Comparative Examination of Different Change Detection Methods for

Remote Sensing Imagery

Jambukeshwar Pujari

1,2 a

, Javed Wasim

1 b

and Aprna Tripathi

3 c

Department of Computer Engineering and Applications, Mangalayatan University, Aligarh-202146, Uttar Pradesh, India

Department of Computer Science and Business Systems, Kolhapur Institute of Technology’s College of

Engineering(Autonomus), Kolhapur, Maharshtra, India

Department of Data Science and Engineering, Manipal University, Jaipur, India

Keywords:

Remote Sensing (RS),Change Detection (CD).

Abstract:

Land cover change analysis relies heavily on Change Detection (CD), which is the process of identifying

semantic changes in satellite images taken at different dates. Several methods have been devised to identify

changes in satellite photos. Simple to compute and apply, image differencing doesn’t necessitate data collected

from the ground.This study employs performance for change detection to examine multiple strategies that have

been created by different researchers. Detecting changes in high-resolution satellite pictures is crucial for a

better understanding of land surfaces. Using existing methods, reliably detecting changes in satellite pictures

is a tough undertaking. Datasets utilized in the trials were the OSCD dataset and the SECOND dataset. In

comparison with other techniques, RSCDNet achieved higher accuracy, precision, recall, and F1-score.

1 INTRODUCTION

Change detection, often known as CD, is a method

that is utilized to recognize variations in events, fea-

tures, and patterns on the surface of the ground

throughout the course of time(Qiu et al., 2013). CD

is utilized extensively in a variety of geoscientiﬁc do-

mains, including but not limited to urban development

also change monitoring, forest monitoring, environ-

mental disaster prevention and map updating. Be-

cause of their extensive coverage, relevant temporal

resolution, availability at various times and places,

high spectral, spatial, and radiometric resolution, dig-

ital format, and the ability to be processed by com-

puters, remote sensing (RS) data are extremely useful

for the study of temporal and spatial changes in land

cover and land use. In remote sensing photos, changes

in land usage and land cover can be seen as variations

in texture, shape, or gray levels(Bao and Guo, 2004).

Accurate change detection is crucial because of

CD’s signiﬁcance in land cover/use. So, the efﬁcacy

and precision of the ﬁndings are heavily dependent on

the CD method’s capability as well as the quality of

https://orcid.org/0000-0003-3422-4515

https://orcid.org/0000-0001-8528-4556

https://orcid.org/0000-0002-4714-3098

the data utilized to identify changes. In order to pro-

cess the data and generate correct information layers

and change maps, suitable and efﬁcient procedures

are necessary. There have been a number of inves-

tigations into RS CD algorithms(Fatemi Nasrabadi,

2019).

Finding new ways to detect changes in satellite

images is motivating researchers to develop better CD

techniques.In recent times, satellite images have been

enhanced with CD techniques that possess strong

discriminative abilities, leading to improved perfor-

mance(Raza et al., 2022).Supervised and unsuper-

vised CD are the two main categories. Although su-

pervised methods are known to produce better out-

comes, labeling training data manually is a real pain.

However, unsupervised approaches are more com-

monly utilized in real-world applications since they

automatically recognize changes(Fang et al., 2022).

In the present study, recently proposed techniques

for RS CD were selected and investigated their per-

formance on OSCD dataset and SECOND dataset.

344

Pujari, J., Wasim, J. and Tripathi, A.

Comparative Examination of Different Change Detection Methods for Remote Sensing Imagery.

DOI: 10.5220/0013592100004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 344-350

ISBN: 978-989-758-763-4

2 CHANGE DETECTION

CHALLENGES

Change detection (CD) is a crucial approach for com-

prehending and identifying signiﬁcant urban trans-

formations through the utilization of available ob-

served Earth data. Change detection is a signiﬁ-

cant application of remotely sensed data and it in-

volves in analyzing more than two repetitive satel-

lite images captured from same geographical earth

area at different time intervals. Change detection

is applied in urban planning assessments, monitor-

ing deforestation in agricultural areas, and managing

disasters, among other applications. In addition to

the temporal variations observed across a land sur-

face that contemporary techniques seek to identify

in remotely sensed images, the spectral characteris-

tics of bi-temporal images correspond to alterations

within the geographical area(Chughtai et al., 2021).

These modiﬁcations can be integrated in various man-

ners and provide a detailed examination of the pro-

cedures involved in identifying the permissible alter-

ations. Utilizing these remotely sensed images for

CD presents certain challenges. Initially, there is a

discrepancy in the data acquisition parameters result-

ing from repetitive coverage at brief intervals. Sec-

ondly, the vast array of spectral and statistical fea-

tures present in these images, along with the imbal-

anced and integrated pixel data, represent signiﬁcant

challenges. The current approaches employ spec-

tral and statistical characteristics that effectively iden-

tify changes. Nonetheless, variations in incident an-

gle lead to changes in the spectral features during

the image acquisition process. The statistical fea-

tures remain constant; thus, data selection is crucial to

mitigate the unwanted disturbances in the variations.

Certain unnecessary alterations may also be associ-

ated with pre-processing techniques like thresholding,

which involves the transformation of images.

3 REVIEW OF LITERATURE

The V-BANet deep learning technique was used to

split landscapes and extract information from photos.

It integrated the V-net with a Bilateral Attention Net-

work. Independent processing of each bi-temporal

image allows V-Net to detect objects in them. Ad-

ditionally, features were retrieved from these seg-

mented pictures utilizing the channel and spatial at-

tention techniques of BANet (Prasad et al., 2023).

An approach to remote sensing image change de-

tection using supervised deep learning called multi-

scale CD. Deep networks monitored by an adminis-

trator were able to identify modiﬁcations to remote

sensing images. According to (Alshehhi and Marpu,

2023), it employed dice correlation between reference

change maps on several scales and multi-scale fore-

cast probability change and error functions and also

used the Domain Knowledge-guided Self-Supervised

Change Detection (DK-SSCD) method, which com-

bines domain knowledge of remote sensing indices

during training and inference, to enable unsupervised

CD capabilities. It improved CD accuracy, reduced

quality spikes, and provided a good feature represen-

tation space that highlighted changed information for

bitemporal images by using contrastive learning and

domain knowledge (Yan et al., 2023). The TD-SSCD

method, which stands for Temporal and Differential

information for Self-Supervised Contrastive Learn-

ing Change Detection, employed as an alternating

iteration learning technique to progressively under-

stand potential connections between bitemporal and

their differential images. Internal to a self-supervised

learning framework, this was executed (Qu et al.,

2023).

A CD deep learning network that makes use of

satellite pictures is called Urban CD Network (UCD-

Net). There was no loss of shape in the altered ar-

eas’ geometry while using UCDNet to forecast their

borders (Basavaraju et al., 2022). The Remote Sens-

ing Change Detection Network is an all-inclusive DL

architecture for CD using RSCDNet data. Problems

with objects touching their borders and with objects

of varying sizes and shapes were handled (Barkur

et al., 2022). Weighted binary cross-entropy loss

function is used and updated U-net with binary cross-

entropy is used in identifying changes, particularly

tiny ones, in complex urban settings. Databases in

northern urban regions, where changes occur quickly

and are monitored with often updated information,

were helped by this network. This was due to its

lightning-fast processing time, which allows it to pro-

duce change maps with pinpoint accuracy in a ﬂash

(Gomroki et al., 2022). One method for detecting CD

in satellite pictures is the Sibling Regression for Opti-

cal Change detection (SiROC) approach, which looks

at how CD pixels change over time (Kondmann et al.,

2021). Deep Learning (DL) approach for CD that is

highly advanced. An effective method for feature ex-

traction and semantic segmentation, EffCDNet makes

use of a pre-trained architecture.

4 METHODOLOGY

Change Detection of satellite images by various tech-

niques(algorithms) are done as shown in the Figure 1.

Comparative Examination of Different Change Detection Methods for Remote Sensing Imagery

345

The dataset OSCD and SECOND are trained by each

considered technique, then trained model are tested

through test sample images.

Figure 1: Change Detection Process

4.1 Data Source

The dataset is derived from two separate sources:

the Onera Satellite Change Detection dataset (OSCD)

and the SEmantic Change DetectiON Dataset (SEC-

OND). OSCD is composed of multispectral images

obtained from Sentinel-2 satellites. This dataset is

utilized to tackle the challenge of identifying changes

between satellite images. The SECOND dataset em-

phasizes the identiﬁcation of changes in land cover

and the accurate categorization of these changes with

precise pixel-level delineations. The approach iden-

tiﬁes semantic changes through the utilization of fea-

ture pairs obtained from modules with varied struc-

tures, which consider different spatial extents and

quantities of parameters. The images from this

dataset undergo processing to eliminate noise and cor-

rect errors during the initial pre-processing stage.

4.1.1 Onera Satellite Change Detection dataset

The dataset comprises 24 collections of images ac-

quired from Sentinel-2 satellites during the period

from 2015 to 2018. Locations are selected from var-

ious regions across the globe. Each site is provided

with pairs of Sentinel-2 satellite images captured in

13-band multispectral mode. The OSCD dataset con-

tains images with varying spatial resolutions of 10m,

20m, and 60m(Caye Daudt et al., 2019).

4.1.2 Semantic Change Detection Dataset

The dataset comprises 4662 pairs of images sourced

from multiple platforms. Each image is characterized

by a resolution of 512 x 512 and includes pixel-level

details. The second annotation is conducted by a spe-

cialized group focused on earth vision applications,

ensuring a high level of label accuracy(Yang et al.,

2020).

4.2 Data Preprocessing

Images undergo pre-processing to address pixel size,

resolution issues, and to eliminate unwanted noise

and errors. Enhance the overall clarity of the origi-

nal image. Through the analysis of neighboring pix-

els and their likeness, it executes adaptive ﬁltering,

thereby improving the overall quality of the image.

The method successfully minimizes noise in images

while preserving their structural integrity.

5 CHANGE DETECTION

TECHNIQUES

5.1 EffCDNet

It is enhanced Convolution Neural Network (CNN)

with preserving the efﬁciency and precision of seg-

mentation within the network. An optimized archi-

tecture known as EffCDNet utilizes a siamese-based

pre-trained encoder pair attached with an Attention-

based UNet decoder,which performs semantic seg-

mentation. The network utilizes a pre-trained Efﬁ-

cientNet architecture, incorporating shared weights to

enhance feature extraction capabilities. The UNet de-

coder utilizes attention mechanisms and adopts the

attention-gate layer just prior to the concatenation op-

eration. This acquires more distinct relevant features

to enhance the segmentation performance. The re-

construction of ﬁne-grained feature maps, leverag-

ing substantial context information for enhancement

in the change map, employed the Undecimated Dis-

crete Wavelet Transform (UDWT) fusion as a post-

processing technique. This approach facilitated spa-

tial and temporal analysis of multi-resolution images,

resulting in a signiﬁcantly improved information dif-

ference map(Patil et al., 2021).

5.2 V-BANet

Deep learning techniques that employ V-Net and Bi-

lateral Attention Network (V-BANet) are utilized for

the segmentation of landscapes and the extraction of

features from images. The bi-temporal images are ini-

tially segmented using V-Net to independently iden-

tify the objects present in each image. The Bilateral

INCOFT 2025 - International Conference on Futuristic Technology

346

Attention Network employs spatial and channel atten-

tion blocks to enhance the extraction of discriminative

features from images. The relationships among the

features are elucidated through a comparison of the

original feature map in one image with the modiﬁed

feature map in another.(Prasad et al., 2023).

5.3 UCDNet

A deep learning model known as the urban CD net-

work (UCDNet) has been developed for urban change

detection using bi-temporal multispectral Sentinel-2

satellite images. The architecture of the model is

founded on an encoder–decoder framework that in-

corporates modiﬁed residual connections along with

the new spatial pyramid pooling (NSPP) block. The

encoder unit consists of two streams that have iden-

tical structures and share weights, as demonstrated in

Every input image is assigned to one of these identi-

cal structures. Similar to the FC-Siamconc network,

the encoder component is made up of convolutional

and pooling layers. Each stream incorporates three

pooling layers. The primary goals of the CD tech-

niques involve detecting variations between two im-

ages, with the disparity in the learned features serv-

ing as the input for the adjusted residual connection

at each stage of the encoder component.The input to

the NSPP block consists of features that have been

learned from the encoder section. This process aids

in extracting features across different ranges, provid-

ing insight into the global context(Basavaraju et al.,

2022).

5.4 RSCDNet

The RSCDNet-integrated Modiﬁed Self Attention

(MSA) module and Gated Linear Atrous Spatial Pyra-

mid Pooling (GL-ASPP) block’s extensive functional-

ity. In order to remove unnecessary channel informa-

tion from features at many scales, a GL-ASPP assem-

bly uses a channel-wise descriptor in conjunction with

a gated module. In order to successfully ﬁlter out ab-

normal information ﬂow from the encoder to the de-

coder, the GL-ASPP block also accounts for channel

dependencies. Careful design went into the Modiﬁed

Self Attention (MSA) block so it could combine the

strengths of the channel attention operation with the

backbone self-attention unit. The input feature vec-

tor’s spatial channel dependency is used by this op-

erator. In contrast to the conventional self-attention

module, the computed channel-self attention is passed

via an attention gate to ﬁlter out extraneous data and

emphasize the important parts.(Barkur et al., 2022).

5.5 IU-Net

This approach employed an enhanced IU-Net con-

volutional network for the purpose of change detec-

tion. The IU-Net architecture consists of two distinct

pathways. The initial pathway, referred to as the en-

coder, is responsible for identifying the background

of images, while the subsequent pathway, known as

the decoder, determines the precise location of fea-

tures through transposed convolution. It employs 64

3×3 double convolution kernels across ﬁve blocks, in-

corporating the ReLu activation function, batch nor-

malization layers, and four 2×2 Max Pooling opera-

tors during the encoding phase. During the decoding

stage, a transposed convolution with a stride of two,

along with concatenation and two convolution layers

utilizing a 3×3 kernel size, was employed to upsam-

ple the multiscale feature maps. Upon completion of

the decoding stage, a singular 1×1 convolution layer

featuring a Softmax activation function, along with

weighted binary cross-entropy, is employed for the

purpose of change detection. The change detection

was conducted initially using RGB bands and subse-

quently with both RGB and NIR bands. In the analy-

sis of the datasets, 67% of the data was allocated for

training purposes, while 33% was designated for test-

ing(Gomroki et al., 2022).

Figure 2: OSCD Dataset: A) Input Image B) Filtered Image

C) Detected Image

Comparative Examination of Different Change Detection Methods for Remote Sensing Imagery

347

Figure 3: SECOND Dataset: A) Input Image B) Filtered

Image C) Detected Image

Samples of input images, preprocessed images

and change detected output images are shown for

OSCD data images and SECOND data images in Fig-

ure 2 and Figure 3 respectively

6 RESULTS AND DISCUSSION

6.1 Performance Analysis for OSCD

Dataset

Table 4 illustrates several performance metrics of

OSCD dataset. EffCDNet model achieved an accu-

racy of 92.6%, with corresponding precision, recall,

IoU and F1-score of 92.6%, 87.7%, 98% and 89.2%,

respectively. V-BANet transformer model demon-

strated high performance across all metrics, attaining

precision of 98.93% along with accuracy, IoU, recall

and F1-score of, 99.29%, 98.31%, 98.96%, 98.87%.

Similarly, UCDNet model showed recall at 86.16%,

with precision, accuracy, kappa co-efﬁcient and F1-

score of 95.53%, 99.30%, 88.85% and 89.21%.

IU-Net model achieved F1-score of 98.65% along

with precision, recall, IoU and accuracy of 98.64%,

98.64%, 97.34% and 97.38%. RSCDNet achieved

an accuracy of 99.2% with recall, F1-score, IoU,

kappa co-efﬁcient and precision of 99%, 99.1%, 96%,

98.10% and 99% respectively.

The performance analysis of various techniques is

shown in the graph ﬁgure 4 for OSCD. here compari-

son is shown for accuracy parameter.

Figure 4: Analysis on OSCD Data

6.2 Performance Analysis for SECOND

Dataset

Table 5 illustrates several performance metrics of

SECOND dataset. EffCDNet model achieved an ac-

curacy of 91.5%, with corresponding precision, re-

call, IoU and F1-score of 94.6%, 85.5%, 95% and

88.6%, respectively. V-BANet transformer model

demonstrated high performance across all metrics,

attaining precision of 96.83% along with accuracy,

IoU, recall and F1-score of, 97.29%, 96.31%, 97.6%,

98.5%. Similarly, UCDNet model showed recall at

85.25%, with precision, accuracy, kappa co-efﬁcient

and F1-score of 96.5%, 98.78%, 87.8% and 89.30%.

IU-Net model achieved F1-score of 97.6% along with

precision, recall, IoU and accuracy of 97.4%, 98%,

97.5% and 96.80%. RSCDNet achieved an accu-

racy of 98.25% with recall, F1-score, IoU, kappa

co-efﬁcient and precision of 98.50%, 98.7%, 97%,

98.30% and 98% respectively. The performance anal-

ysis of different techniques is shown in the Table 1.

The performance analysis of various techniques is

shown in the graph ﬁgure 5 for SECOND. here com-

parison is shown for accuracy parameter.

7 CONCLUSIONS

The discussion focuses on prevalent techniques for

detecting changes for satellite images and the ap-

proach of change detection following classiﬁcation.

This paper proceeds from the principles of them, ana-

lyzing their results and comparing them with respect

to performance parameters such as accuracy, preci-

sion, recall, F1-Score, IoU (Intersection over Union),

INCOFT 2025 - International Conference on Futuristic Technology

348

Table 1: Performance Analysis on OSCD Data

Techniques Accuracy (%) Precision (%) Recall (%) F1-score (%) IoU (%) Kappa co-efﬁcient (%)

EffCDNet 92.5 92.6 87.7 89.2 98 -

V-BANet 99.29 98.93 98.95 98.87 98.31 -

UCDNet 99.30 95.53 86.16 89.21 - 88.85

RSCDNet 99.5 98.40 98.30 98.20 96 98.10

IU-Net 97.38 98.64 98.64 98.65 97.34 -

Table 2: Performance Analysis on SECOND Data

Techniques Accuracy (%) Precision (%) Recall (%) F1-score (%) IoU (%) Kappa co-efﬁcient (%)

EffCDNet 91.5 94.6 85.5 88.6 95 -

V-BANet 96.83 97.29 96.31 97.6 98.5 -

UCDNet 96.5 98.78 87.8 89.30 - 88.85

RSCDNet 98.5 98.7 97 98.30 96 98

IU-Net 97.4 98 97.5 96.8 97.6 -

Figure 5: Analysis on SECOND Data

and Kappa. Two data sets are selected for experi-

mentation: OSCD and SECOND. The different CD

techniques proposed by various experts are imple-

mented and analyzed on the chosen dataset. Among

EffCDNet, V-BANet, UCDNet, RSCDNet, and IU-

Net, RSCDNet demonstrated superior performance,

achieving an accuracy rate of 99.5%.

REFERENCES

Alshehhi, R. and Marpu, P. R. (2023). Change detection

using multi-scale convolutional feature maps of bi-

temporal satellite high-resolution images. European

Journal of Remote Sensing, 56(1):2161419.

Bao, Q. and Guo, P. (2004). Comparative studies on sim-

ilarity measures for remote sensing image retrieval.

In 2004 IEEE International Conference on Systems,

Man and Cybernetics (IEEE Cat. No. 04CH37583),

volume 1, pages 1112–1116. IEEE.

Barkur, R., Suresh, D., Lal, S., Reddy, C. S., Diwakar, P.,

et al. (2022). Rscdnet: A robust deep learning ar-

chitecture for change detection from bi-temporal high

resolution remote sensing images. IEEE Transactions

on Emerging Topics in Computational Intelligence,

7(2):537–551.

Basavaraju, K., Sravya, N., Lal, S., Nalini, J., Reddy, C. S.,

and Dell’Acqua, F. (2022). Ucdnet: A deep learning

model for urban change detection from bi-temporal

multispectral sentinel-2 satellite images. IEEE Trans-

actions on Geoscience and Remote Sensing, 60:1–10.

Caye Daudt, R., Le Saux, B., Boulch, A., and Gousseau, Y.

(2019). Oscd - onera satellite change detection.

Chughtai, A. H., Abbasi, H., and Karas, I. R. (2021). A

review on change detection method and accuracy as-

sessment for land use land cover. Remote Sensing Ap-

plications: Society and Environment, 22:100482.

Fang, H., Du, P., and Wang, X. (2022). A novel unsu-

pervised binary change detection method for vhr op-

tical remote sensing imagery over urban areas. In-

ternational Journal of Applied Earth Observation and

Geoinformation, 108:102749.

Fatemi Nasrabadi, S. B. (2019). Questions of concern in

drawing up a remote sensing change detection plan.

Journal of the Indian Society of Remote Sensing,

47(9):1455–1469.

Gomroki, M., Hasanlou, M., and Reinartz, P. (2022). Iunet-

ucd: Improved u-net with weighted binary cross-

entropy loss function for urban change detection of

sentinel-2 satellite images.

Kondmann, L., Toker, A., Saha, S., Sch

olkopf, B., Leal-

Taix

e, L., and Zhu, X. X. (2021). Spatial context

awareness for unsupervised change detection in opti-

cal satellite images. IEEE Transactions on Geoscience

and Remote Sensing, 60:1–15.

Patil, P. S., Holambe, R. S., and Waghmare, L. M. (2021).

Effcdnet: Transfer learning with deep attention net-

work for change detection in high spatial resolu-

tion satellite images. Digital Signal Processing,

118:103250.

Prasad, J., Sreelatha, M., and SuvarnaVani, K. (2023).

V-banet: Land cover change detection using effec-

tive deep learning technique. Ecological Informatics,

75:102019.

Qiu, L., Gao, L., Ding, Y., Li, Y., Lu, H., and Yu, W. (2013).

Change detection method using a new difference im-

age for remote sensing images. In 2013 IEEE Inter-

national Geoscience and Remote Sensing Symposium

- IGARSS, pages 4293–4296.

Comparative Examination of Different Change Detection Methods for Remote Sensing Imagery

349

Qu, Y., Li, J., Huang, X., and Wen, D. (2023). Td-sscd:

A novel network by fusing temporal and differential

information for self-supervised remote sensing image

change detection. IEEE Transactions on Geoscience

and Remote Sensing.

Raza, A., Huo, H., and Fang, T. (2022). Eunet-cd: Efﬁcient

unet++ for change detection of very high-resolution

remote sensing images. IEEE Geoscience and Remote

Sensing Letters, 19:1–5.

Yan, L., Yang, J., and Wang, J. (2023). Domain knowledge-

guided self-supervised change detection for remote

sensing images. IEEE Journal of Selected Topics

in Applied Earth Observations and Remote Sensing,

16:4167–4179.

Yang, K., Xia, G.-S., Liu, Z., Du, B., Yang, W., Pelillo, M.,

and Zhang, L. (2020). Semantic change detection with

asymmetric siamese networks.

INCOFT 2025 - International Conference on Futuristic Technology

350