Spatial Classification of Sentinel-2 Satellite Images with Machine Learning Approach

Abstract

This study aims to classify buildings and non-buildings from Sentinel-2 Satellite Images using a Machine Learning approach. The limitations of the machine learning method for classification used in this study are the Support Vector Machine (SVM), Logistic Regression (LR), and Decision Tree (DT) methods. The three methods' results are compared to find the best method in the classification process. Furthermore, the proportion between buildings and non-buildings around Universitas Islam Indonesia was calculated from the best method’s results. The results are in the form of a classification with four indicators, namely the level of accuracy, sensitivity, specificity, and Area Under the Curve (AUC). We found that the best performing method in this study is the SVM method based on the average accuracy results, the smallest average variance difference in the variance of training and testing data, and three other indicators from the number of iterations accomplished. In the density proportion, we concluded that the closer the distance to UII campus, the greater the percentage of buildings. As for non-buildings, the farther from the center point, the higher the rate of non-buildings