Journal of Theoretical and Applied Information Technology

15th January 2025. Vol.103. No.1

© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

A NOVEL DEEP LEARNING METHOD FOR DETECTING

CHANGES IN SATELLITE IMAGERY
1
DIAA HAFEZ IBRAHIM, 2 REDA A. EL-KHORIBI, 3 FARID ALI MOUSA

1
Electrical and Electronics Communication Department, Faculty of Engineering, October University for
Modern Sciences and Arts (MSA University), Giza, Egypt
2
Information Technology Department, Faculty of Computers and Artificial Intelligence, Cairo University,
Giza, Egypt.
3
Information Technology Department, Faculty of Computers and Artificial Intelligence, Beni-Suef
University, Beni-Suef, Egypt

Email: 1 dhafez@[Link], 2 ralkhoribi1@[Link], 3 [Link]@[Link]

ABSTRACT

The identification of change information has been crucial to the use of satellite imagery, the monitoring of
land cover and use, the estimation of damage from natural catastrophes and the detection of military targets.
There are numerous conventional techniques for detecting changes in multispectral remote sensing images,
but they frequently fall short of our needs for durability, accuracy, and precision. This paper introduces a
novel deep learning method for identifying changes in satellite data, with an emphasis on environmental
changes and urban expansion. Using 24 pairs of Sentinel-2 satellite image data collected from 2015 to 2018,
of which 10 pairs were used for testing and 14 pairs for training. Thirteen spectral bands with different spatial
resolutions (10 m, 20 m, and 60 m) make up each multispectral image pair. The paper evaluates changes in
urban and rural environments using the visible spectrum bands (2, 3, and 4) at a resolution of 10 m. The
collection contains manually annotated changes.
The paper compares the results obtained from the proposed solutions by Siamese Network and U-Net to
address this change detection problem. With an accuracy of 0.86, the Siamese Network is used to detect high-
level structural changes between pre- and post-event images by learning similarities between paired images.
With an accuracy of 0.84, U-Net, which is intended for semantic segmentation, offers pixel-level predictions
that improve change detection detail while the hybrid method for pixel-level change detection that combines
the Siamese Network and U-Net in order to increase accuracy even further. This approach, which uses the
Siamese Network for patch-wise similarity comparison and U-Net for fine-grained pixel segmentation, yields
the maximum accuracy of 0.91. An efficient framework for applications in urban planning, crisis
management, and environmental monitoring is suggested by the suggested hybrid technique, which shows
great promise for accurate and thorough change detection in satellite imagery
Keywords Satellite Images, U-net, Siamese, Urban development, Change detection

1. INTRODUCTION Deep learning has been introduced for change

detection in remote sensing and has demonstrated
One of the classic problems in satellite imaging
good performance, driven by the afore mentioned
is change detection. Cloud detection, road and
observations.
building appearance, deforestation monitoring, and
Numerous reviews that concentrate on deep
agricultural crop monitoring are only a few examples
learning for data from remote sensing have recently
of the changes that need to be examined. [1]
been released. The deep learning methods used in all
Numerous applications make considerable use of
of the main sub-areas of remote sensing, such as
change detection techniques, such as urban change
classification, restoration, denoising, target
analysis [2], environmental monitoring [3], land
recognition, scene understanding, and other tasks,
management [4], and disaster assessment [5]. The
have been compiled in this research.
frequency of major climate change-related disasters,
including heat waves, storms, floods, and drought,
Satellite image processing poses a number of
has simultaneously exposed a new research issue and
unique issues, mostly because of the images' unusual
the need for more efficient automated change
features in comparison to traditional images.[6]
detection techniques.

96
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

First, a major constraint is frequently the size of specifically used in the agricultural industry for
the items that are detected. A football stadium, for disaster assessment, shifting cultivation monitoring,
instance, may take up hundreds of pixels in a and deforestation monitoring. It is currently used in
traditional photograph, enabling in-depth the military to gather data on battlefield location,
examination. A satellite image with a resolution of enemy military force movements, new military
10 meters per pixel, on the other hand, would only facilities, and damage assessments [8]. Change
use a small number of pixels to depict the same detection is employed in the civil field to regulate the
stadium, reducing spatial information and making growth of urban areas and the expansion of cities [8].
object detection and categorization more difficult. Additionally, it is used to track the effects of climate
Second, the spectral composition of satellite change, which are typically linked to rising
photographs is very different from that of greenhouse gas (GHG) emissions in the atmosphere.
conventional images. Satellite photos can have Examples of these effects include changes in sea
hundreds of spectral bands, but conventional level and glacier facies and mass balance.
photographs usually have three color channels (red, Traditional Techniques for Change Detection in
green, and blue). This high dimensionality makes Satellite Imagery
data processing more difficult and calls for Several fundamental methods have been
sophisticated feature extraction and analysis developed to detect changes in two satellite images
methods. captured at different times. These methods primarily
Third, further unpredictability is introduced by focus on analyzing pixel-level variations between
the sun's angle at the moment of image acquisition. spatially registered images and have been widely
When comparing two photographs, shadows cast by used in remote sensing applications.[9]
large objects like buildings and mountains can show 1. Image Differencing
up as abnormalities or alterations, especially when Image differencing involves subtracting the
utilizing simple image processing methods. This pixel values of two spatially registered images from
fluctuation complicates the detection of true changes different acquisition dates. The resulting difference
over time. image highlights regions where significant
Lastly, satellite imagery is greatly influenced by variations occur, indicating areas of change. One of
the ambient circumstances at the moment of the advantages of this method is its simplicity and
acquisition. Features in the image may be obscured speed, making it an efficient way to quickly identify
by elements like snow, fog, and clouds; therefore, areas that have undergone change. However, the
advanced preprocessing techniques are necessary to primary disadvantage is that it is highly sensitive to
counteract these effects and retrieve relevant noise and can struggle to differentiate between actual
information. changes and minor variations caused by atmospheric
These difficulties demonstrate the necessity of conditions or sensor discrepancies. Equation 1
sophisticated algorithms and reliable processing and describe image difference
analysis techniques for satellite images, especially 𝐼𝐷 = |𝐼𝑚𝑎𝑔𝑒 − 𝐼𝑚𝑎𝑔𝑒 | (1)
for applications like environmental monitoring, Such that 𝐼𝑚𝑎𝑔𝑒 is image 2 / Number of bands and
land-use classification, and disaster management. 𝐼𝑚𝑎𝑔𝑒 is image 1 / Number of bands
A fundamental technique in earth observation,
this procedure aims to differentiate between altered 2. Image Rationing
and unaltered pixels in bi-temporal or multi- Image rationing calculates the ratio of pixel
temporal remote sensing images taken from the same values from two registered images on a band-by-
geographic region or area but at various periods, band basis. In areas where change has occurred, the
respectively [7]. The primary goal of the change ratio values will significantly deviate from 1, either
detection system is to assign a binary label to each being higher or lower depending on the type of
pixel based on a pair or sequence of co-registered change. This method as show in equation 2 is
images. Accordingly, a null label indicates an area advantageous because it offers a simple and rapid
that has not changed, but a positive label indicates way to detect changes. However, it has been
that the area of that pixel has changed. In reality, criticized for producing non-normal histogram
change detection is an effective technique for multi- distributions, which can limit its effectiveness in
temporal analysis, urban mapping, and video detecting subtle changes and make interpretation
monitoring. more challenging, especially in areas with low
contrast or gradual transitions.
Numerous applications have effectively
employed change detection. Change detection is 𝐼𝑅 = | | (2)

97
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

3. Normalized Difference Built-Up Index the pixel values of each corresponding pair of
The NDBI technique involves calculating a images. The second image’s pixel values are then
Built-Up Index for each image date using the Near estimated using the regression function, and the
Infrared (NIR) and Shortwave Infrared (SWIR) difference between the regressed values and the
bands. Equation 3 describes the mathematical actual pixel values is used to detect changes as
foundation of NDBI. shown in equation 6.
𝑁𝐷𝐵𝐼 = (3) 𝐼𝑅 = 𝑎 𝐼 + 𝑏 (6)
Such that I is the first image, a is the slope and b in
The difference between these indices across the two
intercept.
dates helps highlight changes in urban environments,
One of the key advantages of this method is
particularly the expansion of built-up areas. One of
that it helps reduce the impact of atmospheric,
the major advantages of this method is that it
sensor, and environmental differences between the
emphasizes the spectral differences associated with
two images, leading to more reliable change
urban features, reducing the influence of topographic
detection. However, a major limitation is that it
effects and varying illumination. However, it can
requires the development of accurate regression
also enhance random or coherence noise, which can
functions for the selected spectral bands, which can
complicate the interpretation of changes and lead to
be time-consuming and complex, particularly when
false positives in non-urban areas.
the images come from different sensors or have
4. Average Intensity
different radiometric characteristics.
The AI method normalizes the images by
These traditional techniques have their strengths and
calculating the average intensity of pixel values
weaknesses, often facing challenges related to noise
between two images as shown in equation 4.
sensitivity, environmental variability, and sensor
𝐴𝐼 = (4) differences. While they offer straightforward and
It is particularly effective in identifying changes in computationally efficient methods for detecting
shaded areas where the average brightness is lower. changes, their limitations have driven the
The advantage of this method lies in its ability to development of more advanced machine learning
detect changes in regions with low average and deep learning models, which are capable of
brightness, such as areas under shadow or in dense overcoming many of these challenges by leveraging
vegetation. However, when environmental more sophisticated feature extraction and analysis
conditions vary significantly, this method requires techniques.[9]
additional operators, often based on statistical The structure of this publication is as follows:
models, to adapt and produce accurate results, Section 2 summarized earlier research. Section 3
making it more complex to implement under varying describes the proposed method. Section 4 discusses
conditions. the results and their implications. The conclusion is
5. Euclidean Distance ultimately discussed in section 5.
The Euclidean Distance method measures
the change by calculating the spatial distance 2. PREVIOUS WORK
between pixel values in a multi-dimensional spectral Pixel-based and object-based techniques
space as shown in equation 5. are the two categories of traditional change detection
techniques. Studies mostly concentrate on the
𝐸𝐷 = ∑ (𝐼𝑚𝑎𝑔𝑒 − 𝐼𝑚𝑎𝑔𝑒 ) (5)
analysis of difference images (di) using pixel-based
Larger distances indicate a higher likelihood of approaches find in [10] by Kelly,. et al.
change, while smaller distances suggest no change. Object correlation analysis is used in image
The advantage of this method is its ability to quantify
segmentation by Yuan et al. [11]. Information about
changes in a way that is straightforward to interpret.
However, the main disadvantage is that it may not the segmented objects is also used to detect changes.
perform well when there are subtle changes or when Wang et al. [12] segment multitemporal images
spectral overlaps exist between different features in using a modified seed-region growing technique.
the images, leading to ambiguity in the change Geostatistical characteristics are extracted via
detection results. multiresolution segmentation in [12].
These approaches require distinct processes
6. Image Regression
including feature extraction and classification, and
Image regression establishes relationships segmentation is one of several steps.
between two images by fitting a regression model to

98
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

Deng et al. [13] propose an architecture that Convolutional Neural Networks (CNNs), which
combines a Mask R-CNN network for object-based predicted the label of the central pixel in each image
instance segmentation with a multi-scale fully patch. However, this approach overlooked the
convolutional network for pixel-based semantic context surrounding the central pixel, which is
segmentation. This hybrid approach leverages the crucial for detecting small objects. In a subsequent
strengths of both networks to improve the accuracy study, Daudt et al. [18] proposed a fully
of both instance-level and semantic-level change convolutional encoder-decoder network, where the
detection tasks. encoder utilized a Siamese network, and for the first
The field of remote sensing image time, skip connections were integrated within the
processing has become interested in deep learning in Siamese network to improve feature sharing and
recent years. Numerous studies on change detection context preservation.
based on deep learning have been conducted. For An encoder and a decoder make up an encoder-
change detection, Arosio et al. [14] use a fully linked decoder model. The decoder retrieves the details and
deep neural network. One pixel's neighbourhood produces the segmentation map after the encoder
records high-level information along the spatial
serves as input, and the pixel's classification
dimension of feature maps. Deconvolution is used as
outcome serves as output. A bipartite differential the high-level feature recovery in. [19]. Using skip
neural network is proposed by Zhao et al. [15], connections, U-Net [20] connects the encoder
wherein the input image pair is superimposed with features to the same-level decoder output.
two change disguise maps (cdm) that identify the This part offers a thorough analysis of both
altered regions. The network output is used to build conventional and contemporary change detection
an objective function that is optimized. Small-scale methodologies, clearly showing the transition from
object-based and pixel-based approaches to deep
datasets are used to train these models.
learning-driven strategies. While the inclusion of
However, our model uses a Siamese hybrid approaches, such as the integration of Mask
architecture, where the two pictures are passed R-CNN and fully convolutional networks,
through in turn, with the bottom levels sharing demonstrates the potential of combining
weights. The procedure of learning similar features methodologies for improved performance, the
through shared weights makes sense because the two thorough discussion of traditional methods, such as
photos were taken in the same scene at separate segmentation and feature extraction, highlights their
complexity and limitations. Additionally, the
times. A Siamese convolutional network for change
examination of deep learning methods, specifically
detection is suggested in [16]. The model of [16] the use of encoder-decoder frameworks like U-Net
uses a straightforward thresholding segmentation and Siamese architectures, shows a notable
based on Euclidean distance that is distinct from the advancement in the use of contextual and spatial
network while combining the data acquired by the information for increased accuracy. The insights in
Siamese CNN. We create deeper modules in our this section could be strengthened, nevertheless, by
a more thorough evaluation of these methods'
model to improve information fusion and
computational effectiveness, scalability, and
segmentation. Due to its powerful ability to leverage performance on other datasets.
contextual information, the fully convolutional
network [17] is the most widely used convolutional 3. DATA SET AND PROPOSED MODEL
model for semantic image segmentation. 3.1 Dataset
Change detection tasks involve generating The Open and Standardized Change
a change map from a pair of input images and Detection (OSCD) dataset was created to provide a
trustworthy standard by which to evaluate different
assigning pixel-wise semantic change labels. In
change detection techniques. The dataset, which was
recent years, numerous deep learning-based methods created with an emphasis on urban regions, solely
have been proposed for change detection in satellite designates structural changes and urban growth as
imagery. These methods are primarily derived from "Change," ignoring natural oscillations like tide
classification or semantic segmentation techniques. shifts or vegetation growth. Researchers and
Daudt et al. [18] were among the first to introduce an practitioners interested in change detection can use
end-to-end framework for change detection using this dataset, which supports the development and
assessment of single-band, color, and multispectral

99
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

change detection algorithms by providing pixel-level algorithm performance. When using the OSCD
change labels for supervised learning models. dataset for change detection investigations, several
Sentinel-2 satellite photos, which include aspects should be taken into account.
wavelengths from ultraviolet to short-wave infrared, 3.2 Proposed Model
and 13 spectral bands at resolutions ranging from 10 Change detection in satellite imagery is
m to 60 m, are used in the OSCD collection. To critical for understanding urban dynamics,
include locations with discernible urbanization monitoring natural disasters, and managing
shifts, twenty-four regions from around the world— resources effectively. the paper explores and
each measuring roughly 600 by 600 pixels at a evaluates three distinct methodologies—Siamese
resolution of 10 m—were chosen. The Medusa network, u-net, and a hybrid approach combining
toolkit is used to trim each image region's 26 images both—for identifying structural changes in urban
(13 bands for each pair) to precise coordinates.[21] environments using the open and standardized
Challenges and Limitations change detection (OSCD) dataset. these methods
The OSCD dataset has a number of leverage the rich spectral and spatial information
significant drawbacks, despite being a useful tool for provided by sentinel-2 multispectral satellite
applying supervised learning techniques and imagery to achieve pixel-level accuracy in change
methodically evaluating change detection systems. detection tasks.
First, the Sentinel-2 satellite can record important The following sections detail the
structural changes, such as the appearance of experimental workflows for each methodology,
massive buildings, even though its photos have a emphasizing the preprocessing techniques, network
relatively poor quality. Finer alterations, including architectures, and evaluation strategies employed.
adding lanes to roads, extending old structures By comparing these approaches, this study aims to
slightly, or adding new buildings, are frequently identify the most effective method for urban change
harder to spot. As a result, analysts' subjective detection in high-resolution satellite imagery.
interpretations of these minute changes may cause 3.2.1 Change Detection Using a Siamese Network
differences in manually created change maps. A multi-step process of data pre-processing,
Using OpenStreetMap data from various network architecture design, and feature comparison
dates was one method that was used to automate the is used in the methodology for applying a Siamese
creation of change maps. This technique sought to Network to change detection in the OSCD (Open
identify changes in an organized, automated manner and Standardized Change Detection) dataset. This
by comparing map data that corresponded to the approach is optimized for identifying changes in
dates of each image pair. However, a number of urban structure. Sentinel-2 multispectral satellite
drawbacks made it ineffectual. First off, a lot of map images covering wavelengths from the visible to
modifications indicated additions that weren't short-wave infrared spectrum are included in the
necessary constructed between the dates of image OSCD dataset. The bands were taken at different
acquisition, which led to erroneous temporal spatial resolutions (10 m, 20 m, and 60 m). All
alignment. Furthermore, the accuracy of dating photos are resampled to a standard 10 m resolution
earlier OpenStreetMap data was limited because, in order to guarantee spatial consistency between
until 2017, maps frequently only offered one version these bands, enabling consistent pixel alignment
annually, making accurate temporal matching across channels. Pixel-by-pixel comparisons inside
difficult. the network are made possible by this pre-processing
Additionally, the collection is limited to phase, which also guarantees compatibility across
photos taken starting in June 2015 because the spectral bands. Critical spectral information for
Sentinel-2 spacecraft was launched in 2015. This urban characteristics is then captured by structuring
restricts the range of discernible changes by limiting the resampled images into multi-channel inputs
the dataset's temporal scope to roughly two and a using certain band combinations, such as Red,
half years or fewer for many image pairs. Green, Blue (RGB), and Near-Infrared (NIR). Given
Furthermore, there are more "no change" pixels in the dataset's emphasis on urban evolution, this band
the dataset than "change" pixels, which results in an selection prioritizes structural components while
unbalanced label distribution that could impair reducing vulnerability to natural variability.

100
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

Figure 1. Change detection using Siamese Network

As shown in Figure 1. The application of a capturing both spatial and spectral characteristics.
Siamese Network for change detection involves a The shared weights ensure that both images are
systematic process divided into three key phases: processed identically, enabling meaningful feature
data preprocessing, network encoding and comparisons. Following the encoding process, a
decoding, and feature vector comparison for decoding stage refines the extracted feature vectors,
change detection. This methodology ensures reconstructing contextual information while
accurate identification of structural changes between preserving spatial relationships, which is essential
pre-event and post-event satellite images. for precise analysis.
In the data preprocessing phase, paired In the final feature vector comparison
pre-event and post-event images undergo several phase, the pre-event and post-event feature vectors
preparatory steps to enhance their quality and generated by the Siamese Network are compared
maintain consistency. First, all images are resized to using a distance metric, such as the Euclidean
a uniform resolution to ensure spatial alignment distance. This metric quantifies the dissimilarity
between the paired inputs. This alignment is crucial between the feature vectors, reflecting the extent of
for pixel-to-pixel comparisons during the subsequent changes between the two images. A high Euclidean
stages. Next, noise removal techniques are applied to distance indicates significant differences between
mitigate distortions caused by environmental factors the feature representations, suggesting the presence
such as clouds, shadows, or atmospheric of structural changes, while a low distance implies
interference, thereby enhancing the clarity of minimal or no changes. Based on the similarity
structural details. Additionally, image enhancement scores, a binary classification is performed for each
methods, such as contrast stretching and histogram image patch, identifying it as either "changed" or
equalization, improve the visibility of key features "unchanged." The results are aggregated to produce
critical for detecting urban changes. To further a change map that visually highlights the regions
enhance the robustness of the model, data where changes have occurred.
augmentation techniques, including rotations, flips, This structured approach leverages the
and scaling, are applied. These augmentations computational efficiency and feature comparison
expand the diversity of the dataset and improve the capabilities of the Siamese Network, making it well-
model's ability to generalize to unseen data, reducing suited for change detection tasks in urban
overfitting risks. environments. By focusing on feature-level
The network encoding and decoding differences, the methodology enables precise
phase is the core of the Siamese Network identification of structural changes while
architecture. The network consists of two identical minimizing processing demands.
sub-networks (encoders) with shared weights that The Siamese Network uses the Euclidian
independently process the pre-event and post-event distance metric to compare feature vectors from the
input images. Each sub-network, typically a pre- and post-event patches. By quantifying the
convolutional neural network (CNN), extracts high- degree of change in feature space, this distance
level feature representations of the input patches, metric—which is the absolute difference between

101
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

the feature vectors—allows the network to gauge Using a Siamese Network to efficiently
how much has changed between patches. The record and compare structural changes in high-
network can distinguish between stable and changed resolution satellite pictures, this methodology offers
regions by using the Euclidian distance to identify a thorough approach to urban change detection. The
both significant and subtle structural changes. After model is a useful tool for urban monitoring and
passing the resultant distance metric to a fully analysis jobs since it produces dependable findings
connected layer, a softmax classifier produces a through meticulous data pre-processing, feature
binary classification of "Change" or "No Change." extraction, and post-processing.
This layer directly addresses the goal of detecting
3.2.2 Change Detection Using U-Net
structural changes by allowing the network to link
variations in feature vectors to a change detection In order to accurately detect structural
decision. changes in urban areas using high-resolution satellite
Binary cross-entropy loss, a loss function imagery, the methodology for implementing U-Net
that works well for binary classification tasks— for change detection on the OSCD dataset consists
where the goal is to reduce the error between the true of a meticulously planned workflow that includes
and predicted labels—is used to train the model. The data pre-processing, U-Net architecture setup, and
network's prediction is specifically compared post-processing. Multispectral Sentinel-2 photos
against ground truth labels, which are given as with resolutions ranging from 10 m to 60 m over 13
binary indications of change in the OSCD dataset, spectral bands, covering a range of wavelengths
using the cross-entropy loss. The Siamese Network from visible to short-wave infrared, make up the
may learn discriminative characteristics that OSCD collection. All photos are resampled to a
differentiate natural variations, like seasonal standard resolution of 10 m in order to preserve
changes in flora, from urban alterations by training spatial alignment and uniformity across bands. In
on labeled patch pairings. While maintaining order to enable precise pixel-level comparison
insensitivity to incidental environmental variables, during segmentation, these resampling guarantees
this training procedure guarantees that the network consistency across spectral channels.
gains a solid comprehension of structural changes. As shown in Figure 2. The implementation
The Siamese Network creates change maps of U-Net for change detection follows a structured
for fresh pre- and post-event picture pairs following workflow divided into three main phases: data
training. The resulting change maps are subjected to preprocessing, localization using U-Net, and
Gaussian smoothing techniques and morphological change detection. This approach leverages the
filters (such as closing and opening operations) in segmentation capabilities of U-Net to achieve
order to improve these predictions and lower noise. accurate pixel-level identification of structural
By lowering artifacts and improving interpretability changes in satellite imagery.
for additional analysis, these post-processing
procedures raise the change maps' visual quality and
accuracy.

Figure 2. Change detection using U-Net

102
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

In the data preprocessing phase, the input convolutional and max-pooling layers in the
images undergo several preparatory steps to ensure contracting route, each of which captures
uniformity and enhance data quality. All images are progressively abstract information. The compressed
resized to a standard resolution to maintain spatial feature maps are then upsampled using the expansive
consistency across the dataset, facilitating pixel- path, which uses skip connections from matching
level alignment. Noise removal techniques are layers in the contracting path to recover fine-grained
applied to address potential distortions caused by features and spatial resolution. Because it allows U-
environmental conditions such as clouds, shadows, Net to preserve contextual information while
or atmospheric interference, ensuring the visibility generating high-resolution change maps, this
of key structural details. Additionally, image symmetrical design is crucial for change detection.
enhancement methods, such as contrast adjustment For training, the model takes two multi-
and histogram equalization, improve the clarity and channel patches—representing the same region from
contrast of the images, making features more pre-event and post-event times—and stacks them as
discernible. Data augmentation techniques, input. This input format allows the network to
including random rotations, flips, and scaling, are analyze temporal differences within each patch and
employed to increase the diversity of the training learn features indicative of structural changes.
dataset, enhancing the model's generalizability and During the upsampling phase, U-Net merges these
robustness against overfitting. temporally derived features, producing a pixel-wise
The localization phase utilizes the U-Net probability map where each pixel represents the
architecture to generate a preliminary segmentation likelihood of change or no change. The output layer
map that highlights potential regions of change. The is a softmax or sigmoid activation, depending on the
U-Net model, based on an encoder-decoder configuration, to generate binary labels (change/no
structure, localizes changes by capturing both high- change) for each pixel, enabling the model to
level contextual information and fine-grained spatial produce highly localized change maps.
details. During the training process, U-Net generates The model is optimized using binary cross-
standard weights optimized for the specific dataset, entropy loss, as this loss function is well-suited for
enabling consistent and reliable feature extraction. binary pixel-wise segmentation tasks. During
These weights are subsequently used to process the training, each output pixel is compared with a
input data, localizing changes by segmenting areas labeled ground truth from the OSCD dataset, with
that exhibit notable structural variations. The the loss function minimizing discrepancies between
model’s ability to focus on relevant features ensures predicted and actual change labels. Training on
accurate delineation of regions of interest. labeled patch pairs from pre- and post-event times
In the detection phase, the results from the allows U-Net to develop a nuanced understanding of
U-Net localization are analyzed to confirm and change patterns, capturing urban growth and
refine the identified changes. The segmented output structural transformations while ignoring natural
is post-processed to eliminate false positives and variations like vegetation changes.
noise, ensuring that only genuine changes are This approach is quite successful in
retained. This step includes thresholding techniques detecting structural changes in urban areas because
and morphological operations to refine the it makes use of U-Net's segmentation strengths for
boundaries of detected regions. The final output is a accurate, pixel-level change identification. This
detailed change map that highlights structural method produces precise and comprehensive change
differences between pre-event and post-event maps by combining meticulously planned pre-
images at a pixel level. processing, U-Net architecture, and post-processing,
This U-Net-based methodology combines offering important insights for jobs involving urban
the advantages of robust preprocessing, precise monitoring and spatial analysis.
localization, and refined detection to deliver highly
3.2.3 Hybrid Approach
accurate change detection results. Its ability to
capture spatial and contextual features makes it In order to improve accuracy, the hybrid
particularly effective for identifying subtle structural method that combines a Siamese Network with U-
changes in urban and natural environments. Net for pixel-level change detection makes use of
Because the U-Net architecture combines both fine-grained pixel-level segmentation and
an expansive path for accurate localization and a patch-wise similarity comparison. By using the
contracted path for feature extraction, it is especially Siamese Network to pinpoint areas where changes
well-suited for pixel-level segmentation. The input are most likely to have taken place, this technique
patch is downsampled via a sequence of lowers processing demands by letting the U-Net

103
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

concentrate just on these potential areas for in-depth environments while lowering sensitivity to non-
segmentation. This method is computationally structural variables like vegetation.
efficient and offers more accurate change detection
by integrating the advantages of both systems. As shown in Figure 3. pairs of patches taken
from pre-event and post-event photos are compared
To guarantee consistency and relevancy of for patch-wise similarity using the Siamese Network
inputs, pre-processing procedures are carried out on stage. High-level feature representations are
the Sentinel-2 photos in the first step of this produced by each patch pair passing through
workflow. In order to facilitate pixel-to-pixel identical sub-networks (such as CNN-based
alignment between spectral channels, all images are architectures with shared weights). The network
resampled to a single resolution because Sentinel-2 calculates a similarity score for every pair of patches
records multispectral data at several resolutions (10 using a distance metric, like L1 or Euclidean
m, 20 m, and 60 m). Each satellite image is distance, which represents the extent of change
resampled and then split into tiny patches, usually between the two images. A coarse change map is
measuring 64 by 64 pixels, which serve as the produced by designating patches with high similarity
Siamese Network's input. Red, Green, Blue (RGB), scores as potential change zones. This map provides
and Near-Infrared (NIR) are examples of relevant a more targeted and computationally efficient input
spectral bands that are chosen as input channels for further research by highlighting areas with
because they offer crucial information for possible changes and removing stable regions.
identifying structural changes in metropolitan

Figure 3. Proposed Method For Change Detection In Satellite Imagery

U-Net is used to fine-tune these regions at map that designates every pixel as either "Change"
the pixel level after the initial candidate or "No Change." This approach allows U-Net to
identification. By omitting sizable regions with identify small changes with pixel-level precision by
negligible alterations, the Siamese Network's preserving spatial information and capturing context
candidate regions are the only ones sent to the U-Net, through skip links between the encoder and decoder
lowering the computational load. U-Net learns the layers.
precise variations between pixels over time by Post-processing is the last stage of the
receiving a stacked input of pre-event and post-event workflow, which improves the output change map's
patches for each candidate region. With its encoder- quality and usefulness. By filling in the gaps and
decoder structure, the U-Net architecture carries out eliminating tiny noisy areas, morphological filtering
thorough segmentation, generating a binary change techniques like erosion and dilation lower the

104
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

frequency of false positives and enhance continuity Table 1: Experiments Results For OSCD Dataset
in the segmented sections. A smooth view of the
Method Accuracy
identified changes throughout the entire image is
then provided by the reconstruction of the complete
change map using the segmented patches. Siamese Network
0.86
The hybrid approach offers several
U-Net 0.84
advantages. By narrowing down candidate regions,
the Siamese Network minimizes the area that U-Net
needs to process, reducing computational load. The Proposed Method 0.91
U-Net, in turn, provides high-resolution
segmentation, allowing the method to handle The Siamese Network proves highly
complex changes and detect small, localized effective for identifying changes by leveraging its
alterations that may not be apparent with patch-wise capability to compare pre-event and post-event
comparison alone. Post-processing filters also make image patches through feature similarity. Its
this approach robust to noise, resulting in more architecture, which uses paired inputs and shared
reliable and accurate change maps. weights, ensures robust feature extraction and direct
comparison of pixel-level changes. This approach is
particularly advantageous in scenarios where
The architecture of this hybrid method
localized changes must be identified efficiently, as
leverages two key components: the Siamese
the Siamese Network minimizes computational
Network, with twin convolutional sub-networks for
overhead by focusing on pairwise feature
patch-wise similarity assessment, and U-Net, an
comparison. However, while it excels in detecting
encoder-decoder model that refines candidate
coarse changes, its performance may be limited in
regions with precise segmentation. Post-processing
capturing complex spatial details due to its focus on
steps further enhance the final change map by
patch-level similarity rather than detailed
filtering out noise, thereby improving overall
segmentation.
accuracy and reliability.
The U-Net, on the other hand, demonstrates
With this combination, the hybrid Siamese-U- its strength in pixel-level segmentation, making it
Net model offers a well-rounded solution for highly suitable for capturing fine-grained spatial
change detection in high-resolution satellite details and subtle structural variations in satellite
imagery, attaining great accuracy and images. Its encoder-decoder architecture efficiently
extracts high-level features and refines them into
computational efficiency.
detailed segmentation maps, enabling precise
delineation of change regions. Despite its ability to
4. EXPERIMENTAL RESULTS localize changes accurately, U-Net may encounter
challenges in handling noise or ambiguous features
The experimental results demonstrate the in large-scale satellite datasets, which could slightly
performance of three methodologies—Siamese impact its overall accuracy compared to Siamese
Network, U-Net, and the proposed hybrid model Network.
combining these two approaches—for satellite
image change detection. The accuracy values
achieved by each method highlight in Table 1. their By integrating the strengths of both
individual strengths and the advantages of their approaches, the proposed hybrid model achieves
integration in the hybrid model. Specifically, the superior accuracy, leveraging the Siamese Network
Siamese Network achieved an accuracy of 0.86, the for effective localization and U-Net for detailed
U-Net recorded 0.84, and the proposed hybrid model segmentation. In this hybrid setup, the Siamese
outperformed both with an accuracy of 0.91. Network first identifies regions likely to have
undergone change, reducing computational demands
by narrowing the focus to relevant areas.
Subsequently, the U-Net processes these regions
with its segmentation capabilities, capturing intricate
spatial details and producing a refined change map.
This synergy between the two models mitigates their

105
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

individual limitations, allowing the hybrid approach REFERENCES

to achieve the highest accuracy of 0.91.
[1]. Ma, Y., Chen, S., Ermon, S., & Lobell, D. B.
(2024). Transfer learning in environmental
The results underline the benefits of remote sensing. Remote Sensing of
combining complementary deep learning Environment, 301, 113924.
architectures to enhance change detection in satellite [2]. Stilla, U., & Xu, Y. (2023). Change detection of
imagery. The hybrid model's superior performance urban objects using 3D point clouds: A
demonstrates its potential as an effective and review. ISPRS Journal of Photogrammetry and
scalable solution for monitoring urban dynamics, Remote Sensing, 197, 228-255.
environmental changes, and disaster impacts. [3]. Yuan, S., Li, Y., Bao, F., Xu, H., Yang, Y., Yan,
5. CONCLUSIONS Q., ... & Lin, J. (2023). Marine environmental
monitoring with unmanned vehicle platforms:
This study demonstrates the efficacy of Present applications and future
deep learning models—Siamese Network, U-Net, prospects. Science of The Total
and a hybrid approach combining the two—for
Environment, 858, 159741.
change detection in satellite imagery, using the
[4]. AbdelRahman, M. A. (2023). An overview of
OSCD dataset as a benchmark. The experimental
results reveal that while the Siamese Network excels land degradation, desertification and sustainable
in efficiently identifying localized changes with an land management using GIS and remote sensing
accuracy of 0.86 and U-Net delivers robust pixel- applications. Rendiconti Lincei. Scienze Fisiche
level segmentation achieving an accuracy of 0.84, e Naturali, 34(3), 767-808.
the proposed hybrid model outperforms both with an [5]. Mansour, R. F., & Alabdulkreem, E. (2023).
accuracy of 0.91. Disaster Monitoring of Satellite Image
The analysis highlights the complementary Processing Using Progressive Image
strengths of the two models: the Siamese Network’s Classification. Computer Systems Science &
ability to localize change areas through pairwise Engineering, 44(2).
feature comparisons and U-Net’s capability to [6]. Guerri, M. F., Distante, C., Spagnolo, P.,
capture intricate spatial details through its encoder- Bougourzi, F., & Taleb-Ahmed, A. (2024).
decoder architecture. The hybrid model effectively Deep learning techniques for hyperspectral
integrates these strengths, enabling precise change image analysis in agriculture: A review. ISPRS
localization followed by refined segmentation, Open Journal of Photogrammetry and Remote
which addresses the limitations of each individual Sensing, 100062.
method. [7]. Dimitrovski, I., Kitanovski, I., Kocev, D., &
Simidjievski, N. (2023). Current trends in deep
learning for Earth Observation: An open-source
The proposed hybrid approach demonstrates
benchmark arena for image
significant potential for enhancing urban
classification. ISPRS Journal of
monitoring, disaster assessment, and environmental
change detection. Its ability to provide detailed and Photogrammetry and Remote Sensing, 197, 18-
accurate change maps makes it a valuable tool for 35.
practical applications in remote sensing and [8]. Zhu, Q., Guo, X., Li, Z., & Li, D. (2024). A
geospatial analysis. Future work could extend this review of multi-class change detection for
research by addressing dataset limitations, such as satellite remote sensing imagery. Geo-spatial
label imbalance and temporal constraints, and Information Science, 27(1), 1-15.
exploring the integration of additional spectral bands [9]. Cheng, G., Huang, Y., Li, X., Lyu, S., Xu, Z.,
or spatial contexts to further improve model Zhao, H., ... & Xiang, S. (2024). Change
performance. This study underscores the importance detection methods for remote sensing in the last
of leveraging complementary deep learning decade: A comprehensive review. Remote
architectures to advance the field of satellite image Sensing, 16(13), 2355.
change detection.
[10]. Blaschke, T., Kelly, M., & Merschdorf, H.
(2024). 8 Object-Based Image
Analysis. Remote Sensing Handbook, Volume

106
 Journal of Theoretical and Applied Information Technology
15th January 2025. Vol.103. No.1
© Little Lion Scientific

ISSN: 1992-8645 [Link] E-ISSN: 1817-3195

II: Image Processing, Change Detection, GIS, [19]. Gomroki, M., Hasanlou, M., & Chanussot,
and Spatial Data Analysis, 145. J. (2023). Automatic 3D Multiple Building
[11]. Yuan, M., Xia, Y., Dong, H., Chen, Z., Change Detection Model based on Encoder-
Yao, J., Qiu, M., ... & Zhang, L. (2023). Devil Decoder Network using Highly Unbalanced
is in the queries: advancing mask transformers Remote Sensing Datasets. IEEE Journal of
for real-world medical image segmentation and Selected Topics in Applied Earth Observations
out-of-distribution localization. In Proceedings and Remote Sensing.
of the IEEE/CVF Conference on Computer [20]. Vasavi, S., Somagani, H. S., & Sai, Y.
Vision and Pattern Recognition (pp. 23879- (2023). Classification of buildings from VHR
23889). satellite images using ensemble of U-Net and
[12]. Yu, Y., Wang, C., Fu, Q., Kou, R., Huang, ResNet. The Egyptian Journal of Remote
F., Yang, B., ... & Gao, M. (2023). Techniques Sensing and Space Sciences, 26(4), 937-953.
and challenges of image segmentation: A [21]. Daudt, R. C., Le Saux, B., Boulch, A., &
review. Electronics, 12(5), 1199. Gousseau, Y. (2018, July). Urban change
[13]. Deng, R., Zhou, M., Huang, Y., & Tu, W. detection for multispectral earth observation
(2023). Hybrid Task Cascade-Based Building using convolutional neural networks.
Extraction Method in Remote Sensing In IGARSS 2018-2018 IEEE International
Imagery. Remote Sensing, 15(20), 4907. Geoscience and Remote Sensing
[14]. Arosio, R., Hobley, B., Wheeler, A. J., Symposium (pp. 2115-2118). IEEE.
Sacchetti, F., Conti, L. A., Furey, T., & Lim, A.
(2023). Fully convolutional neural networks
applied to large-scale marine morphology
mapping. Frontiers in Marine Science, 10,
1228867.
[15]. Zhao, X., Zhang, K., Zhang, F., Sun, J.,
Wan, W., & Zhang, H. (2024). Multi-View
Hypergraph Fusion Network for Change
Detection in High-Resolution Remote Sensing
Images. IEEE Journal of Selected Topics in
Applied Earth Observations and Remote
Sensing.
[16]. Long, X., Zhuang, W., Xia, M., Hu, K., &
Lin, H. (2023). SASiamNet: Self-adaptive
Siamese Network for change detection of
remote sensing image. IEEE Journal of Selected
Topics in Applied Earth Observations and
Remote Sensing.
[17]. Yin, M., Chen, Z., & Zhang, C. (2023). A
CNN-Transformer Network Combining CBAM
for Change Detection in High-Resolution
Remote Sensing Images. Remote
Sensing, 15(9), 2406.
[18]. Daudt, R. C., Le Saux, B., Boulch, A., &
Gousseau, Y. (2018, July). Urban change
detection for multispectral earth observation
using convolutional neural networks.
In IGARSS 2018-2018 IEEE International
Geoscience and Remote Sensing
Symposium (pp. 2115-2118). IEEE.

107