ResNet Classiﬁer Using Shearlet-Based Features for Detecting Change in

Satellite Images

Emna Brahim

, Sonia Bouzidi

1,2

and Walid Barhoumi

1,3

Universit

e de Tunis El Manar, Institut Sup

erieur d’Informatique d’El Manar, Research Team on Intelligent Systems in

Imaging and Artiﬁcial Vision (SIIVA), LR16ES06 Laboratoire de Recherche en Informatique,

Mod

elisation et Traitement de l’Information et de la Connaissance (LIMTIC), 2080 Ariana, Tunisia

Universit

e de Carthage, Institut National des Sciences Appliqu

ees et de Technologie,

1080 Centre Urbain Nord BP Tunis Cedex, Tunisia

Universit

e de Carthage, Ecole Nationale d’Ing

enieurs de Carthage,

45 Rue des Entrepreneurs, 2035 Tunis-Carthage, Tunisia

Keywords:

Change Detection, CNN, ResNet152, Shearlet Transform.

Abstract:

In this paper, we present an effective method to extract the change in two optical remote-sensing images.

The proposed method is mainly composed of the following steps. First, the two input Normalized Difference

Vegetation Index (NDVI) images are smoothed using the Shearlet transform. Then, we used ResNet152 ar-

chitecture in order to extract the ﬁnal change detection image. We validated the performance of the proposed

method on three challenging data illustrating the areas of Brazil, Virginia, and California. The experiments

performed on 38416 patches showed that the suggested method has outperformed many relevant state-of-the-

art works with an accuracy of 99.50%.

1 INTRODUCTION

In this work, we deal with the detection of the change

in two remote sensing images acquired in the same

area and at different times. In fact, change detection

in remote sensing images is used in many approaches:

Classiﬁcation based on difference, difference based

on classiﬁcation, or only classiﬁcation. In this frame-

work, deep learning is among the most efﬁcient tools

in remote sensing and image interpretation, and it has

shown its effectiveness in several connected domains

such as change detection (Brahim et al., 2021), im-

age classiﬁcation (Mittal et al., 2022), and object de-

tection (Bortoloti et al., 2022). Within the context

of the ﬁrst approach (classiﬁcation based on differ-

ence), there are many methods used in change detec-

tion in optical remote sensing images such as: (Shi

et al., 2021) who utilized a Deeply Supervised Atten-

tion Metric-based Network (DSAMNet). This tech-

nique allowed the treatment of large amounts of data

with a recorded accuracy rate of 93.69%. However,

(Brahim et al., 2021) employed the same techniques

introduced by (Shi et al., 2021), but the input data

were the subtract of the preprocessing images. Then,

the authors used the Convolutional Neural Network

(CNN) to detect the change. This method gave better

results for all types of data compared to those pro-

vided by (Shi et al., 2021). Likewise, (Zhang et al.,

2020) generated the Resnet to determine the features

of the two input images, before identifying the dif-

ference between them in order to detect the change.

This method did not give a better result (=96%), com-

pared to that employed by (Brahim et al., 2021), and

this is mainly due to the fact that (Zhang et al., 2020)

did not utilize a performant Resnet type. However,

(Wang et al., 2022) used a dual-path denoising net-

work (DPDNet). The authors generated the difference

image to extract the labels. Subsequently, they uti-

lized the difference image to extract the features and

applied a supervised classiﬁer to extract the change

in the studied images. This method showed higher

efﬁciency and provided better accuracies (=98, 50%)

than those presented in (Shi et al., 2021), (Zhang et

al., 2020), and (Chen et al., 2022) where a multi-

scale supervised fusion network (MSF-Net) was in-

vestigated in order to detect the change. In fact, de-

spite the precise generation of the difference between

the features, the detection of the change in the two

input images is quite inaccurate. Indeed, the used

network gave an accuracy equal to 71% to detect the

change. Regarding the methods used in the second

approach (difference based on classiﬁcation), (Dong

Brahim, E., Bouzidi, S. and Barhoumi, W.

ResNet Classiﬁer Using Shearlet-Based Features for Detecting Change in Satellite Images.

DOI: 10.5220/0011781200003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

427-434

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

427

et al., 2020) utilized the network DADNN to classify

the two input images and generate the difference be-

tween the obtained results. Although the execution of

this method was very complex, it gave a 96% satisfac-

tory accuracy rate. But, (Seyd et al., 2022) used the

DNN-Net deep neural network (DNN-Net). This net-

work gave a 91.15% accuracy which is low compared

to that provided by the methods used in the ﬁrst ap-

proach because in the ﬁrst approach when we gener-

ated the difference before the classiﬁcation, the infor-

mation about the two images was detailed. Regarding

the methods applied in the last approach (Only clas-

siﬁcation), (Zhang et al., 2022) employed SMD-Net

to detect the change in the two input images. This

method gave a 97.15% accuracy result, which is not

better than that provided by the ﬁrst approach.In this

paper, we try to ameliorate the efﬁciency of of the

ﬁrst approach technique proposed by (Brahim et al.,

2021), through developing a more efﬁcient method

and determining the efﬁcacity of the subtract features

compared with the subtract of the two input images.

The method is applied following the two steps de-

scribed below: Smoothing the two input NDVI im-

ages using the Shearlet transform decomposition and

K-means. Then, detecting the binary change by us-

ing ResNet152 to detect the ﬁnal change. This paper

is organized as follows: Section 2 deﬁnes the sug-

gested method. Section 3 presents the conducted ex-

periments and illustrates the obtained results. Finally,

Section 4 concludes the paper and presents some fu-

ture perspectives.

2 PROPOSED METHOD

The proposed method is presented in (Figure 1). The

Shearlet transform was used to smooth the two input

images by dividing them into high and low frequen-

cies. Then, classifying them using K-means and re-

constructing them to obtain the two smoothed images.

After that, ResNet152 was used to extract the ﬁnal re-

sult of the change.

2.1 Shearlet Transform

The shearlet transform, which is the 2D-Multiscale

Geometric Decomposition (MGD) employed to

smooth the two input NDVI images and detect the low

and high coefﬁcients of the two input images. In 2 di-

mension (ξ = (ξ

, ξ

) ∈

), the Shearlet transform

represents a frequency domain with dilations using

the following function:

j,s,k

(x) = a

−3

ϕ[U

−1

(x −k)]

(1)

Figure 1: Flowchart of the proposed method.

Where ϕ

j,s,k

is a shearlet, j is the scale, s ∈ R is the

shear direction, k is the translation parameter, V



a 0

√



is the dilation matrix, U



1 1

0 1



is the

shear matrix, with a≥ 0.

The coefﬁcients were computed using several

functions stated in the research work of (Brahim et al.,

2021). The steps of image decomposition are as fol-

lows: First, applying the Laplacian Pyramid and di-

rectional ﬁlter to decompose the two images into low

and high frequencies (Brahim et al., 2021) at scale j

(Figure 2).

Figure 2: Exemple of illustration of the decomposition in

two scales using Shearlet transform.

Classifying the low and high frequencies of each

image and in each scale, using K-means with this al-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

428

gorithm:

1. Identifying the number of clusters K. repeat (2, 3)

when the cluster K does not change.

2. Assigning each coefﬁcient to the closest partition.

3. updating the mean of each cluster.

After classifying the coefﬁcient, we employed the

inverse Fast Fourier Transform (FFT) to reconstruct

the two smoothed input images.

Finally, ResNet152 was used to extract the bi-

nary change (0: Unchanged, 1: Changed) of the two

smoothed images.

2.2 Residual Neural Network

The Residual neural network (ResNet152) includes

several block layers and pooling layers used to detect

the change (Figure 3).

Figure 3: Flowchart of Resnet.

In fact, the adopted Resnet152 contains many con-

volutional layers. The ﬁrst convolutional layer uti-

lizes a kernel size equal to (5,5), while the other con-

volutional layers are composed of four blocks. Each

block is made up of four convolutional layers having

kernel sizes equal to (3,3) and the outputs of the chan-

nel in each block are respectively equal to 64, 128,

256, 512. After extracting the features of the ﬁrst con-

volutional layers, the pooling layers were generated

into (2, 2) sizes. Then, the four blocks were utilized

to detect and subtract the features of the two images.

Then, we generated the Average Pooling function into

(2, 2). Finally, Softmax was used to detect the change

in the two images.

3 RESULTS

In this section, we describe the data areas and the

choice of the parameters of each technique applied

in the introduced method. We also demonstrate and

evaluate the obtained results and we show the efﬁ-

ciency of the developed method.

The treated data are multispectral NDVI images in

order to facilitate the analysis and the treatment of the

two images.

In the ﬁrst data, the two images represented the re-

gion of Brazil having the following coordinates: (17

° 10’S, 53 ° 10’N) and (19 ° 45’S, 55 ° 10’W). They

were captured by the Landsat satellite between Au-

gust 2000 and September 2001. The second data,

are the two images representing the region of Vir-

ginia having the following coordinates: (31 ° 07’S,

23’N) and (081 ° 37’S, 44’W). They were captured

by the Modis satellite between June 2021 and July

2022 and characterized by low spatial resolution. The

third data, represent two images captured by Landsat

satellite between March 2021 and April 2022. They

show the region of California with the following co-

ordinates: (34 ° 45’S, 29’N) and (084 ° 07’S, 51’W).

The dimensions of the two images of the three data

were 400×400. These images had low spatial resolu-

tion and were captured using two different satellites,

as shown in Figure 4.

Figure 4: NDVI original images: (a) and (b) NDVI images

captured in two periods 2000 and 2001. (c) and (d) NDVI

images captured in two periods 2020 and 2021. (e) and (f)

NDVI images captured in two periods 2021 and 2022.

The input original images were smoothed using

the Shearlet transform. The latter has 5 scales and 5

orientations (0, 10, 10, 10,18), which allow the good

smoothing of the two images. To show the high per-

formance of the used Shearlet, a part of the two im-

ages (Figure 5) was framed and the Shearlet trans-

form was decomposed. Then, the obtained coefﬁ-

cients were classiﬁed using k-means. The latter is bet-

ter than the OTSU because it does not yield data loss.

After classiﬁcation, the inverse Shearlet was gener-

ResNet Classiﬁer Using Shearlet-Based Features for Detecting Change in Satellite Images

429

ated to reconstruct the two smoothed images in each

data (Figure 5). To enhance the results given by the

Shearlet transform, we examined an area framed in

the two input images.

Figure 5: Smoothing area of the two images captured in

August 2000 and September 2001 using Shearlet transform:

(a) and (c) the area in August 2000 and September 2001,

(b) and (d) Shearlet transform. (e) and (g) the area in June

2021 and July 2022, (f) and (h) shearlet transform. (i) and

(k) the area captured in March 2021 and April 2022, (j) and

(l) Shearlet transform.

Figure 5, shows that the smoothing of the images

using Shearlet transform without employing other

methods, such as Wavelet, etc., provided very satis-

factory results. This method generated the multi-scale

decomposition to smooth the images.

After smoothing, the two images were classiﬁed

by ResNet152 using the CNN result of each pixel

(changed or unchanged labels) as a ground truth. The

convolutional neural network was employed to extract

the labels. It contains nine convolutional layers where

the ﬁrst kernel size (3, 3) and other kernel sizes (5, 5)

were used. CNN used the value 10 in the ﬁrst chan-

nel because the studied data is an optical multispec-

tral remote sensing. When we augmented the chan-

nel of the feature, the results became detailed. For

the two ﬁrst convolutional layers, the pooling func-

tion was transformed into (2, 2) size to extract the

features. Then, the subtraction features and the three

fully connected layers were determined to calculate

the probability of the labels. Afterward, Softmax was

applied to extract the labels in a binary form (0: Un-

changed, 1: Changed). This classiﬁer performed bet-

ter than the others because it gave better classiﬁcation

results. The ground truth of the network CNN is the

subtraction between the two smoothing input images.

If the subtraction is equal to 0, the pixel is unchanged

otherwise it is changed. The network was created us-

ing 200 epochs to detect the label of each pixel. In

this step, we extracted in each data (Brazil, Virgina,

California) a single area in each two input smoothing

images with the same position which contains sev-

eral changes in the ground truth. This area includes

38416 patches to better perceive the efﬁciency of the

network. To evaluate our network, we tested three pa-

rameters: The training data, testing data, and the val-

idation data. The three data are different. Concerning

the training data, we used this data to know the hid-

den features of the network. However, we used the

testing data to test the network after training. Then,

the validation data was used to validate our proposed

network. So, from the 38416 patches, we extracted

(60% = 23050) in the training data, (20% = 7683) in

the validation data, and (20% = 7683) in the testing

data with the size (5, 5) for each patch.

The results obtained using CNN are the ground

truth of the ResNet152 network. The latter was

trained using the ‘Adam’ optimizer because it was the

most widely used in the state of the art and gave good

results. The ﬁndings provided by each data are illus-

trated in Figure 6, Figure 7, Figure 8, and Figure 9.

In Figure 6, the results illustrated in (A), (B), and

the three data. In fact, (a), (b), (c), and (d) repre-

sent respectively the ﬁrst image, the second image,

the benchmark, and the result of the CNN. The ac-

curacy results obtained by the ﬁrst, second, and third

data were equal to 99.50%, 99.42%, and 99.46%, re-

spectively. This network was validated using differ-

ent types of data to validate the efﬁciency of the net-

work by employing many convolutional layers to de-

tect more details about the features. The second data

was of Modis type, where the network gave the best

results. The labels (changed and unchanged) of each

pixel were determined by this network utilized in our

previous work (Brahim et al., 2021). After this step,

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

430

Figure 6: Extraction the labels of each data (Brazil, Vir-

ginia, California) using CNN where white and black pixels

are respectively (Changed, Unchanged).

the results given by CNN were used in ResNet152. It

showed high efﬁciency in detecting the ﬁnal change

using two smoothed images. Four layers were em-

ployed in this network (Resnet152), which contained

blocks to demonstrate the features of these input im-

ages. Different values of kernel size of each convolu-

tional layer were also obtained. The input kernel size,

which was equal to 64 was chosen because it pro-

vided satisfactory results using the input multispectral

optical images of the Resnet152 network. The lat-

ter was trained by the same optimizer and epochs of

CNN. Then, the performance of the proposed method

was validated by comparing it to that of the methods

utilized by (Shi et al., 2021), (Brahim et al., 2021),

(Zhang et al., 2020), (Dong et al., 2020), and (Chen et

al., 2022) (Figure 7, 8 and 9). In each epoch, the accu-

racy and the loss were calculated and the parameters

of this network were updated to improve the learning

optimization.

In Figure 7, 8, and 9, the obtained results reveal

that the ﬁndings provided by the suggested technique

are better than those given by the other methods. The

importance of the proposed method was observed in

the subtraction between the features of the blocks in

ResNet152 to detect the change and the network is

made of several blocks of convolutional layers, which

allow better detection of the change in the two im-

ages. When there were high features, the characteris-

tics of each image were detailed and better classiﬁca-

tion ﬁndings were provided.

Figure 7: Change detection map results of Brazil data: (a),

(b), (c), (d), (e), (f), (g), (h), (i), (j) represent respectively

the ﬁrst image, the second image, the ground truth, the re-

sults of the change detection of (Shi et al., 2021), (Brahim

et al., 2021), (Zhang et al., 2020), (Dong et al., 2020), and

(Chen et al., 2022), the method Shearlet+CNN, and the pro-

posed method. Where white and black pixels are respec-

tively (Changed, Unchanged).

Figure 8: Change detection map results of Virgina data: (a),

(b), (c), (d), (e), (f), (g), (h), (i), (j) represent respectively

the ﬁrst image, the second image, the ground truth, the re-

sults of the change detection of (Shi et al., 2021), (Brahim

et al., 2021), (Zhang et al., 2020), (Dong et al., 2020), and

(Chen et al., 2022), the method Shearlet+CNN, and the pro-

posed method. Where white and black pixels are respec-

tively (Changed, Unchanged).

The Recall, Accuracy, and F1-Score were com-

puted using the following formula:

Recall =

T P

T P + FN

(2)

Accuracy =

(T P + T N)

(T P + FN + TN + FP)

(3)

F1 −score = 2 ×

(

T P

T P+F P

) ×Recall

(

T P

T P+F P

) +Recall

(4)

ResNet Classiﬁer Using Shearlet-Based Features for Detecting Change in Satellite Images

431

Figure 9: Change detection map results of California data:

(a), (b), (c), (d), (e), (f), (g), (h), (i), (j) represent respec-

tively the ﬁrst image, the second image, the ground truth,

the results of the change detection of (Shi et al., 2021),

(Brahim et al., 2021), (Zhang et al., 2020), (Dong et al.,

2020), and (Chen et al., 2022), the method Shearlet+CNN,

and the proposed method. Where white and black pixels are

respectively (Changed, Unchanged).

Where,TP, TN, FP and FN represent respectively

true-positive, true-negative, false-positive and false-

negative (Brahim et al., 2021). The data results are

presented in Table 1 and the TP, TN, FP, FN are illus-

trated in Table 2.

In table 1, the difference values between the ac-

curacy result given by our method and other existing

methods in the ﬁrst data are equal to 4.51%, 0.18%,

1.76%, 0.03%, 1.41%, and 5.2%, respectively. How-

ever, in the second data, the difference values are

respectively equal to 2.52%, 0.29%, 1.7%, 0.05%,

1.39%, and 6.74%. As regards the third data, the

difference values are respectively equal to 3.99%,

0.44%, 0.04%, 2.74%, 1.49%, and 3.75%. Using Ta-

ble 1, we conclude that the value of the difference be-

tween the accuracy result given by the two methods

DSAMNET (Shi et al., 2021), MSF-Net (Chen et al.,

2022), and the proposed method in each data is im-

portant because in the two works (Shi et al., 2021) and

(Chen et al., 2022) the input images were not prepro-

cessed so the result of change detection was not bet-

ter. For the two methods based on Shearlet+DI+CNN

(Brahim et al., 2021) and Shearlet+CNN, the value

of the difference between the accuracy result given

by the two methods and the proposed method was

low. However, in the ﬁrst method based on Shear-

let+DI+CNN (Brahim et al., 2021), the input data of

the CNN were the subtraction of the two smoothed

images using the Shearlet transform. In this shearlet,

the authors used the OTSU to classify all the coefﬁ-

cients but in the second method, Shearlet+CNN, we

utilized K-means to classify all the coefﬁcients with-

out losing the data and we generated the subtraction

of the features for more convolutional layers of CNN.

When the number of convolutional layers increases,

the features become more detailed. So the result ac-

curacy is increased compared to the result of Shear-

let+DI+CNN (Brahim et al., 2021). The following

Recall results obtained in the ﬁrst and second data of

MSF-Net (Chen et al., 2022) and DSAMNET (Shi

et al., 2021) were low compared to the other meth-

ods and the proposed method with a value difference

respectively equal to 5.17%, 0.34%, 0.89%, 0.18%,

0.67%, and 5.19% in the ﬁrst data. In the second

data, the value difference was respectively equal to

1.46%, 0.19%, 1.067%, 0.07%,0.85%, and 2.46%. In

the third data, the following Recall results of Shear-

let+DI+CNN (Brahim et al., 2021) was better com-

pared to the proposed method because in the Shear-

let+DI+CNN method (Brahim et al., 2021) the clas-

siﬁer of the coefﬁcients in shearlet is OTSU so the

data change can be lost. Then, The result of Shear-

let+DI+CNN (Brahim et al., 2021) increased com-

pared to the proposed method with a value equal to

0.4%, which is not a very important difference. The

F1-score result in the ﬁrst data, the second data and

the third data was low compared to the proposed

method. Obviously, the results improved the values,

which were respectively equal, in the ﬁrst data, to

5.05%, 0.17%, 0.9%, 0.01%, 0.44%, and 5.78%. In

the second data, the difference values were respec-

tively equal to 1.55%, 0.17%, 1.06%, 0.037%, 0.85%,

and 2.83%. However, in the third data, they were

equal to 17.05%, 2.59%,10.88%, 1.23%, 7.2%, and

15.58%. The accuracy result and the F1-score re-

sult of the proposed method increased compared to

the result of method Shearlet+CNN. Therefore, The

Resnet152 network played an important role in the

change detection. Table 1 shows that, in each data, the

accuracy result of the proposed method was consider-

ably higher compared to that of the other methods.

In Table 2, the result of FN of the proposed

method in the ﬁrst data decreased compared to other

methods. For example, the difference value pro-

vided by the method Shearlet+DI+CNN (Brahim et

al., 2021) and the introduced method was equal to

0.35% in the ﬁrst data, and equal to 0.19%, in the

second data. However, in the third data, the FN re-

sult increased, because the Recall result of the Shear-

let+DI+CNN method (Brahim et al., 2021) was su-

perior to the Recall result of the proposed method.

Consequently, the TP result of the Shearlet+DI+CNN

method (Brahim et al., 2021) was important compared

to the result of the proposed method. The degrada-

tion value was not important so the accuracy value

of the proposed method was better. Using the pro-

posed method, the FP result in three data was better

and increased compared to the result of the proposed

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

432

Table 1: Change detection results obtained by the proposed method with those provided by some relevant methods applied in

the state of the art (best values are in bold).

Data Methods Recall(%) F1-score(%) Accuracy(%)

Brazil DSAMNET (Shi et al., 2021) 94.43 94.35 94.99

Shearlet+DI+CNN (Brahim et al., 2021) 99.26 99.23 99.314

ResNet (Zhang et al., 2020) 98.72 98.51 97.74

Shearlet+Difference Features+Classiﬁers 99.43 99.40 99.47

DADNN (Dong et al., 2020) 98.93 98.96 98.09

MSF-Net (Chen et al., 2022) 94.42 93.63 94.26

The proposed method 99.61 99.41 99.50

Virginia DSAMNET (Shi et al., 2021) 98.30 98.09 96.90

Shearlet+DI+CNN (Brahim et al., 2021) 99.57 99.47 99.13

ResNet (Zhang et al., 2020) 98.69 98.57 97.72

Shearlet+Difference Features+Classiﬁers 99.69 99.61 99.37

DADNN (Dong et al., 2020) 98.91 98.79 98.03

MSF-Net (Chen et al., 2022) 97.30 96.81 94.75

The proposed method 99.76 99.64 99.42

California DSAMNET (Shi et al., 2021) 83.87 81.78 95.50

Shearlet+DI+CNN (Brahim et al., 2021) 98.38 96.24 99.05

ResNet (Zhang et al., 2020) 87.90 87.95 96.75

Shearlet+Difference Features+Classiﬁers 98.21 97.60 99.46

DADNN (Dong et al., 2020) 93.95 91.63 98.00

MSF-Net (Chen et al., 2022) 84.37 83.25 95.74

The proposed method 97.98 98.83 99.49

Table 2: Obtained TN, FP, FN, and TP rates given by the proposed method, CNN without Shearlet and SCNN (best values

are in bold).

Data Methods TN(%) FP(%) FN(%) TP(%)

Brazil DSAMNET (Shi et al., 2021) 95.44 4.55 5.5 94.43

Shearlet+DI+CNN (Brahim et al., 2021) 99.35 0.64 0.73 99.26

ResNet (Zhang et al., 2020) 93.45 6.5 1.3 98.69

Shearlet+CNN 99.49 0.5 0.56 99.43

DPDNET (Wang et al., 2022) 94.39 5.6 1 98.93

MSF-Net (Chen et al., 2022) 94.10 5.8 5.5 94.42

The proposed method 99.38 0.61 0.38 99.61

Virginia DSAMNET (Shi et al., 2021) 90.90 9 1.6 98.30

Shearlet+DI+CNN (Brahim et al., 2021) 97.27 2.7 0.42 99.57

ResNet (Zhang et al., 2020) 93.24 6.7 1.2 98.72

Shearlet+CNN 97.91 2.04 0.3 99.69

DPDNET (Wang et al., 2022) 94.20 5.79 1 98.91

MSF-Net (Chen et al., 2022) 81.87 18.12 2.6 97.30

The proposed method 97.99 2 0.23 99.76

California DSAMNET (Shi et al., 2021) 99.70 2.8 16 83.87

Shearlet+DI+CNN (Brahim et al., 2021) 99.14 0.85 1.6 98.38

ResNet (Zhang et al., 2020) 98.00 1.9 12 87.90

Shearlet+CNN 99.61 0.38 1.7 98.21

DPDNET (Wang et al., 2022) 98.57 14 6.04 93.95

MSF-Net (Chen et al., 2022) 97.35 2.6 15 84.37

The proposed method 99.71 0.28 2 97.98

ResNet Classiﬁer Using Shearlet-Based Features for Detecting Change in Satellite Images

433

method, because the TN result of the other methods

was very low compared to the result of the proposed

method. Accordingly, the FP result of the proposed

method was better.

4 CONCLUSION AND

PERSPECTIVES

In this work, we presented a method of change

detection based on Shearlet transform, CNN and

ResNet152. The Shearlet transform allowed decom-

posing the two input images into low and high fre-

quencies. Then, reconstructing them to smooth the

two input images. After that, the labels (changed, un-

changed) were detected using the CNN network to use

them in Resnet. Afterward, the ResNet was generated

to extract the features using many blocks and generat-

ing their subtraction. Subsequently, the classiﬁer was

employed to detect the change. Despite its good re-

sults and high efﬁciency compared to the methods in-

troduced in the ﬁrst approach. This method is charac-

terised by the shearlet and two networks where each

networks is composed by many convolutional layers.

As far as the shearlet transform is concerned, it allows

to facilitate the treatement of the two input images.

But, this method can cause problems in the treatment

of the hyperspectral images.

Therefore, in future works, we will further explore

our method to enhance our network or create other

networks to detect the change variance.

REFERENCES

Bortoloti, F. D., Tavares, J., Rauber, T. W., Ciarelli, P.

M., and Botelho, R. C. G. (2022). An annotated im-

age database of building facades categorized into land

uses for object detection using deep learning. Machine

Vision and Applications, 33(5), 1-16.

Brahim, E., Bouzidi, S., and Barhoumi, W. (2021). Change

detection in optical remote sensing images using

shearlet transform and convolutional neural networks.

In 2021 IEEE/ACS 18th International Conference on

Computer Systems and Applications (AICCSA), (pp.

1-7).

Chen, J., Fan, J., Zhang, M., Zhou, Y., and Shen, C. (2022).

MSF-Net: A Multiscale Supervised Fusion Network

for Building Change Detection in High-Resolution

Remote Sensing Images. IEEE Access, 10, 30925-

30938.

Dong, H., Ma, W., Wu, Y., Zhang, J., and Jiao, L.

(2020). Self-supervised representation learning for re-

mote sensing image change detection based on tem-

poral prediction. Remote Sensing, 12(11), 1868.

Han, M., Li, R., and Zhang, C. (2022). LWCDNet: A

Lightweight Fully Convolution Network for Change

Detection in Optical Remote Sensing Imagery. IEEE

Geoscience and Remote Sensing Letters, 19, 1-5.

Li, X., Du, Z., Huang, Y., and Tan, Z.(2021). A deep trans-

lation (GAN) based change detection network for op-

tical and SAR remote sensing images. ISPRS Journal

of Photogrammetry and Remote Sensing, 179, 14-34.

Liu, M., Shi, Q., Liu, P., and Wan, C. (2020). Siamese

Generative Adversarial Network for Change Detec-

tion Under Different Scales. In IGARSS 2020-2020

IEEE International Geoscience and Remote Sensing

Symposium (pp. 2543-2546). IEEE.

Mittal, S., Srivastava, S., and Jayanth, J. P. (2022). A Survey

of Deep Learning Techniques for Underwater Image

Classiﬁcation. IEEE Transactions on Neural Networks

and Learning Systems.

Shi, Q., Liu, M., Li, S., Liu, X., Wang, F., and Zhang,

L. (2021). A deeply supervised attention metric-based

network and an open aerial image dataset for remote

sensing change detection. IEEE Transactions on Geo-

science and Remote Sensing, PP(99):1-16.

Shi, K., Bai, L., Wang, Z., Tong, X., Mulvenna, M. D., and

Bond, R. R. (2022). Photovoltaic Installations Change

Detection from Remote Sensing Images Using Deep

Learning. In IGARSS IEEE International Geoscience

and Remote Sensing Symposium (pp. 3231-3234).

Wang, J., Gao, F., Dong, J., Du, Q., and Li, H. C. (2022).

Change Detection From Synthetic Aperture Radar Im-

ages via Dual Path Denoising Network. IEEE Journal

of Selected Topics in Applied Earth Observations and

Remote Sensing, 15, 2667-2680.

Zhang, H., Liu, J., and Xiao, L. (2020). Bipartite Resid-

ual Network for Change Detection in Heterogeneous

Optical and Radar Images. In IGARSS IEEE Inter-

national Geoscience and Remote Sensing Symposium

(pp. 332-335).

Zhang, X., He, L., Qin, K., Dang, Q., Si, H., Tang, X.,

and Jiao, L. (2022). SMD-Net: Siamese Multi-Scale

Difference-Enhancement Network for Change Detec-

tion in Remote Sensing. Remote Sensing, 14(7), 1580.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

434