Hyperspectral Image Space-spectrum Joint Supervised Classification Algorithm And Software System

Hyperspectral image(HSI)as a new remote sensing technology has been widely applied in many research areas,such as military reconnaissance,environment survey,disaster evaluation,mineral exploration.HSI classification is very important in HSI processing.Conventional HSI supervised classification methods only utilize information from spectral domain,but ignore rich information in the spatial domain.Thus,these methods cannot achieve very good classification performances.Besides,the conventional classification methods have high computational complexities,and thus perform very slow on the commonly-used large-scale HIS datasets.Therefore,it is non-trivial to develop a new HSI classification method that can consider the rich information from both spectral and spatial domains with high accuracy and efficiency.Extreme learning machine(ELM)is a new single hidden layer feed forward neural network algorithm.ELM randomly generates single hidden layer parameters,and thus has fast learning ability and good generalization ability.In the thesis,we first analyze the advantages and disadvantages of ELM applying in HSI classification,and then improve ELM in terms of feature extraction and kernel function.Accordingly,we propose a flexible HSI classification framework by fully considering information from spectral and spatial domains on the problem of pixel-wise ELM classification method's high error rate.With several feature extraction methods and classification methods integrated together,we develop a HSI classification prototype system that includes several functions,i.e.,feature extraction,classification,performance evaluation.The main contributions in this thesis include the following aspects:(1)The hyperspectral image Supervised classification two-step method combining different kinds of feature extraction methods with normal ELM is proposed.Experiment results show:different feature extraction methods,e.g.,principal component analysis(PCA),kernel principal component analysis(KPCA),independent component analysis(ICA),locality preserving projections(LPP)can all help improve normal ELM's classification accuracy in hyperspectral image.By contrast,KPCA combined with ELM performs better.(2)Different kernel functions' influence on hyperspectral image classification in the kernel ELM(KELM)framework is compared and analyzed.And hyperspectral image supervised classification method via KELM combined with composite kernel which is called as KELM-CK is proposed.This method colligates information of spacial and spectral domain which makes it improves the supervised classification accuracy in small sample conditions and overcomes the noise impact.Every datasets validates the method's effectiveness.(3)A multi-view learning supervised classification method based on norm-based 2DPCA(N-2DPCA)and KELM is proposed.This method makes use of N-2DPCA to extract spectral-spacial feature from different view of hyperspectral image and several KELM classifiers for ensemble learning and major voting so that it can achieve quite high supervised classification accuracy.Experiments on HSI datasets demonstrate that the proposed framework outperforms the other HSI classification methods.(4)A HSI classification prototype system is developed.This prototype system integrates several commonly-used feature extraction and supervised classification methods.This system is can realize functions including feature extraction,classification,and performance evaluation.