Research On Spatial-Spectral Feature Extraction And Classification Of Hyperspectral Imagery

Hyperspectral imagery(HSI)is a three-dimensional imagery containing tens or even hundreds of consecutive spectral bands acquired by imaging spectrometers.Since it contains rich spatial and spectral information on surface matter,it has been attracted attention of many scholars at home and abroad and widely used in many fields such as environmental management,agricultural monitoring,military target discrimination and urban classification.With the development of imaging spectroscopy,the spatial resolution and spectral resolution of objects observed have been improved,which also brings some difficulties to HSI classification.On the one hand,the high spectral dimensionality and limited training samples(high cost of labeling samples manually)will result in ”Hughes phenomenon”.On the other hand,the high nonlinearity makes the HSI classification performance poor when the classification algorithms are based on spectral information.In order to solve these problems effectively,based on the characteristics of hyperspectral imagery,this paper focuses on how to fully exploit the spectralspatial feature of hyperspectral imagery and further to promote the small-sample classification accuracy of surface materials.Superpixel segmentation is a process of segmenting the spatial image into many regions and the spatial distribution of surface materials is usually regular and local continuous.Therefore,in order to make full of the spatial information of hyperspectral imagery,this paper presented a superpixel-level 2-D Gabor feature Classification approach.Firstly,a set of predefined2-D Gabor filters are applied to hyperspectral imagery to extract sufficent features.Meanwhile,a classic superpixel segmentation method,called simple linear iterative clustering(SLIC),is adopted to divide the original hyperspectral image into disjoint superpixels.Secondly,the Support Vector Machine(SVM)is applied on each extracted 2D Gabor feature cube,and the majority voting strategy is adopted to combine the classification results.Finally,the superpixel map obtained by SLIC is used to regularize the classification map.Experimental results show that the classification performance can be substantially improved under superpixel guidance,especially for hyperspectral images with more evenly distributed surfaces.Furthermore,in order to take advantage of the contextual information in the spatial-spectral structure of hyperspectral imagery and obtain discriminative three-dimensional spatial-spectral feature,a spatial-spectral feature extraction method of hyperspectral imagery based on threedimensional surface feature is proposed.This method directly deals with the raw hyperspectral imagery data and utilizes its first order derivative magnitude on the spatial and spectral orientation to jointly represent hyperspectral imagery.Then,the spatial-spectral feature is obtained by calculating histogram at the spatial neighborhood of each pixel.Experimental results show that three-dimensional surface feature extraction method can take full use of the characteristic of three dimensinal data and the classification results is properly well.Finally,based on the previous two researh methods,this paper presented a superpixel-level three-dimensional Gabor surface feature based Classification approach.Firstly,Gabor features data and their first order derivative magnitude on the spatial and spectral orientation are encoded jointly and histogram features are obtained.Secondly,principal component analysis is adopted to reduce the dimensionality of the histogram features.Meanwhile,simple linear iterative clustering(SLIC),is adopted to divide the original hyperspectral image into disjoint superpixels.Thirdly,the Support Vector Machine(SVM)is applied on each reduced histogram feature cube,and the majority voting strategy is adopted to combine the classification results.Finally,the superpixel map obtained by SLIC is used to regularize the classification map.Experimental results show that the proposed method can extract much more discriminative feature and can obtain much higher classification accuracy.