Journal of Critical Reviews

ISSN- 2394-5125 Vol 7, Issue 8, 2020

A STUDY OF DEEP LEARNING DATASETS AND MODEL IN REMOTE SENSING IMAGES

CLASSIFICATION
N. Subraja1, D. Venkatasekhar2
1Research scholar, Department of ECE, 2Professor Department of IT
1,2Annamalai University, Tamilnadu, India

Email id: [email protected], [email protected]

Received: 08.04.2020 Revised: 06.05.2020 Accepted: 10.06.2020

Abstract
The remote sensing Images Classification plays a vital role of real time applications and Deep Learning massive growth in dissimilar
domains such as NLP, Computer Vision and medical fields. Compared to the machine Learning algorithms, deep networks provide higher
accuracy and also strong ability to learn for data extraction. Geographical satellite pictures that are utilized for the investigation of
environmental and geological fields are acquired through remote sensing techniques. The crude pictures gathered from the satellites are
not appropriate for factual examination and precise report arrangement so raw images undergo the traditional image classification
techniques such as data preprocessing, segmentation, data feature extraction and classification. The old image classification methods
have spatial and spectral resolution problems. This problem can be solved by deep learning techniques. A new image classification
method ,namely Convolutional Neural Networks(CNN) techniques . A noteworthy endeavors have been taken to build up an assortment
of datasets, Deep learning Models and furthermore to show various methodologies that utilization these datasets for remote sensing
images classification.

Keywords--- CNN, Classification, Deep Learning Models, Deep Learning Datasets, Remote Sensing images

© 2020 by Advance Scientific Research. This is an open-access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
DOI: http://dx.doi.org/10.31838/jcr.07.08.297

INTRODUCTION request over the latest couple of years, with symbolism so far
The Remote sensing is the way toward checking a remote item accessible just to not many research bunches turning out to be
without having a physical contact with that objects. The objects significantly more openly available.
are gathering information utilizing the Artificial satellites that are
propelled to spin around the earth. Earth Observation is a Learning Types
procedure of social event the data about planet Earth through The Machine Learning have 3 major types
remote sensing. The area, where we can gather the most 1. Supervised.
information about our planet, whether estimating, horticulture, 2. Unsupervised.
natural disaster, petroleum derivative and minerals recognizable 3. Reinforcement.
proof, mapping of the land use, etc. The heavenly body satellites
create the top notch pictures of the whole earth in a less measure
of time. The pictures delivered by the topographical satellites
have a lot of commotion and unessential information because of
the interruptions caused in the space. Remote detecting is
routinely depicted by complex data properties as heterogeneity
and class inconsistency, and covering class unforeseen
allocations. Together, these points of view comprise genuine
challenges for making land spread maps or recognizing and
limiting things, making a significant level of powerlessness in
obtained results, despite for the best performing models.

The term Hyper spectral imaging goes under Spectral imaging.

Hyper spectral imaging is the social occasion and handling of
data from over the electromagnetic range. The other worldly
design is the principle factor that separates hyper spectral
symbolism from multispectral imagery. Hyper spectral
symbolism contains groups with slender wavelengths while
multispectral symbolism contains groups with wide
wavelengths. The upside of utilizing hyper spectral information
over. Multispectral information is the capacity to characterize
surface highlights with a higher unearthly resolution. Recent Figure 1. ML and DL Sample
mechanical advances in microelectronics have likewise spiraled
into the satellite assembling industry. The scaling down of space 1) Supervised learning: The algorithms uses the training data
grade segments has brought about the ascent of little satellites, and feedbacks its helps to learn a relationship of inputs and
including an extraordinary number of remote detecting satellites. desired outputs. Supervised algorithms help to predicts the
With diminished dispatch and assembling costs, this has incited a output from the given information. it grouped into classification
democratized access to space. Thusly, satellite imaging (a subset and regression problems.
of remote sensing) has encountered an expansion in intrigue and

Journal of critical reviews 1495

 A STUDY OF DEEP LEARNING DATASETS AND MODEL IN REMOTE SENSING IMAGES CLASSIFICATION

a) Classification: result variable is a category, such as “red” or

“green” or “dog” and “no dog”.
b) Regression: result variable is a real value, such as “dollor” or
“coin”.
2) Unsupervised learning: All given information is unlabeled and
the algorithms learn to the intrinsic structure from the given
information. Identified the pattern and structure. Types.
Clustering and Association
3) Reinforcement learning: Trained to the map action-to-
situation, classifier is not programmed directly , trial and error.

Deep Learning Works

1. Data Collection
Assortment of data is one of the major and most significant
assignments of any AI ventures. Since the data we feed to the
calculations is data. Thus, the calculations productivity and
accuracy relies on the rightness and nature of data gathered. So
as the data same will be the yield.

2. Data preprocessing
Gathering the data is one undertaking and making that data
helpful is an-other crucial assignment. Data gathered from
different methods will be in a sloppy arrangement and there Figure 2. Deep learning working sample
might be part of invalid qualities, in-legitimate data esteems and
undesirable data. Cleaning every one of these data and CNN
supplanting them with proper or inexact data and evacuating The neural systems exploit the hidden structure in images.
invalid and missing data and supplanting them with some fixed Topological information, i.e., spatial information about the
interchange esteems are the essential strides in pre handling of structure in an image, for example, contiguousness and cycles are
data. Indeed, even data gathered may contain totally trash additionally considered. presently investigate the subtleties of
esteems. It may not be in careful arrangement or way that is how the various layers of a convolution neural system interact
intended to be. Every single such case must be confirmed and with each other.
supplanted with interchange esteems to make data meaning
significant and valuable for further preparing. Data must be kept
in a composed organization.

3. Testing and Training

At last in the wake of handling of data and preparing the
following undertaking is clearly trying. This is the place
execution of the calculation, nature of data, and required yield all
shows up out. From the enormous data collection gathered 80
percent of the data is used for preparing and 20 percent of the
data is held for testing. Preparing as talked about before is the
way toward making the machine to learn and giving it the
capacity to make further forecasts dependent on the preparation
it took. Though testing methods previously having a predefined
data collection with yield likewise recently named and the model
is tried whether it is working appropriately or not and is giving
the correct expectation or not. On the off chance that most
extreme number of forecasts are correct, at that point model will
have a decent accuracy rate and is dependable to proceed with
generally better to change the model.

4. Deep Learning Algorithm

The subsequent stage is calculations are applied to data and
results are noted and watched. The calculations are applied in Figure 3. Deep Learning Flow diagram
the design notice in the graph to improve accuracy at each stage.

LITERATURE SURVEY

YEAR
AUTHOR(s) TECHNIQUES /ACCURACY LEVEL
S. NO

1 Cansu et.al [2] 2010 Viola-Jones Approach Accuracy (Ac) 0:9930

Object-Based, Multitemporal
Nesrine Chehata , Classification.
2 2011
et.al[3]
Overall Accuracy Was Of 87.2%

Journal of critical reviews 1496

 A STUDY OF DEEP LEARNING DATASETS AND MODEL IN REMOTE SENSING IMAGES CLASSIFICATION

2012 Object Based Method.

3 Daxiang Zhang, et.al[4]
Overall Accuracy Was Of 93.33%

The Orfeo Toolbox (Otb).

4 Sophie Sauvagnargues 2013
, et.al[5]
Overall Accuracy About 87%

Resource
Selection Analysis (RSA),
5 Mariano R. et.al[6] 2013
Accuracy : 81.1%
Image Processing Technique.
Minakshi Kumar 2013
6 Accuracy 95%.
et.al[7]

Charles Joseph Hanley Object Based Extraction

7 2015
et.al[8] Accuracy 89%

Road Extraction Techniques

Reshma Suresh Babu, 2016
8
et.al[9]

Gram–Schmidt Method
Sandra Heleno, 2016
9
et.al[10]
Accuracy :77 %

L.Chaouche Ramdane, 2016

10 Multi-Resolution Approach Accuracy : 98%
et.al [11]

K. Joshil Raj Bio Inspired Algorithms (BIA)

11 2016
et.al[12] Accuracy 94.5 %

MLC, NBC, SVM,

12 Zhiyong Lv 1, et.al[13] 2017
Accuracy : Above 90 %

Fuzzy Synthetic Evaluation (FSE) Approach.

Bakhtiar Feizizadeha,
13
et.al[14] 2017
Accuracy :93.87 %

Two-Stage Optimization Strategy

Maher Ibrahim
14
Sameen et.al[15] 2017
Accuracy : 76 %

Segmentation Algorithm.
Valentine Lebourgeois
15 2017
,Stéphane et.al
Accuracy : 91.7 %

DEEP LEARNING CONVOLUTIONAL NEURAL NETWORK maximum pooling layer is utilized to maximal estimations of
MODELS AND DATSETS spatially progressive nearby locales on the component maps to
1) AlexNet: The convolution layer is actualized by the extricate the yield.
information which is tangled with a lot of channel. The non-
straight capacity is used to create a gathering of highlight. The

Journal of critical reviews 1497

 A STUDY OF DEEP LEARNING DATASETS AND MODEL IN REMOTE SENSING IMAGES CLASSIFICATION

2) CaffeNet: The Caffe model gives a total toolbox to preparing, Table 2. Accuracy level comparison of Optimization Techniques
testing, fine-tuning, and sorting out models, with well Algorithm/tools Classification Accuracy Level
documented. Simultaneously, it's conceivable the snappiest Cuckoo search Per Pixel 0.94
accessible execution of these calculations, making it in a flash algorithm
important for mechanical organization ACO2/PSO 25 Per Pixel 0.97
. Rough Set ,BBO Object Based 0.67
3) Google Net Going Deeper with Convolutions additionally ACO2/PSO , BBO Per Pixel and Object 0.98
called Google Net. I) by utilizing channels of various sizes at each Fuzzy , BBO Object 0.69
layer, it reviews increasingly precise spatial information. ii) it ACO , SOM Object 0.70
altogether diminishes the quantity of free parameters of the
ACO,BBO Object 0.76
network, making it less arranged to overfitting and enabling it to
FPAB , BBO Object 0.67
be more profound. Beginning modules are utilized within the all
Ant Miner , BBO Object 0.97
convolutions layer by utilizing corrected straight actuation.
ABC , BBO Object 0.91
4) VGGNet-16 VGGNet has two surely understood models: AIN Per Pixel 0.86
VGGNet-16 and VGGNet-19. In this estimation[15], its utilized the
previous one due to its more straightforward structure and to
some degree better execution. This model had three stages first
stage 13 convolutional layers, second stage 5 pooling layers, and
third stage 3 completely associated layers. The VGGNet-16 CNN
highlight was likewise take out from the second completely
associated layer to get a component vector of 4,096
measurements.

5) Places Net, has a comparable design with Caffe Net. The

profound highlights from Places Net are more powerful in
perceiving common scenes than profound highlights from
convolution neural network prepared on Image Net. We will
compute Places Net to affirm whether it brings about exceptional
execution.

6) VG GNet To evaluate the presentation of various profound

convolution neural network structure and partner them on a
shared opinion, created three convolution neural network
models dependent on the Caffe toolbox, every one of which finds
a not a similar speed/accuracy.

7) Inception-v3 Inception-v3 is extra Image Net-improved Figure 5. Sample Remote sensing Images
structure. It is built up by Google and has a solid significance on
making scaling to profound networks computationally proficient. Table 3. Sample Dataset for Remote sensing Images
The Inception model takes in 299 x 299 pictures for this Classification
methodology.
Image
Total Scene Image
8) ResNet-50 ResNet-50 network is set up by Microsoft Research S.No Data Set Per
Image Class Per Size
utilizing a structure that utilizations lingering capacities to which Class
help to add impressive security to profound networks. It assess
18-layer and 34-layer leftover nets (ResNets). 1 SAT-4 50,000 100,00 4 28X28
2 IMAGENET 14000 2 2 256X256
The gauge structures associations are added to each match of
3 UC 2100 100 21 256X256
3X3 filters .Res Net diminishes the best 1 blunder by, coming
MERCED
about because of the adequately decreased preparing
LAND USE
error.ResNet-50 is the 50-layer variant of ResNet. ResNet utilizes
224 x 224 pictures for this network. 4 EUROSAT 27000 2000 10 64X64
5 WHU-RS 50 50 19 600X600
6 RSSCN7 400 400 7 400X400
7 SPACENET 17355 50 1 650X650
8 SAT-6 405,00 32400 6 28X28
9 BRAZILIAM 2876 1438 2 64X64
COFFEE
SCENE
10 SIRI-WHU 2400 200 12 200X200
11 RSC11 1232 10 11 512X512

QUANTITATIVE EVALUATION METRICS FOR DEEP LEARNING

Quantitative evaluation based on three matrices
1. Accuracy
Figure 4. Satellite Image Sample Classes 2. Precision

Journal of critical reviews 1498

 A STUDY OF DEEP LEARNING DATASETS AND MODEL IN REMOTE SENSING IMAGES CLASSIFICATION

2. Recall
TP + TN
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
TP + FP + TN + FN

TP
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
TP + FP

TP
𝑅𝑒𝑐𝑎𝑙𝑙 =
TP + FN

CONCLUSION
An wide comparative survey given on various methodologies
available in deep learning techniques and models. Also in this
paper analyzing the difference literature gap for deep learning
techniques and models. In this paper provides quantitative
metrics like accuracy, precision, recall for evaluate the satellite
images. All the supplementary information will be extremely
much useful for image classification and identification of images
.Its help the users to learn significant feature representations.

REFERENCES
1. M. P. Vaishnnave et.al ., "A Study on Deep Learning Models
for Satellite Imagery" IJAER vol 14. 887.
2. F. Hu, et.al., "Transferring deep convolutional neural
networks for the scene classification of high-resolution
remote sensing imagery". Page no 14707.
3. C. Tao, et.al., "Unsupervised spectral–spatial feature
learning with stacked sparse auto encoder for
hyperspectral imagery classification". IEEE geo science.
4. X. Yao, et.al. "Semantic annotation of high resolution
satellite images via weakly supervised learning". ,3660–
3671.
5. Y. Yang et.al ., "Bag-of-visual-words and spatial extensions
for land-use classification". Syst:270-279.
6. Y. Wang, L. et.al., "Mathiopoulos A Three Layered Graph-
Based Learning Approach for Remote Sensing Image
Retrieval". Page no 6020-6034.
7. L. Gueguen et.al., "Classifying compound structures in
satellite images: A compressed representation for fast
queries". IEEE 1803-1818.
8. G. Sheng, W. et.al., "High resolution satellite scene
classification using a sparse coding based multiple feature
combination".
9. B. Zhao, et.al., “The Fisher Kernel coding framework for high
spatial resolution scene classification". Remote Sensing:
157.
10. B. Zhao, et.al. “Dirichlet-derived multiple topic scene
classification model for high spatial resolution remote
sensing imagery". IEEE Trans. Geosci. Vol 1. 448.
11. H. Wu, et.al., “ Hierarchical Coding Vectors for Scene Level
Land-Use Classification.” Remote Sensing:8(5), 436.
12. K. He, et.al "Deep residual learning for image recognition".
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition: 770–778.
13. O. A. Penatti, K. et.al "Do deep features generalize from
everyday objects to remote sensing and aerial scenes
domains" in Proc. IEEE Int. Conf. Comput. Vision Pattern
Recognit. Workshops: 44-51.
14. M. Papadomanolakia∗, et.al. "Bench Marking Deep Learning
Frame Works For The Classification Of Very High
Resolution Satellite Multispectral Data". Remote RISIS
volume 7 ,1-19.
15. Zhong Ma, et.al. “Satellite Imagery Classification Based on
Deep Convolution Network". IJCIE page no 1155-1159.

Journal of critical reviews 1499