Extraction of Spectral-Spatial 3-Dimensional Homogeneous Regions from Hyperspectral Images and Its Application to Fast Classification

摘要

Hyperspectral images have been widely applied to various fields due to the high spectral and spatial resolution. However, the vast amounts of spectral and spatial information also bring difficulties in hyperspectral image processing, where the efficiency is one of the biggest challenges. To address this challenge, we propose a method to extract the spectral-spatial 3-dimensional homogeneous regions (SS3DHRs) from hyperspectral images. First, highly correlated neighbor spectral bands are selected based on the correlation coefficients between adjacent bands; Based on the sub-band selection, a superpixel segmentation method is improved for hyperspectral images to gather the spatial information; Combining the spectral sub-bands and spatial superpixels, the SS3DHRs are collected from the 3-deminsion hyperspectral data cube. The SS3DHR can be processed as a unit for the subsequent applications, which may significantly reduce the redundant data and thus raise the efficiency. In experiment part, the extracted SS3DHRs are applied for hyperspectral image classification, where the experimental results demonstrate the effectiveness and efficiency of the proposed method.