A method to Derive Optimal Decision Boundary in SVM Method for Forest and non-Forest Classification in Indonesia

摘要

SVM (Support Vector Machine) is a new and promising classification method that performs structural risk minimization to obtain the optimal separating hyperplane from a given training data.The basic of training set selection of SVM is founded on the desire to describe each class in feature space which can discriminate between cases of the different classes through selecting training samples (SVs) that lie at the edge of the class distributions or closest to the hyperplane that partitions feature space. The objective of this research is to find an optimum decision hyperplane or boundary in SVM method for forest and non-forest classification derived from ALOS PALSAR data. The C parameter and parameters of kernel functions such as gamma (γ) of radial basis function in SVM were produced by using a grid search followed by manual methods. Sungai Wain in Balikpapan, East Kalimantan province was selected and a set of ALOS PALSAR data with HH (Horizontal-Horizontal), HV (Horizontal-Vertical), and VV (Vertical-Vertical) polarizations acquired during 2010 - 2011 were used. Field survey was conducted in April 5-7, 2010. In the grid search method, generalization capability of SVM for predicting pixels in ALOS PALSAR data as forest or non-forest classes is 79 +/- 2% and in the final result the method produced an accuracy of 80 +/- 1%.