Similarity Report ID: oid:3618:52022925

PAPER NAME

IEEE_Conference_Template.pdf

WORD COUNT CHARACTER COUNT

2275 Words 13300 Characters

PAGE COUNT FILE SIZE

3 Pages 78.2KB

SUBMISSION DATE REPORT DATE

Feb 15, 2024 11:35 AM GMT+5:30 Feb 15, 2024 11:35 AM GMT+5:30

28% Overall Similarity

The combined total of all matches, including overlapping sources, for each database.
16% Internet database 23% Publications database
Crossref database Crossref Posted Content database
10% Submitted Works database

Summary
 A Comparative Study of Satellite Image
Segmentation
*

Chandril Adhikary Prof. Monalisa Dey

11
Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering
Institute of Engineering and Management Institute of Engineering and Management
Kolkata, India Kolkata, India

Abstract—Satellite Image Segmentation assumes a pivotal role • Methodological Diversity: This criteria ensures that the
in various sectors like defence , finance , disaster management , selected papers showcase a wide range of segmentation
agriculture , urban planning , etc. The aim of this paper is to techniques such as rule-based segmentation , traditional
survey various advancements in satellite image segmentation and
provide a comprehensive study after exploring the advantages image processing , machine learning , deep learning and
and disadvantage of the various methodologies used. semantic segmentation.
13
Index Terms—satellite, image segmentation • Comparative Studies: This criteria ensures that the se-
lected papers has multiple segmentation techniques eval-
I. I NTRODUCTION uated . This allows a direct comparison of performance.
Satellite Image Segmentation makes use of the vast number • Citations and Impact Factor: This criteria takes into
21
of images which satellites capture from any given point in account the number of citation the paper received and the
space and discern and categorize the information to draw impact factor of the journal or the conference in which
meaningful insights from it. Early segmentation techniques the paper was published . This helped us to establish the
where rule based and hand crafted limited in handling diverse effect the paper had in this particular field.
satellite imagery. Advancements came with the rise of remote
sensing applications , fuelling the demand for advanced image III. A COMPARATIVE STUDY
analysis techniques . The revolution in satellite imagery began In this section we compare different methods of satellite
with the rise of machine learning particularly deep learning segmentation starting from the year 2011 all the way till
algorithms. The enhanced capabilities of Convolutional Neural 2023 and try to figure out the advancements made in this
Networks (CNNs) prompted the focus to shift towards se- particular field of work with respect to the application areas
mantic segmentation, which allowed detailed classification of and efficiency figures.
each pixel in satellite images, enabling applications like land
cover mapping. With the advancements in techniques like deep A. T-Clustering Self Organizing Map [1] Based Technique
learning , transfer learning and domain adaptation , state-of- Self-organizing map and threshold technique work hand in
the-art models continue to evolve leveraging these techniques. hand one after another to complete the segmentation process.
7
Self-Organizing Map is a technique which falls under artificial
II. M ETHODOLOGY neural network method. It identifies patterns and converts the
7
A. Criteria used for selecting the techniques under review patterns which are of arbitrary dimensions into the responses
12
For the inclusion of papers in this comparative review , an of a two-dimensional neuron array. The feature map helps
appropriate criteria needs to be selected . This is a crucial in preserving the neighborhood relations of the given input
step for ensuring the comprehensiveness and relevance of our pattern which is one of the major features of SOM. After
comparative review. The selected criterion are : SOM is completed threshold the T-Cluster technique is used
• Relevance: This criteria ensures that the selected pa-
to decrease over segmentation and remove tiny clusters. These
pers address satellite image segmentation and verify the two techniques need to work together to complete the whole
7
methodologies used are directly applicable to the field process. First the satellite image is provided to the SOM to
arrange the pixel in different groups. The T-Cluster method
• Publication Date: This criteria ensures that the selected 7
is then used to obtain the final number of cluster centers
papers are of a publication date of different years and all
from the organized pixels. The problems of over and under
of the papers capture the all advancements made in the
segmentation which is generated by the SOM is handled
field. 14
separately by the TSOM. A crucial step in image processing is
19
Identify applicable funding agency here. If none, delete this. segmentation. Lack of unsupervised neural network methods
 14
lead to this model where a sequential methods for satellite information of different branches by using the principle of
3
image processing is implemented using SOM and T-Cluster stacking and achieve multi-scale feature fusion. The depth
Segmentation used one after another. The threshold value of a branch in a parallel structure is the same and every
3 3
and number of iterations compute the efficiency of the given one of them has only one single convolution layer. The
selected process. However individual values can be determined importance of different effective feature layers is first weighed
easily for individual images and these values can be used and then their usage is enhanced. This is called the attention
further in segmenting the next set of images. mechanism. In conclusion this model refines segmentation
3
results through network design and loss function design. In
B. Decision Tree, SVM and K-NN [2]Based Technique 3
order to add AD-Link to the central portion of AD-LinkNet,
3
K – Nearest Neighbor Algorithm also known as KNN this procedure builds a combination module (AD-Link), which
20
algorithm is a machine learning algorithm used to classify consists of a series-parallel combination dilated convolution
18
objects into classes based on the closest training examples in and two channel-wise Attention mechanism.
the feature space. Among the machine learning algorithms, this
one is the most basic. The label of the unlabeled data point is D. Deep Learning Segmentation Using U-Net [4] Based Tech-
simply assigned through the classification procedure based on nique
12
its k-nearest neighbors. Based on the divide and conquer tactic, Semantic Segmentation is a method of labelling each pixel
4
the decision tree classifier is one of the inductive learning of an image and then labelling each pixel into different
algorithms in machine learning that creates a classification classes.The encoder-decoder architecture, on which U-Net is
tree using the training data / sample. The process begins built, is highly well-liked because of its excellent performance,
when high resolution and low-resolution images are chosen versatility, and efficiency. The basic operations of U-Net are
4 2
for classification. Then these images are segmented based (1) the convolutional layers which are implemented for feature
on correlation between neighborhood pixels. In the next step extraction through the use of multiple 3 × 3 convolution
classification data is selected where 70 percent of the total kernels (denoted by Convolution); (2) the application of the
4
pixels are selected to yield proper classification results. There batch normalizing layer, which is represented by batch nor-
are two methods used in this process, KNN and SVM, which malization and speeds up convergence during training; (3)
need to select a lot of parameters to run. Lastly the ground the layer of the activation function that handles the nonlinear
2
level truth data is compared to the segmented images to get modification of the feature maps; for this layer, we chose the
the accuracy result. In conclusion KNN, SVM and decision commonly used rectified linear unit (ReLU), also known as
tree data mining algorithms are used to classify the the activation; (4) the max-pooling layer (represented by max-
input images. The accuracy in the SVM method is higher as pooling) for feature map downsampling; (5) the application of
17
compared to the KNN and Decision Tree algorithms. High the up-sampling layer, which restores the size of the feature
resolution images have a accuracy of 78.60% and the low- maps that the max-pooling layer (represented by up-sampling)
2
resolution images have an accuracy of 82.34% respectively. had previously down-sampled; and (6) the concatenation layer
The kappa coefficient in SVM method is quite higher than in which was used for combining the up-sampled feature map
the other two methods. in the deep layers along the the corresponding feature map
22 obtained from the shallow layers. The U-Net architecture can
C. Attention dilation-LinkNet (AD-LinkNet) [3] Neural Net- 2
be briefly summarized as: it receives the input images, and
work Based Technique the input image runs through two convolution operations with
Attention Dilation-LinkNet, also known as AD-LinkNet, ReLu activation function. The image is encoded then and
2 13
uses a pretrained encoder, a serial-parallel combination dilated passes through the pooling layer. The size of the feature maps
convolution, a channel-wise attention mechanism, and an is decreased as a result of this several-time iterative approach.
encoder-decoder structure to achieve semantic segmentation. The model initiates sampling after obtaining all of the sample
Dilated convolution in serial–parallel combination broadens maps. The feature map’s layers are concatenated during the
3
the receptive field and builds multi-scale characteristics for downsampling procedure. After sampling up the feature maps,
multi-scale objects, such a little pool and a long-span road. the segmentation maps are ultimately displayed as an output
3
Based on our previous D-LinkNet34, AD-LinkNet is a neural with the same dimensions as the original picture.
network that brings together the benefits of several networks. 16
To produce a more comprehensive semantic segmentation E. Active Learning for Improved Semi-Supervised Semantic
network, AD-LinkNet incorporates an attention mechanism Segmentation in Satellite Images [5] Based Technique
3 1
and a series parallel combination dilated convolution into Active learning uses the most informative and representative
the network. A structure that is a series-parallel combination instances of the data and uses them to label from the unlabelled
1
is also produced by the residual network’s ”parallel expan- pool of data based on some information measure.A conditional
sion” properties and the dilated convolution formed by using Generative Adversarial Network (GAN) is used in which
1
the short-cut connection.The variety of convolutional struc- a sample of the subset of images and their corresponding
tures are enabled to be used by parallel dilated convolution labels at random from the dataset is used as an input. Using
3 1
structures with different dilated ratios. It then combines the an active learning-based sampling strategy the sampling is
 1
done again to provide a more diverse set of training data research methodologies which would serve as an stepping
which shows a relative performance improvement. The use of stone for advancements and applications in sectors such as
1
active learning techniques is proposed to select a labelled and defense,disaster management,agriculture,urban planning,etc.
1
informative data subset that would help the semi-supervised
ACKNOWLEDGMENT
semantic segmentation model use the representative pool of
labelled data more wisely.This work presents an approach that I extend my sincerest gratitude to my professor and mentor,
1
uses active learning-based modeling techniques to improve Prof. Monalisa Dey, whose guidance and support throughout
1
the performance of semi-supervised semantic segmentation the entire process, along with her valuable insights and con-
tasks for land cover classification in satellite images. The structive feedback, significantly contributed to the enhanced
1 15
samples which are labelled pixel wise are obtained using quality of my work. I would also like to thank my institution
coarse image classification based active-learning strategies and for giving me an opportunity to work in such an interesting
then selected .Semi-supervised semantic segmentation network subject and providing me with the necessary facilities required
1
helps to obtain appropriate amount of initial data to learn to complete this project.
suitable representatives starting from the best registered mod- R EFERENCES
1
els.The tagged images were chosen using pool-based active 10
1 [1] Awad, M., 2010. ”An Unsupervised Artificial Neural Network Method
learning algorithms in a GAN-based semi-supervised semantic for Satellite Image Segmentation”. Int. Arab J. Inf. Technol., 7(2),
segmentation network prototype.Next, we provide quantitative pp.199-205.
9
and qualitative evidence of our method’s effectiveness for two [2] Nurwauziyah, I., UD, S., Putra, I.G.B. and Firdaus, M.I., 2018. ”Satellite
image classification using Decision Tree, SVM and k-Nearest Neigh-
datasets of satellite images, showing significant performance bor”. no. July.
improvements. This study introduces a novel methodology 8
[3] Wu, M., Zhang, C., Liu, J., Zhou, L. and Li, X., 2019. ”Towards accurate
that utilizes active learning-based sampling techniques to en- high resolution satellite image semantic segmentation”. Ieee Access, 7,
pp.55609-55619.
hance the performance of semi-supervised semantic segmen- 5
[4] Pan, Z., Xu, J., Guo, Y., Hu, Y. and Wang, G., 2020. ”Deep learning
tation in the realm of land cover classification for satellite segmentation and classification for urban village using a worldview
imagery. The key innovation involves strategically selecting satellite image based on U-Net”. Remote Sensing, 12(10), p.1574.
6
[5] Desai, S. and Ghose, D., 2022. ”Active learning for improved semi-
samples for pixel-wise labeling, guided by active-learning supervised semantic segmentation in satellite images”. In Proceedings
strategies rooted in coarse image classification.The objective of the IEEE/CVF winter conference on applications of computer vision
is to provide an ideal set of labeled instances to the semi- (pp. 553-563).
supervised semantic segmentation network, which will aid
in the acquisition of essential preliminary data for learning
an appropriate representation. .Implemented as a prototype
for a semi-supervised semantic segmentation network based
on Generative Adversarial Networks (GANs), our approach
selects labeled images through pool-based active learning
strategies.Two distinct satellite image datasets is used to show-
case the method’s effectiveness,leveraging both quantitative
metrics and qualitative assessments to showcase significant
performance improvements. Our findings suggest that the
intelligent selection of labeled samples using active learning
1
strategies substantially contributes to the success of the GAN-
based semi-supervised semantic segmentation network in the
context of satellite image analysis.
C ONCLUSION
By comparing different methods and techniques in the field
of satellite image segmentation , a comparative study was
presented and the comprehensive analysis was provided .
Through meticulous examination of different methodologies
we evaluated the evolution of satellite image segmentation
through different approach throughout the time period. Each
methodology presented and solved different set of challenges
through their different approaches and manifested unique
strengths and weakness. The comparative analysis highlighted
the contextual nuances that researchers and practitioners
should weigh in. As this comparative study is concluded ,
it is imperative to acknowledge the inherit complexity of
Satellite Image Segmentation and the evolving nature of the
 Similarity Report ID: oid:3618:52022925

28% Overall Similarity

Top sources found in the following databases:
16% Internet database 23% Publications database
Crossref database Crossref Posted Content database
10% Submitted Works database

TOP SOURCES
The sources with the highest number of matches within the submission. Overlapping sources will not be
displayed.

Shasvat Desai, Debasmita Ghose. "Active Learning for Improved Semi-...

1 5%
Crossref

ijrpr.com
2 4%
Internet

Ming Wu, Chuang Zhang, Jiaming Liu, Lichen Zhou, Xiaoqi Li. "Towards...
3 4%
Crossref

de.slideshare.net
4 2%
Internet

University of Derby on 2024-01-18

5 2%
Submitted works

University of Hertfordshire on 2022-07-22

6 1%
Submitted works

mts.intechopen.com
7 1%
Internet

University of Liverpool on 2020-09-15

8 1%
Submitted works

Sources overview
 Similarity Report ID: oid:3618:52022925

biorxiv.org
9 1%
Internet

University of Leeds on 2020-01-27

10 <1%
Submitted works

researchgate.net
11 <1%
Internet

espace.curtin.edu.au
12 <1%
Internet

Sina Ghassemi, Attilio Fiandrotti, Gianluca Francini, Enrico Magli. "Lear...

13 <1%
Crossref

Mohamad Awad. "<![CDATA[Multicomponent Image Segmentation Usi...

14 <1%
Crossref

The University of Wolverhampton on 2023-08-29

15 <1%
Submitted works

ijarcce.com
16 <1%
Internet

res.mdpi.com
17 <1%
Internet

dokumen.pub
18 <1%
Internet

Ahsanullah University of Science and Technology on 2020-11-06

19 <1%
Submitted works

University of Bedfordshire on 2013-05-24

20 <1%
Submitted works

Sources overview
 Similarity Report ID: oid:3618:52022925

University of Sheffield on 2021-05-19

21 <1%
Submitted works

Zhanyu Ma, Sunwoo Kim, Pascual Martinez-Gomez, Jalil Taghia, Yi-Zh...

22 <1%
Crossref

Sources overview