THE EFFECTS OF TRAINING SET SIZE ON THE ACCURACY OF MAXIMUM LIKELIHOOD, NEURAL NETWORK AND SUPPORT VECTOR MACHINE CLASSIFICATION

Abstract

In this paper, we assess the accuracy of maximum likelihood, neural network and support vector machine classification with changing training set size. The data come from Landsat-5 TM satellite covering the area of Klang, located in Selangor, Malaysia. Initially, single or multiple region of interest (ROI) are drawn on each of the land cover classes identified in order to extract the training sets. The size of the training pixels are then varied from 10% to 90% by resampling the pixels within the ROI using stratified random sampling technique, where nine training sets are generated. Landsat bands 1, 2, 3, 4, 5 and 7 are then used as the input for the maximum likelihood, neural network and support vector machine classification by making use all the nine training sets. The accuracy of the classifications are then assessed by comparing the classifications with a reference set using a confusion matrix. The result reveals that support vector machine classification has a more stable increase in accuracy than maximum likelihood but neural network shows a decreasing trend as the size of training set increases.