Research On Deep Subspace Clustering Algorithm For Large-scale Hyperspectral Images Classification

Hyperspectral images are characterized by high spectral resolution,many wavebands and rich information.Land-cover classification using hyperspectral images is one of the most classical applications in remote sensing.Subspace clustering,as a clustering method suitable for processing high-dimensional data,has been widely used in land-cover classification of hyperspectral images.With the rapid development of deep neural networks,deep subspace clustering methods have made a breakthrough in the field of natural images.However,there are two major problems that limit the application of existing deep subspace clustering for landcover classification of large-scale hyperspectral images: First,the hyperspectral dataset has high dimensionality,complex spatial-spectral structure,and usually contains a large number of unlabeled "background information" samples,which makes it difficult to extract discriminative null spectral features from the feature extraction structure in the field of natural images.The feature extraction structure in the natural image domain is difficult to extract discriminative spatial-spectral features,resulting in unsatisfactory land-cover classification.Second,deep subspace clustering requires all samples to be input for feature extraction at one time.Subsequent spectral clustering requires singular value decomposition of the affinity matrix of all samples,which causes a sharp increase in computational and storage costs with the increase in the number of samples,and makes it difficult to unsupervised land-cover classification on large-scale hyperspectral data sets.Based on the above problems,the following researches are carried out in this thesis:(1)In this thesis,a deep low-rank graph convolutional subspace clustering based hyperspectral image land-cover classification method is proposed to enhance the discriminability of the spatial-spectral features.Firstly,a 1D convolutional auto-encoder is designed to extract spectral band features,which is used to maintain the proportion of original band information.Secondly,a feature fusion strategy and a low-rank self-representation layer are designed in the network to extract more discriminative joint spatial-spectral features.Finally,the joint spatial-spectral features are mapped from Euclidean space to graph space by graph convolution technique for graph convolution subspace clustering,which is used to alleviate the interference of "background information" samples of hyperspectral images on land-cover classification.Experiments on three small-scale hyperspectral datasets demonstrate the effectiveness of the proposed method.(2)In this thesis,a joint framework of deep subspace clustering and graph convolutional based method for hyperspectral images is proposed to further improve the discriminability of the spatial-spectral features and unsupervised land-cover classification performance on largescale hyperspectral datasets.Firstly,a graph attention joint auto-encoder network is designed for small-sample clustering to fuse the extracted graph structure features and spatial patch features to further improve the quality of joint spatial features,and then improve the accuracy of small-sample clustering.Secondly,a mini-batch graph convolutional network is designed to select results with high confidence and low uncertainty as pseudo labels using an uncertainty pseudo label selection strategy,and self-supervised training on a complete large-scale hyperspectral dataset to obtain land-cover classification results.Experiments demonstrate that the proposed method can achieve good results on both small-scale and large-scale hyperspectral datasets.(3)In this thesis,a masked autoencoder based superpixel contrast subspace clustering method is proposed to enable more accurate unsupervised land-cover classification on largescale hyperspectral datasets.Firstly,a spatial patch mask auto-encoder is designed based on the characteristics of hyperspectral spatial patches to improve the discriminability of the null spectral features,while self-attention feature extraction is performed on the constructed hyperspectral spatial patches with low computational consumption.Secondly,by preprocessing the complete large-scale hyperspectral image dataset with superpixel segmentation,the number of input data is reduced to alleviate the high time and space complexity problem when processing large-scale hyperspectral data by deep subspace clustering.In addition,a unique positive and negative sample construction method is designed to learn two self-representation coefficient matrices focusing on different dimensions using contrast learning and weighted fusion to improve their quality,and the final land-cover classification results are obtained by applying spectral clustering to the fused self-representation coefficient matrices.Experiments on three large-scale hyperspectral datasets demonstrate that the accuracy of the proposed method has reached a leading level.