CHAPTER 3

RESEARCH METHODOLOGY

Salem city has been taken for this work it is situated in Salem District of Tamilnadu,

India. It is bounded between the latitude 11.609 decimal degrees to 11.728 dd and longitude

between 78.0901 dd to 78.2114 dd and it is located about 160 kilometers northeast of

Coimbatore, 186 kilometers southeast of Bangalore and about 340 kilometers southwest of the

state capital, Chennai. It is divided by the river Thirumanimuthar in the main division. The name

Salem appears to have been derived from Sela or Shayla or Sayilam which are found in

inscriptions referring to the country around the hills. It is 278m above Mean Sea Level.

The soil types of the study area are red non-calcareous and red calcareous soils. The

average annual rainfall is 363.5 mm. If we considering the temperature is very high during

summer and it ranges from 20.0 to 37.9° degree Celsius. The mining areas covers in the northern

part of Salem city and western part are covered with kalvarayan hill of Yercaud. Southern part

and western part covered by small hills. Following Figure will show the geographical feature and

extend of Salem city. The satellite image for the study area has been taken from (USGS) Earth

Explorer. The Lansat satellite Image for Salem city has been taken and involve in the following

steps

Phase 1:

Proposed Methodology

The flow of proposed vegetation area segmentation methodology in satellite images is

shown in Fig.1.
 Fig. 1 Proposed vegetation area segmented methodology in satellite images

3.1 Preprocessing

The objective of pre-processing is an improvement of the image data that suppresses unwanted

distortions or enhances some image features important for further processing. It is used to

conduct steps that will reduce the complexity and increase the accuracy of the applied algorithm.

The input satellite images are in Red (R) Green (G) Blue (B) format with low pixel resolution.

This low resolution RGB satellite images are converted to YCbCr format which is used to

achieve more efficient representation of input images. The following equations state the

conversion of RGB to YCbCr process.

 Y  0.299 * R  0.587 * G  0.114 * B (1)

U  0.1478  R  0.289 * G  0.4368B  (2)

V  0.615 * R  0.5158  G  0.100 * B (3)

After conversion from RGB to YUV the chrominance and luminance component of the

satellite image is the luminance component is highlighted then the Chroma component

bandwidth is reduced .without any perceptual appearance this type of conversion useful in image

compression were required. Histogram equalization technique is applied in order to get the

detailed information about the required vegetation area feature of the images.

3.2 Feature Extraction

Features correlate the linear relationship between each pixel with its surrounding pixels in

satellite images. Texture feature is the compound statistical distribution of combination of

various intensities of pixel position in the image. These Features plays an important role to

define our objective in this paper. GLCM features along with grey level feature and local

derivative pattern LDP features are used for vegetation pixel classifications in satellite images.

Gray Level Co-Occurrence ( GLCM) features

Gray level co-occurrence matrices have been used extensively in remote sensing

applications for land-cover vegetation classification and feature analysis and it is very useful in

texture analysis. It calculates the second order statistics related to image properties by

considering the spatial relationship of pixels. GLCM depicts how often different combinations of

gray levels co-occur in an image. The Spatial Relationship can be specified in different ways, the

default one is between a pixel and its immediate neighbor to its right. However it can specify this

relationship with different offsets and angles.

 The pixel at position (i, j) in GLCM is the sum of the number of times the (i, j)

relationship occurs in the image. The following features are extracted from the GLCM matrix.


N
Contrast : i , j 1
pd (i  j )^2 (4)


N
Entropy : i , j 1
Pd ( ln Pd ) (5)

(i  i )( j  j ) Pd

N
Correlation : (6)
i , j 1
ij

N
Pd
Homogeneity : 1
i , j 1 i j
(7)

Grey level features

Considering a small pixel region in luminance component satellite image, with the

described pixel at the centre, five different grey level features are extracted and these features are

taken for the classification process. These feature sets for a candidate pixel (s, t) in a sub-image J

(square sized window w*w) are given as follows,

F1(s, t )  I (s, t )  min{J } (8)

F 2(s, t )  max{ j}  I (s, t ) (9)

F 3(s, t )  abs( I (s, t )  mean{J } (10)

F 4(s, t )  std{J } (11)

F 5(s, t )  I (s, t ) (12)

LOCAL DERIVATIVE PATTERN FEATURES

The first order Local Binary Pattern (LBP) feature is extended to higher order feature in

four different directions as 0°, 45°, 90°, and 135° developed by Subramanian and Murala [16]. is

called as Local Derivative Pattern feature for a specific orientation is computed in n-order is

given as,

LDR ( n1) (rc )   0 ,45 ,90 ,135 (13)

Where, ‘k’ represents the order of the feature and it is set to 2 in this paper. ‘ is the center

pixel in 3*3 sub window.

The nth order LDP feature is computed using the following equation.

p
LDR K (rc )   2 ( p 1)  f 1( I ( K 1) (rc ), I ( K 1) (rp )) p  8 (14)
p 1

Where,

1, for rp and rc  0

 
f1 ( x, y )   0 else 
 
 

‘P’ represents the number of surrounding pixels and ‘p’ varies from 1 to 8.

3.3 Classifications

In general, classification plays an important role in classifying each object in an image

into either class 1 or class 2. In this paper, classification role is to classify each pixel in the

satellite image into either the pixel belonging to vegetation area or the pixel belonging to non-

vegetation area.
 The conventional classification methods such as SVM and Neural Network did not

provide optimum pixel classification due to its non-stabilized weight function behavior of their

internal layers. In order to overcome such non-behavior functionalities, ANFIS classification

methodology is adopted in this method. This has linear weighting functional behavior which also

improves the classification rate of the vegetation pixels in satellite images.

The properties of Neural Networks and fuzzy logic integrality are combined in linear

manner which forms the closed loop as adaptive manner known as ANFIS classifier. It has five

layers as layer1 to layer 5, layer 1 act as input layer, whereas layer 5 acts as output layer. The

input layer receives the extracted features from the satellite image and the output layer produces

binary response based on the input feature set. By proper training with desired value of

vegetation pixels and non-vegetation pixel in satellite images, the system is trained as if the pixel

belong to the vegetation area the response produce by the system is high and for non-vegetation

pixel the response is low.

Phase 2:
Satellite image Preprocessing Feature Classification
using Mean Extraction
Shift Filter

Segmentation
Performance analysis
using K-Means
clustering

Figure 1 Proposed work flow

The input is the satellite image which is initially pre-processed by using Mean shift filter used to

enhance the image and for region smoothing. After that K-means clustering is used for

segmentation of images into Land Use and Land Cover regions. Proposed method is using a

median algorithm based on estimating the optimal direction used to measure if the tested pixel is

noisy or noise-free pixel. if the accurate or optimal direction of the edge is determined then more

edge pixels can be detected. The noisy pixel that has small deviations with the pixels in the

optimal direction is defined an original pixel. Noisy pixel is calculate MSE values for various

noise values.

Pre-processing

Pre-processing is important and necessary step for the image segmentation. It is

performed to get uniform image for further processing. This step includes enhancement of input

satellite images and for region smoothing by applying mean shift filter.

Mean Shift Filter

A mean shift filter is an edge-preserving commonly used in object tracking and image

segmentation. It also used as data clustering algorithm commonly in computer vision and image

processing based on unsupervised learning techniques. Mean shift is a nonparametric iterative

algorithm used to generalize the kernel approach and density gradient estimation. Mean shift

filtering can work with binary, gray scale, RGB and arbitrary multichannel images. Mean shift

image analysis has wide usages in,

 Region segmentation

 Point feature tracking

 Equalize image
Preliminaries of Mean Shift Filter

This filter replaces each pixel in the image with the mean values given as colour and

spatial radius. Mean shift treats the points as a probability density function. So the stationary

points obtained via gradient ascent represent the modes of density function and stationary point

belongs to same cluster. For each point xi, the Mean shift procedure is given as

1. To compute mean shift vector by m(xit)

2. Move the density estimation by m(xit)

3. Repeat till convergence

Applications of Mean Shift Filtering

 Mean shift does not assume any predefined shape on data clusters.

 It is an application-independent tool suitable for real data analysis.

 It is capable of handling arbitrary feature spaces.

 It used for visual tracking with single parameter bandwidth.

Image segmentation

Images get segmented into Land Use and Land Cover (LULC) regions and study the land

use changes using K-Means clustering.

K-Means Clustering

K-means is the unsupervised learning algorithms that solve the well-known clustering

problem. It is a method of vector quantization from signal processing and it is popular in cluster

analysis. It is a partitioning method and the k-means is to partitions the data into k mutually

exclusive clusters. The objective of K-Means clustering is to minimize total intra-cluster

variance, or, the squared error function

 𝐾 𝑛 2

𝐽= 𝑋𝑖 − 𝐶𝑗
𝑖=1 𝑖=1

Where, J-objective function, xi- Case i, ||xi-cj||2 - Distance function

Algorithmic steps for K-Means clustering

1. Clusters the data into k groups where k is predefined.

2. Select k points at random as cluster centres.

3. Assign objects to their closest cluster centre according to the Euclidean distance function.

4. Calculate the centroid or mean of all objects in each cluster.

Repeat steps 2, 3 and 4 until the same points are assigned to each cluster in consecutive

rounds. K-means is a time complexity O(KnT) where k is the number of clusters, n is the

number of points and T is the number of iterations.

K-Means Clustering Application

K-Means is fast, robust and easier to understand. It gives best result when data set are

well separated from each other. It is used in many applications as follows

Vector quantization:

Vector quantization is used in computer graphics, colour quantization to reduce the

colour palette of an image into a fixed number. Then it includes non-random sampling, k-means
can easily use to choose k different prototypical objects from a large data set.

Cluster Analysis:

In cluster analysis, the k-means algorithm can be used to partition the input data set into k

partitions with some limitations. Then the clustering cannot be used with arbitrary distance

functions or on non-numerical data, so the k-means produce better result in image segmentation.
Feature Extraction

Features correlate the linear relationship between each pixel with its surrounding pixels in

satellite images. Texture feature is the compound statistical distribution of combination of

various intensities of pixel position in the image. These Features plays an important role to

define our objective in this paper. GLCM features along with grey level feature and local

derivative pattern LDP features are used for vegetation pixel classifications in satellite images.

Gray Level Co-Occurrence (GLCM) features

Gray level co-occurrence matrices have been used extensively in remote sensing

applications for land-cover vegetation classification and feature analysis and it is very useful in

texture analysis. It calculates the second order statistics related to image properties by

considering the spatial relationship of pixels. GLCM depicts how often different combinations of

gray levels co-occur in an image. The Spatial Relationship can be specified in different ways, the

default one is between a pixel and its immediate neighbor to its right. However it can specify this

relationship with different offsets and angles. The pixel at position (i, j) in GLCM is the sum of

the number of times the (i, j) relationship occurs in the image. The following features are

extracted from the GLCM matrix.

𝑁
𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡 = 𝑖,𝑗 =1 𝑝𝑑(𝑖 − 𝑗)2 (4)
𝑁
𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = 𝑖,𝑗 =1 𝑃𝑑(−𝑙𝑛𝑃𝑑) (5)
𝑁 𝑖−𝜇𝑖 𝑗 −𝜇𝑗 𝑃𝑑
𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 = 𝑖,𝑗 =1 (6)
𝜎𝑖 𝜎𝑖

𝑁 𝑃𝑑
𝐻𝑜𝑚𝑜𝑔𝑒𝑛𝑒𝑖𝑡𝑦 = 𝑖,𝑗 =1 1+|𝑖−𝑗 | (7)

Grey level features

Considering a small pixel region in luminance component satellite image, with the

described pixel at the centre, five different grey level features are extracted and these features are
taken for the classification process. These feature sets for a candidate pixel (s, t) in a sub-image J

(square sized window w*w) are given as follows,

F1(s, t) = I (s, t) - min{J} (8)

F2(s, t) = max{j} - I (s, t) (9)

F3(s, t) = abs(I (s, t) - (10)

F4(s, t) = std {J} (11)

F5(s, t)= I(s, t) (12)

LOCAL DERIVATIVE PATTERN FEATURES

The first order Local Binary Pattern (LBP) feature is extended to higher order feature in

four different directions as 0°, 45°, 90°, and 135° developed by Subramanian and Murala [16] is

called as Local Derivative Pattern feature for a specific orientation is computed in n-order is

given as,

𝐿𝐷𝑅 ∅ 𝑛 −1 𝑟𝑐 | ∅ = 0°, 45°, 90°

Where, ‘k’ represents the order of the feature and it is set to 2, ‘𝑟𝑐 ’ is the center pixel in 3*3 sub

window.

The nth order LDP feature is computed using the following equation.

𝑝
𝐿𝐷𝑅∅𝐾 𝑟𝑐 = 𝑝=1 2
(𝑝−1)
∗ 𝑓1 𝐼∅𝑘−1 𝑟𝑐 , 𝐼∅𝑘−1 𝑟𝑝 |𝑝=8 (14)

Where,

1, 𝑓𝑜𝑟 𝑟𝑝 𝑎𝑛𝑑 𝑟𝑐 ≥ 0
𝑓1 𝑥, 𝑦 =
0 𝑒𝑙𝑠𝑒

‘P’ represents the number of surrounding pixels and ‘p’ varies from 1 to 8.
Classifications

In general, classification plays an important role in classifying each object in an image

into either class 1 or class 2. In this paper, classification role is to classify each pixel in the

satellite image into either the pixel belonging to vegetation area or the pixel belonging to non-

vegetation area. The conventional classification methods such as SVM and Neural Network did

not provide optimum pixel classification due to its non-stabilized weight function behavior of

their internal layers. In order to overcome such non-behavior functionalities, ANFIS

classification methodology is adopted in this method. This has linear weighting functional

behavior which also improves the classification rate of the vegetation pixels in satellite images.

The properties of Neural Networks and fuzzy logic integrality are combined in linear

manner which forms the closed loop as adaptive manner known as ANFIS classifier. It has five

layers as layer1 to layer 5, layer 1 act as input layer, whereas layer 5 acts as output layer. The

input layer receives the extracted features from the satellite image and the output layer produces

binary response based on the input feature set. By proper training with desired value of

vegetation pixels and non-vegetation pixel in satellite images, the system is trained as if the pixel

belong to the vegetation area the response produce by the system is high and for non-vegetation

pixel the response is low.