ONE-AGAINST-ALL REMOTE SENSING IMAGE CLASSIFICATION USING SUPPORT VECTOR MACHINE

Abstract

This research presents a new method of extending a binary support vector machine algorithm to a multi-class remote sensing image task using a one-against-all technique. A Landsat image is used for the experiment. The land use classes of interest are: developed, undeveloped and water. The spectral bands are extracted in ArcGIS while MATLAB programming software is used for the modelling. The selection of support vector machine kernel functions and parameters are based on the k-fold cross-validation. The initial classification result yields four land use classes: developed, undeveloped, water and unclassified; while the final classification result is resolved to three land use classes: developed, undeveloped and water. For the final result, the Kappa statistic was obtained for 20 iterations; the highest Kappa statistic is obtained at the 9th iteration while the least Kappa statistic is obtained at the 20th iteration. The initial result yields a Kappa value of 0.7787 which indicates a substantial agreement with the ground truth data; while the final image yields a Kappa value of 0.8671 which indicates an almost perfect agreement with the ground truth data.