Evaluation of Image Classification Algorithms on Hyperion and ASTER Data for Land Cover Classification

Abstract

Land use and land cover (LULC) mapping is one of the widely adopted applications of satellite data. With the advent of new technologies and sensor improvements, many classification algorithms are being developed. However, there are rarely studies on comparison of these classifiers using identical classification scheme and training data over different sensors and their products. In this article, we tested the effect of improved spectral and spatial resolution on classification performance of ASTER data (15 m), Hyperion data (30 m) and their fused product (15 m). For this purpose, we have used five supervised classification algorithms -three spatial classifiers, namely, Maximum Likelihood (MLC), Support Vector Machines (SVM), Artificial Neural Network (ANN) and two spectral classifiers, namely, Spectral Angle Mapper (SAM) and Spectral Information Divergence (SID). The performance of image classification algorithms was assessed using overall accuracy (OA) and kappa coefficient. MLC and SVM performed the best on all the three datasets. OA and kappa values for almost all the classifiers were comparable for higher spatial resolution ASTER and fused product and were higher by nearly 10% than that for higher spectral resolution Hyperion data.