Research Of Hyperspectral Image Classification Based On Stacked Wasserstein Auto-encoder And Mixture Generative Adversarial Networks

Hyperspectral remote sensing is a typical earth observation technology,the classification and recognition technology is one of the core technologies in hyperspectral image(HSI)processing.How to extract the advanced features of HSI and establish a robust classification model under small samples are an urgent problem to be solved.Two HSI classification algorithms based on deep learning theory are proposed in this thesis.The main research contents can be summarized in the following two parts:In order to improve the HSI classification performance,an HSI classification algorithm based on Stacked Wasserstein Auto-Encoder(SWAE)is proposed in this thesis.The spectral and spatial features of the HIS sample are extracted simultaneously by the SWAE.The original sample is encoded by the encoder to obtain the encoding of the sample.Afterward,the decoder maps the encoding into a reconstructed sample.Finally,the algorithm measures the discrepancy between the original sample and the reconstructed sample,which forces the distribution of the feature to be consistent with the prior distribution and obtains a deep feature with strong decision-making ability.The algorithm addresses the curse of dimensionality and nonlinear data structures problem and enhances the classification performance.Furthermore,in order to solve the problem of scarcity of labeled samples and the high labeling cost of HSI,an HSI classification algorithm based on Mixture Generative Adversarial Networks(MGAN)is proposed in this thesis.Firstly,through the adversarial training of the multi-generators and the discriminator of the MGAN,multi-modality and identically distributed samples are generated by MGAN,and the mode collapse has been effectively alleviated.Then,the MGAN with category multi-classifier is applied to classification tasks by the step-by-step adjustment strategy,the classification model of the HSI is constructed.The algorithm greatly prevents over-fitting problems and improves the robustness of the classification in small sample HSI classification.