Multimedia Tools and Applications

https://doi.org/10.1007/s11042-019-07872-y

Multi-model fusion based satellite image classification

using versatile unsupervised vector zone (VUVZ) fusion
and intensive pragmatic blossoms (IPB) technique

R. Swathika 1 & T. Sree Sharmila 1

Received: 4 May 2018 / Revised: 1 May 2019 / Accepted: 10 June 2019

# Springer Science+Business Media, LLC, part of Springer Nature 2019

Abstract
To get the result if satellite image classification in precisely the image fusion method is widely
used. The satellite images at different spectral and spatial resolutions with the guide of image
handling strategies can improve the idea of data. For the most part image fusion is important to
expel the spatial data from two images of different spatial, spectral and worldly images of a
comparable zone. A task of image examination, for instance, image order on fused images
gives better results in contrast with one of a kind data. In this work versatile unsupervised
vector zone (VUVZ) method have been used for image fusion. The resultant images have been
classified using the supervised classification with Intensive Pragmatic Blossoms (IPB) rule for
information extraction and comparison between them regarding their accuracy. In This work
describes a case study of micro seepage signals detection in the land, using multi-sensor
satellite time series based on linear iterative segmentation (LIS) and Intensive Pragmatic
Blossoms (IPB). Results show that the spectral anomalies identified from a satellite are closely
correlated to the known oilfields and that the micro seepage maps can produce new high-
quality data to reduce exploration risk. The simulation work classification accuracy of 94.12%
which is much better than past outcomes in this engaged field of research. The execution of the
usage is analyzed, an examination is likewise influenced concerning the clustering precision,
time complexity, and false proportion are introduced.

Keywords Fusion . Temporal filter . Linear iterative clustering . Intensive pragmatic blossoms

* R. Swathika
swathikaphd123@yahoo.com

1
Department of Information Technology, SSN College of Engineering, Kalavakkam, Chennai 603110,
India
 Multimedia Tools and Applications

1 Introduction

Satellite Image classification is the way of adjusting at least two pictures of a similar scene
taken at various times, from multiple perspectives, by different sensors. It is broadly utilized as
a part of PC vision applications. The fundamental thought of 3D image registration is to figure
the relative transformation between an arrangements of 3D images and change all images into
a standard organizing framework [20]. References give the thorough study of image registra-
tion techniques. Satellite image enlistment is a vital and testing image preparing task to arrange
and modify diverse images which are gotten at various climatic conditions or by using
particular sensors or at different situations. In remote detecting, Image registration is done
mainly for climate determining, image mosaicing, change revelation, image mix and environ-
mental checking [25]. In an extensive segment of the traditional image registration systems, the
control centers are picked physically and applying the change illustrate. In any case, this
procedure is time-devouring and less precise. So extraction of observing concentrates conse-
quently tackles the execution issues of manual choice of control focuses. In programmed
image registration, the expulsion of checking concentrations in the two images is the under-
lying advance, and after that, the correspondence between the images are assessed by utilizing
coordinating system criteria lastly the transformation parameters are evaluated.
All in all, satellite image registration algorithms are arranged into pairwise registration and
multiview registration. Pairwise Registration is the way of computing an ideal Transformation
between a couples of images [11]. Multiview registration is to register the global changes of an
arrangement of images and incorporate all images. Pairwise Registration right off the bat looks
correspondence between two images and furthermore ascertains the transformation of a present
image on a reference image. The notable Iterative Closest Point (ICP) is a productive device
for pairwise registration [5]. The conventional technique for multiview registration is to figure
the adjusting transformation between two successive images right off the bat and afterward to
chain these pairwise transformations together to compute the worldwide difference in each
photo. Since just a couple of continuous images is enrolled autonomously, neighborhood
blunders from pairwise registration tend to aggregate, to cause an oversight float and imme-
diate conflicting worldwide record result.
Along these lines, a few systems of spreading the amassed mistake among all
perspectives or playing out a general improvement transformations have proposed.
Metaview system for multiview registration is to enroll and consolidate each image
into the Metaview incrementally. Its hindrance is that the registration mistake that has
joined into Meta see isn’t remedied later. Synchronous minimization approach enlists
and incorporates all images at the same time. It is a sure answer for the issue of
blunder engendering, yet it experiences high memory and computational prerequisite.
Chart approach builds a diagram in light of the pairwise transformations. It tries to
locate the ideal way or recognize a cycle to limit worldwide registration blunder. This
algorithm, be that as it may, depends on comprehensively enrolling and looking at
any two images. Measurable procedure considers the issue of image registration as
Simultaneous Localization and Map Building issue. It can deal with the vulnerability
proficiently. However, the control of the huge state vector, which incorporates vast
scale natural highlights, brings about high computational complexity [7]. Likewise,
because of different changes, the examining images are related to clamor and blun-
ders. Image registration ought to measure the hazard, utilize location and examination
of mistakes naturally and check the worldwide consistency of record.
Multimedia Tools and Applications

The paper proposes a satellite image grouping strategy by assessing different vulnerabil-
ities. The framework state incorporates position and introduction of all perspectives to enhance
the worldwide consistency. The area and presentation of point have been evaluated with Linear
Iterative Clustering algorithm and hyper graph structure strategies used to process the global
transformation in the comparing image. This procedure demonstrates a programmed,
accomplishing better accuracy and dynamically reconfigurable image registration in dimin-
ished complexity.
This work comprises of section 1 manages the introduction part of the work, Section 2
discuss the research background of satellite image classification, Section 3 manages proposed
materials and method, section 4 manages results and discussion of the proposed system.
Finally, conclusion and future enhancement are discussed in chapter 5.

2 Related works

Classification of hyper spectral images (HSIs) assumes noteworthy parts in a few applications,
including natural checking, land explorations, accuracy cultivating, and domestic resistances
[10, 21, 26]. In the hyper spectral images, every pixel is a special-dimensional vector. The
special dimensional vector entrances to speak the phantom reaction in several unreasonable
groups, spreading over obvious to the ultraviolet wavelengths [24]. In [18] projected an
alternate positioning built band choice technique of hyper spectral image arrangement. As of
late, a little demonstrating procedure has been broadly. In [15] projected a broad structure to
understand an enormous measure and test grouping issues of portrayal established subspace
clustering. In [13] recommended a unique unsubstantiated subspace learning technique, it’s
naturally decides ideal measurement of highlight planetary and acquires the small dimensional
portrayal is given informational index. Underneath the structure of chart implanting, in [14]
created dual algorithms using inserting the line two -chart into a small dimensional planetary
for hearty subspace learning and clustering.
The essential thought of test pixel is meagerly approached as a straight mix of the couple of
particles from whole preparing lexicon using the 0-standard or 1-standard regularization. The
last mark of the test pixel is allocated to the class whose sub-word reference gives the littlest
portrayal blunder. To get the nonlinear likeness of tests, a bit type of Kernel Sparse Repre-
sentation classifier (KSRC) has proposed in [1]. KSRC maps the cases into a high-dimensional
segment incited incorporate space first, and a while later Sparse Representation classifier
(SRC) is completed in this new component space by utilizing a suction trap. To facilitate the
high computational multifaceted nature of SRC utilizing 1-standard regularization, Zhang et al.
[23] proposed collective representation based classifier (CRC) by 2-standard regularization.
In the hyper spectral remote detecting group, Chen et al. [2] presented a joint sparse
representation classification (JSRC) structure through simultaneous orthogonal matching
pursuit (OMP). Besides, the kernel variant of JSRC is misused in [28]. Liu et al. [16] proposed
a spatial-spectral KSRC using 1-standard minimization to enhance the execution of SRC. Li
and Du [4] suggested a joint inside class CRC, examining the relationship of the hyper spectral
neighbor. Moreover, they proposed a fused representation based classifier (FRC) by joining the
CRC and the SRC. The goal of the FRC is to incorporate the benefits of the CRC and the SRC
to enhance the grouping capacity. Yuan et al. [27] proposed the spectral-spatial order conspire,
which centers around multitasking joint sparse representation (MJSR) and a stepwise Markov
random field structure.
 Multimedia Tools and Applications

Data territory has been an essential issue in numerous machine learning applications. Wang
et al. [17] proposed a zone prerequisite separate metric learning for development blockage
area. In [6, 9, 17, 22], a multistage clustering procedure was relied upon to discover total
developments with both neighborhood and overall consistency along time.
It should be focused on that the general SRC or CRC show does not consider the
partition or difference association between the test pixel and each word reference
particles. Independently, word reference particles, which are closer or more like the
test pixel, expect a more basic part in addressing the test pixel. Once such a
distinction between a vocabulary particle and the test pixel is settled, the vitality of
these particles in a addressing the test pixel should be known. To update the
discriminative limit of SRC, in [19] proposed weighted SRC (WSRC) by using the
weighted 1-minimization issue to consider the distinction data. A partition weighted
CRC, to be a particular nearest regularized subspace (NRS) [3, 12], is furthermore
proposed to manhandle the apparition likeness [i.e., Euclidean detachment (ED)] to
control the regularization term. In [8] proposed Image fusion is a strategy for joining
the Multispectral (MS) and Panchromatic (PAN) images into one image contains more
data than any of the information.
It is discovered that the framework has an issue of satellite image classification in all the
above methods So that another propelled fusion based technique has proposed in this work.
This work discloses how to improve the viability of satellite image classification in various
systems. The simulation work does with the MATLAB condition by using Versatile Unsuper-
vised Vector Zone fusion technique, LIC segmentation, and IPB classification, the simulation
comes about affirm the proficiency of the proposed system.

3 Materials and method

The proposed technique has five stages specifically preprocessing, image fusion, LIC seg-
mentation, feature extraction, and IPB classifier. To take care of the satellite image character-
ization issue recommended work thought of the arrangement of a Versatile Unsupervised
Vector Zone fusion technique, Linear Iterative Clustering segmentation and Intensive Prag-
matic Blossoms classification techniques. In this method need to take care of that difference of
pattern. Before examining the patterns, the system first needs to bring all the images to a
certain level where the patterns are much more visible for the usage without noisy and
unwanted data then we extract the particular features like area features color functions that
could be used for creating the classification model. From this classification model system
finally, can predict the multi-view hyper spectral satellite image for area and surface classifi-
cation and again based on this predicted area will ask question form the user and based on
answer system will decide area and surface and it is a condition. The Block diagram of
proposed system is shown in Fig. 1.

3.1 Preprocessing using temporal filter

Preprocessing of on image is a basic factor for advanced image preparing. The point
of preprocessing is to enhance the image quality before any handling. These prepro-
cessing of an image have different procedures for image improvement. In this work,
the improvement is finished with three stages. The first step is to remove RGB Salt
Multimedia Tools and Applications

INPUT SATELLITE IMAGE

FEATURE IPB
PREPROCESSING VUVZ Fusion LIC
EXTRACTION CLASSIFICATION
image Segmentation

OUTPUT IMAGE

Fig. 1 Block diagram of proposed system

and Pepper noise in color image, the second step is to make edge enhancement, and
the third step is image enhancement which is done by a temporal filter. Temporal
filter sorts all the pixel values in the window and then change the middle value with
the pixel value being considered. It preserves the sharp edges. Since the temporal
quality is computed from the area pixel, it is more robust to anomalies and does not
make another sensible pixel esteem, which preserves edge blurring and loss of image
detail. It protects the sharp high-frequency subtle elements. It is appropriate to expel
salt and pepper commotion from the image. The impact of a temporal filter in
evacuating noise is expanded as the window estimate increment. The mathematical
function for a temporal filter is written below in equation.
f ðx; yÞ ¼ medianfgði;ÞgJi; jϵN ð1Þ

Where f(x, y) determines the filtered image & g (I, j) indicates the first image.

a) Land sat I Input Image b) Land sat I Preprocessed

c) ALI Input Image d) ALI Preprocessed Image

Fig. 2 ALI and Land Sat I Preprocessed Image
 Multimedia Tools and Applications

3.1.1 Algorithm

The above figure 2 shows the preprocessed phase results using a temporal filter for Land sat I
and ALI dataset images, as compared with input image the preprocessed image is look more
energetic view.

3.2 Image fusion using versatile unsupervised vector zone (VUVZ)

The fusion is the process for merging two images to get a better result regarding their accuracy.
The Images acquired by a different sensor with their different resolution are fused together to
extract more information about their land use and land cover. Versatile unsupervised vector
zone fusion method for merging of two images. ALI and Land sat I images have been selected
for fusion. The fusion techniques depend on the pixel value, decision level and features of the
satellite image. The VUVZ has been used to increase the visualization of contrast of the low
and high ends of the image histogram.
The above Fig. 3 shows the workflow diagram of versatile unsupervised vector zone fusion
technique. The algorithm steps are discussed in following.

3.2.1 Algorithm

Step 1: Resampling of the ALI image with the motivation behind conveying its measure-
ment to the measurement of the Land sat I image.
Step 2: Symmetrical and radiometric adjustment of the ALI and Land sat I images.
Step 3: Contrast and diagram remedy of the ALI and Land sat I images.
Step 4: VUVZ image fusion technique.
Multimedia Tools and Applications

Fig. 3 Workflow Diagram of Versatile Unsupervised Vector Zone

Step 5: Adapting of the got multispectral image ([MUL]_MOD) into the HSV shading
model ([HSV]_MUL).
Step 6: Changing the R, G and B groups of the multispectral image into HSV parts
([HSV]_MUL).
 Multimedia Tools and Applications

Step 7: The following stage of an arrangement of another image is substitution the

brilliance part V of the [HSV]_MUL image with the splendor segment of
[HSV]_MOD and changing over the outcome from the HSV shading model to
the RGB shading model.
Step 8: DE- correlation of the novel multispectral image in accordance with the chosen
algorithm:

f RGB ðrÞ→f xyz ðrÞ ð2Þ

Step 9: A pyramidal decay with the given wavelet premise to the given disintegration level
(L) of the brightness filter (X):

f RGB ðrÞ ¼ C XL ðrÞ þ ∑Li¼1 d xi;1 ðrÞ; d xi;2 ðrÞ; d xi;3 ðrÞ ð3Þ

Step 10: A pyramidal consumption with the given wavelet premise to the given breakdown
level (L) of the Land sat I picture:

f p ðrÞ ¼ C XL ðrÞ þ ∑Li¼1 d xi;1 ðrÞ; d xi;2 ðrÞ; d xi;3 ðrÞ ð4Þ

Step 11: Improvement of new parts of the disintegration as per the accompanying guidelines
of unification of coefficients:

L
APP ðrÞ ¼ C XL ðrÞ;
2
X 3
Ai;1 d xi;1 ðrÞ þ Bi;1 d pi;1 ðrÞ ð5Þ
6 7
Detp ðrÞ ¼ ∑Li¼1 4 Ai;2 d xi;2 ðrÞ þ Bi;2 d pi;2 ðrÞ 5
Ai;3 d xi;3 ðrÞ þ Bi;3 d pi;3 ðrÞ

Where.

APP XL approximating wavelet components of the new multispectral image

Detp detailing wavelet components of the new multispectral image.
The above Fig. 4 displays the fusion image steps for two images. In Fig. 4a and b is the
input image from ALI and Land sat I Datasets from Salem region in Tamil Nadu. Figure 4c
shows the image obtained by layer stacking of fusion image.

3.3 Linear iterative clustering segmentation

Linear Iterative Clustering (LIC) which has been demonstrated to outflank different
techniques on account of its effortlessness, adherence to limits, computational speed,
and memory effectiveness. LIC has just a single parameter, Q, which is the coveted
number of similarly estimated super pixels to be created. A discretionary parameter m
can be set up, which controls the minimization of the segmentation. The calculation
Multimedia Tools and Applications

a) Preprocessed Land sat I Image b) Preprocessed ALI Image

c) Fusion Image
Fig. 4 Fusion Image phases

begins by establishing k starting bunch focuses on a logical network divided S pixels

separated. To avoid from placing the seed on an edge or a noisy pixel, group
discussions are bothered in a 3 × 3 neighborhood to the least inclination position.
An iterative methodology begins by appointing every pixel to the closest bunch focus
utilizing a separation measure D (Eq. (6)), which consolidates separation of shading
closeness (Eq. (7)) and separation of spatial

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2  2ffi
dc ds
D¼ þ ð6Þ
m S

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
    2
dc ¼ ∑S p ∈B I xi; yi; S p −I x j; y j; S p ð7Þ

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ffi
 2 
ds ¼ x j −xi þ y j −yi ð8Þ

Input: The number of levels L, beginning square size w (1) and covariance matrices PolSar.
Output: land, water, forest etc.…
Consolidate the segmentation to be a normal network of size w (1) and h (1), and figure the
mean covariance grid Ci and mean location vi for each segmentation i. The last Lth level is
pixel-level.
 Multimedia Tools and Applications

For l = 1 to L do

Present every piece B (l) i with a collective framework on level l, and let xj be the segmentation
that the square B (l) i has a place.

For iteration = 1 to mailers do

For every square B (l) i on level do

If B (l) i is a limit square and related segment then

xs takes values of adjacent segmentation if xs is restructured then

{
Compute the mean covariance matrix and mean position in an incremental manner for
two convoluted super pixels.
}
}
}
}
}
% Pixel Update
For item = 1 to maxItems do
{
For each limit pixel do
{
Update the pixel label if the energy reductions.
If pixel updated then
Update the mean covariance matrix for the two involved super pixels
}
}
}
Multimedia Tools and Applications

The above Fig. 5 shows the clustering segmentation process steps for the fusion image. In
this work, five clusters are taken for linear iterative clustering segmentation by which the
essential information is extracted and then identifying the different components such as water
dense region, roads, vegetation, etc. from the detailed image.

3.4 Feature extraction using collective learning zone feature vector (CLZFV)

After fusion and segmentation work completed the features are extracted. During feature
extraction, more than ten categories of elements were extracted. They are geometric features,
color features, texture features, fractal features and satellite image related features. Most of the
studies related to satellite image recognition use only leaf, color and texture features. In this

a) Fused Image b) Segmentation of Cluster_1

c) Segmentation of Cluster_2 d) Segmentation of Cluster_3

e) Segmentation of Cluster_4 f) Segmentation of Cluster_5

Fig. 5 Clustering steps for Segmentation Image
 Multimedia Tools and Applications

study, the geometric and fractal characteristics were also considered. The list of the various
elements extracted under each category is presented in Table 1.
The five feature sets were first analyzed for their efficiency through satellite image
recognition. The experiments discovered that the satellite image feature shows improved
identification accuracy when compared with other four feature categories. Motivated by this
fact and in search for further improvement possibilities towards precision, the satellite image
features were fused with the different four types to obtain combined feature sets, namely,
GAFS (Area + Geometric), CAFS (Area + Color), TAFS (Area + Texture) and FAFS (Area +
Fractal).

3.4.1 Algorithm

Step 1: Begin with a choice on the estimation of Q = number of bunches

Step 2: Set any underlying bundle that groups the information into Q clusters.
1. Take the principal k information as single-component bunches
2. Assign each of the rest of the (N-Q) information to the group with the closest centroid.
After each task recomputed from the centroid of the separating up cluster.
Step 3: Take every information in arrangement and figure its separation from the centroid of
each of the groups. On the off chance that an example is not at present in the group
with the nearest centroid, change this specimen to that bunch and refresh the
centroid of the cluster picking up the new example and the group losing the sample.
The separation metric utilized is Euclidean separation.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2
d ij ¼ ∑nQ¼1 xiQ −xjQ

Step 4: Repeat step 3 until the meeting is accomplished, that is, until a go through the
preparation test, causes no new assignments.
Step 5: Transform picture from RGB shading space to L*u*v shading space and break down
each shading channel L, u and v.
Step 6: Obtain dark scale image from RGB shading space.
Step 7: Use wavelet outline change to break down the dim level picture into sub-groups (LL,
LH, HL, and HH) and utilize the LH and HL subgroups for further handling.
Step 8: Texture Feature Extraction

(a) Calculate the middle vitality of LH and HL subgroups coefficients in a neighborhood

window. The extent of the window ought to be sufficiently huge to catch the block
surface attributes.
(b) Use Q-implies bunching calculation to characterize the vitality esteems in two classes for
each sub band.
Table 1 Features Extracted From Satellite Image

Feature Category Feature Details

Geometric Features Eccentricity, Extent, Orientation

Texture features Energy, Entropy, Homogeneity, Variance
Color features Mean, Standard Deviation, Skew, Kurtosis
Fractal features Average Fractal Dimension, Standard Deviation Fractal Dimension, Lacunarity
Multimedia Tools and Applications

Fig. 6 GUI Screen with Feature Extraction

(c) Using these data, describe the pixels in finished locale into four gatherings to be distinct,
smooth, vertical, level and complex.

Step 9: Generate the component vector with the end goal that each pixel in the picture has
dimensional element vector which incorporates spatial (x, y), shading (dark level or
L*u*v esteems) and surface (horizontal, vertical, flat or complex) data.
Step 10: Filter the picture utilizing mean move calculation in higher dimensional element space
which incorporates spatial, shading and surface data. The shifting operation can be
controlled by setting the area hiding sweep (hs) and hiding range (hr). The channel
yield (meeting point in mean move calculation) is dictated by shading and additionally
surface data, not at all like in standard mean move separating. This gives better
segregation between districts where hues are comparative, yet the cover is distinctive.
Step 11: Segment the yield picture utilizing linear iterative clustering.

By using versatile zone unsupervised vector totally two features were extracted for proposed
technique, few of them are described as follows.

3.4.2 Energy

The energy is the extent of consistency between the pixels extend = [0, 1].the consistency of
the pixel has communicated

Energy ¼ ∑i; j pði; lÞ2 ð9Þ

3.4.3 Contrast

Contrast is the measure of the distinction in luminance to make question discernable. Range = [0, 1].

Contrast ¼ ∑i; j ji−lj2 pði; jÞ ð10Þ

 Multimedia Tools and Applications

3.4.4 Correlation

A correlation is the measure of the connection between the neighbor pixels. Range = [−1, 1].

ði; jÞp−σμx μy
Correlation ¼ ∑i; j ð11Þ
σx σy

3.4.5 Homogeneity

The homogeneity measures of the closeness of the component dissemination in MGRLBP to

MGRLBP diagonals. Range = [0, 1]
1
Homogeneity ¼ ∑i; j ð12Þ
1 þ ði−l Þ2

Where i, j are pixel and p (i, j) is the pixel value. Thus to categorize the image as normal or abnormal,
the intermediate images are acquired, preprocessed, the feature extracted by using diagnosis zone
unsupervised vector technique. The below Fig. 6 demonstrate GUI screen with feature extraction
parameters of satellite images by using collective learning zone feature vector method.

3.5 Intensive pragmatic blossoms (IPB) classification

The classifier is used to classify zone types from original images. For simplicity, strong
augmented blossom classifier is used here. It takes set of pictures and predicts for each input
image belongs to which of the two categories of fusion image and without fusion image. The
purpose of IPB is to create hyper spectral images to separates two classes with a maximum gap
between them in our proposed system output of segmentation is given as input to intensive
pragmatic blossoms classifier which takes training data, testing data and grouping information
which classifies whether given input image.
To quantify the execution of grouping, we directed five unique arrangements of tests. In the
first round of tests, we utilized 30% of the examples of each class of the dataset to make class
agent vectors (preparing) and the staying 70% of the examples for testing reason. The special
arrangement of examinations is with the quantity of preparing and testing tests in the
proportion 40:60, 50:50, 60:40, and 70:30 separately. In each set, tests are rehashed 20 times
by picking the preparation tests haphazardly. Amid experimentation, we break down 20 trials
for each arrangement of preparing and testing tests arbitrarily.
The execution of the classification has assessed concerning arrangement exactness, accu-
racy, review, and F-measure from the confusion network of grouping. The estimations have
registered by utilizing the conditions portrayed underneath with the accompanying
traditions.TP (True Positive) = Positive units classified as positive. True Negative (TN) =
Negative samples classified as negative. FP (False Positive) = Negative samples classified as
positive. FN (False Negative) = Positive samples classified as negative.

Precision It is the proportion of number of positive examples accurately arranged to the

aggregate number of tests in a class
Multimedia Tools and Applications

Fig. 7 Clustering segmentation steps

 Multimedia Tools and Applications

Fig. 8 Clustering segmentation steps

Multimedia Tools and Applications

Tp
precision ¼
T p þ Fp

Recall It is the proportion of some positive examples accurately grouped to the aggregate
number of tests delegated positive.
Tp
Recall ¼
T p þ Fn

F-measure It is the harmonic mean of precision and recall given by the equation below.
2*precision*recall
F measure ¼
precision þ recall

Accuracy It is the aggregate number of tests accurately arranged to the aggregate number of
tests delegated given by the condition underneath.
Accuracy ¼ ðTP þ TN Þ=ðTp þ Tn þ Fn þ fnÞ

The satellite image classification is performed using both linear iterative segmentation and
versatile unsupervised zone vector technique. The method first performs the region based
satellite image estimation and then captures the image to preprocess and segment the image.
Then the process computes the classification algorithm to classify the results given.

Algorithm:
Input: Satellite Image
Output: Class S.
Start
Region S= Perform Fusion based absorption Estimation.
Image Img = Preprocess Image from two datasets (Temporal Filter)
Image Img = Fusion image (Versatile Unsupervised Vector Zone Technique)
Image Img = segmentation Image(Linear iterative clustering)
KSDF = Intensive Pragmatic Blossoms classification.
If KSDF<Th1> then
Region A = S
Else KSDF <Th2> then
Region B = S
Else if KSDF <Thn> then
Region N = S
End
Stop.
 Multimedia Tools and Applications

Table 2 Evaluation metrics for proposed and existing method

Methods Sensitivity of Tracking Specificity of Tracking Accuracy of Tracking

(%) (%) (%)

Linear Iterative Clustering 93.27 92 94.12

segmentation
Fuzzy Interference System 89.45 90 90.74

4 Results and discussion

The planned approach has been implemented with the Matlab2013a software. MATLAB
is an leading programming language for particular enrolling and imaginative condition
for advanced algorithm movement. It facilitates algorithm, portrayal, and programming
in a complete to-use situation where issues and game plans have conveyed in the first
numerical documentation.
The above Fig. 7 shows the clustering segmentation process steps for the first fusion image,
the input image taken from Land sat I and ALI dataset. In this work, five clusters are taken for
linear iterative clustering segmentation by which the essential information is extracted and then
identifying the different components such as water dense region, roads, vegetation, etc. from
the detailed image.
The above Fig. 8 shows the clustering segmentation process steps for the second fusion
image, the input image is taken from Land sat I and ALI datasets. The same Procedure
repeated for the second input and as the number of clusters increases leading to better accuracy
in the segmentation results.
To assess the execution of our proposed strategy to examined different assessment
measurements like accuracy, sensitivity, and specificity. The acquired assessment mea-
surements are then contrasted with a current algorithm with demonstrating the efficiency
of our framework. In the Table 2 exhibits above demonstrates the assessment parameters
got utilizing our proposed technique and existing strategy.
The above mentioned Fig. 9 demonstrates the measurements like sensitivity, specificity, and
accuracy obtained for our proposed and previous technique. From the figure, obviously our
proposed technique conveys better accuracy for satellite image applications when compared
with that of existing.

Performance comparision
Evaluvaon metrics%

96
94
92
90
88
86
Tracking Sensivity Tracking Specificity Tracking accuracy

Linear Iterave Clustering segmentaon Fuzzy Interferance System

Fig. 9 Graphical representation of various evaluation metrics for proposed and existing method
Multimedia Tools and Applications

Performance Comparision
150

100 98.1
89.3 81.3
67.12
50
20.23
10.2
0
Posive Predicve Value Neagave Predicve Value FDR

Linear Iterave Clustering segmentaon Fuzzy Interferance system

Fig. 10 Graphical representation of different evaluation metrics for different methods

The above mentioned Fig. 10 shows the measurement of positive predictive, negative
predictive and FDR got for our proposed and existing technique. The outcomes clarify
that the proposed system in essential point here is the processing time, which is recorded
to histogram includes as it is minimal complex than the other.
The above mentioned Fig. 11 shows the examination of results got from existing
and proposed strategies mistaken classification. As a compared existing framework the
proposed strategy has a less false classification rate which is insignificant.
Figure 12 shows the comparison of exsiting and proposed method on time complexity, and
it depicts that the proposed method has less time complexity than others. From the most
importantly examination played out, the proposed technique has been assessed with different
parameters and has delivered powerful outcomes with all elements of optical properties and
order.

5 Conclusion

In This work implementation adopts fusion technique based satellite image classification. The
input image is preprocessed, removes the features like gray esteem and unkind gray distribu-
tion. It output higher efficient segmentation excellence and consumes a smaller amount space
and time complexity. The segmented image is utilized to achieve area established grayscale

30
False Rao
25
False Rao in %

20

15

10

0
Fuzzy IPB

Fig. 11 Comparison of false classification ratio

 Multimedia Tools and Applications

Time Complexity

Time Complexity in Seconds

70
60
50
40
30
20
10
0
Fuzzy Proposed IPB
Land Sat I AND ALI DATASET 59 49

Fig. 12 Comparison of time complexity of different methods

estimate, and then the strategy computes the grayscale distributional feature for diverse zones
of the segmented image. To conclude founded on the grayscale distributional feature and the
approached area value the technique classifies the satellite image below a diverse class of
regions. This technique has produced useful results in image classification and improved the
accuracy and reduces the false classification ratio also. The work output classification accuracy
of 94.12% which is far better than earlier results in this engrossed arena of research. The
presentation of the execution is examined, a comparison is also made regarding the clustering
accuracy, time complexity, and false classification ratio is presented.

References

1. Alvera-Azcárate A, Barth A (2016) Reconstruction and analysis of long-term satellite-derived sea surface
temperature for the South China Sea. J Oceanogr 72(5):707–726
2. Chen C, Li Y (2015) SIRF: Simultaneous Satellite Image Registration and Fusion in a Unified Framework.
IEEE Trans Image Process 24(11)
3. Fiorucci F, Santangelo M (2013) Very-High Resolution Stereoscopic Satellite Images for Landslide
Mapping. Landslide Science and Practice 125:95–101
4. Gharbia R, El-Baz AH (2015) Remote Sensing Image Fusion Approach Based on Brovey and Wavelets
Transforms. In: Proceedings of the Fifth International Conference on Innovations in Bio-Inspired
Computing and Applications IBICA 2014 pp 311–321,(AISC, volume 303)
5. Guo Y, Sohel F, Bennamoun M, Wan J, Lu M (2014) An Accurate and Robust Range Image Registration
Algorithm for 3D Object Modeling. IEEE Transactions on Multimedia 16(5):1377–1390
6. He T, Liang S (2014) Fusion of Satellite Land Surface Albedo Products Across Scales Using a
Multiresolution Tree Method in the North Central United States. IEEE Trans Geosci Remote Sens 52(6)
7. Kalay A, Ulusoy I (2009) Vision-based Simultaneous Localization and Map Building: Stereo and mono
SLAM. In: 2009 IEEE 17th Signal Processing and Communications Applications Conference, pp. 125–128
8. Kumaraswamy S (2016) Satellite image fusion using fuzzy logic. Acta Univ Sapientiae, Informatics 8(2):
241–253
9. Liang L, Schwartz MD (2014) A Cross-Comparison of Spatiotemporally Enhanced Springtime
Phenological Measurements From Satellites and Ground in a Northern U.S. Mixed Forest. IEEE Trans
Geosci Remote Sens 52(12)
10. Liu Q, Hang R (2018) Learning Multi scale Deep Features for High-Resolution Satellite Image Scene
Classification. IEEE Trans Geosci Remote Sens 56(1)
11. Luo N, Wang Q (2018) Effective outlier matches pruning algorithm for rigid pairwise point cloud
registration using distance disparity matrix. IET Comput Vis 12(2):220–232
12. Mukherjee S, Chakrabarti A (2014) Satellite Image Fusion Using Window Based PCA. Advances in
Intelligent Systems and Computing book series (AISC, volume 249), pp. 293–301
Multimedia Tools and Applications

13. Ren H, Ye X (2017) Improving Land Surface Temperature and Emissivity Retrieval From the Chinese
Gaofen-5 Satellite Using a Hybrid Algorithm. IEEE Trans Geosci Remote Sens PP(99)
14. Saul’skii VK (2017) Analysis of satellite systems for the periodic survey of the Earth. Cosm Res
55(4):275–289
15. Sawada Y, Koike T (2017) Fusing Microwave and Optical Satellite Observations to Simultaneously
Retrieve Surface Soil Moisture, Vegetation Water Content, and Surface Soil Roughness. IEEE Trans
Geosci Remote Sens 55(11)
16. Singh PP, Garg RD (2015) Classification of High-Resolution Satellite Images Using Equivariant Robust
Independent Component Analysis. Advanced Computing, Networking, and Informatics 1:283–290, (SIST,
volume 27)
17. Singh D, Kumar S (2015) Application of Fusion Technique in Satellite Images for Change Detection.
Advances in Intelligent Informatics pp 97–105, (AISC, volume 320)
18. Son S, Wang M (2017) Ice Detection for Satellite Ocean Color Data Processing in the Great Lakes. IEEE
Trans Geosci Remote Sens 55(12)
19. Song H, Huang B (2013) Spatiotemporal Satellite Image Fusion Through One-Pair Image Learning. IEEE
Trans Geosci Remote Sens 51(4)
20. Tao L, Bui T, Hasegawa H (2017) Global ray-casting range image registration. IPSJ Transactions on
Computer Vision and Applications 9:14. https://doi.org/10.1186/s41074-017-0025-4
21. Tao X, Liang S (2017) Improving Satellite Estimates of the Fraction of Absorbed Photosynthetically
Active Radiation Through Data Integration: Methodology and Validation. IEEE Trans Geosci
Remote Sens PP(99)
22. Tian J, Cui S (2014) Building Change Detection Based on Satellite Stereo Imagery and Digital Surface
Models. IEEE Trans Geosci Remote Sens 52(1)
23. Vanjare A, Omkar SN, Senthilnath J (2016) Multi-temporal Satellite Image Analysis Using Unsupervised
Techniques. Advances in Computing and Information Technology (AISC) 177:757–765
24. Wei J, Wang L (2017) Spatiotemporal Fusion of MODIS and Landsat-7 Reflectance Images via
Compressed Sensing. IEEE Trans Geosci Remote Sens 55(12)
25. Xiang Y, Wang F, You H (2018) OS-SIFT: A Robust SIFT-Like Algorithm for High-Resolution Optical-to-
SAR Image Registration in Suburban Areas. IEEE Trans Geosci Remote Sens PP(99):1–13
26. Yang B, Pi Y (2018) Relative Geometric Refinement of Patch Images Without Use of Ground Control
Points for the Geostationary Optical Satellite GaoFen4. IEEE Trans Geosci Remote Sens 56(1)
27. Yu X, Zhang Y (2015) Multi-source Remote Sensing Image Fusion Method Based on Sparse
Representation. Geo-Informatics in Resource Management and Sustainable Ecosystem pp 252–265,
(CCIS, volume 482)
28. Zeng T, Ao D (2015) Multiangle BSAR Imaging Based on BeiDou-2 Navigation Satellite System:
Experiments and Preliminary Results. IEEE Trans Geosci Remote Sens 53(10)

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

R. Swathika Assistant Professor in the Department of Information Technology has 15 years of teaching
experience. She received her B.E (CSE) degree with first class from Bharathiyar University, M.Tech (Computer
Science and Engineering) degree first class with distinction from National Institute of Technology (NIT), Trichy.
 Multimedia Tools and Applications

She is pursuing Ph.D. (part-time programme) at Anna University, Chennai in the area of Image Processing. Her
research interests include Image processing, Feature Extraction and Satellite image analysis.

T. Sree Sharmila Associate Professor in the Department of information Technology has 14 years of teaching
experience. She obtained her Ph.D. degree from Anna University, Chennai under the Faculty of Information and
Communication Engineering in July 2013. She completed her Master of Engineering in Computer Science and
Engineering from Annamalai University, Chidambaram during June 2005 and her Bachelor of Engineering in
Information Technology from Manonmaniam Sundaranar University, Tirunelveli during June 2003. She joined
SSNCE in June 2005 as Lecturer in the department of Information Technology. Her current research areas of
interest include image preprocessing techniques, texture analysis, feature extraction, satellite and medical image
analysis.