Klasifikasi Tutupan Lahan Berbasis Rona dan Tekstur dengan Menggunakan Citra ALOS PRISM

Abstract

Land cover classification is important issue in remote sensing, so its accuracy needs to be increased. The aim of this research was to analyze ability of tone and texture in differ land cover classes, to measure classification accuracy of each methods, and to analyze pattern of accuracies. This research was using ALOS PRISM imagery. Texture processing was done by using co-occurence matrix starting at kernel size 3x3 up to 41x41, with some extensions at 53x53, 73x73, and 141x141. Classification was performed by using Minimum Distance to Mean, Decision Tree QUEST, and SVM. Results showed that tone was unable to differ land cover properly so there has to be texture in addition to classify. At the beginning, accuracy increased and then reached a peak and fairly stable at that level. It decreased when kernel size was too large. It indicated the law of deminishing return phenomenon, because of excessive complexity within statistical analysis. Texture or the combination of tone and texture based classification's accuracy were higher than tone based classification, except to small or too large kernels, i.e. 3x3 and 141x141. Minimum Distance to Mean showed the lowest accuracy. Accuracy of multi-scale based classification was higher than single scale based classification. The choice of textures also affects the classification accuracy. Accuracy of Decision Tree QUEST were the most difficult to be modeled by rational equation.