COMPARISON OF MAXIMUM LIKELIHOOD AND SUPPORT VECTOR MACHINE CLASSIFIERS FOR LAND USE/LAND COVER MAPPING USING MULTITEMPORAL IMAGERY

Abstract

Most of the previous research was related to non-parametric classification methods, states that the Maximum Likelihood (ML) and Support Vector Machine (SVM) methods are the leading classification methods in producing high accuracy. However, only a small proportion of studies have compared the performance of these two methods using multitemporal remote sensing imagery, particularly on Sentinel-2 and Landsat data. This study tries to evaluate and compare the performance of ML and SVM classifiers to mapping land use/cover using Sentinel-2 and Landsat multitemporal imagery data. The Tabunio watershed with an area of 62.586 ha has been mapped with ten types of land use/cover are water body, forest, bare land, residential, plantations, agriculture, swamps, shrubs, pond, and mining. The confusion matrix and the Kappa coefficient was used to assess classification accuracy. All classification results show a high overall accuracy (OA) ranging from 86% to 95%. Among the two classifiers, four data series with different images and sample sizes, SVM produced the highest OA than ML.