Hyperspectral Image Classification Combining Multi-Scale Dilated Convolution and Residual Feature Extraction

The high-dimensional information of hyperspectral images makes the classification task more complicated, especially in fine-grained feature extraction and deep network training. In this paper, a multi-scale dilation convolution and Residual Feature Extraction (MSDCRFE) method for hyperspectral image classification is proposed with designing multi-scale dilation convolution module (MSDCM) and residual feature extraction module (RFEM). MSDCM captures rich spatial features through dilated convolutions with different scales, and RFEM mitigates gradient disappearance through residual connections to enhance feature transfer ability and fine-grained feature extraction. Experimental results show that the proposed method can yields improved performance compared with several state-of-the-art methods for hyperspectral image classification.