LDS2MLP: A Novel Learnable Dilated Spectral-Spatial MLP for Hyperspectral Image Classification

The multilayer perceptron (MLP) has gained widespread popularity and demonstrated outstanding performance in hyperspectral image (HSI) classification in recent years. However, the native MLP architecture and its variants are insufficient in expressing fine spatial structural information and long-range dependencies. To this end, a learnable dilated spectral-spatial MLP (LDS$^{2}$MLP) is proposed for HSI classification. LDS$^{2}$MLP applies the learnable dilated receptive field and grouped MLP to extract more discriminative spectral-spatial features with fewer computational costs, which improves the classification performance. Specifically, a plug-and-play spectral-spatial mixing (S$^{2}$Mixing) block is designed to aggregate grouped spectral detail information and fine spatial structural features. The S$^{2}$Mixing block consists of two feature extraction modules operating on the spectral and spatial domains. The spectral grouping mixer module captures subtle spectral differences through grouped MLP. The spatial dilating with learnable Spacing mixer module employs a dilated receptive field with learnable spacing to enhance the refined expression of spatial structural features. Extensive experiments on four public HSI datasets illustrate that the proposed LDS$^{2}$MLP outperforms state-of-the-art deep learning models in classification performance. In addition, the proposed model is shown to be efficient and generalizable in HSI classification with limited samples.