Research On Hyperspectral Image Classification Based On Graph Neural Network

With a large amount of high-resolution spectral information and spatial context information,hyperspectral images(HSIs)can provide rich feature attributes of land covers,and have been widely used in various fields,such as military reconnaissance,environmental monitoring,and resource exploration.Classification is one of the research hotspots in the field of HSI processing.In recent years,graph neural networks(GNNs)have shown great potential in HSI classification with their ability to capture the spatial-spectral correlations using graph structure and geometric attribute information.As such,by focusing on the structure characteristics of HSI data,this thesis carries out research on GNN-based classification methods,and the major research work is summarized as follows:Aiming at the influence of spectral variability and noise interference,this thesis designs a novel Propagation-Regularization(P-reg)constrained graph attention auto-encoder(PCGAAE)for HSI classification.The superpixel graph attention auto-encoder model is firstly constructed to solve the limitation of fixed weights in the graph convolution process.And then,the constraint of P-reg graph regularization term is introduced into the encoder to provide additional graph structure supervision information for the whole model,so as to effectively filter out the noise interference and pay more attention to the key node information,thus extracting more discriminative spatial-spectral features and improving the classification performance of HSIs.Experimental results illustrate the effectiveness of the proposed PCGAAE model.To solve the problems of single convolution scale and insufficient utilization of spatial information in existing GNNs,a new dual-channel multiscale hypergraph convolution network(DMHCN)is constructed for HSI classification.The spectral hypergraph convolutional network and the spatial multiscale hypergraph convolutional network are respectively designed as the spectral and spatial channels to exploit the higher-order correlation between hypergraph nodes,extract the spectral-and multiscale-enhanced spatial-spectral features,and improve the classification accuracy of the whole model.Experimental results on three widely used HSI datasets show that the proposed model is superior to other deep learning algorithms.