
classification groups together the pixels of the image into number
of different defined classes. The pixels are grouped together based
on the digital values extracted from the satellite images. The pixel
values extracted from the satellite images could be a single value
as in case of gray scale image or multivariate value for multi spec-
tral, temporal or multi-modal image [2]. The classification helps in
extracting the information contained in different bands [3] of the
satellite sensor and the information is extracted in terms of digital
numbers which is then converted to a category.
Traditionally the method of classification can be supervised or un-
supervised. The unsupervised classification [4] also referred to as
clustering attempts for an unclear grouping when no sample sets
are available. Supervised classification [5] requires input from an-
alyst and identifies different classes based on the sample training
sets. Supervised classification is more advantageous over unsuper-
vised classification in most of the applications. A wide range of
classical classification algorithms and different classifying methods
exists for supervised classification. This paper provides a compar-
ative analysis on the accuracy of different supervised classification
algorithms and techniques.
2. LITERATURE SURVEY
Some of the frequent researches on different supervised clas-
sification methods for satellite images are discussed in the
survey and a comparative analysis is done. Satellite Image
classification has different approaches. Classical algorithms as
parallelepiped[5][6], minimum distance[5][6], maximum likeli-
hood[5][6], non-parametric classifiers and machine learning tech-
niques as decision tree method[6] , support vector machine [4], Ar-
tificial neural networks[5] and genetic algorithms[5] which refines
the learning process, were employed for efficient Image classifica-
tion. All these methods have their strengths and limitations. Listed
below are few problems related to one or the other of classical clas-
sification algorithms
(1) In some algorithms which classify the input image with high
degree of heterogeneity, the pixels may be uncertain, i.e. a pixel
can belong to more than one class
(2) Some other algorithm may misclassify a pixel
(3) Some may leave tiny areas of the image unclassified
3. SUPERVISED CLASSIFIERS
3.1 Parallelepiped classification
Parallelepiped classification [7] is a simple classification based on
a decision rule for multispectral data. Decision boundaries for the
parallelepiped algorithm are formed based on a standard deviation
threshold which is chosen from the mean of each selected class in
the training set. The decision boundaries form an interval between
two pixel values with a hyper rectangle region in feature space. .A
pixel is classified based on whether the value of that pixel lies above
1
International Journal of Computer Applications (0975 - 8887)
Volume 134 - No.16, January 2016
the lower threshold value and below the higher threshold value of
the interval.
The mean value Mt of all the pixels for a class C for band M
is taken for all the N classes of the training set and the variation
(standard deviation) of the training data class C of band M of all
the N classes be St. The mean and the standard deviation forms the
parallelepiped boxes as decision boundaries or intervals for assign-
ing the pixels. A pixel will be assigned to a particular class if the
digital value Dv of the pixel lie inside the parallelepiped decision
boundaries.
M t − St ≤ Dv < M t + St (1)
The pixel will be assigned to the class if its value lies in between
the lower and the upper threshold value.
3.2 Minimum Distance Classifier
This is also a simple supervised classifier which uses the center
point (average of all pixels of sample class) to represent a class
in training set. This technique uses the distance measure, where
the Euclidean distance is considered between the pixel values and
the centroid value of the sample class. The pixel with the shortest
distance with the class is assigned with that class.
The classifier is fast in execution, computation time is minimum
as it depends mainly on the training dataset and all pixels will be
classified, but the algorithm may be prone to errors resulting in mis-
classification of pixels as it will classify a pixel even if the shortest
distance is far away. Spectral distance is calculated for all values of
a class mean, the unclassified pixel is assigned to the class with the
lowest spectral distance resulting in classification of all pixels.
The minimum distance algorithm is based on the minimum distance
from the mean value M t of each class of the training data to the dig-
ital value Dv of each pixel in the imagery. The minimum distance
is calculated by using the Euclidean distance measurement.
sqrt(Dv − M t)
2
(2)
The class mean with the minimum distance with the pixel will be
assigned as the class of the pixel.
Fig. 1. Parallelepiped and Minimum distance Classification
3.3 Maximum Likelihood Classification
This method of classification calculates the probability for a given
pixel to each class and then the pixel will be allocated to a partic-
ular class with the highest probability. It calculates the mean and
covariance matrix for the training samples and assumes that the
pixel values are normally distributed. A class can be characterized
by the mean value and the covariance matrix. A probability density
function is defined and the input pixels are mapped based on the
likelihood that the pixel belongs to that particular class. The likeli-
hood expressed for normal distribution can be calculated as below
L
k
(X) =
1
2Π
n
2
|Σ
k
|
1
2
exp−
1
2
(X − m
k
)Σ
−1
k
(X − m
k
)
t
(3)
Here X indicates the image data of n bands. L
k
(X) represents the
likelihood of X belonging to class k, m
k
is the mean vector of class
k, Σ
k
is the variance covariance matrix of class k.
This classifier is a sophisticated classifier with good separation of
classes, but the training set should be strong to sufficiently describe
mean and covariance structure. Also the algorithm is computation-
ally intensive.
Fig. 2. Maximum Likelihood Classification
3.4 Decision Tree Method
In this method of classification a tree structure is built with root
node and leaf nodes. The leaf nodes represent the classes of the
features. Every interior node in the tree consists of a decision crite-
rion. The attributes are partitioned based on spectral characteristics.
The partition takes place based on the homogeneity (similarity of
pixel values) until a leaf node is assigned with a particular class.
A group of pixel values will be classified into two groups based
on the separability with respect to a feature. This method uses the
hierarchical rule and use Non-Parametric approach.
This method does not require any extensive design or training.
Computational efficiency is good but requires complex calculations
and the accuracy may depend fully on the design and selection of
features.
3.5 Support Vector Machines
SVM classifier is based on decision planes that define decision
boundaries. It builds a hyperplane from the training data which
separates pixels with different class membership. The hyperplane
gives the largest minimum distance to the training samples. Larger
the margin, lower will be the generalization error.The method gives
2
International Journal of Computer Applications (0975 - 8887)
Volume 134 - No.16, January 2016
Fig. 3. Sample Decision Tree Method
a good separation achieved by the hyperplane. Here the computa-
tional complexity is reduced but training is time consuming as se-
lection of optimum hyperplane is necessary for improved classifica-
tion. Structure of SVM is difficult to understand. The performance
and the accuracy of SVM depends on the hyperplane selection and
the number of classes.
The hyperplane is normally represented by a normal vector v and a
bias b, b an element in real numbers, v is an element of the feature
space.
v.x + b = 0 (4)
Here x corresponds to the digital values. SVMs usually maximizes
the margin between data values of opposite classes.
Fig. 4. Support vector machine
3.6 K-nearest neighbor algorithm
This is a non-parametric mining technique which uses K nearest [8]
training samples to determine the pixel class. Here the K samples
are chosen based on a similarity measure. Commonly used simi-
larity measure is the distance function. The classification uses Eu-
clidean, Manhattan or other distance measure to calculate the dis-
tance between a pixel and the different training samples. A class is
assigned to a pixel based on the majority voting from the K training
samples which would be to assign the most common class among
the training samples. The technique is simple to process, but com-
putationally expensive to select K nearest neighbors when the train-
ing dataset is large.
Euclideandistance =
p
(Dv − M t)
2
(5)
M anhattandistance = |Dv − M t| (6)
Here Mt measures the digital value of each training sample and
Dv represents the digital value of each pixel in the imagery
3.7 Artificial Neural Network Approach
The algorithms with artificial neural network approach simulate the
human learning process for assigning meaningful labels to images.
ANN imitates few functions of a persons mind. It consists of a se-
quence of layers with a set of neurons in each layer and the pre-
ceding and succeeding layers are joined by weighted connections.
The accuracy and performance of artificial neural network highly
depend on the network structure. ANN networks are data driven
and self adaptive technique. The computation rate is high and han-
dles noisy input. But ANN training is time taking and the network
type architecture is difficult to choose. The papers [9] [10] discuss
the satellite image classification using neural network approach and
different classification algorithms. Per the research ANN has high
classification accuracy rate but the time taken for training the clas-
sifier is huge. The approach is suitable for applications where study
of dynamic data is needed as they have the capability to model non-
linear processes [11] and to identify unknown patterns and images
based on their learning model.
3.8 Genetic Algorithm Approach
GA approach [12] is a search heuristic that mimics the process
of natural evolution. It generates useful solutions to optimization
and search problems using techniques inspired by natural evolu-
tion. Genetic algorithm based approaches are used for land cover
classification[13] and an advantage of using the method is that it
extracts classification rules that are easy for users to realize. GAs
are techniques for optimization and finds the minimum or max-
imum of some arbitrary function. Paper [14] represents a model
using the GA approach to extract classification rules for land use
classification predictions in remote sensing imagery.
The weakness of a GA based classifier is time consumption when
the training dataset contains large instances.
4. COMPARISON OF DIFFERENT SUPERVISED
CLASSIFICATION ALGORITHMS
Comparison of Different classification methods are shown in Table
1 -Summarized comparison between different classification algo-
rithms.
5. POST CLASSIFICATION OF SATELLITE
IMAGES
The short comings of classification algorithms can be eliminated by
applying the post classification techniques. The techniques of post
classification improve the accuracy of the classified image.
5.1 Comparing Different Post classification Techniques
The problems related to classical classification algorithms like un-
classified or misclassified pixels led to the application of post clas-
sification algorithms/techniques. Commonly used post classifica-
tion techniques are Majority filter and probability label relaxation
technique [15]. Here we discuss the accuracy comparison of dif-
ferent post classification technique with the technique of cellular
automata
3
International Journal of Computer Applications (0975 - 8887)
Volume 134 - No.16, January 2016
Table 1. Summarized comparison between different classification
algorithms
Algorithm Advantages Disadvantages
Parallelepiped
• Fast Execution
• Computationally Ef-
ficient
• Pixels not classified
• Pixels in several
classes
Minimum dis-
tance
• All Pixels classified
• Fast Execution
Prone to commission
errors
Maximum Likeli-
hood
Provides good separa-
tion between classes
• Requires well trained
training set
• Computationally in-
tensive
Decision tree
method
• No extensive design
and training
• Reduced Computa-
tional time
• Complex calculation
• Accuracy depends
fully on feature selec-
tion
Support Vector
Machine
Reduced Computa-
tional complexity
• Training is time con-
suming
• Structure is difficult
to understand
• Accuracy may de-
pend on the number of
classes
K Nearest Neigh-
bor Classifier
Low cost and effort
for learning process
• Computationally ex-
pensive to find the K
neighbors when sam-
ple dataset is large
• Performance de-
pends on the number
of dimensions
Artificial Neural
network
• Handles noisy input
• Self adaptive tech-
nique
• Training is time tak-
ing
• Difficult to handle
network type architec-
ture
Genetic Algorith-
mic approach
Extracts rules that are
easier to realize
Time consumption is
large for large training
instances
5.2 Majority Filter
This technique of post classification relabels the center pixel when
it is not a member of the majority class. The method improves the
overall accuracy of classification but merges some land covers to-
gether. If p is the center pixel, then the pixel would be relabeled
as
If (c
i
> c
j
and c
i
> c
t
) then p is assigned the class of w
i
Here c
i
and c
j
refers to the count of pixels belonging to class i and
j.
5.3 Probability Label Relaxation
This is an iterative technique which considers the probabilities of
the neighboring pixels for updating the probability of the center
pixel. The method is based on relation among pixel labels specified
by compatibility coefficients describing the context of the neighbor.
The PLR method of post classification provides higher accuracy
than the majority filter method, but it requires lot of computation
and a wise choice of the compatibility coefficient.
5.4 Cellular Automata approach
The approach of cellular automata [16][17] consists of regular grid
of cells. Each cell is associated with a particular state from a set of
possible states. The state depends on the states of the neighboring
pixels/cells and a set of rules. A Pixel changes its state based on a
transition function and a set of rules.
The Post classification based on cellular automata reassigns a class
of the pixel based on the class of the neighboring pixels based on
defined rules and function
On comparing different post classification techniques, the cellular
automata approach provides a better accuracy than other two filters
6. CONTEXTUAL CLASSIFICATION
APPROACHES
6.1 Spectral and contextual classification
Common classification methods use the spectral properties of the
satellite image pixels and the use of supervised or unsupervised
algorithm depends on the analysts knowledge about the area under
study.
The above defined techniques of classification works well if the
images are non-noisy and the spectral properties define the classes
sufficiently well. If wide variation in class pixel properties are
present or in case of noisy image the image classification may not
be correct and there would be small pixels that are not classified. To
avoid this misclassification we have different techniques applied as
contextual information [18] in addition to spectral data. Contextual
algorithms uses mean values, variances, texture description from a
pixel neighborhood to improve the pixel spectral classification. The
methods usually reduce the error rates considerable as compared to
non-contextual rules [19].The uncertainty of classes arising in the
contextual method can very well used as information for indicating
border zones.
6.2 Contextual classification through texture
extraction
Texture describes the placement and spatial arrangements of rep-
etition of tones and quantifies the variability of pixels in a neigh-
borhood. The texture metrics can improve the classification accu-
racy through mitigating spectral confusion among spectrally simi-
lar classes. Texture can be considered the key visual criterion [20]
for the information from imagery for forest and vegetation applica-
tions. One limitation of texture extraction is the existence of unre-
liable classification results near the edges of classes. In the paper
[21] a characterization of the texture of images by using cellular
automata approach has been explained.
6.3 Contextual classification based on spectral values
The spectral classifiers are the dominant approach for classifying
remote sensing imagery due to their conceptual simplicity and easy
implementation. The contextual information compliments the spec-
tral classifiers. High resolution images are having higher within-
class spectral variability [22]. Classification for images with high
spectral variability provides less satisfactory results. The contex-
tual information can address such problems and can attain better
results.
4
International Journal of Computer Applications (0975 - 8887)
Volume 134 - No.16, January 2016
7. RECOMMENDED APPROACH OF CELLULAR
AUTOMATA WITH SPECTRAL AND
CONTEXTUAL CLASSIFICATION
Different classification methods analyzed in the paper has its sig-
nificance in different applications.A satellite image classification
technique using the classical classification algorithm and apply-
ing an efficient post classification process as cellular automata ap-
proach with contextual classification which helps in eliminating the
problems related to the traditional classification algorithm can pro-
vide an improved accuracy for classification. A method based on
parallelepiped and minimum distance classifiers which undergoes
a spectral classification followed by a contextual classification with
a cellular automata approach of the image pixels can be an efficient
method for accurate satellite image classification.. The approach
can provide a better classification accuracy for satellite images with
high degree of heterogeneity. The classical classification algorithms
fails in case of classifying images with diversified characteristics as
in case of green houses[23][24].
The uncertain and misclassified pixels disappear as contextual clas-
sification is applied based on neighborhood states if the pixel be-
longs to several class. Different problems related to the classical
classification algorithm we discussed here would be eliminated by
the use of contextual classification. The technique also provides a
hierarchical classification based on the degree of membership of
each pixel to a class. The post classification process of the ACA
approach leads to uncertain pixel refinement.
8. EXPERIMENTAL ANALYSIS
The experimental analysis of the proposed approach of applying
cellular automata and contextual post classification[ACA] to clas-
sical classification algorithms showed a more efficient system in
terms of accuracy. The ACA approach applied to the classical par-
allelepiped algorithm showed an increase of 4.8% for low clas-
sification complexity field images and 15.7% for high classifica-
tion complexity images. The classification complexity of images
is considered based on the heterogeneity of images.For the mini-
mum distance algorithm the ACA modification led to an increase
of 3.3% for low complexity images and 9.7% for high complexity
images[25].
The Table 2 shows a comparative analysis of different classification
algorithms with the ACA approach in terms of accuracy.
Table 2. Experimental Analysis of Proposed ACA Approach
Algorithm
Low Complexity
Field Images
High Complexity
field images
ACA Paral-
lelepiped
89.15% 82.01%
ACA Minimum
Distance
88.36% 78.87%
Naive Bayes 87.87% 72.57%
K-NN 85.95% 71.20%
9. CONCLUSION
This paper analyses different supervised classification approaches
and methods, post classification algorithms and the concept of ap-
plying cellular automata and contextual classification for satellite
image classification. Satellite image classification is a field which
has great significance for different socio-economic, environmental
applications. Through classification of satellite imagery, the infor-
mation as cadastral information, land cover type, vegetation type,
soil properties could be obtained. Different methods discussed in
the literature review emphasize on different techniques and has its
own advantages and limitations. But they can be used for different
specific applications. The classical classification algorithms with
other learning techniques were discussed. The researches specify-
ing different post classification techniques were also discussed.The
role of contextual classification in addition to spectral classifica-
tion and its significance for classifying images with high degree of
heterogeneity were also analyzed.
The comparative study concluded with the high accuracy rate
of classification for the method of classical classification algo-
rithm combined with cellular automata with contextual classifica-
tion which combines the classification and post classification tech-
niques.
10. REFERENCES
[1] E. C. Barret and L. F. Curtis, Introduction to Environmen-
tal Remote Sensing. Cheltenham, U.K.: Cheltenham Stanley
Thornes Publishers
[2] Bhabatosh Chanda,Dwijesh Dutta Mjumder, Digital Image
processing and analysis.
[3] W. G. Rees, Physical Principles of Remote Sensing, 2nd ed.
Cambridge,U.K.: Cambridge Univ. Press, 2001
[4] P. Mather and B. Tso, Classification Methods for Remotely
Sensed Data,2nd ed. Boca Raton, FL, USA: CRC Press, 2009.
[5] Sunitha Abburu , Suresh Babu Golla Satellite Image Classifi-
cation Methods and Techniques: A Review, International jour-
nal of computer applications, Volume 119 No.8, June 2015
[6] Pooja Kamavisdar, Sonam Saluja, and Sonu Agrawal, A Sur-
vey on Image Classification Approaches and Techniques, in
Proc. IJARCCE.
[7] Minakshi Kumar DIGITAL IMAGE PROCESSING Indian
Institute of Remote Sensing, Dehra Dun
[8] B.K. Mayanka; Classification of Remote Sensing data us-
ing KNN method, Journal Of Information, Knowledge And
Research In Electronics And Communication Engineer-
ing(Volume 02, Issue 02)
[9] Smriti Sehgal, Remotely Sensed LANDSAT Image Classifi-
cation Using Neural Network Approaches, International Jour-
nal of Engineering Research and Applications, Vol. 2, Issue 5,
September- October 2012
[10] Aarsi Saini, Er. Sukvinder Kaur,Er. DV Saini Satellite Image
Classification using Artificial Neural Network, SSRG Inter-
national Journal of Electronics and Communication Engineer-
ing (SSRG-IJECE) EFES April 2015
[11] Priyanka Sharma, Urvashi Mutreja Analysis of Satellite Im-
ages using Artificial Neural Network, International Journal of
Soft Computing and Engineering (IJSCE) ISSN: 2231-2307,
Volume-2, Issue-6, January 2013
[12] Moje Ravindra K, Patil Chandrashekhar G Classification of
Satellite images based on SVM classifier Using Genetic Al-
gorithm, INTERNATIONAL JOURNAL OF INNOVATIVE
5