Multilayer feature fusion using covariance for remote sensing scene classification


ACTA IMEKO 
ISSN: 2221-870X 
March 2022, Volume 11, Number 1, 1 - 8 

 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 1 

Multilayer feature fusion using covariance for remote sensing 
scene classification 

S. Thirumaladevi1, K. Veera Swamy2, M. Sailaja3 

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  
3 ECE Department, Jawaharlal Nehru Technological University, Kakinada - 533003, Andhra Pradesh, India  

 

 

Section: RESEARCH PAPER  

Keywords: Feature extraction; pre-trained convolutional neural networks; support vector machine; scene classification  

Citation: S. Thirumaladevi, K. Veera Swamy, M. Sailaja, Multilayer feature fusion using covariance for remote sensing scene classification, Acta IMEKO, 
vol. 11, no. 1, article 33, March 2022, identifier: IMEKO-ACTA-11 (2022)-01-33 

Section Editor: Md Zia Ur Rahman, Koneru Lakshmaiah Education Foundation, Guntur, India 

Received December 25, 2021; In final form February 20, 2022; Published March 2022 

Copyright: This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 License, which permits unrestricted use, 
distribution, and reproduction in any medium, provided the original author and source are credited. 

Corresponding author: S. Thirumaladevi, e-mail: thirumaladeviece1@gmail.com  

 

1. INTRODUCTION 

Remote sensing scene categorization has gotten a lot of 
attention recently, and it may be utilized in a variety of practical 
applications like urban planning, defence, space applications, in 
which measurement technology plays a key role [1]. On the other 
hand, it is a difficult challenge, since scene images often have 
complicated spatial structures with great intra-class and slight 
inter-class variability. To solve this problem, numerous strategies 
for scene classification have been advised in recent years [2]. 
Recently, inspired by the tremendous achievement of 
convolutional neural networks (CNNs) relating to the computer 
vision field [3]. Deep neural networks have gained prominence 
in the remote sensing community due to their exceptional 
performance, particularly in scene classification and computer 

vision applications [4]. Developing a deep CNN model from 
scratch, on the other hand, frequently necessitates a large amount 
of training data, whereas commercially available scene image data 
sets of remote sensing scene image data sets are typically tiny. 
Deep CNN models have a high degree of generalization on a 
wide range of tasks because they are commonly trained on 
ImageNet [5], which contains millions of images (e.g., scene 
classification and object detection [6]). In this context, the idea 
of using off-the-shelf pretrained CNN models for example 
AlexNet [7], Visual Geometry Group (VGG) 16 [8], and VGG19 
as feature extractors for scene categorization using remote 
sensing has gained attraction. The success is due to these models 
representing images using hierarchical architecture and can 
extract more representative Features. While these models can 
achieve categorization performance is excellent. Hu et al. [9] 
looked into two scenarios for using a pretrained CNN model 

ABSTRACT 
Remote sensing images are obtained by electromagnetic measurement from the terrain of interest. In high-resolution remote sensing 
imageries extraction measurement technology plays a vital role. The scene classification is one of the interesting and challenging 
problems due to the similarity of image structure and the available HRRS image datasets are all small. Training new Convolutional Neural 
Networks (CNN) using small datasets is prone to overfitting and poor attainability. To overcome this situation using the features 
produced by pre-trained convolutional nets and using those features to train an image classifier. To retrieve informative features from 
these images we use the existing Alex Net, VGG16, and VGG19 frameworks as a feature extractor. To increase classification performance 
further makes an innovative contribution fusion of multilayer features obtained by using covariance. First, to extract multilayer features, 
a pre-trained CNN model is used. The features are then stacked, downsampling is used to stack features of different spatial dimensions 
together and the covariance for the stacked features is calculated. Finally, the resulting covariance matrices are employed as features 
in a support vector machine classification. The results of the experiments, which were conducted on two difficult data sets, UC Merced 
and SIRI-WHU. The proposed Staked Covariance method consistently outperforms and achieves better classification performance. 
Achieves accuracy by an average of 6 % and 4 %, respectively, when compared to corresponding pre-trained CNN scene classification 
methods. 

mailto:thirumaladeviece1@gmail.com


 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 2 

(VGG16). The final few fully connected layers are portrayed as 
final image attributes for scene classification in the first scenario. 
In the second case, the final convolutional layer's feature maps 
are encoded to represent the input image using a standard 
method of feature encoding, such as the improved Fisher kernel 
[10]. The Vector Support Machine (SVM) is used as the final 
classifier in both cases. To improve the efficacy of the proposed 
method, the features were extracted from multiple CNNs of the 
same image combined by Xue et al. [11] for classification. For 
feature fusion, Sun et al. [12] have used the gated bidirectional 
connection method. In [13], the image is represented by 
combining the last two fully connected (FC) layers of a CNN 
model. 

Here we propose an innovative method, called the Stacked 
Covariance strategy to fuse features from different layers of a 
pre-trained CNN to classify remote sensing scenes. In the first 
phase, a pre-trained CNN model is used to extract multi-layered 
features and concatenate them. The covariance approach is used 
to aggregate the concatenated multiple feature vectors extracted 
from different layers. In contrast to traditional strategies, which 
only use first-order statistics to integrate feature vectors, the 
proposed strategy allows for the use of second-order statistics 
information. More representative features can thus be learned as 
a result. Then, the features are stacked, and covariance is 
calculated. Finally, for classification using an SVM classifier and 
improved the classification performance. 

This is how the rest of the paper is organized. Section 2 
explains the intended scene classification framework, novel 
aspects of our proposed technique. Section 3 contains the full 
experimental results for two data sets, Section 4 of this work 
concludes with some observations. 

2. PROPOSED TECHNIQUE DESCRIPTION 

The process of transforming the raw image into numerical 
features that can be processed while retaining the original 
information is referred to as feature extraction. With the upsurge 
of deep learning, the first layers of deep networks have largely 
replaced feature extraction, particularly for image data. Pre-
trained networks with hierarchical architecture can extract a large 
number of features from an image, which is thought to transmit 
additional information that can be put to much better use to 
increase categorization accuracy. The Learned image 
characteristics are first retrieved from a pre-trained convolutional 
neural network and then used to train an image classifier. All pre-
trained CNNs requisite fixed-size input images and specify the 
desired image size, as well as create augmented image data stores 
and use these data stores as input arguments to activations to 
automatically resize the training and test images before they are 
submitted to the network. Remove the pre-trained CNN's last 
FC layer FC8 and consider the rest as a fixed feature extractor. 
We feed an input image scene into the CNN and using the vector 
as a representation of global features of the input image, generate 
in advance a dimensional activation vector from the first or 
second FC layer. Finally, use the dimensional features to train a 
linear SVM classifier for scene classification. Figure 1 shows an 
illustration of this. 

To improve the classification accuracy further proposed a 
modified pre-trained network design by combining information 
from several convolution layers. The shallower levels of the 
CNN model are more likely to represent low-level visual 
components (such as edges), whereas the deeper layers exemplify 
more abstract information in the images. Furthermore, in certain 

computer vision applications, combining different levels, from 
shallow to deep, can provide state-of-the-art performance, 
meaning that merging different layers of CNN can be very 
helpful. 

Our proposed approach uses a similar strategy to take 
advantage of the information held by multiple layers in this way. 
This is represented in Figure 2. Here Convolutional layers of the 
last three blocks of pretrained networks are adopted and 
concatenated the features extracted from these layers. Namely 
conv3, conv4, conv5 in case of AlexNet and "conv3-3", "conv4-
3", and "conv5-3" in case of VGG16, VGG19. Different 
convolutional layers predominantly have distinctive spatial 
dimensions, they cannot be directly concatenated. To address 
this issue, downsampling with bilinear interpolation is used in 
conjunction with channel-wise average fusion. Obtained features 
that were reformed into a matrix along the channel dimension 
and aggregated using covariance. The proposed technique is 
described below. 

CNN Model is a collection of functions in which each 

function 𝑓𝑛 takes data samples 𝑋𝑛 and a filter bank 𝑏𝑛  as inputs 
and outputs 𝑋𝑛+1 , where 𝑛 = 1, 2, … 𝑁 and 𝑁 is the number of 
layers represented as a number (1) 

𝐹(𝑋) = 𝑓𝑁(… 𝑓2(𝑓1(𝑋; 𝑏1); 𝑏2)  …  𝑏𝑁 ) . (1) 

The filter bank 𝑏𝑛  for a pretrained CNN model was learned 
from a large data collection. The multiplayer characteristics are 

retrieved from an input image 𝑋 as follows: 𝐿1 = 𝑓1(𝑋; 𝑏1), 
𝐿2 = 𝑓2(𝑋; 𝑏2), and so on. As pretrained models, AlexNet, 
VGG16, and VGG19 are employed in this paper. The features 
produced from the convolutional layers of the last three blocks 
of pretrained networks are adopted and utilized. Different 
convolutional layers typically have different spatial dimensions; 
therefore, they can't be concatenated directly. Direct 

concatenation is not allowed when conv3 may have 𝐿1 ∈

𝑅1
ℎ×𝑤×𝑑1 , conv4 𝐿2 ∈ 𝑅2

ℎ×𝑤×𝑑2 , and conv5 has 𝐿3 ∈

𝑅3
ℎ×𝑤×𝑑3 . Downsampling with bilinear interpolation, as well as 

channel-wise average fusion, are used to solve this problem. 

Using downsampling with a 𝑑 number of dimensions three 
convolutional layers that have been pre-processed have been 
obtained and channel-wise average fusion is performed then the 

stacked feature set is acquired as follows: 𝐿 = [𝐿1, 𝐿2, 𝐿3] ∈

𝑅𝑖
𝑆×𝑆×𝐷

, where 𝐷 = 7 𝑑 and 𝑆 is the predefined down-sampled 
spatial dimension. The covariance-based pooling can be written 
as [14] 

 

Figure 1. Classification using single-layered pre-trained CNN as a feature 
extractor.  

 

Figure 2. Classification using stacked multilayer pre-trained CNN as a feature 
extractor. 



 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 3 

𝑃 =
1

𝑁 − 1
∑(𝑌𝑖 − 𝜇) (𝑌𝑖 − 𝜇)

T

𝑁

𝑖=1

 ∈ 𝑅𝑖
𝐷×𝐷 , (2) 

where [𝑌1, 𝑌2, … 𝑌𝑁 ] ∈ 𝑅𝑖
𝐷×𝑁

 is the vectorization of 𝐿, 𝑁 = 𝑆 2 

and 𝜇 = (1 𝑁⁄ ) ∑ 𝑌𝑖 ∈ 𝑅
𝐷𝑁

𝑖=1 , the covariance between the two 
separate features makes is represented by P, while the variance of 
each map is represented by diagonal entries. This method 
incorporates covariance (i.e., second-order statistics) to produce 
a more dense and discriminative exemplification. Second, the 
correlation between two distinct feature maps is represented by 
each entry in the covariance matrix. This is an easy approach to 
merge data from different feature maps that complement each 
other. 

The suggested method varies from existing CNN-based 
algorithms that are pre-trained. Concatenating the CNN's unique 
convolutional features (from shallow to deep layers), feature 
maps from several layers are merged. As a result, the suggested 
technique performs much better in terms of categorization. 
Furthermore, because the covariance matrices do not lie in 
Euclidean space, they cannot be processed by the SVM. The 
covariance matrix, on the other hand, can be mapped into 
Euclidean space using the matrix logarithm operation [15]. 

�̂� = logm(𝑃) = 𝑈 log ∑ 𝑈T  ∈ 𝑅𝐷×𝐷  , (3) 

where, 𝑃 = 𝑈 ∑ 𝑈T is the Eigen decomposition equation of 𝑃. 
The preceding operations are carried out both on the samples of 

training and testing. {𝑉𝑖 , 𝑆𝑖 }, 𝑖 = 1, 2, … 𝑛 and the testing sets are 

now taken into account where 𝑆𝑖  represents the number of 
training samples and 𝑛 represents the number of corresponding 
labels.  

{𝑉𝑖 , 𝑆𝑖 }, 𝑖 = 1, 2, … 𝑛 is exploited to train an SVM model as  

Min𝑎,ζ ,𝑏  {1 2⁄  ‖𝑎‖2 + 𝑃 ∑ ζ𝑖 } 

𝑆𝑖 (∠φ(𝑉𝑖 , 𝑎 + 𝑏))  ≥ 1 −  ζ𝑖  

𝜀𝑖 > 0, 𝑉𝑖 = 1, 2, … 𝑛  

(4) 

where 𝑎 and 𝑏 are the parameters of a linear classifier (.) is the 
mapping function and 𝜀𝑖 are positive slack variables to assert 
with outliers in the training set.  

With k(𝑉𝑖 , 𝑉𝑗 )  = 𝑉𝑖
T𝑉𝑗  

𝑓(𝑥) = sgn (∑ 𝑆𝑖  𝜆𝑖  𝑘(𝑣𝑖 , 𝑣)

𝑛

𝑖=1

+ 𝑏 ). (5) 

3. EXPERIMENTAL RESULTS ANALYSIS DISCUSSION 

3.1 Experimental Data Sets  

We run tests on two tough Image data sets related to remote 
sensing scene images to see how well the suggested approach 
performs. 1) Land Use Data Set from UC Merced [16]: 2100 
pictures are classified into 21 scene groups in the UC Merced 
Land Use (UC) [17] data set. Each class consists of 100 images 
in the RGB space with a size of 256 × 256 pixels. Each image 
has a one-foot pixel resolution. Figure 3. depicts sample images 

a)  b)  c)  d)  e)  f)  g)  

h)  i)  j)  k)  l)  m)  n)  

o)  p)  q)  r)  s)  t)  u)  

Figure 3. Land-use categories of 21 Example classes representation of UC Merced data set : a) agricultural13, b) airplane19, c) baseball diamond3, d) beach33, 
e) buildings21, f) chaparral13, g) denseresidential40, h) forest23, i) freeway23, j) golfcourse41, k) harbour31, l) intersection3, m) mediumresidential12, n) 
mobilehomepark12, o) overpass45, p) parkinglot32, q) river32, r) runway26, s) sparse residential64, t) storagetanks54, u) tenniscourt16.  



 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 4 

from each class, some categories (forest and sparse residential, 
for example) exhibit a significant level of interclass similarity, 
making the UC data set a difficult one to work with. 

SIRI-WHU was obtained from Google Earth (Google Inc.) 
and covers urban regions in China, as well as SIRI-WHU [18]. 
There are 12 items in the data set. Each class has 200 photos, 
each cropped to 200 × 200 pixels and with a spatial resolution of 
2 meters. In this study, 80 % of the training samples were chosen 
from the SIRI-WHU [19] Google data set, while the remaining 
amount of samples was kept for testing. The sample images of 
the SIRI-WHU data set are shown in Figure 4. 

3.2 Experimental Setup 

In our approach, multilayer features are extracted using three 
well-known CNN pretrained models: AlexNet [7], VGG-19 [8], 
and VGG-16 [8]. VGG-16 and VGG-19 three convolutional 
layers (e.g., "conv3–3," "conv4–3," and "conv5–3") , AlexNet's 
three convolutional layers (i.e., "conv3," "conv4", and "conv5") 
are used. For feature extraction, the scene images are preset to 
the size of the input layer such as 227 × 227 × 3 in the case of 
Alex Net and 224 × 224 × 3 forVGG-VG16 and 19. ImageNet 
is used to train both models. For illustration purposes, the UC 
data set and the SIRI-WHU data set are used, with 80 % training 
samples and 20 % testing samples selected. The most frequently 
used image classification assessment criteria are OA and 
confusion matrix, F1-score. 

Confusion Matrix: This is a special matrix that is commonly 
used g to visualize the output. Each column in this matrix 
represents the estimated value, while each row signifies an 
authentic category. As a result, evaluating is relatively simple. 

Overall Accuracy (OA): The number of appropriately 
categorized images Divide by the total number of images in the 
data set, regardless of which class they belong to. 

F1 score: The F1 score is a metric used to determine how 
accurate a test is. The harmonic mean of recall and precision is 
calculated using the test's precision and recall. 

Based on the combination of real and anticipated categories, 

classification problem with several 𝑀, which comprises 𝑃 

positive instances and 𝑁 negative instances, There are four types 
of cases: true positives (𝑇𝑃), false positives (𝐹𝑃), true negatives 
(𝑇𝑁), and false negatives (𝐹𝑁). The positive sample 𝑃 = 𝑇𝑃 +
𝐹𝑁 provides a positive sample that is expected to be positive, 
while 𝑇𝑃 represents a positive sample that is forecast to be 
negative. Similarly, 𝑇𝑁 denotes the number of negative cases that 
are identified as negative, while 𝐹𝑃 denotes the number of 
negative incidences that are predicted to be positive; thus, 𝑁 =
𝑇𝑁 + 𝐹𝑃 denotes the total number of negative samples. 

The fraction of correct instances is the accuracy, and the 
calculation equation is 

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃

𝑇𝑃 + 𝐹𝑁 + 𝑇𝑁 + 𝐹𝑃
. (6) 

The fraction of actually positive instances in all cases 
projected to be positive elements is called precision. The formula 
is as follows: 

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃

𝑇𝑃 + 𝐹𝑃
. (7) 

The recall calculation equation is the fraction of all positive 
samples that are projected to be positive. 

𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃

𝑇𝑃 + 𝐹𝑁
. (8) 

The F1-score is a precision and recall evaluation indicator 
with a comprehensive calculation equation. 

𝑅𝑒𝑐𝑎𝑙𝑙 =
2 ∙ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∙ 𝑅𝑒𝑐𝑎𝑙𝑙

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙
 . (9) 

The confusion matrix, along with accuracy, is shown in Figure 
5. The first column depicts a single fully connected layer is used 
as the final feature extractor for scene classification, whereas the 
second column depicts the proposed SC-based network 
classification. Experiments on the UC data set revealed that the 
OAs of pre-trained AlexNets is 79.76 %, VGG-19 is 81.19 %, 
and VGG-16 is 83.81 % while using SC and combining the last 

a)  b)  c)  d)  

e)  f)  g)  h)  

i)  j)  k)  l)  

 

Figure 4. Example class representation of the SIRI-WHU dataset: a) agriculture1, b) commercial50, c) harbor64, d) idle_land76, e) industrial111, f) meadow120, 
g) overpass15, h) park29, i) pond37, j) residential1, k) river106, l) water97.  



 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 5 

three Conv layers increases accuracy by 85 %, 87.14 %, and 
88.33 %, respectively. The proposed method accomplishes 
perfect classification performance on the majority of classes, 
such as Agricultural, beach, chaparral, forest, harbour, parking 
lot, runway and improved classes are buildings, dense residential, 
baseball diamond, tennis court and on average 6 % accuracy is 
increased in the case of UCM. 

Similar results can be obtained in the experiments conducted 
on the SIRI-WHU data set, confusion matrix for the fully 
connected layer is treated as a final feature extractor for scene 
classification, proposed SC-based classification are shown in 
Figure 6, where the pre-processed single layer of AlexNet is 
86.52 %%, VGG-19 is 87.6 %, and VGG-16 is 88.04 % while 
using the proposed strategy increases the accuracy by 90 %, 

a)  d)  

b)  e)  

c)  f)   

Figure 5. Confusion matrix of UC Merced Dataset using three pre-trained networks. First column corresponding to single-layered, a) Alex Net, b) VGG19, c) 
VGG16, second column corresponding to multi-layered fusion, d) SC-Alex Net, e) SC-VGG19, f) SC-VGG16.  

Table 1. Comparison results for the two data sets UCM and SIRI-WHU. 

 
UCM Dataset with 80% training 

(Overall Accuracy %) 
SIRI-WHU Dataset with 80% training  

(Overall Accuracy %) 

Network/Method 
Pre-trained network FC7  

as a feature extractor 
Proposed SC network  
as a feature extractor 

Pre-trained network FC7  
as a feature extractor 

Proposed SC network  
as a feature extractor 

AlexNet 79.76 85 86.52 90 

VGG-VD19 81.19 87.14 87.60 91.08 

VGG-VD16 83.81 88.33 88.04 92.60 



 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 6 

91.08 %, and 92.60 %, respectively. In most classes, the 
suggested technique achieves optimal classification performance 
like Agriculture, commercial, harbour, meadow, meadow and 
improved classes are Industrial, overpass, pond, overall 4 % 
accuracy is increased and comparison graphs are shown in Figure 
7. 

The related comparison results for the two data sets are 
shown in Table 1. The proposed scenario shows a clear 
improvement in OA when several Conv layers are combined. 

As illustrated in Figure 8, F1 scores of improved classes of 
UCM data set. By using the proposed strategy, a considerable 
number of classes show noticeable improvement such as 
agricultural, beach, harbour, runway reached 100 % and dense 
residential-class improves approximately 40 %. 

Likewise, Figure 9, shows the corresponding F1 scores of pre-
trained single-layered and proposed networks on the SIRI-WHU 
data set. As can be witnessed the proposed strategy, exhibits 
obvious improvements in most of the classes For example, on 
the SIRI-WHU data set water reached 100% and harbour, 
idle_land, industrial, overpass, park classes reach above 90%. 

4. CONCLUSION 

In this research, we portray Stacked covariance, a new 
technique for fusing image features from multiple layers of a 
CNN for scene categorization using remote sensing data. Feature 
extraction is performed initially with a pre-trained CNN model, 
followed by feature fusion with Covariance in the presented SC-
based classification framework. More dense features are 
recovered for classification because the proposed scenario takes 
into account second-order data. Each feature represents the 
covariance of two distinct feature maps and these features are 
applied to SVM for classification. Our extensive suggested SC 
method's effectiveness is validated by comparison with state-of-
the-art methodologies using two publicly accessible remote 
sensing image scene categorization data sets. We recognize that 
utilizing the proposed SC technique, the accuracy attained for 
most classes shows obvious enhancements, indicating that this is 
a viable improvement strategy. 

a)  d)  

b)  e)  

c)  f)  

Figure 6. Confusion matrix of SIRI-WHU Dataset using three pre-trained networks. First column corresponding to single-layered, a) Alex Net, b) VGG19, c) 
VGG16, second column corresponding to multi-layered fusion, d) SC-Alex Net, e) SC-VGG19, f) SC-VGG16.  



 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 7 

REFERENCES 

[1] L. Fang, N. He, S. Li, P. Ghamisi, J. A. Benediktsson, Extinction 
profiles fusion for hyperspectral images classification IEEE Trans. 
Geosci. Remote Sens., vol. 56, no. 3, 2018, pp. 1803–1815. 
DOI: 10.1109/TGRS.2017.2768479 

[2] X. Bian, C. Chen, L. Tian, Q. Du, Fusing local and global features 
for high-resolution scene classification, IEEE J. Sel. Topics Appl. 

Earth Observ. Remote Sens., vol. 10, no. 6, Jun. 2017, pp. 2889–
2901.  
DOI: 10.1109/JSTARS.2017.2683799  

[3] X. Lu, B. Wang, X. Zheng, X. Li, Exploring models and data for 
remote sensing image caption generation, IEEE Trans. Geosci. 
Remote Sens., vol. 56, no. 4, Apr. 2018, pp. 2183–2195.  
DOI: 10.1109/TGRS.2017.2776321  

 
 a) UCM Dataset  b) SIRI-WHU Dataset 

Figure 7. Comparison of pre-trained networks and proposed SC framework-based networks of UC Merced (UCM), SIRI-WHU Datasets.  

 

Figure 8. Comparison between F1 scores of UC Merced Data Set improved classes with pre-trained, proposed framework networks. 

 

Figure 9. Comparison between F1 scores SIRI-WHU Data Set improved classes with pre-trained, proposed framework networks. 

0

20

40

60

80

100

 AlexNet Proposed
AlexNet

VGG 19 Proposed
VGG19

 VGG16 Proposed
VGG16

agricultural

beach

denseresidential

golfcourse

harbor

intersection

river

60

70

80

90

100

 AlexNet Proposed
AlexNet

VGG 19 Proposed
VGG19

 VGG16 Proposed
VGG16

agriculture

harbor

idle_land

industrial

overpass

park

pond

https://doi.org/10.1109/TGRS.2017.2768479
https://doi.org/10.1109/JSTARS.2017.2683799
https://doi.org/10.1109/TGRS.2017.2776321


 

ACTA IMEKO | www.imeko.org March 2022 | Volume 11 | Number 1 | 8 

[4] B. M. Reddy, M. Zia Ur Rahman, Analysis of SAR images using 
new image classification methods, International Journal of 
Innovative Technology and Exploring Engineering, vol.8, no.8, 
2019, pp. 760-764. 

[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, ImageNet: 
A large-scale hierarchical image database, in Proc. IEEE Conf. 
Comput. Vis. Pattern Recognit., Miami, FL, USA, 20-25 June 
2009, pp. 248–255.  
DOI: 10.1109/CVPR.2009.5206848  

[6] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature 
hierarchies for accurate object detection and semantic 
segmentation, in Proc. IEEE Int. Conf. Comput. Vis., Columbus, 
OH, USA, 23-28 June 2014, pp. 580–587.  
DOI: 10.1109/CVPR.2014.81  

[7] Y. Wang, C. Wang, L. Luo, Z. Zhou, Image classification based on 
transfer learning of convolutional neural network, Chinese 
Control Conference (CCC), Guangzhou, China, 27-30 July 2019, 
pp. 7506-7510.  
DOI: 10.23919/ChiCC.2019.8865179  

[8] K. Simonyan, A. Zisserman, Very deep convolutional networks 
for large-scale image recognition, 3rd IAPR Asian Conference on 
Pattern Recognition (ACPR), Kuala Lumpur, Malaysia, 3-6 
November 2015, pp. 1-13.  
DOI: 10.1109/ACPR.2015.7486599  

[9] S. Tammina, Transfer learning using VGG-16 with deep 
convolutional neural network for classifying images, International 
Journal of Scientific and Research Publications (IJSRP), vol. 9, 
2019, no. 10, pp. 143-150.  
DOI: 10.29322/IJSRP.9.10.2019.p9420  

[10] S. Putluri, M. Z. Ur Rahman, S. Y. Fathima, Cloud-based adaptive 
exon prediction for DNA analysis, Healthcare Technology 
Letters, vol. 5, no. 1, 2018, pp. 25-30.  
DOI: 10.1049/htl.2017.0032  

[11] Y. Bi, B. Xue, M. Zhang, Genetic programming with image-related 
operators and a flexible program structure for feature learning in 
image classification, IEEE Transactions on Evolutionary 

Computation., vol. 25, no. 1, 2020, pp. 87–101.  
DOI: 10.1109/TEVC.2020.3002229  

[12] H. Sun, S. Li, X. Zheng, X. Lu, Remote sensing scene classification 
by gated bidirectional network, IEEE Trans. Geosci. Remote 
Sens., vol. 58, no. 1, pp. 82–96, 2019.  
DOI: 10.1109/TGRS.2019.2931801  

[13] G.Fiori, F.Fuiano, A.Scorza, J.Galo, S.Conforto, S.A Sciuto, A 
preliminary study on an image analysis based method for lowest 
detectable signal measurements in Pulsed Wave Doppler 
ultrasounds, Acta IMEKO, vol. 10, no.2, 2021, pp. 126-132. 
DOI: 10.21014/acta_imeko.v10i2.1051  

[14] L. Fang, N. He, S. Li, A. J. Plaza, J. Plaza, A new spatial– spectral 
feature extraction method for hyperspectral images using local 
covariance matrix representation, IEEE Trans. Geosci. Remote 
Sens., vol. 56, no. 6, Jun. 2018, pp. 3534–3546.  
DOI: 10.1109/TGRS.2018.2801387  

[15] V. Arsigny, P. Fillard, X. Pennec, N. Ayache, Geometric means in 
a novel vector space structure on symmetric positive-definite 
matrices, SIAM J. Matrix Anal. Appl., vol. 29, no. , 20071, pp. 328–
347.  
DOI: 10.1137/050637996  

[16] UC Merced Data Set. Online [Accessed December 2019]  
http://weegee.vision.ucmerced.edu/datasets/landuse.html  

[17] I. M. E. Zaragoza, G. Caroti, A. Piemonte, The use of image and 
laser scanner survey archives for cultural heritage 3D modelling 
and change analysis, Acta IMEKO, vol. 10, no.1, 2021, pp. 114-
121.  
DOI: 10.21014/acta_imeko.v10i1.847  

[18] SIRI-WHU Data Set. Online [Accessed August 2020]  
http://www.lmars.whu.edu.cn/prof_web/zhongyanfei/Num/G
oogle.html  

[19] Y. Liu, Y. Zhong, F. Fei, Q. Zhu, Q. Qin, Scene classification 
based on a deep random-scale stretched convolutional neural 
network, Remote Sensing, vol. 10, no. 3, Apr. 2018.  
DOI: 10.3390/rs10030444  

 

 

https://doi.org/10.1109/CVPR.2009.5206848
https://doi.org/10.1109/CVPR.2014.81
https://doi.org/10.23919/ChiCC.2019.8865179
https://doi.org/10.1109/ACPR.2015.7486599
https://doi.org/10.29322/IJSRP.9.10.2019.p9420
https://doi.org/10.1049/htl.2017.0032
https://doi.org/10.1109/TEVC.2020.3002229
https://doi.org/10.1109/TGRS.2019.2931801
https://doi.org/10.21014/acta_imeko.v10i2.1051
https://doi.org/10.1109/TGRS.2018.2801387
https://doi.org/10.1137/050637996
http://weegee.vision.ucmerced.edu/datasets/landuse.html
https://doi.org/10.21014/acta_imeko.v10i1.847
http://www.lmars.whu.edu.cn/prof_web/zhongyanfei/Num/Google.html
http://www.lmars.whu.edu.cn/prof_web/zhongyanfei/Num/Google.html
https://doi.org/10.3390/rs10030444