Research On Hyperspectral Image Classification Algorithm Based On Graph Neural Network

Classification and recognition of hyperspectral image are one of the most widely used techniques in the areas of remote sensing.The abundant spectral and spatial information in hyperspectral image helps to achieve pixel-level ground object recognition and classification,but this information also leads to data redundancy issues,and cause dimensional disasters for traditional machine learning and deep learning methods,so that there are problems such as complex algorithm models,large computational complexity,poor real-time performance,and reduced resolution.Hyperspectral image can also be considered as graph data,and Graph Neural Network(GNN)can use graph structures to extract,aggregate,and transfer node features of graph data.At this point,GNN has provided a new solution for hyperspectral image classification.This article focuses on the research of hyperspectral image classification methods based on GNN,aiming at the problems of high computational complexity and large classification errors in existing algorithms.The main contents are as follows:(1)Aiming at the problem of small acceptance neighborhood range in traditional Graph Convolutional Network(GCN),a hyperspectral image classification model based on improved GraphSAGE is proposed.Firstly,the SLIC algorithm is applied to segment the original image into hyperpixels,and then a graph architecture is established according to the spatial correlation of the hyperpixels to reduce the amount of data input by the algorithm.Secondly,a random walk graph embedding strategy is used to vectorize graph structure data to sum the mean values of target node features and neighbor node features and solve the problem of local important information loss caused by limited sampling numbers while enhancing the ability to express context information.Simulation experiments are conducted on three publicly available hyperspectral datasets,and performance analysis and comparison with other algorithms showed that the improved GraphSAGE model has high classification accuracy.(2)In order to further improve the classification accuracy,a hyperspectral image classification model based on B-GraphSAGE is proposed.Firstly,partial labeled samples and unlabeled samples are selected to construct an initial image.Secondly,a new biased sampling strategy is adopted to explore more representative feature information by setting hyperparameters p and q controlling the direction of random walk.In addition,by replacing the primary aggregation function with a newly designed edge aggregation function,context information fusion can be more effectively implemented,which to some extent solves the problems of missing edge information and insufficient labeled samples.Simulation experiments were conducted on three publicly available hyperspectral datasets,and performance analysis and comparison with other algorithms showed that the classification results of the B-GraphSAGE model are closer to the real circumstance of the ground.(3)In order to ensure the real-time and efficiency of the hyperspectral image classification,a hyperspectral image classification model based on adaptive graph convolution network is proposed.Firstly,the principal component analysis method is used to lower the dimensions and retain the main band information.Secondly,appropriate attention coefficients are selected to construct a graph structure to achieve adaptive allocation of different neighbor weights and enhance the expression ability of remote distance nodes in the two-dimensional space,so as to dynamically adjust the adjacency matrix.Finally,the adjacency matrix is transferred to the graph convolution layer to train the network model to capture global feature information.Simulation experiments are conducted on three publicly available hyperspectral datasets,and performance analysis and comparison with other algorithms showed that the adaptive graph convolutional network model is competitive in both qualitative and quantitative analysis.