Hyperspectral Image Classification Based on Adaptive Neighborhood of Combined Kernel

Abstract

The traditional combined kernel classification method adopts fixed neighborhood in combining spatial information and ignores the case of the neighboring pixels at edge of object not belonging to the same classification. The paper suggests two methods of adaptive neighborhood. The first one segments the hyperspectral image, then defines neighborhood of each pixel according to the boundary of segmented region which each pixel belongs to, and extract spatial information. While the another one directly extracts the spatial information by holding edge filter which can inhibit the weight of outliers in the neighboring pixels. The experimental results show that the proposed methods effectively reduce the misclassification on the edge and hence improve the classification precision.