Hyperspectral Image Classification Based on Three-Dimensional Dilated Convolution and Graph Convolution

Hyperspectral images with high dimensionality, strong inter-band correlation and high spectral resolution make the research of existing classification methods a great challenge. Typical convolutional neural network models cannot capture the feature information of irregular or inhomogeneous regions of feature classes in images, and the features extracted by using a single network model lack diversity and cannot provide the best classification results. To address the above problems, a hyperspectral image classification method that fuses 3D dilated convolution and graph convolution is proposed. Firstly, a 3D dilated convolution network model is constructed to extract multi-scale null-spectral features using dilated convolution with different dilated parameter sizes; secondly, a neighborhood relationship-based graph convolution neural network model is established to obtain spatial structure contextual features by aggregating the neighborhood feature information of graph nodes; then, to improve the feature representation capability, the extracted deep null-spectral features are fused with spatial contextual features Finally, the proposed method is compared and analyzed with seven related classification methods on two hyperspectral datasets, Indian Pines and Pavia University.