A Kernel Spectral Angle Mapper algorithm for remote sensing image classification

A Kernel Spectral Angle Mapper (KSAM) algorithm is proposed to deal better with the nonlinear classification problem of the remote sensing image. The so-called KSAM algorithm is achieved by introducing the kernel method into the standard Spectral Angle Mapper (SAM) algorithm. Experimental results indicate that the classification accuracy of the KSAM algorithm is superior to one of the SAM algorithm in the remote sensing image classification. However the kernel parameters of the polynomial and sigmoid kernel functions of the algorithm are excessively sensitive. A narrow bound of the kernel parameters in the polynomial and sigmoid kernel functions can be chosen for the optimal classification of the remote sensing image. The classification performance of the Radial Basis Function (RBF) kernel function is superior to one of the polynomial and sigmoid kernel functions. A wide bound of the kernel parameter in the RBF kernel function can be chosen for the optimal classification of the remote sensing image in the KSAM algorithm.

I. Introduction

Over the last decade, remote sensing image classification algorithms have been improved with the development of the pattern recognition methods. In recent research of the remote sensing image classification, the kernel method is successfully used for the nonlinear classification as discussed in the former presentation of [1] to [5]. The Spectral Angle Mapper (SAM) algorithm has been widely utilized for remote sensing image [6]. Pixel with minimum or zero spectral angles in comparison to the reference spectrum is assigned to the class defined by reference vector. However, when threshold for classification based on spectral angle is modified, the probability of incorrect object detection may increases. The SAM algorithm is a linear model which does not work well when these classes are overlapped with each other as first described in [7]. In this paper, we discussed the capability of the Kernel Spectral Angle Mapper (KSAM) algorithm in dealing with the nonlinear classification problem of remote sensing image. The KSAM algorithm is achieved by introducing kernel method into the standard SAM algorithm. Thus, the KSAM algorithm is generally the nonlinear extension of the standard SAM algorithm. With the kernel method presented in [8], the input data is mapped implicitly into a high-dimensional feature space in which the nonlinear pattern now appears linear.