Spatio-Spectral Remote Sensing Image Classification With Graph Kernels

This letter presents a graph kernel for spatio-spectral remote sensing image classification with support vector machines (SVMs). The method considers higher order relations in the neighborhood (beyond pairwise spatial relations) to iteratively compute a kernel matrix for SVM learning. The proposed kernel is easy to compute and constitutes a powerful alternative to existing approaches. The capabilities of the method are illustrated in several multi- and hyperspectral remote sensing images acquired over both urban and agricultural areas.

I. Introduction

Classification of remote sensing images constitutes a challenging problem because of the potentially high dimensionality of images and low number of training samples, the spatial variability of the spectral signature, and the presence of noise and uncertainty in the data [1]. In this context, the use of classifiers that are robust to the dimensionality and noise is strictly necessary. This is typically guaranteed by using regularized classifiers that not only minimize a cost function but also control the complexity of the classification function. Kernel methods, in general, and support vector machines (SVMs), in particular, are a family of methods that nicely implement these properties. The methods have been successfully used in pixel-based (spectral-based) classification and are among the state-of-the-art remote sensing image classifiers [2], [3].