Microsoft Word - ETASR_V12_N5_pp9351-9356
Engineering, Technology & Applied Science Research Vol. 12, No. 5, 2022, 9351-9356 9351
www.etasr.com Slimani & Hedjam: A Hybrid Metaheuristic and Deep Learning Approach for Change Detection in …
A Hybrid Metaheuristic and Deep Learning Approach
for Change Detection in Remote Sensing Data
Yacine Slimani
Department of Computer Science
Laboratory of Intelligent Systems
University of Ferhat Abbas Setif 1
Setif, Algeria
slimaniy09@univ-setif.dz
Rachid Hedjam
Department of Computer Science
Sultan Qaboos University
Muscat, Oman
rachid.hedjam@squ.edu.om
Received: 6 August 2022 | Revised: 17 August 2022 | Accepted: 20 August 2022
Abstract-This study aimed to adapt Convolutional Neural
Networks (CNN) to solve the problem of change detection using
remote sensing imagery. Specifically, the goal was to investigate
the impact of each CNN layer to detect changes between two
satellite images acquired on two different dates. As low-level
CNN layers detect fine details (small changes) and higher-level
layers detect coarse details (large changes), the idea was to assign
a weight to each layer and use a genetic algorithm based on a
training dataset to generalize the detection process on the test
dataset. The results showed the effectiveness of the proposed
approach based on two real-life datasets.
Keywords-change detection; remote sensing; deep learning;
convolutional neural networks; genetic algorithms
I. INTRODUCTION
The aim of binary Change Detection (CD) in remote
sensing is to compare two images acquired at two different
dates to detect meaningful differences [1]. Usually, two CD
approaches are used: supervised or unsupervised. Supervised
CD requires temporal reference data for the training phase [2,
3], while unsupervised CD is based on a direct comparison of
input images without using labeled data [4-6]. In general,
unsupervised binary CD techniques consist of two steps: i)
compute the difference between the features of the two input
images to generate a difference image or change index, and ii)
generate the binary change map by segmenting (thresholding)
the difference image computed in the first step into change and
no-change regions. However, traditional CD methods that use
handcrafted features are not effective in complex situations,
because the designed features cannot accurately capture high-
and medium-level image representations [7]. Recently, deep
learning has emerged and has become a state-of-the-art
approach for CD. Deep learning is very effective in extracting
representative features from low, middle, and higher image
representation levels. The advantages of deep learning are that,
it learns discriminant features and computes them
automatically without relying on the involvement of an expert.
Several deep-learning methods for CD have been proposed.
In [8], a method was proposed to compute the difference image
using a backpropagation algorithm and a deep belief network.
A deep belief network learns low and high-level features
around a pixel neighborhood and the backpropagation
algorithm builds the difference image using training samples.
Finally, a simple segmentation algorithm was used to compute
the binary change map. A method was proposed in [9] that
combined deep features, saliency detection, and Convolutional
Neural Networks (CNNs) to compute the change. A patch-
based Siamese Neural Network was presented in [10], where
external images, whose textures resembled the changing area,
were used to generate genuine and imposter pair samples for
the training process. A method that combined CNN features to
create a single higher feature vector was proposed in [7], using
the pixel-wise Euclidean distance between the extracted feature
vectors after having been transformed into matrices to compute
the change map. A review of the deep learning-based CD
methods can be found in [11]
This study extends [12] and is related to [7]. The difference
between this study and [7] lies in the combination process of
the CNN layers. In [7], all CNN layers were combined into a
single feature vector, but this study proposes the assignment of
weights to the CNN layers before combining them. In [12], the
weights were binary (i.e. 1 for considering a layer and 0 for
not) and assigned manually. This study used a Genetic
Algorithm (GA) to automatically learn the weights based on
training data. The GA aims to find the best weights that lead to
the best match between the CD reference (ground truth) and the
change map detected by the proposed algorithm. Therefore, the
weight vector can be seen as a mask, where the goal is to
demonstrate that the layers can be assigned different weights
before being combined to detect different regions that represent
the changes between the two input images. The assumption
was that, to detect large changes, high-level layers are assigned
higher weights, i.e. considered more.
II. THE PROPOSED APPROACH
Usually, the spatial changes in remote sensing images are
specific patterns with special features in terms of color, shape,
and texture. Therefore, their CNN characteristics are different
from those of the same location in the image before the change.
In other terms, an unchanged spatial area between the two
Corresponding author: Yacine Slimani
Engineering, Technology & Applied Science Research Vol. 12, No. 5, 2022, 9351-9356 9352
www.etasr.com Slimani & Hedjam: A Hybrid Metaheuristic and Deep Learning Approach for Change Detection in …
inputs should have almost similar CNN features, whereas the
changed areas have different. Thus, it seems reasonable to
compute the features of the two input images using the same
CNN structures and then compute their difference to generate
the difference image, where the brighter pixels represent the
changed areas due to the larger difference values. To detect the
changed areas, the difference image can be segmented by a
thresholding method into "changed" and "unchanged" classes.
In a CNN, the lower layers capture low-lever image features
such as edges, color, and gradient orientation, while the mid-
high level layers capture coarse patterns of the images that can
represent whole objects in the images [13]. Since the changes
are almost a random natural process, they may affect areas with
different sizes from fine to coarse. Therefore, detecting changes
between the two input images based on the difference between
their last CNN feature is not effective. To overcome this
limitation, a new change detection method was proposed,
which is an improvement over [7]. The procedure obeys the
following: i) extract low, mid and higher features from each
image using a pre-trained CNN (e.g. VGG19 [14]), ii) resize
the layers to the same size and combine them into a single
feature vector for each image, iii) reshape the two feature
vectors into square matrices, and finally iv) compute the
Euclidean distance between the two matrices to generate the
difference image, which will be segmented into two classes,
"changed" and "unchanged".
In [12], a binary weight was assigned manually to each
layer to include it or not in the combination process of the
layers. In other terms, the binary vector of weights played a
mask role that allowed or prevented some layers from the
combination process. If the weight of a given layer was equal
to one (1), it meant that this layer was included in the
combination process, while if it was equal to zero (0) it was
not. Assigning weights to layers is application dependent. In
practice, it is best to consider high-level layers when detecting
large changes, while low-level layers should be taken into
account when detecting minor changes.
Fig. 1. Convolutional feature-based change detection with VGG19.
In the training phase, N image tuples were used {(Im1; Im2;
Tr)}i, i=1…N, where Im1 and Im2 are the images before and
after a change and Tr is the corresponding ground truth. Each
image is divided into M patches of size d×d. In other terms,
there is a set of S=N×M patches before the change, i.e. {Pj1,
j=1...S}, a set of patches after the change, i.e. {Pj2, j=1...S}, and
the same number of ground truth patches. Each patch is fed to a
VGG19 CNN to extract 5 feature maps from 5 different layers.
This study used the 3
rd
, 6
th
, 10
th
, 14
th
, and 18
th
layers of the
VGG19. From each input patch, 5 layers (feature maps) were
extracted and resized to the same size.
Formally, let [Xj11, Xj12, Xj13, Xj14, Xj15] be the list of the
feature maps extracted from Pj1, and [Xj21, X j22, X j23, X j24, X j25]
be the list of the feature maps extracted from Pj2. Thus, the
corresponding weighted feature maps are:
��� � ��� � ����, �
� ���
, �� � ����, �
� ���
, �� � ���� �
��
� ��� � ��
�, �
� ��
, �� � ��
�, �
� ��
, �� � ��
� �
where W=[w1, w2 …, w5] are the continuous weights learned by
a GA. The difference image for the j
th
patch-pair is then
computed as follows (see Figure 2):
��� � dist���1, ��2�
� �∑ �� � ��� ! � � ��
"
� #�
(1)
Fig. 2. CNN-based features vectors process.
The goal is to learn the weight vector W using a GA that
maximizes the fitness (f-score) between the detected change
and the corresponding ground truth for all the patches used in
the training phase [15]. Once the optimal weights are learned,
they will be used in the test phase. The overall GA for learning
the weights, shown in Figure 3, was:
• Step 1 (Initialize population): The first step of GA is to
randomly create and initialize the chromosomes of the
initial population. Nb_Ind vectors $ � %$ �&�|( �
1. . *+_�-./ are generated. Each $ �&� is a chromosome
with five (5) real-valued genes�� �&�|0 � 1. . .5�.
Engineering, Technology & Applied Science Research Vol. 12, No. 5, 2022, 9351-9356 9353
www.etasr.com Slimani & Hedjam: A Hybrid Metaheuristic and Deep Learning Approach for Change Detection in …
• Step 2 (Evaluation): The fitness function measures the
fitness of the change map between each pair of the patch
and the corresponding ground truth patches as follows:
�(2-344�$ �&�� � ∑ 56789:�;<= ,>9=�
?=@A
B (2)
Fig. 3. GA-based learning process algorithm.
• Step 3 (Selection): The elitism selection method selects the
best W
(i)
chromosomes from the previous population to
integrate them into the next population. According to the
best fitness function values Fitness(W
(i)
), a portion of
(ProbSelect%) from the precedent population was selected
to breed a new generation.
• Step 4 (Crossover): The crossover method creates a portion
of (ProbCross%) from the precedent population. A one-
point crossover method was used.
• Step 5 (Mutation): The goal of this function was to
introduce diversity into the population. A portion of
(ProbMut%) was chosen and a random value was assigned
to one randomly chosen gene.
• Redo steps 2, 3, 4, and 5 until stability (no change in the
fitness) (see Algorithm 1). Finally, the best W is chosen to
be used in the test phase.
III. EXPERIMENTATION AND EVALUATION
A. Dataset Description
Two datasets were used to evaluate the proposed change
detection framework, namely the SZTAKI AirChange
Benchmark set [16] and the Onera Satellite Change Detection
dataset [17]. The SZTAKI AirChange Benchmark set contains
13 aerial image pairs of 952×640 pixels with a resolution of
1.5m/pixel, and binary change masks (a ground truth defined
by experts). The Onera Satellite Change Detection dataset
consists of 24 pairs of multispectral images taken using the
Sentinel-2 satellites between 2015 and 2018. The locations
were picked from all over the world, Brazil, the United States,
Europe, the Middle East, and Asia. For each location,
registered pairs of 13-band multispectral satellite images are
required. The images vary in spatial resolution between 10m,
20m, and 60m. The pixel-level change ground truth is provided
for the image pairs. The annotated changes focus on urban
changes, such as new buildings or new roads.
B. GA Parameter Setting
In the training phase, a pre-trained VGG19 was used to
extract the feature maps. The GA requires several parameters to
search for the optimal layer weights:
• Number of generations Nb_Gen=200
• Number of Individuals Nb_Ind=100
• Probability of selection ProbSelect=40%
• Probability of crossover ProbCross=40%
• Probability of mutation ProbMutation=20%
C. Results, Evaluation, and Comparison to Other Methods
The proposed method was compared with two classes of
existing change detection methods, traditional and deep
learning based. The traditional methods were the Iteratively
Reweighted Multivariate Alteration Detection Method (IMAD)
for change detection [18], Slow Feature Analysis (SFA)
algorithm for change detection [19], Principal Component
Analysis and k-means clustering (PCA-Kmeans) [20], and
Change Vector Analysis (CVA) [21]. The deep learning-based
methods were two simple CNN-based without layer weighting:
the VGG19 [7], and the ResNet50 [22].
Table I shows the change detection results using different
methods in terms of f-score, recall, precision, and accuracy,
based on the SZTAKI dataset. Based on f-score and accuracy,
the proposed method gave the best results with the test images
Szada3, Tiszadob2, and Archive (accuracy was 0.90, 0.85, and
0.85 respectively). In the case of the Szada4 image, the
proposed method gave better results than the original VGG19
algorithm (accuracy was 0.70 versus 0.69) but had lower
accuracy than IMAD, which can be justified by the poor
quality of the ground truth.
Engineering, Technology & Applied Science Research Vol. 12, No. 5, 2022, 9351-9356 9354
www.etasr.com Slimani & Hedjam: A Hybrid Metaheuristic and Deep Learning Approach for Change Detection in …
TABLE I. CHANGE DETECTION RESULTS OF THE SZTAKI AIRCHANGE DATASET
Images Methods f-score Recall Precision Accuracy White pixels detected
Szada3
IMAD 0.45 0.63 0.35 0.86 4261
ISFA 0.35 0.38 0.33 0.89 2553
PCA-Kmeans 0.37 0.50 0.29 0.84 3368
CVA 0.16 0.60 0.09 0.42 4062
VGG 19 0.31 0.32 0.30 0.87 2166
ResNet50 0.41 0.66 0.29 0.82 4462
Proposed 0.45 0.45 0.44 0.90 3078
Szada4
IMAD 0.57 0.50 0.67 0.78 14438
ISFA 0.42 0.39 0.45 0.67 11350
PCA-Kmeans 0.59 0.57 0.61 0.76 16548
CVA 0.38 0.57 0.29 0.45 16418
VGG 19 0.27 0.20 0.45 0.69 5646
ResNet50 0.50 0.45 0.57 0.73 12884
Proposed 0.30 0.24 0.48 0.70 6865
Tiszadob2
IMAD 0.39 0.48 0.32 0.70 7608
ISFA 0.36 0.47 0.29 0.66 7555
PCA-Kmeans 0.48 0.63 0.39 0.74 10011
CVA 0.32 0.67 0.21 0.45 10734
VGG 19 0.23 0.17 0.36 0.78 2664
ResNet50 0.36 0.36 0.37 0.75 5769
Proposed 0.44 0.35 0.60 0.85 4945
Archive
IMAD 0.40 0.45 0.36 0.76 6392
ISFA 0.31 0.22 0.56 0.83 3118
PCA-Kmeans 0.40 0.46 0.35 0.76 6544
CVA 0.31 0.71 0.20 0.45 10092
VGG 19 0.31 0.20 0.65 0.84 2863
ResNet50 0.45 0.48 0.42 0.80 6776
Proposed 0.32 0.21 0.70 0.85 2958
(a) (b) (c) (d) (e)
Fig. 4. Subjective results. From left to right: (a) image (1) before change, (b) image (2) after change, (c) ground-truth, (d) CD map, (e) overlay of CD on image
(2). Images from SZTAKI, from top to bottom, Szada3, Szada4, Tiszadob2, Archive.
Engineering, Technology & Applied Science Research Vol. 12, No. 5, 2022, 9351-9356 9355
www.etasr.com Slimani & Hedjam: A Hybrid Metaheuristic and Deep Learning Approach for Change Detection in …
TABLE II. CHANGE DETECTION RESULTS OF THE ONERA SATELLITE DATASET
Images Methods f-score Recall Precision Accuracy White pixels detected
Beirut
IMAD 0.46 0.58 0.39 0.90 2262
ISFA 0.42 0.94 0.27 0.80 3682
PCA-Kmeans 0.54 0.71 0.43 0.90 2804
CVA 0.17 0.82 0.09 0.35 3231
VGG 19 0.26 0.17 0.53 0.92 670
ResNet50 0.43 0.55 0.36 0.89 2165
Proposed 0.55 0.57 0.67 0.94 2252
Chongqing
IMAD 0.56 0.75 0.45 0.94 1897
ISFA 0.52 0.77 0.39 0.93 1943
PCA-Kmeans 0.42 0.53 0.35 0.93 1340
CVA 0.07 0.50 0.04 0.36 12682
VGG 19 0.22 0.19 0.27 0.93 492
ResNet50 0.36 0.45 0.31 0.92 1125
Proposed 0.41 0.43 0.44 0.94 1095
Las Vegas
IMAD 0.60 0.61 0.59 0.92 3183
ISFA 0.33 0.55 0.24 0.77 2858
PCA-Kmeans 0.54 0.49 0.60 0.91 2558
CVA 0.17 0.65 0.10 0.37 3346
VGG 19 0.37 0.42 0.32 0.85 2196
ResNet50 0.64 0.94 0.48 0.89 4853
Proposed 0.40 0.46 0.49 0.90 2362
Montpellier
IMAD 0.67 0.65 0.69 0.92 4141
ISFA 0.62 0.54 0.72 0.92 3451
PCA-Kmeans 0.69 0.76 0.62 0.91 4825
CVA 0.26 0.75 0.16 0.47 4725
VGG 19 0.51 0.40 0.71 0.90 2554
ResNet50 0.63 0.63 0.62 0.91 4014
Proposed 0.44 0.35 0.60 0.85 4945
(a) (b) (c) (d) (e)
Fig. 5. Subjective results. From left to right: (a) image (1) before change, (b) image (2) after change, (c) ground-truth, (d) CD map, € overlay of CD on image
(2). Images from Onera. From top to bottom, Beirut, Chongqing, Las Vegas, Montpellier.
Table II shows the change detection results of the different
methods in terms of f-score, recall, precision, sensitivity, and
accuracy, based on the Onera Satellite dataset. The proposed
method was again better than VGG19 in any case, and it was
the best for Beirut and Chongqing test images, but weaker than
IMAD for Las Vegas and Montpellier.
Engineering, Technology & Applied Science Research Vol. 12, No. 5, 2022, 9351-9356 9356
www.etasr.com Slimani & Hedjam: A Hybrid Metaheuristic and Deep Learning Approach for Change Detection in …
Figures 4 and 5 show the results of change detection on the
test image from the SZTAKI and the Onera datasets
respectively. The last column shows the overlay of the change
detection map on the image after the change, where the
changes are highlighted in red and the ground truth is
highlighted in cyan blue. It can be noted that most of the
changes that occurred were detected. Unfortunately, the
method detected unwanted changes, which means that it
triggered more false alarms. Finally, the results obtained show
that the proposed method gave better results than [7] and [12],
unveiling the problem with the weight of the layer. Moreover,
the GA learning phase made it possible to detect fine and
coarse details, by finding the best weight vector.
IV. CONCLUSION
This paper presented an artificial intelligence-based change
detection approach for remote sensing. The challenge of
finding the best weighted CNN layers to the change detection
problem was solved using a genetic algorithm. The proposed
approach combines a CNN (pre-trained VGG19) and genetic
algorithm to build a near-optimal weight vector. This vector
was combined with the feature maps to compute their
difference and produce the change map. The purpose of this
study was to investigate the adaptation of CNNs to detect
changes in remote-sensing images. Two datasets were used to
show that it is recommended to specify the CNN layers to be
used to detect different changes. More precisely, detecting
small changes requires the use of the first CNN layers, whereas
large changes require the use of the last layers. Therefore, a
natural way to solve this problem is to use a meta-heuristic
optimization method to find the best weights for the layers to
be combined to generate the relevant feature maps. Future
work could investigate other CNN architectures and
metaheuristic search methods and extend the work on a larger
number of datasets.
REFERENCES
[1] P. Coppin, E. Lambin, I. Jonckheere, and B. Muys, "Digital change
detection methods in natural ecosystem monitoring: a review," in
Analysis of Multi-Temporal Remote Sensing Images, vol. Volume 2,
Singapore: World Scientific, 2002, pp. 3–36.
[2] B. Demir, F. Bovolo, and L. Bruzzone, "Classification of Time Series of
Multispectral Images With Limited Training Data," IEEE Transactions
on Image Processing, vol. 22, no. 8, pp. 3219–3233, Dec. 2013,
https://doi.org/10.1109/TIP.2013.2259838.
[3] J. Collomb, P. Balland, P. Francescato, Y. Gardet, D. Leh, and P. Saffré,
"Thermomechanical Optimization and Comparison of a Low Thermal
Inertia Mold with Rectangular Heating Channels and a Conventional
Mold," Advances in Materials Science and Engineering, vol. 2019, May
2019, Art. no. e3261972, https://doi.org/10.1155/2019/3261972.
[4] T. L. Dammalage and N. T. Jayasinghe, "Land-Use Change and Its
Impact on Urban Flooding: A Case Study on Colombo District Flood on
May 2016," Engineering, Technology & Applied Science Research, vol.
9, no. 2, pp. 3887–3891, Apr. 2019, https://doi.org/10.48084/etasr.2578.
[5] S. Liu, L. Bruzzone, F. Bovolo, and P. Du, "Unsupervised
Multitemporal Spectral Unmixing for Detecting Multiple Changes in
Hyperspectral Images," IEEE Transactions on Geoscience and Remote
Sensing, vol. 54, no. 5, pp. 2733–2748, Feb. 2016, https://doi.org/
10.1109/TGRS.2015.2505183.
[6] X. Zheng, X. Chen, X. Lu, and B. Sun, "Unsupervised Change Detection
by Cross-Resolution Difference Learning," IEEE Transactions on
Geoscience and Remote Sensing, vol. 60, pp. 1–16, 2022, https://doi.org/
10.1109/TGRS.2021.3079907.
[7] A. M. E. Amin, Q. Liu, and Y. Wang, "Convolutional neural network
features based change detection in satellite images," in First
International Workshop on Pattern Recognition, Jul. 2016, vol. 10011,
pp. 181–186, https://doi.org/10.1117/12.2243798.
[8] G. Cao, B. Wang, H.-C. Xavier, D. Yang, and J. Southworth, "A new
difference image creation method based on deep neural networks for
change detection in remote-sensing images," International Journal of
Remote Sensing, vol. 38, no. 23, pp. 7161–7175, Dec. 2017,
https://doi.org/10.1080/01431161.2017.1371861.
[9] D. Peng and H. Guan, "Unsupervised change detection method based on
saliency analysis and convolutional neural network," Journal of Applied
Remote Sensing, vol. 13, no. 2, May 2019, Art. no. 024512,
https://doi.org/10.1117/1.JRS.13.024512.
[10] R. Hedjam, A. Abdesselam, and F. Melgani, "Change Detection from
Unlabeled Remote Sensing Images Using SIAMESE ANN," in IGARSS
2019 - 2019 IEEE International Geoscience and Remote Sensing
Symposium, Yokohama, Japan, Jul. 2019, pp. 1530–1533,
https://doi.org/10.1109/IGARSS.2019.8898672.
[11] L. Khelifi and M. Mignotte, "Deep Learning for Change Detection in
Remote Sensing Images: Comprehensive Review and Meta-Analysis,"
IEEE Access, vol. 8, pp. 126385–126400, 2020, https://doi.org/10.1109/
ACCESS.2020.3008036.
[12] Y. Slimani and R. Hedjam, "Which Cnn Layer For Which Change? A
Cnn Adaptation Approach For Change Detection In Remote Sensing
Data," in 2020 Mediterranean and Middle-East Geoscience and Remote
Sensing Symposium (M2GARSS), Tunis, Tunisia, Mar. 2020, pp. 5–8,
https://doi.org/10.1109/M2GARSS47143.2020.9105168.
[13] M. D. Zeiler, G. W. Taylor, and R. Fergus, "Adaptive deconvolutional
networks for mid and high level feature learning," in 2011 International
Conference on Computer Vision, Aug. 2011, pp. 2018–2025,
https://doi.org/10.1109/ICCV.2011.6126474.
[14] S. Sahel, M. Alsahafi, M. Alghamdi, and T. Alsubait, "Logo Detection
Using Deep Learning with Pretrained CNN Models," Engineering,
Technology & Applied Science Research, vol. 11, no. 1, pp. 6724–6729,
Feb. 2021, https://doi.org/10.48084/etasr.3919.
[15] B. K. Alsaidi, B. J. Al-Khafaji, and S. A. A. Wahab, "Content Based
Image Clustering Technique Using Statistical Features and Genetic
Algorithm," Engineering, Technology & Applied Science Research, vol.
9, no. 2, pp. 3892–3895, Apr. 2019, https://doi.org/10.48084/etasr.2497.
[16] C. Benedek and T. Sziranyi, "Change Detection in Optical Aerial Images
by a Multilayer Conditional Mixed Markov Model," IEEE Transactions
on Geoscience and Remote Sensing, vol. 47, no. 10, pp. 3416–3430, Jul.
2009, https://doi.org/10.1109/TGRS.2009.2022633.
[17] B. L. Saux, R. C. Daudt, A. Boulch, and Y. Gousseau, "OSCD - Onera
Satellite Change Detection." IEEE, Oct. 09, 2019, Accessed: Aug. 21,
2022. [Online]. Available: https://ieee-dataport.org/open-access/oscd-
onera-satellite-change-detection.
[18] A. A. Nielsen, "The Regularized Iteratively Reweighted MAD Method
for Change Detection in Multi- and Hyperspectral Data," IEEE
Transactions on Image Processing, vol. 16, no. 2, pp. 463–478, Oct.
2007, https://doi.org/10.1109/TIP.2006.888195.
[19] C. Wu, B. Du, and L. Zhang, "Slow Feature Analysis for Change
Detection in Multispectral Imagery," IEEE Transactions on Geoscience
and Remote Sensing, vol. 52, no. 5, pp. 2858–2874, Feb. 2014,
https://doi.org/10.1109/TGRS.2013.2266673.
[20] T. Celik, "Unsupervised Change Detection in Satellite Images Using
Principal Component Analysis and k-Means Clustering," IEEE
Geoscience and Remote Sensing Letters, vol. 6, no. 4, pp. 772–776, Jul.
2009, https://doi.org/10.1109/LGRS.2009.2025059.
[21] W. A. Malila, "Change Vector Analysis: An Approach for Detecting
Forest Changes with Landsat," in Symposium on Machine Processing of
Remotely Sensed Data and Soid Information Systems and Remote
Sensing and Soil Survey Proceedings, Jun. 1980, pp. 326–335.
[22] K. L. de Jong and A. Sergeevna Bosman, "Unsupervised Change
Detection in Satellite Images Using Convolutional Neural Networks," in
2019 International Joint Conference on Neural Networks (IJCNN),
Budapest, Hungary, Jul. 2019, pp. 1–8, https://doi.org/10.1109/
IJCNN.2019.8851762.