Multi-Scale Feature Attention and Transformer for Hyperspectral Image Classification

In hyperspectral image (HSI) classification, convolutional neural networks (CNNs) have shown great potential, but they often overlook multi-scale information and the relationship between features at different scales. To address these problems, HSI classification based on multi-scale feature attention and transformer (MSFAT) is proposed in this paper. Specifically, the proposed MSFAT first extracts multi-scale features by using convolutional kernels of different sizes. Then, the squeeze-and-excitation (SE) module is used to get the attention weight of features at each scale. Next, a simple but effective cross-scale attention module is used to enhance informative features at different scales. Furthermore, to better extract more discriminative features, a transformer encoder is incorporated to capture long-range dependencies between features at different scales. According to experimental results on two common hyperspectral scenes, our proposed MSFAT has demonstrated favorable classification performance when compared with several advanced methods.