ISSN 2278-3091

Volume
BEKKOUCHE Ibtissem et al., International Journal 11, No.6,Trends
of Advanced November - December
in Computer Science and2022
Engineering, 11(6), November - December 2022, 241 - 248
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse041162022.pdf
https://doi.org/10.30534/ijatcse/2022/041162022

Automatic Change Detection on Satellite Images using

Principal Component Analysis, ISODATA and Fuzzy
C-Means Methods
BEKKOUCHE Ibtissem1, FIZAZI Hadria2
1
Department of Computer Sciences, University of Science and Technology of Oran, Algeria,
[email protected]
2
Department of Computer Sciences, University of Science and Technology of Oran, Algeria,
[email protected]

Received Date: October 17, 2022 Accepted Date: November 23, 2022 Published Date: December 06, 2022

System (GIS) as Integrated Method of GIS and Remote

ABSTRACT Sensing, GIS Approach…; Visual analysis using Visual
interpretation; Hybrid approach in Combination of methods;
Change detection is the process of comparing two or more Advanced models like Spectral Mixing Model, Li-Strahler
images and identifying the parts where a change has occurred. Reflectance Model… [2][3]
Difference detection processing between simple digital
images, such as photographic images, is easy to implement. In recent years, this technique has become one of the most
Whereas for satellite images, which compose of several interesting subjects in the extraction of information from
images’ grayscale and bands, this requires a methodological satellite images and several researchers have opted, These last
approach to image processing appropriate to the exploitation years, for the hybridization of methods such as: in 2020,
of these data because this will allow to follow the evolution Neelam Ruhil et al, have suggested an unsupervised change
over time of a region of interest through change detection detection method based on wavelet fusion and the Kohonen
techniques, so these images are a tool of choice in the Hybrid FCM-σ [4], in the same year, Mohan Singh et al, have
management of natural resources. So, in this paper, we proposed an image fusion using image fusion using image
propose a hybrid automatic change detection approach for normalization and radiometric calibration and Particle Swamp
multi-temporal satellite images. It is based on several Optimization Fuzzy C-Means (PSOFCM). In this article, an
algorithms: ISODATA for automatic thresholding, Principal unsupervised change observation technique based on the
Component Analysis as transformation technique, Fuzzy PSOFCM [5], and in 2022, Abdelkrim Maarir et al, have
C-Means as classification technique. Experiments were proposed an unsupervised method of detecting change in
performed and assessed by their overall accuracy and results satellite images by following two main steps: The first step
validated the effectiveness and efficiency of the proposed focuses on data reduction using the Independent Component
approach, named ISOFAP. Analysis (ICA) algorithm to improve the efficiency of the
classifier. The second stage for processing uses the Fuzzy
Key words : Change detection, Fuzzy c-means clustering, C-Means classification method to find specified clusters [3].
ISODATA, Principal component analysis.
So, for our paper, since there are several change detection
1. INTRODUCTION techniques, the most interesting attitude would be to try to
Change detection is a digital process that can be performed by combine these techniques and develop a hybrid method, so for
traditional methods and using remote sensing technologies. that we used: ISODATA for automatic thresholding, Principal
The basic this process is to measure the change on the Earth's Component Analysis (PCA) as transformation technique,
surface by jointly analyzing two or more two temporally Fuzzy C-Means (FCM) as classification technique.
separated images, in order to locate and quantify
(automatically) the changes existing between these images After having tested several thresholding algorithms such as:
[1]. This is a very active subject due to preoccupation about Binary thresholding on the mean, OTSU thresholding and EM
the consequences of global and local changes in the earth. algorithm [6], we chose ISODATA. This is the algorithm that
is used for automatic thresholding, we have chosen it because
There are many change detection techniques in the literature: it is easy to implement more than it gives good results.
Algebraic methods such as Image differencing and Change Principal Component Analysis (PCA) consists of
vector analysis,…; Transformation methods such as Tasseled transforming variables, interconnected, into new variables
Cap Transformation, Principal Component Analysis…; unsquared from each other for dimension reduction. So, we
Classification methods as Artificial Neural Networks, chose because it has been widely used for change detection
Comparison after classification…; Geographic Information

241
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

[7]. It has the ability to project the multi-dimensional original  Calculate the covariance matrix.
and it is less susceptible to image overlapping.  Calculate the eigenvector and the eigenvalue of the
covariance matrix.
We found that the most used methods for change detection  Extract the diagonal of the matrix as a vector.
used a lot the K-means, so we tested it on our data but we  Variance sorting in descending order.
chose to use the improved Fuzzy C-Means (FCM) version.  Choose components and form a feature vector.
FCM has reduced complexity and gives best result for  Derivation of the new data set.
overlapping data sets and comparatively better than k-means
algorithm. At the end, the number of PC is less than the number of
variances in the original image. In CD studies, the
2. METHODS USED consequence of this linearization is that the unchanged pixels
or common information shared by a pair of images are
2.1. Isodata assumed to be in a narrow and elongated space. Cluster along
a principal axis equivalent to the first component (PC1). On
It is an image segmentation technique by clustering and it is the contrary, pixels containing a change would be more
an improved version of the k-means algorithm, which was unique in their spectral appearance and should lie far from this
first introduced by Velasco in 1980 [8] as a classic algorithm axis (PC2). [13][14]
which makes it possible to carry out a very good
categorization and which gives significant results. [9] 2.3. Fuzzy C-Means

The thresholding of an image can be done by manual Fuzzy C-Means (FCM), is an unsupervised fuzzy
thresholding or by automatic thresholding. classification algorithm. Issued from the C-means algorithm,
developed by Dunn in 1973 [15] and improved by Bezdek in
Manual image thresholding involves 4 steps [9]: 1981 [16], it introduced the notion of fuzzy set in the
1. Observing the histogram of the image definition of classes: each point in the data set belongs to each
2. Choice of thresholds in the valleys cluster with a certain degree, and all clusters are characterized
3. Definition of the classes of the regions by color range by their center of gravity [17].
4. Pixel classification The goal of Fuzzy C-Means clustering is to find the minimum
of the following function:
ISODATA thresholding is global thresholding, where a single
threshold is used across the entire image to divide it into two (1)
clusters. It allows to find the value of the sought threshold in where m is any real number greater than 1, uij is the degree of
an automatic way following its steps [10]: membership of xi in the cluster j. As a rule, for each pixel, sum
1. Choose an initial threshold T, for example T = the of all membership value belonging to all classes must be 1.
average intensity. [3][14]
2. Divide the image into two groups G1 and G2 using T.
3. Calculate the average values of each region R1 and R2 3. METHODOLOGY
4. Calculate the value = ( 1+ 2)/2 The Figure 1 shows and summarizes the organization of the
5. Repeat steps 2 to 4 until the T value does not change. processing steps adopted for the detection of changes on
satellite images by our ISOFAP approach and this according
2.2. Principal Component Analysis to the basic procedure of an image change detection
processing system [18].
Principal component analysis (PCA) is a mathematical
technique used for data redundancy reduction by Jackson and This scheme is composed of several implementation phases:
Bund, 1983 [12]. It is useful when you have obtained data on a
number of variables (perhaps a large number of variables), 1st step: This is a data preparation step and we must:
and there is some redundancy in these variables. [12]
 Data acquisition and preparation: This is the stage of
PCA is one of the most popular multivariate analysis collecting and assembling data, which can be
algorithms for change detection studies and can be performed satellite images and field investigations. Then we
on original or normalized data [7]. With this technique the can go to pre-processing, for example image cutting
digital images acquired by remote sensing, we can reduce its if necessary.
dimensionality such that the multispectral bands are the  Geometric correction: the verification of the
variables to be introduced. geometric accuracy is essential for the detection of
changes besides a bad georeferenced of more than
There are certain steps to follow to implement PCA [7]: one pixel would cause abnormal results for analyzes
 Take an original data set and calculate the mean of the pixel by pixel.
data set.
 Subtract the mean for each dimension.

242
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

Figure 1: General scheme of the proposed ISOFAP method.

 Data normalization: Data must be normalized,  Confusion Matrix: One of the most popular ways to
especially for satellite images, in order to reduce the measure the performance of a classification model.
variability between multi-date images over the same Each line corresponds to an actual class and each
geographical area. column corresponds to an estimated class and it
includes the following values [19]:
2nd step: This is a digital processing step for the detection of o True Positive, TP, when the actual class and the
changes and for this we will: estimated class are both positive
o True Negative, TN, when the actual class and the
 Apply the difference image method estimated class are both negative
 After that, the obtained image is partitioned into H*H o False Positive, FP, when the actual class is
blocks, then the creation of eigenspace space using negative but the estimated class is positive. This
PCA and principal components are achieved. is called a Type 1 error.
 Apply the ISODATA algorithm for thresholding on X. o False Negative, FN, when the real class is
 After applying PCA and ISODATA, create the feature positive but the estimated class is negative. This
vector space using the eigenspace space. is called a Type 2 error.
 To determine the areas that have changed and not
changed, we used the FCM algorithm to generate the It can be used for more in-depth measurements to get a
two classes (k=2) and assign each feature vector to better assessment of the quality of the model. Among the
the class closest to either modified pixels or classification measures used are accuracy, precision,
unchanged pixels. error and specificity. [19]

3th step: This is a generalization of the final product, which is Accuracy is the number of correct predictions made by
a map of land cover changes at a scale equivalent to that of the the model.
input data. (2)

4th step: Once a model has been determined and This measure is used when the number of True Positives
implemented, the last step is to analyze and interpret the and True Negatives are the most important.
results to establish the quality of this model. There are various Error = 1 – Accuracy (3)
evaluation measures that can be used and chosen carefully, Precision is the number of correct elements rendered by
since the choice of measure can influence how performance is the model.
assessed and interpreted. For this we have chosen to use: (4)
This metric is used when the number of False Positives
 Visual interpretation: the use of the human visual is highest.
system as a quality judgment tool is not to be
neglected but necessary to verify the quality of the Specificity is the number of negative classes predicted
images obtained by the classification. To evaluate by the model.
this approach, we also used visual analysis according (5)
to the ground truth available in the area.

243
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

4. IMAGES USED

Images database: A set of RGB images of 650 x 650 found in

a database used for the detection of changes in satellite
imagery using deep-learning, produced by Héloïse
BAUDHUIN and Antoine LAMBOT. [20]

We applied our approach to two examples, presented in

Figure 2 and Figure 3, of this database:

(b)
Figure 3: Images 2 (a) before change and (b) after change.

Images of Boumerdes: Figure 4 shows a high-resolution

satellite image of a Landsat 5 TM earthquake and both
acquired in 2003, provided by the Center National des
Techniques Spatiales d'Arzew and acquired by QuickBird.

The characteristics of the two images are: natural composition

(a) image with three channels: TM1, TM2 and TM3 bands and
their size in pixels is 1002 x 1002. They contain different
classes which are: asphalt, soil, vegetation and shade, and area
of damage to the post-disaster image.

(b)
Figure 2: Images 1 (a) before change and (b) after change
(a)

(a) (b)
Figure 4: Images 3 (a) before disaster and (b) after disaster.

244
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

By the visual interpretation, we notice that PCA-K-means has

5. EXPERIMENT RESULTS ANS DISCUSSION some confusions and that the detection is not so correct.
However, ISOFAP gave us a better result than PCA-K-means
The first thing for our work, we did a study on a known although there are also confusions but it is minimal.
change detection method to understand its concept and we
chose PCA-FCM. After the study our proposal was to change After confirming the correct operation of the two algorithms
the classification algorithm and replace it with another, so on a simple artificial image, we used the images 1 of the
after several tries, we chose to work with the Fuzzy C-means database (presented in Figure 2) and we applied for the two
(FCM) which is an improved version of the K-means methods PCA-K-means and ISODATA. For initialization of
algorithm. Then we did other tests with PCA-FCM but each the PCA-kmeans method, after several tests, the threshold was
time we had to initialize the threshold manually with the manually initialized to 60, the number of iterations to 80 and
variation of the threshold parameters for each image. So, for we varied just the number of blocks (h*h) between h equal to
that we have integrated an ISODATA thresholding algorithm 2 and 5. We notice that he detected the changes for h equal 2.
to have an automatic threshold. In the rest of the article, some
results of our proposed method "ISOFAP" in comparison with The same image applied to our method ISOFAP, the number
PCA-K-means. of iterations at 75, the variation also in the number of blocks
between 2 and 5, and knowing that ISODATA has initialized
To analyze and validate the proposed approach we used three the threshold to 43. from the results we also notice that at h
different data images described in section 4 below. equal to 2 the result is better than the others.

For the PCA-K-means parameters, we varied the parameters

as follows: number of classes is 2 classes, such as class 1 for
changed pixels and class 2 for unchanged ones, number of
iterations between 70 and 100, number of blocks: between 2
and 5, and Threshold: between 10 and 80.

And for ISOFAP we varied the parameters like the

PCA-K-means except the threshold by ISODATA.

Figure 5 shows the first performance test, we created two

examples of simple artificial images to do our tests, but we
reduced the number of iterations to 20 and the threshold to 10
because they are simple images.

(a)

(a) (b)

(b)
Confusion Detection error
(c) (d) Figure 6: Visual interpretation of the best results by (a) ISOFAP and
Figure 5: (a) artificial image, (b) artificial image changed, change (b) PCA-K-means on images 1.
map results (c) with PCA-K-means and (d) with ISOFAP.

245
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

The previous Figures 6 present the best results given by The comparison between the best result given by the two
ISOFAP and PCA-K-means applied to images 1. From our methods approves the previous result because we can say that
visual interpretation for comparing the best results of the two the visual interpretation, presented in Figure 7, is the same and
methods to the original change image, we notice that even if that ISOFAP has better detected the change even if there are
ISOFAP had confusions and detection errors caused by the conflicts and detection errors but they are less than
conflict between the two changed and unchanged classes, it PCA-K-means.
there are some details were better detected compared to
PCA-K-means. For the last tests, we chose the best parameters: number of
blocks (h*h) h at 2, number of iterations at 90 and we applied
For images 2, we did the same tests as images 1. For them to the same part of the images 3 (In Figure 8). We notice
PCA-Kmeans, after several tests we initialized the threshold that the PCA-K-means only gave us two classes: changed and
at 75, the number of iterations at 85 and varied the number of unchanged, but ISOFAP gave a third class of pixels with the
blocks. we note that these results reinforce the change gray color, for the program it is conflicts and at the same time
detection results on images 1 because the best result is given we notice that it is not totally unchanged but the change is not
on the number of blocks equal to 2. great.

For ISODATA the best result for images 2 is the number of

blocks at 2 because it gives more detailed detection and fewer
conflicts, such as for this image the threshold at 20 and the
number of iterations at 93.

(a) (b)

(a)

(c) (d)
Confusion Detection error

Figure 8: (a) image before, (b) artificial after, change map results (c)
with ISOFAP and (d) with PCA-K-means.
From the visual interpretation of the results of our ISOFAP
approach on the three groups of images, we notice that there
are detection errors in the images, which can be caused by
several reasons among them the resolution of the image
(b) because sometimes when the image is of lower quality poses
Confusion Detection error conflicts between the pixels therefore gives errors.
Figure 7: Visual interpretation of the best results by (a) ISOFAP and
(b) PCA-K-means on images 2.

246
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

Table 1: Evaluation metric results for images 1 by ISOFAP and PCA-K-means

Method Accuracy Error Precision Specificity Execution time

ISOFAP 0.8556 0.1444 0.8918 0.1082 65.283269 s

PCA-K-means 0.7624 0.2376 0.6881 0.3119 57.882101 s

Table 2: Evaluation metric results for images 2 by ISOFAP and PCA-K-means

Method Accuracy Error Precision Specificity Execution time

ISOFAP 0.6256 0.3744 0.7488 0.2512 75.243269 s

PCA-K-means 0.5824 0.4176 0.6181 0.3819 71.874101 s

Table 3: Evaluation metric results for images 3 by ISOFAP and PCA-K-means

Method Accuracy Error Precision Specificity Execution time

ISOFAP 0.7356 0.2644 0.7918 0.2082 55.267169 s

PCA-K-means 0.6424 0.3576 0.5841 0.4159 47.817301 s

We finish our study by applying the confusion matrix, to the REFERENCES

original images and the images of the best results, for extract
the information that interests us and here are the results: 1. S. Achour. Détection de changement en imagerie
satellitaire multi-temporelle pour le suivi et la
After the comparison by the evaluation metric (in Table 1, surveillance de l'environnement, Ph.D. dissertation,
Table 2 and Table3) and the visual interpretation, we notice Dept. Elect. Eng., Djillali Liabès Univ, Sidi Bel Abbes,
that our proposed ISOFAP approach gives good results to Algeria, 2022.
detect changes and the values of accuracy and precision are 2. D. Lu and P. Mausel, E. Brondizio, and E. Moran.
high for three different data images, so we conclude that Change detection techniques, Change detection
ISOFAP is better than PCA-K-means with the exception of techniques, International Journal of Remote Sensing, vol.
the execution time, it takes longer than PCA-K-means. 25, pp. 2365-2401, 2004.
3. A. Maarir and E. S. Azougaghe, and B. Bouikhalene.
6. CONCLUSION Automatic Change Detection Based on the
Independent Component Analysis and Fuzzy
In this work, we have addressed one of the image processing C-Means Methods, In: International Conference on
operators which is the detection of changes in satellite images. Business Intelligence, Springer, Cham, pp. 178–187,
We have proposed a hybrid method based on two techniques 2022.
to do this treatment. 4. N. Ruhil and M. Singh, D. Mitra, A. Singh, and K. K.
Singh. Detection of changes from Satellite Images
Our approach is based on a method already used for the Using Fused Difference Images and Hybrid Kohonen
detection of changes (PCA-K-means) and we have tried to Fuzzy C-Means Sigm, Elsevier, Procedia Computer
improve it. We used ISODATA to make the thresholding Science, vol. 167, pp. 431-439, 2020.
automatic and kept the PCA because it is the most used for 5. M. Singh, K. D. Tyagi, A. Singh, and K. K. Singh.
change detection and the least sensitive to image overlap, Detection of changes in Landsat Images using Hybrid
while we chose FCM because it has reduced complexity and it PSO-FCM, Elsevier, Procedia Computer Science,
is an improved version of the k-means algorithm. vol.167, pp. 423-430, 2020.
6. B. Cayla. Traitement d’images (partie 3 : Seuillage
After the tests and the additions, we arrived at the d’image), available at
implementation of our ISOFAP method, which is compared https://www.datacorner.fr/image-processing-3/
with the PCA-K-means, the results allowed us to conclude 7. M. Saravanan and M. A. Santhosh Sivan. Efficient &
that our method can detect the change and that it gives better analysis of PCA based image change detection
results. algorithms, International Journal of Advanced

247
BEKKOUCHE Ibtissem et al., International Journal of Advanced Trends in Computer Science and Engineering, 11(6), November - December 2022, 241 - 248

Technology in Engineering and Science, vol.5, pp.

235-243, 2017.
8. F. R. D. Velasco. Thresholding using the ISODATA
Clustering algorithm, IEEE Transaction on system,
Man and Cybernetics, vol.10, pp. 771-774, 1980.
9. T. Baakek. Segmentation et classification intélligente
des images cérébrales, Ph.D. dissertation, Dept.
Electrical and Electronic Eng., Abou Bekr Belkaid Univ,
Tlemcen, Algeria, 2015.
10. J. Neira, Course Program, Lesson 1: Thresholding,
University of Zaragoza.
11. B. B. Jackson and B. Bund. Multivariate Data Analysis:
An Introduction, McGraw-Hill, 1983.
12. R. D. Jeanson. Détection de changement sur
l’occupation du sol à l’aide de la méthode ACP et la
logique floue, Memory, Antananarivo Univ., R. D.
Madagascar, 2015.
13. S. Panthan, and P. Thakre. Fuzzy Clustering Technique
and PCA Based Unsupervised Change Detection
Method in Multitemporal SAR Images, Vandana
Publications, International Journal of Engineering and
Management Research, vol.7, pp. 345-348, 2016.
14. M. H. Kesikoglu, U. H. Atasever, and C. Ozkana.
Unsupervised change detection in satellite images
using fuzzy C-means clustering and principal
component analysis, International Archives of the
Photogrammetry, Remote Sensing and Spatial
Information Sciences, Vol. 7, P; W2, 2013.
15. J. C. Dunn. A Fuzzy Relative of the ISODATA Process
and Its Use in Detecting Compact Well-Separated
Clusters, Taylor & Francis, Journal of cybernetics, vol 3,
pp. 32-57, 1973.
16. J. C. Bezdek, R. Ehrlich, and W. Full, FCM: The
fuzzy c-means clustering algorithm, Elsevier,
Computers & Geosciences, vol. 10, pp. 191-203, 1984.
17. Z. Wang, N. Zhao, W. Wang, R. Tang, and S. Li. A Fault
Diagnosis Approach for Gas Turbine Exhaust Gas
Temperature Based on Fuzzy C-Means Clustering
and Support Vector Machine, Hindawi, Mathematical
problems in Engieering, vol. 2015, 2015.
18. T. T. H. Pham. Détection des changements de
l'occupation des terres de la zone littorale : cas du
district de Tiên Hai (Viêt-nam), M. S. thesis,
Sherbrooke Univ., Montréal, Canada, 2005.
19. A. A. Al Analyst. Apprentissage Supervisé et
classification, available at
https://fr.linedata.com/apprentissage-supervise-et-classif
ication
20. H. Baudhuin and A. Lambot; LamboiseNet, Master
Thesis about Change Detection in Satellite Imagery using
Deep Learning, available at
https://www.datacorner.fr/image-processing-3/

248