International Journal of Advanced Engineering Research

and Science (IJAERS)

Peer-Reviewed Journal
ISSN: 2349-6495(P) | 2456-1908(O)
Vol-9, Issue-8; Aug, 2022
Journal Home Page Available: https://ijaers.com/
Article DOI: https://dx.doi.org/10.22161/ijaers.98.22

Application of PCA-CNN (Principal Component Analysis

– Convolutional Neural Networks) Method on Sentinel-2
Image Classification for Land Cover Mapping
Ahmad Rizqi Pradana1, Alfian Futuhul Hadi2, Indarto3

1 Departmen of matematic. FMIPA. Universitas Jember, Indonesia

Email: [email protected]
2 Departmen of matematic. FMIPA. Universitas Jember, Indonesia

Email: [email protected]
3 Departmen of Agricultural Engineering. FTP. Universitas Jember, Indonesia

Email: [email protected]

Received: 09 Jul 2022, Abstract— Land cover information based on remote sensing imagery is
Received in revised form: 01 Aug 2022, effective information for land use management. The use of Sentinel-2
imagery is considered to be able to provide better information on land
Accepted: 07 Aug 2022,
cover because it has a spatial accuracy of 10 meters. Convolutional Neural
Available online: 15 Aug 2022 Networks is one of the deep learning methods that can be used for image
©2022 The Author(s). Published by AI interpretation in order to obtain image classification results which will later
Publication. This is an open access article obtain information about land cover. PCA-CNN (Principal Component
under the CC BY license Analysis-Convolutional Neural Network) is a development method of the
(https://creativecommons.org/licenses/by/4.0/). Convolutional Neural Network method which gives special treatment to
the dimension reduction process in the input data. The dimension reduction
Keywords— Land Cover, Sentinel-2, Deep
process is carried out by utilizing the PCA method so that the data
Learning, PCA, CNN.
processing process becomes faster without losing important information so
that better method performance is obtained. The PCA-CNN method is
implemented on a dataset of the Situbondo district which is classified into
five land cover classes. The results of the PCA-CNN method have an
Overall Accuracy of 94.4% and Kappa Indeks 0,92 with 100 epochs of
repeated experiments.

I. INTRODUCTION multispectral and has 13 bands obtained from the

The large area and the mapping of the Situbondo area multispectral imager [11]. Automation methods for
that has not been mapped properly are separate obstacles processing Sentinel-2 satellite imagery include the use of
in the process of developing and planning the area. deep learning. Deep learning is a learning method for
Automation of land cover monitoring and classification is data that aims to create a multilevel data representation
required to monitor existing land use. The technology [1]. The most important thing about deep learning
needed to analyze the earth's land cover automatically and emphasizes that the data representation is not made
cover a large area is by utilizing geospatial data in the explicitly by humans but is generated by an algorithm [5].
form of satellite image data. One of the satellite images According to Heryadi and [5] in the last ten years the
that can be used is the Sentinel-2. Sentinel-2 imagery is application of deep learning shows that models based on
an image generated from remote sensing by the Sentinel- Convolutional Neural Networks (CNN) with deep
2 satellite. The Sentinel-2 satellite is equipped with a structures have excellent performance in the field of

www.ijaers.com Page | 188

Pradana et al. International Journal of Advanced Engineering Research and Science, 9(8)-2022

pattern processing, such as object classification in decreases sharply and generally shows PC with
images. CNN or ConvNet is a deep feed-forward eigen values of more than 1.
artificial neural network that is widely applied in image b. Using the cumulative proportion of variance
analysis. CNN consists of one input layer (input layer), which is formulate as follows
one output layer (output layer), and a number of hidden
layers [10]. ∑𝑘𝑖=1 𝜆𝑖
𝑝𝑃𝐶𝑘 = × 100%
∑𝒏𝒊=𝟏 𝜆𝑖
with 𝜆1 > 𝜆2 > ⋯ > 𝜆𝐷 . The number PCs has at
II. METHODOLOGY
least a cumulative proportion of variance of 80%
2.1 Principal Component Analysis (PCA)
[8].
Dimensional reduction is a process carried out to
8. The new variable resulting from the reduction is
simplify the existing variables to be fewer without losing
obtained by using an eigen vector matrix with an
the information contained in the initial data. One of the
input.
methods used in dimension reduction is Principal
Component Analysis (PCA). The workings of PCA is to 𝑃𝐶1 = 𝑒1′ 𝑋 ′ = 𝑒11 𝑋1′ + 𝑒21 𝑋2′ … +𝑒𝑝1 𝑋𝑝′
change the initial variable as many as n variables are 𝑃𝐶2 = 𝑒2′ 𝑋 ′ = 𝑒12 𝑋1′ + 𝑒22 𝑋2′ … +𝑒𝑝2 𝑋𝑝′
reduced to k new variables called Principal Component ⋮ ⋱ ⋮
(PC). Sum The number of k is less than n but by using a 𝑃𝐶𝑝 = 𝑒𝑝′ 𝑋 ′ = 𝑒1𝑝 𝑋1′ + 𝑒2𝑝 𝑋2′ … +𝑒𝑝𝑝 𝑋𝑝′
number of k(PC) can produce a value that is close to the 2.2 Convolutional Neural Networks (CNN)
same using n variables. PC that is formed is a linear Convolutional Neural Networks (CNN) or ConvNet is a
combination of the initial variables that are independent or deep feed-forward artificial neural network that is widely
not correlated with PC other. The following are the steps applied in image analysis. CNN consists of an input layer
to perform dimension reduction using PCA: (input layer), an output layer (output layer), and a number
1. Compile the input matrix X as one of the k attribute of hidden layers (hidden layer). Hidden layers generally
vector data 𝑥𝑖𝑗 where 𝑖 = 1,2, … , 𝑛 and 𝑗 = 1,2, … , 𝑚. contain convolutional layers, pooling layers, normalization
layers, ReLu layers, full connected layers, and loss layers.
𝑥11 𝑥12 … 𝑥1𝑚
𝑥21 𝑥22 … 𝑥2𝑚 All the layers are arranged in a pile. CNN uses a three-
𝑋=[ ⋮ ⋮ ⋱ ⋮ ] dimensional architecture, namely width, height, and depth.
𝑥𝑛1 𝑥𝑛2 … 𝑥𝑛𝑚 The width and height dimensions on CNN are
representations of the image (texture and morphology)
2. Calculating the mean 𝑋 = 𝑋̅ which statisfies the
while the inner dimensions represent color channels [11].
following equation
The following is the architecture of CNN can be seen in
1 𝑛
𝑋̅ = ∑ 𝑥𝑖 Figure 1 [1].
𝑛 𝑖=1

3. Calculating the covariance matrix C which satisfies

the following equation
1
𝐶= (𝑋 − 𝑋̅ )(𝑋 − 𝑋̅)𝑇
𝑛−1
4. Calculating the eigen values 𝜆 which satisfies the
following equation
|𝐶 − 𝜆𝐼| = 0
5. Calculating the eigen vector 𝑣 which satisfies the Fig.1. CNN Architecture
following equation
[𝐶 − 𝜆𝐼][𝑣] = 0 2.3 Sentinel-2
6. Extract the diagonal values from the eigen values and The Sentinel-2 satellite is a European optical imaging
sort them in descending. satellite that was first launched in 2015 which was
launched as the Europe Space Agency (ESA) Copernicus
7. Here are some ways to determine I column eigen
program. The Sentinel-2 satellite has 13 spectral bands
vector to be selected as PC.
carrying various swaths of high-resolution multispectral
a. Using a scree plot of the proportion of variance, imager. The Sentinel-2 satellite system is often referred to
based on the point of the curve that no longer as a twin satellite, namely Sentinel-2A (S2A) and Sentinel-

www.ijaers.com Page | 189

Pradana et al. International Journal of Advanced Engineering Research and Science, 9(8)-2022

2B (S2B) because it works in sync so that it looks like one

satellite. Each satellite has a revisit frequency (temporal
resolution) every 10 days. Sentinel-2A and Sentinel-2B
satellites have a revisit time offset of 5 days (phase shift
1800), so that the same location on the earth's surface will
be recorded by Sentinel-2A (S2A) and Sentinel-2B (S2B)
every 5 days alternately. The Sentinel-2 satellite has
several sensors, including Visible and Near Infrared
(VNIR) and Near Infrared (NIR) to Short Wafe Infrared Fig.2. Confusion Matrix
(SWIR). The Sentinel-2 satellite can be used for
supporting services such as forest monitoring, land cover
change detection and natural disaster management [2]. According to [8] the following is a suitability category
between the two tools or methods of measuring the kappa
2.4 Evaluation of the model
index. as shown in Table 2.
The evaluation of the model in this study was carried out
Table 2. Strength Of Kappa Index
based on accuracy tests performed using a confusion
matrix to determine the producer's accuracy,user accuracy, Kappa Index (%) (Strength of Agreement)
overall accuracy and kappa index. Producer's accuracy is <0,20 Poor
the accuracy seen from the side of the map producer, while
0,21 – 0,40 Fair
user accuracy is the accuracy seen from the side of the map
user. Overall accuracy is the model's accuracy value, while 0,41 – 0,60 Moderate
the kappa index is a measure that states the consistency 0,61 – 0,80 Strong
between two measurement tools or methods.
0,81 – 0,99 Very strong
Mathematically it can be seen in Table 1.
Table 1. Size of Classification Evaluation Model
III. RESEARCH
No Ukuran Rumus
3.1 Study area and data source
1. Producer's 𝑋𝑖𝑖
100% The research was conducted in January – July 2022.
Accuracy 𝑋+𝑗
The research area covers part of Situbondo Regency. Data
2. User 𝑋𝑖𝑖 collection was carried out based on the Sentinel-2 satellite
100%
Accuracy 𝑋𝑖+ image from the https://scihub.copernicus.eu/. The tools and
materials used in this study are a laptop with specifications
3. Overall ∑𝑛𝑖=1 𝑋𝑖𝑖 Intel® Core™ i5-3337U CPU @ 1.80GHz, 8.00 GB
100%
Accuracy 𝑋𝑚𝑛 RAM, NVIDIA GeForce GT720M with 2GB VRAM and
64-bit OS.Software ESA SNAP8.0 used for preprocessing
4. Indeks ∑𝑛𝑖=1 𝑋𝑖𝑖
− ∑𝑛𝑖=1 𝑋𝑖+ 𝑋+𝑗 dataGoogle Colab Software is used for the data
Kappa 𝑋𝑚𝑛
100% classification process. Sentinel-2 data used in this study is
1 − ∑𝑛𝑖=1 𝑋𝑖+ 𝑋+𝑗
part of the Situbondo district, East Java province. Image
data was taken on July 14, 2021 at 02:25:41 GMT. The
Where 𝑋𝑖𝑖 is the diagonal value of the i-th row and i-th following is a Sentinel-2 image format that was
column matrix. 𝑋+𝑗 is the number of pixels in the j-th successfully downloaded “S2A MSIL2A 20210714 T
column, 𝑋𝑖+ is the number of pixels in the i-th and 𝑋𝑚𝑛 is 022551 N0301 R046T49MHM 20210714 T070327”.
the number of pixels in the example. The following is a 3.2 Model Input Variables and Parameters PCA-CNN
description of the confusion matrix as illustrated in Figure PCA-CNN.
2.
Modeling on satellite imagery for land cover analysis in
Situbondo Regency has several stages. The first stage is
the determination of parameters. The parameters used in
the PCA-CNN model include the determination of the
number of convolutional layers, the selection of the
pooling and the activation function. Parameters on the
PCA-CNN model can be seen in appendix 4. The second
step is to determine the batch_size and the number of

www.ijaers.com Page | 190

Pradana et al. International Journal of Advanced Engineering Research and Science, 9(8)-2022

iterations (epochs) on the model to be run. The PCA-CNN process are shown in the confusin matrix in Figure 4.
model uses batch_size = 20 and the number of iterations
(epochs) = 100. A total of 1000 images are used as training
data for each class and 500 images are used as testing data
for each class.
3.3 Classification Result and Visual Assessment
The following are the results of the classification process
using the PCA-CNN model which are presented in the
“Training and test accuracy” graph and the “Training and
test loss” graph can be seen in Figure 3.a and Figure 3.b.

Fig.4. Confusin Matrix Of PCA-CNN Model

3.4 Classification accuracy assessment

The model test is carried out using testing originating from
the distribution of data sets using the hold-out method. The
model test carried out provides predictive results from the
PCA-CNN method which can be seen in Table 2.
Fig.3.a Graph of “Training and Test Accuracy”
Table 2. PCA-CNN Model Prediction Results
PCA-CNN
Kelas Producer User Accuracy
Accuracy (%) (%)
Kebun 90,5 95,9
Perumahan 100 93,6
Pertanian Lahan
88,54 99,23
Kering
Sawah 96,9 87,5
Fig.3.a Graph of “Training and Test Loss”
Tubuh Air 97,5 95,5
Overall Accuracy (%) Indeks Kappa
Seen from graph 3a. The blue line shows the accuracy of
94,4 0,92
the training. The results that show an increase in accuracy
in each iteration indicate that the model runs well at the
training so that the accuracy results are stable and Values from Table 1 are obtained from the confusion
continue to increase. Different things are shown in the matrix Figure 4 above. Table 1 shows that the highest
orange line which shows the accuracy of the test results. accuracy value for the prediction of the five land cover
The results obtained in the test process indicate the value classes is the Producer Accuracy in the housing class,
of the test accuracy is fluctuating. These results indicate which is 100%. That is, by using the PCA-CNN Producer
that the model experiences heavy learning in each iteration Accuracy on the housing class, each prediction is
of the test results. The test results at the end of the iteration successfully guessed accurately for each existing
show an accuracy value that is not too far from the training data.Overall Accuracy of the PCA-CNN model has a value
so that the model can be said not to be overfitting or fail to of 94.4% with a kappa index of 0.92. This value shows the
guess the results of the predictions.The results obtained in results of the model prediction on the test data are very
graph 3.a will be equivalent to the results that occur in good, which is above 80%.
graph 3.b The results in graph 3.b show the ability of the
model to make errors in the classification process. If in
graph 3.a the results show a high accuracy value, then the IV. CONCLUSION
results in graph 3.b will show a loss in the same iteration. The PCA-CNN method as a whole can be applied to
The detailed results of the PCA-CNN model classification land cover classification using Sentinel-2 imagery with

www.ijaers.com Page | 191

Pradana et al. International Journal of Advanced Engineering Research and Science, 9(8)-2022

five main classes namely kebun, perumahan, Pertanian [18] Zang, C, X. Pan., H. Li., A. Gardiner., I. Sargent., J. Hare.,
lahan kering, sawah, and Tubuh Air. The PCA-CNN P.M. Atkinson. 2018. A Hybrid MLP-CNN classifier for
method has the Overall Accuracy of the PCA-CNN model very fine resolution remotely sensed image classification.
International Society for Photogrammetry and Remote
which has a value of 94.4% with a kappa index of 0.92.
Sensing (ISPRS) 140(1): 133-144.

REFERENCES
[1] Alom, Taha, Yakopic, Westbreg, Sidike, Nasrin, Esesn,
Abdul, Asari. 2018. The History Began from AlexNet: A
Comprehensive Survey on Deep Learning Approaches.
https://arxiv.org/abs/1803.01164v2.
[2] ESA. 2015. Sentinel-2 User Handbook.z ESA Standard
Document User Handbook: Europe Space Agency.
[3] Hakim F. L. 2019. Interpretasi Citra Satelit Landsat 8 untuk
Pemetaan Tutupan Lahan Provinsi Jawa Timur. Jember:
Universiats Jember.
[4] Han J, M. Kembler, dan J. Pei. 2012. Data Mining:
Concepts and Techniques Thrid. Elsevier.
[5] Heryadi Y. dan E. Irwansyah. 2020. Deep Learning dan
Aplikasnya Dibidang Informasi Geosapsial. Depok:
Artifisia Wahana Informa Teknologi.
[6] Indarto. 2017. Pengindraan Jauh Metode Analisis dan
Interpretasi Citra Satelit. Yogyakarta: Andi
[7] Jia, K., Xiangqin, W., Xiangfa, G., Yunjun, Y., Xianhong,
X. Bin, L. 2014. Land Cover Classification Using Landsat 8
Operational Land Imager Data in Beijing, China. Geocarto
International 29(1): 941-951.
[8] Jhonson, R. A. dan D. W. Wichern. 2007. Applied
Multivariate Statistical Analysis. Pearson Prentice Hill.
[9] Munir, R. 2004. Pengolahan Citra Digital. Bandung:
Institut Teknologi Bandung.
[10] Putra I.W.S.E, A.Y. Wijaya., R. Soelaiman. 2016.
Klasifikasi Citra Menggunakan Convolutional Neural
Network (CNN) pada Caltech 101. Jurnal Teknik ITS 5(1):
1-5.
[11] Saadat. H., J. Adamowski, R. Bonnell, F. Sharifi, M.
Namdar, S. Ale-Ebrahim. 2011. Land use and land cover
classification over a large area in Iran based on single.date
analysis of satellite imagery. ISPRS Journal of
Photogrammetry and Remote Sensing 66(5): 608– 619.
[12] Sampurno, R. dan Toriq, A. 2016. Klasifikasi Tututpan
Lahan Menggunakan Citra Landsat 8 operational Land
Imager (OLI) di Kabupaten Sumedang. Jurnal Teknotan
10(2): 61-70.
[13] Sutojo T, P.N Andono, dan Muljono. 2017. Pengolahan
Citra Digital. Yogyakarta: Penerbit Andi.
[14] Suyanto. 2018. Machine Learning Tingkat Dasar dan
Lanjut. Bandung: Informatika.
[15] Suyanto. 2019. Deep Learning Modernisasi Machine
Learning untuk Big Data. Bandung: Informatika.
[16] Wuryandari, M. D dan I. Afrianto. 2012. Perbandingan
Metode Jaringan Saraf Tiruan Backpropagation dan
Learning Vector Quantization pada Pengenalan Wajah.
Jurnal Komputer dan Informatika (KOMPUTA) 1(1) : 45-
51.
[17] Yu, S, S. Jia., dan C. Xu. 2017. Convolutional Neural
Networks for Hyperspektral Image Classification.
Neourocomputing 219(1): 88-98.

www.ijaers.com Page | 192