Hyperspectral Image Classification Based On Graph Embedding And Deep Learning

Hyperspectral images(HSIs)compose of hundreds of spectral bands,which are very useful for describing surface features and have been widely used in land surveying,geological and mineral exploration,ecological environment protection,natural disaster monitoring,military target interpretation and so on.However,the difficulty of data analysis is gradually increasing owing to the few labeled samples,higher dimension and correlation.Therefore,it is very necessary to extract the efficient feature of HSIs with higher correlated and redundancy.According to the characteristics of HSIs,we propose a novel method based on graph embedding and deep learning to analyze hyperspectral images.The main contents of this paper are summarized as follows:(1)A novel dimensionality reduction method based on tensorizational graph embedding framework is proposed to classify HSIs.In order to make full use of the spatialspectral information and keep the inherent high-order characteristics of the data unchanged,we propose to use a tensorizational graph embedding framework to extract the spatial-spectral feature.The obtained low-dimensional projection not only contains the spectral information of the sample itself,but also integrates the spatial neighborhood information,thus it is more discriminative.In addition,the sparse and low-rank graph has the ability to express the local and global structure of data,the Laplacian regularized term can maintain the inherent geometric relation of the original space,thus,the sparse and low rank regularized graphs with richer information are constructed.Experimental results on HSIs show that the proposed method is effective in classifying edges and homogeneous regions.(2)A new feature extraction algorithm based on sample expansion and Generative Adversarial Networks is proposed.It is not sufficient to train the deep network because of the few labeled samples,thus we firstly proposed a sample expansion method based on multi-scale superpixel segmentation.Secondly,Generative Adversarial Networks is adopted to classify HSIs,which includes two parts: discriminator using Convolution Neural Network as the feature extraction and classification framework and generator using Multi-layer Perception to capture the data distribution of samples.The two parts are trained alternately,which can improve the discriminant ability of each network.Besides,in order to make the captured data distribution accurate,the loss function of the generator is optimized.Experimental results on HSIs show that the proposed algorithm can effectively improve the precision of classification and recognition.(3)A new spatial-spectral feature extraction algorithm based on improved Generative Adversarial Networks is proposed.In order to fully consider the spatial and spectral information of HSIs,Generative Adversarial Networks is improved: in the generator,a spatial-spectral feature extraction algorithm based on improved Convolution Neural Network is proposed;in the discriminator,a spectral feature extraction algorithm based on Convolution Neural Network is proposed.Finally,the extracted features from the two models are combined in the discriminator and fed to a classifier for classification.In addition,two convolution models are alternately trained,which can achieve higher generalization performance than single network.Experimental results on HSIs show that the proposed spatial-spectral feature extraction method achieves higher performance.