Competent scene classification using feature fusion of pre-trained convolutional neural networks

TELKOMNIKA Telecommunication Computing Electronics and Control
Vol. 21, No. 4, August 2023, pp. 805~814
ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v21i4.24463  805
Journal homepage: http://telkomnika.uad.ac.id
Competent scene classification using feature fusion of pre-trained
convolutional neural networks
Thirumaladevi Satharajupalli1
, Kilari Veera Swamy2
, Maruvada Sailaja1
1
ECE Department, Jawaharlal Nehru Technological University, Kakinada-533003, Andhra Pradesh, India
2
ECE Department, Vasavi College of Engineering, Ibrahimbagh, Hyderabad-500 031, Telangana, India
Article Info ABSTRACT
Article history:
Received Aug 29, 2022
Revised Dec 10, 2022
Accepted Feb 16, 2023
In view of the fact that the development of convolutional neural networks
(CNN) and other deep learning techniques, scientists have become more
interested in the scene categorization of remotely acquired images as well as
other algorithms and datasets. The spatial geometric detail information may
be lost as the convolution layer thickness increases, which would have a
significant impact on the classification accuracy. Fusion-based techniques,
which are regarded to be a viable way to express scene features, have
recently attracted a lot of interest as a solution to this issue. Here,
we suggested a convolutional feature fusion network that makes use of
canonical correlation, which is the linear correlation between two feature
maps. Then, to improve scene classification accuracy, the deep features
extracted from various pre-trained convolutional neural networks are
efficiently fused. We thoroughly evaluated three different fused CNN
designs to achieve the best results. Finally, we used the support vector
machine for categorization (SVM). In the analysis, two real-world datasets
UC Merced and SIRI-WHU were employed, and the competitiveness of the
investigated technique was evaluated. The improved categorization accuracy
demonstrates that the fusion technique under consideration has produced
affirmative results when compared to individual networks.
Keywords:
Feature extraction
Feature fusion
Pre-trained networks
Scene classification
Support vector machine
This is an open access article under the CC BY-SA license.
Corresponding Author:
Thirumaladevi Satharajupalli
ECE Department, Jawaharlal Nehru Technological University
Kakinada-533003, Andhra Pradesh, India
Email: thirumaladeviece@gmail.com
1. INTRODUCTION
Classifying remote sensing images into various classes depending on image content has gotten a lot of
attention nowadays because of its wide range of applications. Classification of remote sensing image scenes is
primarily a machine learning and computer vision problem. The investigation of scene classification entails
convolutional neural networks (CNNs) were highly successful. However, to train their parameter sets, most of
these models require vast amounts of labeled data and many iterations. Several CNN-based scene categorization
approaches [1]-[3] have arisen as a result of varied tactics for utilizing CNNs. Deep learning models, such as
AlexNet [4], and VGG-Net [5], have achieved considerable success in computer training data [6], ImageNet [7].
In 2012, Krizhevsky took first place in the ImageNet large scale visual recognition challenge. For a specific
purpose, acquiring a large dataset can be expensive. When it comes to remote sensing image scene
categorization, deep neural network model-based methods are becoming increasingly popular [8], [9]. Deep
learning-based scene image categorization has a distinct advantage over typical machine learning methods in
that can be extracted from more complicated and relevant feature structures [10], [11]. Deep layers can then
be used to get crucial and discriminatory feature representation, while irrelevant versions are ignored.

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 4, August 2023: 805-814
806
There are three types of CNN-based techniques utilizing pre-trained utilizing CNNs to act as feature
extractors and fine-tuning already trained models on target datasets and creating new CNN models for scene
categorization from scratch [12], [13]. The way that CNN is affianced as a feature extractor is the simplest of
the three approaches. CNNs were first introduced as feature extractors. In 2015, Penatti et al. [14] applied
CNNs to image scene categorization using remote sensing and investigated the adaptive ability of off-the-shelf
CNNs for remote sensing image categorization. CNNs outperform low-level descriptors in their studies,
according to the researchers. Later, Cheng et al. [15] researched how to fully utilize pre-trained CNNs for scene
categorization by treating them as feature extractors. Marmaris et al. [16] presented a two-stage CNN scene
categorization framework. It utilized CNNs that had been pre-trained to extract a collection of interpretations
from images. Covariance-based multilayer feature fusion is also proposed [17]. Classifiers were then given the
extracted representations.
Various structures of CNNs have varying receptive fields and may capture different types of data
from images. In this paper, AlexNet, VGG-19, and VGG-16 are employed as deep feature extractors
independently, and to improve accuracy, features extracted by two of three are fused using canonical
correlation analysis (CCA) and form three fused networks, which are assessed and compared. The datasets
from UC Merced and SIRI-WHU, both of which are publicly available and created for research purposes,
were used to assess the performance of the fused networks.
The remainder of this work is structured in a succeeding manner. Section 2 shows and explains the
scene classification workflow established on the individual pre-trained CNN model including a flow chart of
three fusion networks. In section 3, we will find typical datasets that exhibit experimental evaluation and
analysis. This paper comes to an end with the section 4 conclusion.
2. WORKFLOW OF THE PROPOSED METHOD
Convolutional, pooling, and fully connected layer are the core components of the pre-trained CNN
model, which only use pre-trained weights [18]. Figure 1 shows a distinctive remote sensing image scene
categorization procedure using a CNN model that has been pre-trained. Samples of the input training images
are loaded into the input layer, which pre-processes and modifies the image dimensions as per the network
model’s input layer. The finishing output is achieved with a fully connected layer after some a succession of
pooling and convolutional computations which are then used as final features and applied to support vector
machine (SVM) for classification.
Figure1. A feature extractor’s flowchart that makes use of pre-trained CNN
For image scene classification, we employed pre-trained deep CNN models, where used AlexNet,
VGG-19, and VGG-16 as feature extractors and selected helpful layers to acquire a good depiction of the image
scene. We are the first to integrate several fully connected layers of the AlexNet, VGG-Net architecture, where
each layer’s output is expected to be a feature descriptor and is fused to produce a final feature illustration of the
input image. Other individual feature representation methods perform less than fused deep feature learning.
The visual scene is well described by fused features, which offer a lot of information.
In very high resolution (VHR) image scene categorization, the purpose of feature fusion is to combine
two correlated scene features into a single feature vector with much more discriminant data than the input
feature vectors. Here novelty of this paper is combining characteristics collected from dense layers of two
different pre-trained networks to create various fusion networks. Finally, a multi-kernel learning approach was
used to train the SVM classifier, as illustrated in Figure 2.

TELKOMNIKA Telecommun Comput El Control 
Competent scene classification using feature fusion of … (Thirumaladevi Satharajupalli)
807
Figure 2. Proposed fused pre-trained CNNs flowchart for scene categorization
Combining two or more characteristics properly is becoming increasingly difficult. We are using the
canonical correlation approach to combine two distinct feature vectors to create a new feature vector that is
significantly more discriminative than the original two. This approach transforms a correlation analysis of
two random vectors into a few uncorrelated pairs of variables. 𝑀 and 𝑁 are two zero-mean arbitrary vector
matrices. Determine a couple of orientations that will improve the projection’s correlation. Choose a
collection of canonical variables. Two features matrices 𝑀 and 𝑁, in addition, 𝑀 ∈ 𝑅𝑎𝑥𝑛
and 𝑁 ∈ 𝑅𝑏𝑥𝑛
where
𝑛 is the training feature vectors containing the matrices. 𝑀’s covariance matrix is epitomized as 𝐶𝑝𝑝 ∈ 𝑅𝑎𝑥𝑎
as well as 𝐶𝑞𝑞 ∈ 𝑅𝑏𝑥𝑏
is the covariance matrix of 𝑁 between-sets the covariance matrix is 𝐶𝑝𝑞 ∈ 𝑅𝑎𝑥𝑏
and
𝐶𝑞𝑝 = 𝐶𝑝𝑞
𝑇
. Overall covariance matrix 𝐶 ∈ 𝑅(𝑎+𝑏)×(𝑎+𝑏)
is deliberate as:
𝐶 = (
𝑣𝑎𝑟(𝑝) 𝑐𝑜𝑣(𝑝, 𝑞)
𝑐𝑜𝑣(𝑞, 𝑝) 𝑣𝑎𝑟(𝑞)
) = (
𝐶𝑝𝑝 𝐶𝑝𝑞
𝐶𝑞𝑝 𝐶𝑞𝑞
) (1)
It’s difficult to deduce the correlations between all these matrices two pairs of feature vectors since the
correlations between two groups of feature vectors might not conform to a predictable pattern. CCA’s goal is
to describe a linear combination [19] that maximizes correlation.
𝑀∗
= 𝑊
𝑝
𝑇
𝑀 and 𝑁∗
= 𝑊
𝑞
𝑇
𝑁 (2)
𝐶𝑜𝑟𝑟 (𝑀 ∗, 𝑁 ∗) =
𝑐𝑜𝑣 (𝑀∗,𝑁∗)
𝑣𝑎𝑟(𝑀∗),𝑣𝑎𝑟(𝑁∗)
(3)
Wherever:
− (𝑀∗) = 𝑊
𝑝
𝑇
𝐶𝑝𝑝𝑊
𝑝
− 𝑣𝑎𝑟(𝑁∗
) = 𝑊
𝑞
𝑇
𝐶𝑞𝑞𝑊
𝑞
− 𝑐𝑜𝑣 (𝑀 ∗, 𝑁 ∗) = 𝑝𝑇
𝐶𝑝𝑞𝑊
𝑞
Exploiting the covariance between 𝑀 ∗ and 𝑁 ∗ considering the constraints 𝑣𝑎𝑟(𝑀 ∗) = 𝑣𝑎𝑟(𝑁 ∗) = 1.
The transformation matrices 𝑊
𝑝, 𝑊
𝑞 is initiated by especially the eigenvalue concerns.
𝐶𝑝𝑝 − 1𝐶𝑝𝑞 𝐶𝑞𝑞 − 1𝐶𝑞𝑝𝑊𝑝^ = ^2𝑊𝑝^
𝐶𝑞𝑞 − 1𝐶𝑞𝑝 𝐶𝑝𝑝 − 1𝐶𝑝𝑞𝑊𝑞^ = ^2𝑊𝑞^ (4)
Where 𝑊
𝑝
^
and 𝑊
𝑞
^
are eigenvectors. ^2
is the diagonal matrix with the highest eigenvalues or correlation
squares. Each equation contains the following number of non-zero eigenvalues:
− 𝑑 = 𝑟𝑎𝑛𝑘(𝐶𝑝𝑞) ≤ 𝑚𝑖𝑛(𝑛, 𝑎, 𝑏) which is to be listed in descending manner 𝛼1 ≥ 𝛼2 ≥ 𝛼3 … ≥ 𝛼𝑑.
− 𝛼1
𝑇
𝑀 and 𝛽1
𝑇
𝑁 (the first pair)
− 𝛼2
𝑇
𝑀 and 𝛽2
𝑇
𝑁 (the second pair)
− 𝛼𝑑
𝑇
𝑀 and 𝛽𝑑
𝑇
𝑁 (the 𝑑𝑡ℎ pair)

 ISSN: 1693-6930
808
𝑀∗
= (𝛼1
𝑇
𝑀, 𝛼2
𝑇
𝑀, … … … , 𝛼𝑑
𝑇
𝑀) = (𝛼1, 𝛼2, … … … , 𝛼𝑑)𝑇
𝑀 = 𝑊
𝑝
𝑇
𝑁∗
= (𝛽1
𝑇
𝑁, 𝛽2
𝑇
𝑁, … … … , 𝛽𝑑
𝑇
𝑁) = (𝛽1, 𝛽2, … … … , 𝛽𝑑)𝑇
𝑁 = 𝑊
𝑞
𝑇
(5)
The ordered eigenvectors analogous to non-zero eigenvalues make up the transformation matrices 𝑊𝑝 and 𝑊𝑞.
A summing of the changed feature vectors is used for feature-level fusion. Canonical discriminant correlation
The following features are included:
𝑍 = 𝑀∗
+ 𝑁∗
= (𝑊
𝑝
𝑇
𝑀 + 𝑊
𝑞
𝑇
𝑁) = (
𝑊
𝑝
𝑊
𝑞
)
𝑇
(
𝑀
𝑁
) (6)
SVM: this categorization is based on the data of choice, which manages space with high-dimensional
features with choice constraints. Using a set of labeled training datasets, SVM can produce linear capacity in
either input space or maximum space [20]. This allows us to distinguish between positive and negative
samples. A data matrix that has been labeled to either a positive or negative class is utilized as input data
used in the SVM’s training phase. One can utilize trained SVM to forecast what the class has predicted in
test samples.
The benefits of CCA are straightforward to apply to two variables. The intermodality relationship is
thought to be linear in CCA, and both modalities are interchangeable and given the same considerations.
Linear feature transforms do not affect canonical correlations.
3. RESULTS AND DISCUSSION
3.1. UC Merced land-use dataset
UC Merced land-use dataset [21]: a lot of work has gone into making datasets available to the
general public, including the first publicly accessible high-resolution remote sensing imagery collected for
scene classification. This University of California Merced (UCM) land-use dataset [22] comprises 2,100
aerial shots of scenes distributed into 21 land-use scenario groups. Every single class contains 100 images
that are 256×256 pixels in dimensions and have a resolution of every pixel, 0.3 meters. This dataset was
created exhausting the United States Geological Survey’s National Map, which was retrieved using aerial
ortho imagery (USGS). This dataset features overlying land-use classifications, such as Figure 3 exhibits
representations of sample images from every class Figure 3(a) agricultural, Figure 3(b) airplane, Figure 3(c)
baseball diamond, Figure 3(d) beach, Figure 3(e) buildings, Figure 3(f) chaparral, Figure 3(g) denseresidential,
Figure 3(h) forest, Figure 3(i) freeway, Figure 3(j) golfcourse, Figure (3k) harbou, Figure 3(l) intersection,
Figure 3(m) mediumresidential, Figure 3(n) mobilehomepark, Figure 3(o) overpass, Figure 3(p) parkinglot,
Figure 3(q) river, Figure 3(r) runway, Figure 3(s) sparse residential, Figure 3(t) storagetanks and Figure 3(u)
tenniscourt and it’s been widely utilized for visual scene categorization and retrieval using remote sensing data.
3.2. SIRI-WHU
SIRI-WHU [23]: it’s a collection of 2,400 remote sensing images that have been categorized into
12 scene types. Every single class comprises 200 images that are 200×200 pixels in size and have a
resolution of 2 meters. It was sent to Wuhan University’s intelligent data extraction and remote sensing
(RS IDEA) group via Google Earth (Google Inc.). In Figure 4, sample images from the SIRI-WHU dataset are
displayed Figure 4(a) agriculture, Figure 4(b) commercial, Figure 4(c) harbor, Figure 4(d) idle land, Figure 4(e)
industrial, Figure 4(f) meadow, Figure 4(g) overpass, Figure 4(h) park, Figure 4(i) pond, Figure 4(j) residential,
Figure 4(k) river, and Figure 4(l) water are among the 12 land-use categories. Even though this dataset has been
studied using a variety of approaches [24], the number of scene classes is very small. It also concentrates mostly
on China’s urban areas.
3.3. AlexNet
AlexNet: the ImageNet large-scale image recognition competition (ILSVRC) was won by AlexNet [4]
in 2012. The first to notice it was Alex Krizhevsky and his coworkers. This model has three fully connected
layers in addition to a pooling layer and five convolutional layers. The initial and second convolutional layers
are constrained by two normalization layers. This model accepts images with a 227×227 input size.
The second fully-connected layer is where the 4096-pixel-long output feature vector is generated. SVM was
used in our study to categorize the AlexNet CNN feature.

809
(a) (b) (c) (d) (e) (f) (g)
(h) (i) (j) (k) (l) (m) (n)
(o) (p) (q) (r) (s) (t) (u)
Figure 3. Example images of 21 classes’ depiction of UC Merced dataset: (a) agricultural, (b) airplane, (c) baseball
diamond, (d) beach, (e) buildings, (f) chaparral,(g) denseresidential, (h) forest, (i) freeway, (j) golfcourse,
(k) harbou, (l) intersection, (m) mediumresidential, (n) mobilehomepark, (o) overpass, (p) parkinglot, (q) river,
(r) runway, (s) sparse residential, (t) storagetanks, and (u) tenniscourt
3.4. VGG-Net
VGG-Net: a proposal was made in [5] and emerged victorious in the ILSVRC-2014 competition’s
localization and classification tasks. Two well-known architectures are VGG-19 and VGG-16. In this
evaluation, the designs and overall performance were improved to a somewhat greater extent. This model
contains 3 fully-connected layers, 5 pooling layers, and 13 convolutional layers. Images with a size of
224×224 are used as input to this model. The output feature vector is 4096 pixels long and derives from the
second fully connected layer. In our evaluation, we employed SVM to classify the VGG-Net CNN feature.
3.5. Fused network
Fused network: two pre-trained networks are used to create this network. The input layers of the
pre-trained networks continue to stay the same for AlexNet 227×227, and for VGG-19, VGG-16 224×224,
up to feature extraction, the procedure is the same as for individual pre-trained networks, and the output
feature is obtained from the second dense layer with dimension 4096, the fusion of these features can be done
using the conical correlation concept already explained in section 2. It is possible to combine these two
feature vectors from two distinct networks to obtain a final feature vector with a dimension of 4096. In our
evaluation, we employed SVM to classify the fused net feature.
We employ the UC Merced dataset and the SIRI-WHU dataset to examine the scene categorization
performance of individual pre-trained, fused networks. We use the experiment setup from [25] for evaluation,
which chooses a total of 80 images from each category that will be used for training and the rest for testing.
Regarding classification accuracy, the technique when fusion is absent is compared to several fusion
methods, with fusion networks showing a significant improvement.

 ISSN: 1693-6930
810
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
Figure 4. Example image depiction of the SIRI-WHU dataset: (a) agriculture, (b) commercial, (c) harbor, (d) idle
land, (e) industrial, (f) meadow, (g) overpass, (h) park, (i) pond, (j) residential, (k) river, and (l) water
3.6. Metrics used for evaluation
Metrics used for evaluation: when it comes to image classification, the average accuracy, overall
accuracy, and confusion matrix are the three metrics that are most frequently employed for evaluation
purposes. The number of samples that have been correctly recognized is calculated by dividing the total
number of samples by the category to which they are assigned. This division is performed irrespective of the
quality of the overall assessment. The average accuracy is determined by adding up the classification
accuracy of each class, regardless of the number of samples that are included in each category. The overall
accuracy value and the average accuracy value are identical this is because the count for each class in the
dataset is thesame. Because of this, we only used overall accuracy and confusion matrix criteria to judge the
performance of the various classification algorithms used in this work. The confusion matrix and overall
accuracy measurements were also looked into to ensure consistency. A total of five iterations of the study
confirmed that an 80 to 20 split between training and testing yielded the best results. The formula for
calculating the accuracy, which is expressed as a percentage of correct occurrences, is:
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃
𝑇𝑃+𝐹𝑁+𝑇𝑁+𝐹𝑃
(7)
The proportion of truly positive occurrences that are expected to occur in all positive situations is known as
precision. The formula is as:
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃
𝑇𝑃+𝐹𝑃
(8)
The recall calculation equation determines the predicted percent of true positive samples.
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃
𝑇𝑃+𝐹𝑁
(9)
An exhaustive mathematical calculation, the F1-score considers both accuracy and recall.
𝐹1 =
2 𝑋 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑟𝑒𝑐𝑎𝑙𝑙
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑟𝑒𝑐𝑎𝑙𝑙
(10)

811
Figure 5 illustrates together the confusion matrix in addition to the accuracy. In the first row, networks
Figure 5(a) AlexNet, Figure 5(b) VGG-19, and Figure 5(c) VGG-16 are the proposed individual pre-trained
CNN-based network classification is shown, while for scene categorization, the final feature extractor is a single
fully-connected layer FC7. In the second row, a fusion-based proposed network confusion matrix with accuracy
is shown Figure 5(d) AlexNet−VGG-19, Figure 5(e) AlexNet−VGG-16, and Figure 5(f) VGG-19−VGG-16.
Pre-trained AlexNets had an overall acuuracy (OA) of 79.76 percent, VGG-19 had an OA of 81.19 percent, and
VGG-16 had an OA of 83.81 percent, according to experiments on the University of California Merced dataset,
but fusion-based pre-trained networks had an OA of 89.28 percent, 90.23 percent, and 91.42 percent,
respectively. The proposed approach provides optimal classification efficiency for the majority of classes in the
case of UCM with an average gain of 8% in classification accuracy.
In the testing of the SIRI-WHU dataset, the confusion matrix is shown in Figure 6. To classify scenes,
the fully connected layer FC7 is employed as the final feature extractor, as illustrated in the first row, Figure 6(a)
AlexNet, Figure 6(b) VGG-19, Figure 6(c) VGG-16 and potential fusion-based classification models are
displayed in the second row as Figure 6(d) AlexNet-VGG-19, Figure 6(e) AlexNet−VGG-16, and Figure 6(f)
VGG-19−VGG-16. Whereas the accuracy of AlexNet’s preprocessed single layer is 86.52 percent, VGG-19 is
87.60 percent, and VGG-16 is 88.04 percent, the proposed technique enhances accuracy by 90.62 percent,
91.87 percent, and 92.91 percent, respectively.
Table 1 illustrates the corresponding performance assessment for the two datasets. The proposed
scenario indicates an increase in OA. When multiple individual pre-trained and fusion-based learning
networks are utilized. As demonstrated in Figure 7 correspondingly Figure 7(a) UCM dataset and the Figure 7(b)
SIRI-WHU dataset the proposed technique achieves optimal Improves categorization performance in the
majority of classes and overall accuracy by 4%.
Table 1. Evaluation results comparison for the individual proposed fused network accuracies for the two
datasets UCM, SIRI-WHU
Method Network used
UCM dataset with 80% training
(overall accuracy %)
SIRI-WHU dataset with 80% training
(overall accuracy %)
Pre-trained individual network FC7
as a feature extractor
AlexNet 79.76 86.45
VGG-VD19 81.19 87.7
VGG-VD16 83.81 88.12
Proposed fusion network using CCA AlexNet −
VGG-19
89.28 90.62
AlexNet −
VGG-16
90.23 91.87
VGG-19 −
VGG-16
91.66 92.91
(a) (b) (c)
(d) (e) (f)
Figure 5. First row consistent to single-layered confusion matrix using three pre-trained networks: (a) AlexNet,
(b) VGG-19, (c) VGG-16 second row resultant to proposed fusion, (d) AlexNet-VGG-19, (e) AlexNet-VGG16,
and (f) VGG-19−VGG-16 of UC Merced dataset

 ISSN: 1693-6930
812
(a) (b) (c)
(d) (e) (f)
Figure 6. First row consistent to single-layered confusion matrix using three pre-trained networks: (a) AlexNet,
(b) VGG-19, (c) VGG-16 second row resultant to proposed fusion, (d) AlexNet−VGG-19, (e) AlexNet−VGG-16,
and (f) VGG-19−VGG-16 of SIRI-WHU dataset
Figure 7. Proposed fusion-based networks versus pre-trained networks comparison of using (a) UC Merced
and (b) SIRI-WHU datasets
Figure 8. Comparison of F1 results when compared to
pre-trained networks the UC Merced dataset improved
classes using the suggested fusion networks
Figure 9. Comparison of F1 scores SIRI-WHU dataset
improved classes using suggested fusion networks
compared to ones that have already been pre-trained

813
From the UCM dataset, Figure 8 depicts the F1 scores. Individual pre-trained networks and
proposed fusion-based approaches are presented with F1 scores of improved classes. The proposed fusion
approach benefits a wide range of classes, including denseresidential, which improves accuracy from 20% to
68%, golfcourse, river, runway, sparseresidential, which improves accuracy from 90% to 100%, and
baseballdiamond, intersection, storagetanks, and tenniscourt.
Figure 9 displays the SIRI-WHU dataset’s recommended fusion networks and single-layered pre-trained
networks F1 scores. As can be observed, the recommended strategy improves scores in the majority of classes.
In the SIRI-WHU dataset, the river and pond improves from 70% to 95%. The classes industrial, overpass are
achived 80 to 90 % and majority of categories scored above 95%.
4. CONCLUSION
Significant advances in remote sensing technology have presented us with a torrent of remote
sensing data for scene categorization using images from remote sensing throughout the previous decade.
Because there is a scarcity of freely available remote-sensing image data, especially for deep learning-based
expertise, which severely restricts the development of new approaches. The purpose of this research was to
investigate in what manner machine learning also fusion network designs performed while categorizing data.
Using the UCM and SIRI-WHU datasets, classification was done on three feasible architectures: AlexNet,
VGG-19, and VGG-16. With accuracy from a pre-trained network of 89 percent, 90 percent, and 91 percent
for the UCM dataset and 91 percent, 92 percent, and 93 percent for the SIRI-WHU dataset, the suggested
methodology improved the state-of-the-art and created a standard. The classification results of the machine
learning concept were compared using SVM. The use of fusion networks has been proven to be an effective
strategy for optimal outcomes. This methodology can only determine the linear correlation between two
features. Future extensions could include handling non-linear feature spaces with more than two dimensions.
REFERENCES
[1] G. Cheng, C. Yang, X. Yao, L. Guo, and J. Han, “When deep learning meets metric learning: Remote sensing image scene
classification via learning discriminative CNNs,” IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 5,
pp. 2811–2821, 2018, doi: 10.1109/TGRS.2017.2783902.
[2] R. Minetto, M. P. Segundo, and S. Sarkar, “Hydra: An ensemble of convolutional neural networks for geospatial land classification,”
IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 9, pp. 6530–6541, 2019, doi: 10.1109/TGRS.2019.2906883.
[3] L. Ma, Y. Liu, X. Zhang, Y. Ye, G. Yin, and B. A. Johnson, “Deep learning in remote sensing applications: A meta-analysis and
review,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 152, pp. 166–177, 2019, doi: 10.1016/j.isprsjprs.2019.04.015.
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,”
Communications of the ACM, vol. 60, no. 6, pp. 84-90, 2017, doi: 10.1145/3065386.
[5] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” ICLR 2015, 2015, pp. 1–14,
doi: 10.48550/arXiv.1409.1556.
[6] Y. LeCun, Y. Bengio, and G. E. Hinton, “Deep learning,” Nature, vol. 521, pp. 436–444, 2015, doi: 10.1038/nature14539.
[7] T. -Y. Lin et al., “Microsoft COCO: common objects in context,” in European Conference on Computer Vision – ECCV 2014,
2014, pp. 740-755, doi: 10.1007/978-3-319-10602-1_48.
[8] J. Deng, W. Dong, R. Socher, L. -J. Li, K. Li, and L. F. -Fei, “ImageNet: a large-scale hierarchical image database,” in 2009 IEEE
Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255, doi: 10.1109/CVPR.2009.5206848.
[9] N. He, L. Fang, S. Li, J. Plaza, and A. Plaza, “Skip-connected covariance network for remote sensing scene classification,” IEEE
Transactions on Neural Networks and Learning Systems, vol. 31, no. 5, pp. 1461–1474, 2020, doi: 10.1109/TNNLS.2019.2920374.
[10] Sutikno, H. A. Wibawa, and P. S. Sasongko, “Detection of Ship using Image Processing and Neural Network,” TELKOMNIKA
(Telecommunication Computing Electronics and Control), vol. 16, no. 1, pp. 259–264, 2018, doi: 10.12928/telkomnika.v16i1.7357.
[11] Y. Yu, Z. Gong, C. Wang, and P. Zhong, “An unsupervised convolutional feature fusion network for deep representation of remote
sensing images,” IEEE Geoscience and Remote Sensing Letters, vol. 15, no. 1, pp. 23–27, 2018, doi: 10.1109/LGRS.2017.2767626.
[12] B. K. O. C. Alwawi and A. F. Y. Althabhawee, “Towards more accurate and efficient human iris recognition model using deep
learning technology,” TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 20, no. 4, pp. 817-824,
2022, doi: 10.12928/telkomnika.v20i4.23759.
[13] Y. Liu, Y. Liu, and L. Ding, “Scene classification based on two-stage deep feature fusion,” IEEE Geoscience and Remote Sensing
Letters, vol. 15, no. 2, pp. 183–186, 2018, doi: 10.1109/LGRS.2017.2779469.
[14] O. A. B. Penatti, K. Nogueira, and J. A. D. Santos, “Do deep features generalize from everyday objects to remote sensing and
aerial scenes domains?,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2015,
pp. 44-51, doi: 10.1109/CVPRW.2015.7301382.
[15] G. Cheng, X. Xie, J. Han, L. Guo, and G. -S. Xia, “Remote sensing image scene classification meets deep learning: challenges,
methods, benchmarks, and opportunities,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,
vol. 13, pp. 3735–3756, 2020, doi: 10.1109/JSTARS.2020.3005403.
[16] D. Marmanis, M. Datcu, T. Esch, and U. Stilla, “Deep learning earth observation classification using imagenet pretrained
networks,” IEEE Geoscience and Remote Sensing Letters, vol. 13, no. 1, pp. 105–109, 2016, doi: 10.1109/LGRS.2015.2499239.
[17] S. Thirumaladevi, K. V. Swamy, and M. Sailaja, “Multilayer feature fusion using covariance for remote sensing scene
classification,” Acta IMEKO, vol. 11, no. 1, pp. 1-8, 2022, doi: 10.21014/acta_imeko.v11i1.1228.
[18] X. Lu, H. Sun, and X. Zheng, “A feature aggregation convolutional neural network for remote sensing scene classification,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 57, no. 10, pp. 7894–7906, 2019, doi: 10.1109/TGRS.2019.2917161.
[19] J. R. Schott, “Principles of multivariate analysis: A user’s perspective,” Journal of the American Statistical Association, vol. 97,
no. 458, pp. 657-658, 2002, doi: 10.1198/jasa.2002.s479.

 ISSN: 1693-6930
814
[20] B. Petrovska, E. Zdravevski, P. Lameski, R. Corizzo, I. Stajduhar, and J. Lerga, “Deep learning for feature extraction in remote
sensing: A case-study of aerial scene classification,” Sensors, vol. 20, no. 14, 2020, doi: 10.3390/s20143906.
[21] Y. Yang and S. Newsam, “Bag-of-visual-words and spatial extensions for land-use classification,” in GIS '10: Proceedings of the
18th SIGSPATIAL International Conference on Advances in Geographic Information Systems, 2010, pp. 270-279,
doi: 10.1145/1869790.1869829.
[22] W. Zhou, S. Newsam, C. Li, and Z. Shao, “PatternNet: A benchmark dataset for performance evaluation of remote sensing image
retrieval,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 145, pp. 197–209, 2018, doi: 10.1016/j.isprsjprs.2018.01.004.
[23] B. Zhao, Y. Zhong, G. -S. Xia, and L. Zhang, “Dirichlet-derived multiple topic scene classification model for high spatial
resolution remote sensing imagery,” IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 4, pp. 2108–2123, 2016,
doi: 10.1109/TGRS.2015.2496185.
[24] Y. Liu, Y. Zhong, F. Fei, Q. Zhu, and Q. Qin, “Scene classification based on a deep random-scale stretched convolutional neural
network,” Remote Sensing, vol. 10, no. 3, 2018, doi: 10.3390/rs10030444.
[25] F. Zhang, B. Du, and L. Zhang, “Saliency-guided unsupervised feature learning for scene classification,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 53, no. 4, pp. 2175–2184, 2015, doi: 10.1109/TGRS.2014.2357078.
BIOGRAPHIES OF AUTHORS
Thirumaladevi Satharajupalli research scholar Jawaharlal Nehru Technological
University Kakinada, Kakinada. Asst. Professor in the Department of ECE, KKR & KSR
Institute of Technology & Sciences, Guntur. Having 10 years of experience in teaching and
Publishing work in various reputed journals national and international. Area of research in
digital image processing. She can be contacted atemail: thirumaladeviece@gmail.com.
Kilari Veera Swamy Dr.K. Veera Swamy is currently working as Professor in
ECE Department, Vasavi College of Engineering, Hyderabad. Earlier he worked as a Professor
and Principal at QIS College of Engineering & Technology, Ongole, India. He received his M.
Tech and Ph.D. from JNTUH, India. He has twenty-five years of experience in teaching
undergraduate students and post graduate students. Two scholars received Ph.D under his
guidance. He published 54 research articles (International 51 and National 3) in several
reputed journals. He attended 44 international and 11 national conferences. He received 2
patent grants. He published 3 patents. He received Best Teacher Award from JNTUK,
Kakinada for the AY:2014-15. He Executed one RPS, one MODROBS, and one Consultancy
Project. His research interests are in the areas of image compression, image watermarking,
machine learning, antennas, and networking protocols. He can be contacted at email:
k.veeraswamy@staff.vce.ac.in.
Maruvada Sailaja Professor, Department of ECE, Jawaharlal Nehru Technological
University Kakinada, Kakinada. Publishing various journals and papers from national and
international. Having 20 years of experience in teaching and industry. She is guiding several
research scholars her current research interests include areas of research in communications,
networks and signal processing. She can be contacted at email: s.maruvada@rediffmail.com.

Competent scene classification using feature fusion of pre-trained convolutional neural networks

Recommended

More Related Content

Similar to Competent scene classification using feature fusion of pre-trained convolutional neural networks (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Competent scene classification using feature fusion of pre-trained convolutional neural networks