Spectral-spatial classification techniques for hyperspectral imagery

Abstract

Hyperspectral image (HSI) classification plays an important role in a variety of applications such as land-use classification, mineral identification, climate change detection, and urban planning. Many classifiers have been developed in recent decades; however, the extraction of efficient features is still a challenging issue because of some problems, such as Hughes phenomenon and limited number of training samples. This thesis investigates four efficient techniques for HSI classification that take advantage of both spectral and spatial information to overcome the limitations of traditional classifiers. This study investigates HSI classification from different perspectives. Firstly, a framework that integrates two promising techniques: a joint sparse model and a discontinuity preserving relaxation algorithm, is proposed to perform the classification task. Secondly, this study develops a novel neighbour selection strategy for joint sparse models, and a multi-level joint sparse model is constructed to fully exploit spectral-spatial information for HSI classification based on different parameters used. This method can overcome the oversmoothing effect of the first technique. Thirdly, an extension of the second approach is developed in this study based on a multi-scale conservative smoothing scheme and adaptive sparse representation. This method can automatically overcome the oversmoothing effect as well as exploit the correlations among features extracted from different perspectives. The last approach solves the HSI classification task with multiple feature learning and a convolutional neural network (CNN). This method not only takes advantage of the CNN capability for enhanced feature extraction, but also fully and jointly exploits the spectral and spatial information.This thesis exploits the spectral-spatial information of HSIs from four different perspectives: integration of different techniques, multi-level-based, multi-scale-based, and multi-number-based. Experimental results demonstrate that exploiting spatial information from multiple perspectives can boost the classification accuracies of single perspective-based methods. This study also suggests that the multiple perspectives-based methods can reduce the negative impacts of limited training samples and Hughes phenomenon of conventional classifiers in HSI classification.