Remote Sensing Image Scene Classification

SUMMARY
Satellite View Scene Classification provides an essential use for broad array of applications
and also at a method and also at a therefore is getting special following. In this project, we
have presented a mechanism for remote sensing and also at an Image Classification of large
dataset Image collections. Bag of Visual Words (BoVW) model is used in first part of the
project. But, the legacy Bag of Visual Words structure just considers the nearby traits inside
Images via accessing proximate traits and also at a sector. The set of traits converted to
describing words from the present Bag of Words Classification are generally a result of hand
also at a selected description classes like the scale static trait transform (SIFT). When the
view Classification task turns out to be highly tough, their ability to describe are then
constrained and also at an impoverished. Apart from previously used feature Classification
based on words which are usually acquired under the use of human aided engineering and
also at a hand also at a selected trait to generalize into some classifiers, the stated method
using convolutional neural network called Bag of Convolutional Featured Bag (BoCF)
creates classifiers using deep learning inspired convolutional traits using off-the-shelf CNN.
Grey Wolf inspired Optimization (GWO) has been emerging as a useful meta-heuristic
technique based on particle swarm optimization. Single or Multi modal, superior and also at a
composite function are explored while excluding localized minimum value. This is the that
part of the project that simulates the Grey Wolf Algorithm for satellite view Image
Classification. The formal methods have been tweaked to perform an automated clustering.
The Images of traffic, terrains, residential plots, farmland also at as, etc are targeted. It is
calculated out with its outcomes of numerical efficient value and also at an also, accuracy is
encouraged considering Davies-Bouldin (DB) indexing, mean of the inter-cluster distance
also, the mean within-cluster distance.
Division of the objects is an uncomplex piece of work, but it is exigent for the system. The
Image categorization encompasses Image pre-defined processing, Image sensing, object

recognition, object division, trait discovery and also at a Classification of the found objects.
The system comprises of a set of data which is a predefine fashion compared with objects to
enrol into proper fields. Image Classification is an important and also at a demand also at
acing task in various significant domains, including biology-medical imaging, biometric
identification, CCTV video surveillance, which includes biological and also at a medicinal
imaging, vehicle GPS navigation, industry visual monitoring, robot direction, industrial
visual monitoring, and also at a remote image sensing. A typical classification model consists
because of camera set high on the interested zone, where Images are captured and also at an
accordingly processed. Classification process consists of following steps:
[a].Pre-processing: Atmospheric rectifying, Image transformation, noise removal, main
component analysis, Image transformation etc.
[b].Discovery and also at an extraction of object: Discovery includes Discovery of
position and also at other characteristics of moving object Image derived from
camera. And also, at a in extraction from the discovered object estimating the path of
the object in the Image plane.
[c]. Training: Selection of the specific attribute which describes the patterns to the best.
[d].Classification of the object: Object Classification categorizes detected objects into
arranged classes by using appropriate method that compares the Image patterns with
the aim/target patterns.
With the development of new age satellite tech, High Spatial Resolution (HSR) remote image
sensing Images has provided detailed space-based information. By a “scenic view” with
respect to interpretation of HSR type Remote Sensing Images, it is usually referred to the
local areas in the Images that are containing the clear semantic information on the ground,
e.g., the residential span, greenery land also at as, commercial business area, farmland also at
as and also at a barren land also at a. Object-based and also at a contextual-based methods are
both used for precise object recognition. Scene Classification methods, automatically labels
an Image from the set of semantic meaning-based categories, have been proposed to obtain
the semantic based information from HSR Imagery.
Image Classification is analysing numerical data of Images that are presented in one feature
or combined various traits to sort Images into different categories. Classifying Images
semantically is of great significance in scientific research as well as in industrial practice.
Since the swift development of Internet and also at a multimedia like tech, the count of digital
Images, videos and also at another multimedia information is expand also at acing

exponentially. In order to organize, manage and also at a retrieve Images accurately and also
at an efficiently, computers are required to understand also at the content of Images
accurately. Image Classification is an important approach to solve the problem of Image
understand also at acing which also plays an important role in promoting the development of
multimedia retrieval technology. In addition, as the basis of Image processing, Image traits
representation is the key study content in this field, as its performance affects the results of
Image Classification and also at an object recognition directly.

1.1 Problem Statement
Division of scenic images are difficult if it is improper and also at a noisy content in image.
The two significant areas of scene Classification problem are: learning and also at a scene’s
models for formal categories. If the Images are disturbed for the occurrence of noises, poor
images, occlusion, jitters or background, it becomes quite a tedious task to generalize an
Image. This difficult gets multiplied whenever an Image consists of many objects. There is an
invariable raise in newer algorithms and also at a technique from the latest time. Accurate and
also at a precise discovery of the traits in an Image is a major objective of doing Image
Classification. The support vector machine (SVM), now is that new inheld system of learning
which is then applied on both of regression and also at a pattern finding or recognition as of
now. In the case of machine learning, Support Vector Machines (SVM) is superseded model
to learn which is working on the associated method of learning algorithms that does analysis
of data and also at an identification of traits. They are used for Classification as well as
regression analysis. During remote-sensing Classification, we have to remember many factors
which make it a complicated task. The essential is the step that is involving Image
Classification and also at a determining a fitting system, trait extraction, selection of worthy
sample for training, Image pre-defined processing and also at a to select appropriate number
of methods to pre and also at a post Classification processing, and also at a then at last
calculate the overall efficiency + accuracy. Imaging sensing devices and also at a remote
satellite technology are upcoming. This gave a jump to newer systems which takes HSR
Images via satellites and also at an airborne sustainable platform.

1.2 Empirical Survey
Since the past few Decades, affordable tasks are done for growth of varied ways for the work
of scene detection via remote satellite or aerial pictures. As classification of scene is
Typically dead within the area of features, the foremost effective feature illustration performs
a vital help in creating high performance ways. We are able to currently divide existing ways
in three best sorts of step with the features they’re using:
• Handcrafted based ways, this way primarily specializes in employing an acceptable quantity
of engineering hand additionally to style varied human engineering traits like, shape, spatial,
colour, texture and at a spectral data or their combination are the foremost characteristic of a
scenic image and thus carries essential data used for scene Classification.
• Unsupervised feature learning-based ways to remedy the constraints of hand additionally at
a crafted feature, learning features mechanically from pictures are thought of as an additional
possible strategy. Unsupervised feature learning aims to be told a group of basic functions (or
filters) used for feature encryption, during which the input of the functions may be a set of
hand crafted and crafted features or juvenile intensity value and output may be a set of
learned traits.
• Deep Feature Learning based ways deep learning models are composed of multiple process
layers will learn additional powerful feature representations of knowledge with multiple
levels of abstraction. Deep feature learning ways have additionally result bent to be terribly
vital at discovering knotty structures and additionally at a discriminative data hidden in high
dimensional knowledge.

1.3 Proposed Approach
Remote sensing Image Classification scene plays a necessary role during huge selection of
applications and hence has been receiving outstanding attention. Throughout the past years,
varied knowledge sets are developed or inaccurate spread of approaches for scene
classification from remote sensing pictures. Scene classification is studied to permit U.S. to
do formal interpretation of high abstraction resolution (HSR) remote sensing imagining. The
Bag of visual words (BoVW) modelling proves to be a great technique for image-based scene
classification. Abstraction traits are accepted to be helpful in rising the illustration of the
Image and conjointly at an increasing classification of accuracies. Bag of visual words
(BoVW) finds the culmination of some nearby traits in the images and a combining look,
abstraction of images. However, the standard BOVW model solely captures the native
patterns of pictures by utilizing native traits despite the actual fact that they are tested in tiny
knowledge sets, their action is variable and less tight. Whereas managing giant datasets is
difficult because of the quality and the diversity of landscape and canopy patterns. Therefore,
it's a challenge to get higher accuracies illustrated at once to manage the augmented
geographical imaging. During this project, a local-global feature bag of visual words scene
classifier is projected for HSR imaging. The shape-based static texture index is intended
because the world texture feature, the mean standard deviation values of area unit are utilized
since, the native spectral feature, and thus the dense scale-invariant feature remodel (SIFT)
feature has been engaged for the structural feature. The model will effectively mix the native
associated degree at the worldwide traits by the acceptable features of fusion strategy at key
chart level. Further, a feature illustration technique for reflex learning uses essential traits
from image knowledge, Convolutional Neural Networks (CNNs) are introduced for HSR
remote sensing image scene classification because of their wonderful performance in natural

image classification. To beat the drawbacks within the ancient technique, we tend to combine
the CNN based (mostly abstraction traits) and at the BoVW based Image interpretation into
remote scene Classification to enforce Bag of Convolutional Features (BOCF) model. On
getting the required accuracy and then projecting Image Classification at a segmentation
model exploits a unique feature choice technique for victimization of gray wolf optimization
(GWO) algorithmic program from Extracted SIFT traits over gravy wolf optimized dataset.
1.4 Tabular Comparison
I. Handcrafted Feature BasedMethod
Table 1. Overall accuracies (%) of three kinds of handcrafted
features under the training ratios of 10% and 20%.[5]
II. Unsupervised Feature Learning Method
Table 2. Overall accuracies (%) of three kinds of unsupervised
features learning methods under the training ratios of 10% and 20%.[5]
III. DeepLearning BasedMethod

Table 3. Overall accuracies (%) of three kinds of deep learning-
based CNN features under the training ratios of 10% and 20%.[5]
Table 4. Overall accuracies (%) of three kinds of fine-tuned deep
CNN features under thetraining ratios of 10% and 20%.[5]
1.1 Significance of Problem
Amongst the best applications of remote satellite sensing is the making of Land also at a Use
/ Land also at a Cover maps from Image Classification is making great rise in the past some
decades are the below four areas:
(1) Mapping of a land at regional and also at a global scale
(2) Development and also the use of some newer and deep algorithms, like the sub division
pixel, pre-field, and also a knowledge inspired classification.
(3) Use of plural remote based sensing traits, including spectral, spatial area, multi temporal,
and also a bi-sensory data
(4) Incorporation of secondary data into classification procedures, including such data as the
topography, soils, roads, and also the census data. Accuracy evaluation is the integral sector
in the classification method.
The triumph of an Image Scene Classification is according to many factors, which are, the
availability of highly found quality of remotely sensed Imagery and also of ancillary type
database, the architecture of a fitting procedure and also the and skills of an analyst and the
past some experiences. The essential rise in this technology over the last few decades resulted
to exponential remote based sensing data for the smart global observations. However, the loss

of the openly found “Big data” of the remote sensing images excellently limits the rise of the
newly found approaches and that the deep learning-based methods especially. But, now,
almost all of the scene Classification methods are based upon only the traditional made
remote based sensing which is the upward Images, to differentiate various type of the areas
and also the covers of a space.
2. Literature Survey
Classification of image makes a major accord in environmental & socioeconomic
applications. Scientist’s has made more improved techniques to make the classification
better. But it’s still typical to make a thematic map from the classification of a distant sensed
data because of the following reasons:
(a)Image sensing & processing
(b)Landscape complexity
(c)Classification approaches
Also views on the current techniques & classification approaches is not available. This
report tells about the current advancements achieved in this field. This report provides the
analysis on the most current advancements in this field. Covering approximately 50
publications our report has analysed the data sets & three important types of ways for
remote sensing image scene classification, including methods based on handcrafted feature,
unsupervised learning of feature & deep feature learning based methods. Many attempts
have been made from past decades to produce the advanced ways for scene classification by
the assistance of aerial & satellite images. Fruitful presentation of feature becomes a
vital role in developing high performance methods for scene classification since it is done in
feature space. The positive outcomes of Bag of Words in classification of text has inspired
Bag of visual words (BoVW) model. In BoVW, every file is presented by a disordered group
of non-
distinctive words represented in the file, without taking grammar & word order into
consideration. File is presented with the assistance of number of times of occurrences
(histogram) of words in vocabulary. From the assistance of these histograms the classification
& retrieval of file is done. Similarly, the image is presented by the disordered group of
non- distinctive unit of visual features. Based on the BoVW many local feature scene
classification methods have also been produced. However, the Scale invariant feature

transform can only brief the local information i.e continuous & it do not has features
produced from global perspective. Also, the visual words that are obtained by grouping the
long feature vector using k-means are not accurate.
Considering the truth that there are many mixed pixels in HSR imagery which is not
appropriate to use in k-means clustering .Training & testing images are used in k-means.
Next step is to arrange test images in accordance with the data grasped by training images.
Some frequently used classification methods include Support Vector Machine
(SVM).Together use of Histogram Intersection Kernel (HIK) with the Support Vector
Machine(SVM) helps in extracting more features from different scenes. Remainder of SVM
will organize the experimental outcomes of the UC Merced & Google Datasets and will
describe details of the BoVW for imagery scene classification. Since the scene classification
job gets to be more difficult, its capability to distinguish gets limited. Currently, many of the
deep learning algorithms, more specifically convolutional neural networks (CNNs), have
proved their power for feature representation. Motivated by the achievements of CNNs, our
paper has further described a method for feature representation named as Bag o
Convolutional Features (BoCF).In BoCF CNN is used to extract features making the features
more strong for scene-level semantic understanding, thus making it different from the
conventional BoVW models where the handcrafted features were extracted. Optimal feature
selection by Grey Wolf Optimization proposed image classification and segmentation model.
Because many features are drawn out by SIFT, thus it becomes important to filter the features
optimally. This becomes possible by the assistance of GWO algorithm because it enables to
optimally select the features in a way that the noisy features can be removed which are
recognized by drawing out the features with minimum correlation. Also if the correlation
between the features is minimum it can help to distinguish many emotions with precision.
Below is the step which should be taken care for SIFT & optimised SIFT based on
GWO. Featured based on GWO principle are drawn out by the optimised SIFT i.e location of
wolves Iα, Iβ, Iδ is discovered out & eventually, best key-point is drawn is drawn that is
leading location of the wolf. To decrease the relation among the SIFT features is the main
objective of this method.

3. Analysis and Design
3.1 Overall description
This project implements, reviews and compares the recent progress of remote sensing image
scene classification using different methods traditional as well as recent methods, proposes a
large-scale benchmark data set, and evaluates a number of state-of-the-art methods using the
proposed data set with their accuracies and efficiency.
Figure 1. Overview of Image Scene Classification using CNN

Figure 2. Overview of GWO architecture
3.2 Requirement Analysis
1. Software Requirements:
[1] Operating System: Windows 10
[2] Anaconda (Mandatory)
[3] Internet browser (Google Chrome recommended).
[4] Tensor Flow GPU Version. (Recommended)
[5] Training dataset
2. Hardware Requirements:
[1] 64-bit distribution capable of running 32-bit applications
[2] 2 GB RAM least, 8 GB RAM recommended
[3] 2 GB of available disk space minimum, 4 GB Recommended (for various
image datasets).
[4] Intel processor with support for Intel VT-x.
[5] Intel EM64T (Intel 64), and Execute Disable (XD) Bit functionality

3. Functional Requirements:
[1] NWPU-RESISC45 Data Set
[2] UC Merced Land-Use Data Set
4. Non-Functional Requirements:
[1] They must be able to run the program and have a background knowledge of all
the libraries used.
[2] The user and developer must know Python language.
[3] Consistent and dependable quality of service.
[4] 80% of the users will be able to complete tasks without requiring assistance.
[5] All users will be satisfied with the usability of the application.

4. IMPLEMENTATION
4.1 Implementation Details
4.1.1 Data Set Analysis: Multiple data sets that are publicly available high-
resolution remote sensing image data sets have been introduced by different groups to
perform research for scene classification and to evaluate different methods in this
field. We will briefly review data sets used in the project implementation.
4.1.2 Bag of Visual Words: This method is applied to classify the image by assuming
image feature data as visual words. Every image is considered as the record and can
show distinguishably with the help of a histogram. The project encompasses
algorithm designing, hard programming, experimental and analysis of results of the
experiments. The principal result of our project is a process which trains and tests
images and returns classification results for evaluation of the algorithm. The
performance is assessed over efficiency and accuracy by experimental analysis with
image data sets. There is related work on the same project and domain bag of visual
words model from last 15 years. Even if most researches have been done across this
process domain, there are four major steps of this method among all of them:
(1) Sampling of image data set: Regions can be sampled consciously or based on
interest points to for overall performance and time efficiency.
(2) Description of sampled parts: Most broadly used descriptor is SIFT in related
work of image processing. Besides them there are more frequently used tools are
including Surf, PCA-sift
(3) Quantisation of vectors: It is the process of generating “visual words” to build
“visual vocabulary” by clustering and transforming data of each image to a sparse
vector by counting occurrence of each visual word in this image to get a histogram.
(4) Application: Operations are being performed on the basis of generation of data
from (step 3)

Flow chart involved in BoVW:
Figure 3: Flow chart of BOVW model in classification of image.
The final result of our project is a programme for classification of a set of test images which
are labelled into several categories that are set for images from the indicated training set and
returning the results for analytical overview.
Key Feature involved in Bag of visual words:
 SIFT-Scale-Invariant Feature Transform has been in good use since it is being put
forward, with several kinds of disparities developed. It is mostly used to match new
images to a Big data base of labelled categories. The number of features that have to

be matched are reasonably very large. Thus, to some range, SIFT depends on data
structures or algorithms to improve th efficiency.
 Codebook Generation: The feature descriptors extracted from all training images are
input to this and outputs is given as visual codebook
 K-means: This is an unsupervised learning algorithm in this we want to minimise the
Euclidean distance between the point 'x' and the nearest cluster centre 'm'. So, the
output of this is cluster centres which determine the visual vocabulary of image.
D( X, M ) = ∑ ∑ ( xi - mk )2
Cluster k point i
in cluster k

Limitations:
This method BoVW has got a number of limitations. Elements are all confounded the
training data, in which spatial information is not found, neither is covariance between
these elements. When differentiating between categories of objects, background
information is not eliminated. And confusion can be caused when scale of objects of the
same kind varies too much in a training set. The algorithm still has much more to
improve.
 Using certain sampling procedures in choosing some of the patches in processing
instead of computation of feature for number of all patches. Sampling from points of
interest will enormously improve time efficiency and throw away most of the noise.
 More efficient classifiers, such as SVM, Naive Bayes are available.
 CNN-convolutional neural networks.
 The accuracy of traditional BoVW method with dense SIFT is 41.72% under training
ratio of 10% and 44.97% under the training ratio of 20% [1].
Figure 4: Overall accuracies of BOVW with SIFT feature extraction, BoCF with performing the training ratios of (a) 10%
and (b) 20%. [1]

4.1.3 Convolutional Neural Networks CNN is one of the most widely used architectures
of deep learning and in computer vision. A classical CNN model is mostly structured
as a series of layers consisting of three layers that are convolutional , pooling, and
fully connected.
1) Convolutional Layers: They are the most important layers for feature extraction.
The apex layers generally extract low-level cues and the deeper layers capture high-
level features. Each unit in a convolutional layer is connected to a local receptive field
in the feature maps of the previous layer through a set of kernels.
There are two main advantages of the convolution operation: the weight sharing
mechanism in the same feature map reduces the number of parameters and the local
connectivity learns correlations among neighbouring pixels.
Figure 5. Image Matrix Figure 6. Filter Matrix
Figure 7. Convolved Feature Matrix

2) Pooling Layers: The layers of pooling are generally placed between two layers of
convolutional used for minimizing the dimensions of featured maps and networking
variables. Also, they create invariance to small shifts and distortions by taking
neighbouring pixels into account. Average pooling and max pooling are two most
widely used strategies.
Figure 8. Max Pooling Operation on Matrix [9]
3) Fully Connected Layers: They are typically used as the last few layers to better
extract the information conveyed by lower layers in view of the final decision.
Figure 9: Convolutional neural network for scene classification.[9]

4.1.4 Bag of Convolutional Features: To overcome the drawbacks, we combined the CNN
based spatial features and the BoVW based image interpretation into image scene
classification. The convolutional neural network (CNN) is a nature inspired process
that can learn multilevel ranking of features, which perfectly meet the demand of
feature learning tasks. Different from the classical BoVW process model in which the
visual words are usually formed from human handcrafted features, Instead our
proposed method BoCF works with generation of visual words directly from
convolutional zed features, and so the new feature depiction is more robust for scene
level semantic understanding.
The accuracy of Bag of CF method is almost increased by 82.55% under training set
ratio of 10% and 84.12% under the training set ratio of 20% [1].

Flow Chart involved in BoCF:
Figure 10: Overview of BoCF model in scene classification.
4.1.5 Grey Wolf Optimization:

Grey wolf optimizer (GWO) is one among the recent meta heuristics and swarm
intelligence algorithms. It has been mostly formed for an extensive diversity of
optimization problems due to its splendid attributes over other nature inspired
intelligence methods. Additionally, it is essential, versatile, adaptable, simple to
utilize and has a unique potential to strike the correct harmony between the
investigation and misuse amid the hunt which prompts ideal intermingling.
Consequently, the GWO has as of late picked up an exceptionally huge research
enthusiasm with huge gatherings of people from a few areas in a brief span.
Figure 11. Grey Wolf Optimisation (GWO) application.[16]
The Hierarchical Ranks
1 Alpha: Alpha wolves are the leaders of the group that can be male wolves and/or
female wolves. The dominant alphas are mostly responsible for making decisions
about hunting, sleeping place, time to wake, and so on. The alpha’s decisions are
dictated to the pack.
2 Beta: The next level in fitness of grey wolves is beta. The betas are wolves that help
the alpha in decision-making or other group activities.
3 Omega The most reduced positioning dim wolf is omega. The omega assumes the
job of a substitute. Omega wolves dependably need to submit to the various
overwhelming wolves.

4 Delta: If a wolf isn't ordered as an omega, alpha, or beta, it is called delta or
subordinate. Delta wolves need to submit to alphas and beta however they rule the
omega. Scouts, sentinels, older folks, seekers, and overseers have a place with this
class
In continuation to the social hierarchy of wolf, hunting in group is another
interesting social trait of the wolves. The main points of grey wolf hunting are
discussed below [17]:
 Tracking the prey, chasing the prey, and approaching the prey
 Pursuing, encircling the prey, and harassment off prey until it stops moving.
 Attack the prey
Figure 12. GWO based feature selection flow chart to obtain optimal features.

The steps that are involved in the metaheuristic grey wolf optimization algorithm have
been discussed below:
1) Initializing of the Search Agents.
2) Assigning of Alpha, Beta, and Gamma to among wolf by respective fitness.
3) Encircling of prey (equation 3.0 and 3.1)
where t represents present iteration, A ⃗ and C ⃗ are coefficient vectors, X ⃗ indicates
the wolf’s position vector and Xp ⃗ is the prey’s position vector.
The A ⃗ and C ⃗ vectors are given by the following (equation 3.2 and 3.3):
where components of a ⃗ are linearly decreased in the range [2, 0] for successive
iterations. The r1 ⃗ and r2 ⃗ are random vectors in the range 0 to 1.
4) Start Hunting-After, the encircling process, comes to the second step involving
hunting being guided by the alpha wolf. The following formulas describe hunting
process (equation 3.4, 3.5, 3.6 respectively)

Followed by Equation 3.7
(5) Attacking prey-The eventual step of attacking the prey is being done by straightway
decreasing the value of a ⃗ from 2 - 0. With this, the fluctuation in the value of A ⃗ is also
minimized.
(6) Steps 2 - 5 are replicated for a limited number of iterations.

Figure 13: Graphical definition of alpha, beta, delta and omega in GWO
Figure 14. Pseudo code of theGWO algorithm

 The GWO has only two main parameters to be adjusted (a and C).
Applications of GWO:
1. Feature selection is one of the significant procedures in AI and information
mining. The objective of highlight choice is to decrease the number of highlights,
select the most agent ones and to wipe out repetitive, boisterous and unimportant
highlights. The issue of hunting down the best arrangement of highlights is
considered as a mind-boggling and troublesome issue because of the incredibly
huge inquiry space when the quantity of highlights is huge.
2. Clustering is a typical AI and information mining task where the objective is to
separate information examples into various gatherings that have comparative
attributes in some sense. Meta heuristic calculations have been generally utilized
and connected for grouping errands.
3. GWO has been conveyed in various methodologies for different restorative and
bioinformatics applications. A twofold GWO for highlight choice with
extraordinary learning machine (ELM) classifier for two restorative analysis
issues

4. FINDINGS & CONCLUSION
5.1 Conclusion
This report has presented BoCF for scene classification which is an effective way for image
feature representation. This method generated visual words with deep convolutional features
whereas in conventional Bag of Visual Words (BoVW) model visual words are drawn by the
use of handcrafted features such as SIFT. In the algorithm, random scale stretching is applied
to make the CNN model to grasp feature representation i.e. rugged to object feature scale
variation. Universal evaluations on an open site available NWPU-RESISC45 and UC Merced
Land scene classification of remote sensing image shows the potency of proposed BoCF
method. In later part of the project Grey Wolf Optimizer (GWO) algorithm is implemented
for satellite image segmentation. The conventional GWO is accurately worked on to make it
work as an instinctive clustering algorithm. Further, the proposed method and the existing
methods were compared to analyse their performances. In future work need to be done to
produce new methods and systems where the fusion of remote sensing data &
information
extracted from social media & other technologies can be used to promote the scene
classification of remote sensing image.

5.2 Future Scope
Future work related to our project mainly involves incorporating more low-level
features such as
colour, spectral information etc. to improve the accuracy of classification & the
efficiency of the
algorithm including the sparse matrix multiplications method to improve the actual
speedup
factor,& the testing of our proposed method for more visual recognition tasks
such as geographic
image retrieval & geospatial object detection.
Future work will mainly be done keeping the following points in mind:-
(a)The training time involved with the proposed method has to be reduced for this it is
mandatory to
further work on BoVW model.
(b)By going through 3D reconstruction of binocular stereo vision, it can be concluded
that the blend
of classification & reconstruction can be applied to the UAV based target
recognition & location of
grasping.
(c)Similar method can be explored in hand written, camera based documents.
(d)More features can be involved with the rest of kernel features to enhance the Bag
of words
model & further some extra encoding techniques can be initiated such as local
soft assignment.
In spite of the acceptance of GWO & other advances, there are many more fields
that need new
work as follows:-
(a)The main drawback of GWO algorithm is that while working for large scale
problems, its ability to

handle a large number of variables & the escape of local solutions needs to be
improved.
(b)One of the most efficient ways to solve benchmark problems is the division of the
population into
four groups that is taken care by GWO. Taking into account more or fewer groups
with a variety of
wolves in each can be considered as an interesting field of research to enhance
GWO's efficiency
when dealing with challenging true-world issues.
(c)Dynamic problems can’t be solved by using GWO algorithm as there is no related
work in the
literature.

Remote Sensing Image Scene Classification

More Related Content

Similar to Remote Sensing Image Scene Classification (20)

More from Gaurav Singh (6)

Recently uploaded (20)

Remote Sensing Image Scene Classification