COmparison of support vector machines, random forest and decision tree methods for classification of sentinel - 2A image using different band combinations

Abstract

Classification of remotely sensed images is a prerequisite for many earth observation studies including change detection, yield forecast and water quality analysis. Recent studies showed that machine learning algorithms used for the classification of satellite image high accurate results. In this study, three popular machine learning algorithms namely, random forest (RF), support vector machines (SVM) and decision tree (DT) classifiers were utilized considering three datasets that comprise different band combinations of a Sentinel-2A image. These datasets consist of five, seven and eleven bands containing an image of normalized difference vegetation index (NDVI). In the classification process, six land use/cover classes covering the bulk of the study area were determined as forest, grass, asphalt road, soil and bare area, urban and water. In the classification stage, 700 pixels for training and 300 pixels for testing were selected for each class to avoid possible bias among the classes. Classification resulted revealed that SVM classifier produced the best accuracy results for all three datasets. The highest accuracy (95.17%) was achieved with SVM classifier using the 11-band combination dataset. The combinations containing high spatial resolution bands provided higher accuracies. Moreover, McNemar’s test was applied to analyze statistical significance of classifier performances for the datasets. Also F-score test was applied for all class to evaluate classification accuracy results. The results indicated that the differences between the performances were statistically significant except for SVM and RF using 7-band and 11-band combinations. To sum up, the efficiency of the machine learning algorithms applied in this study were all found effective in classification of Sentinel-2A imagery. © 2021 Elsevier B.V., All rights reserved.