Research On Unmixing And Classification Algorithm For Hyperspectral Remote Sensing Imagery

In recent years,with the research and development of hyperspectral imaging systems,the spatial,spectral,and temporal resolution of hyperspectral remote sensing images have also been continuously improved.To widely using hyperspectral images in various application fields,spectral unmixing and hyperspectral classification are very important for ground objects identification and analysis of hyperspectral remote sensing images.However,the hyperspectral remote sensing image during acquisition process is susceptible to interference from the external environment and the spectrometer,which brings huge difficulties to unmixing and classification.To improve the accuracy of the results of ground object identification and analysis,fusion of spatial and spectral features in hyperspectral remote sensing images has become an important development trend.Because it is very difficult to comprehensively extract spatial features of hyperspectral images,this article mainly focuses on how to extract comprehensive and effective spatial information and improve the unmixing and classification.The specific work content can be summarized as follows:(1)To solve mixed pixels in hyperspectral remote sensing images,a hyperspectral remote sensing image unmixing method is proposed,which based on a nonnegative matrix fraction with spatial information constraints.All the pixels in the superpixel after the superpixel segmentation process are used to construct the spatial group sparsity constraint.At the same time,superpixels are used as a guide to adaptively set the search area for extracting non-local spatial information,which is used construct non-local spatial constraints.Under the framework of non-negative matrix transformation,based on a linear mixed model,these two constraints are combined to find the endmembers and abundances that are close to truth.(2)To solve the limited labeled pixels in hyperspectral images,a novel semi-supervised classification method for hyperspectral images is proposed,which is based on graph regularization with spatial and spectral features.Based on the anchor graph theory,the superpixel centers are used as the anchor to construct graph,called superpixel-based anchor graph,and the graph regularization model is introduced into the objective function for hyperspectral image classification.To improve the classification accuracy,the local grouped order pattern and the non-local binary pattern are used to extract the finer and more effective local and non-local spatial features of the hyperspectral image,which are concatenated with the spectral features to form new feature data for classification.(3)To solved the shortcomings of hyperspectral remote sensing images such as high complexity,high data dimension,big data and large scale,a based on graph attention network is proposed for hyperspectral image classification.In this network,the graph attention network layer and the depthwise separable network convolutional layer are used to construct the feature extraction module.The graph attention network layer perform convolution on the accurate graph,which is constructed by superpixel,replacing pixels,and guided by labelled pixels.The depthwise separable network convolutional layer is used to extract the local spatial and spectral features of the hyperspectral remote sensing image for improving the classification ability of the network.