A Residual Attention Network with Spectral and Spatial Selective Kernel for Hyperspectral Image Classification

In recent years, convolutional neural networks have been widely used in hyperspectral image (HSI) classification, with spectral-spatial dual-branch network models being particularly popular. However, these networks typically use convolutional kernels with fixed-size receptive fields. This approach is unable to adaptively capture multiscale detailed features, affecting the classification results. To alleviate this issue, this paper proposes a spectral selective kernel (SK) block and a spatial SK block that can adaptively extract multi-scale features on the spectral and spatial dimensions respectively. Following this, we design a residual attention network with spectral and spatial selective kernel (SKRN), which has two branches for spectral and spatial processing respectively. On each branch, we initially use a SK block and a res2 block to simultaneously conduct branch feature extraction. After applying residual summation, we then reconstruct the feature map through global context (GC) attention. Notably, different from the traditional residual block, the res2 block also has the capability of extracting multi-scale features. Hence, this dual multi-scale feature extraction method using both SK block and res2 block makes the local detail feature extraction of the image more perfect. A series of experiments on Indian Pines (IP) and University of Pavia (UP) datasets demonstrate that our model outperforms other related algorithms, and that the combination of modules in our model achieves the best effects.