Cross-Layer Multi-Attention Guided Spectral-Spatial Classification of Hyperspectral Images

Deep learning based methods are very popular for hyperspectral image classification. However, those methods usually ignore the fact that discriminative information lies on specific spatial positions and spectral bands. To solve this problem, we introduce the attention mechanism, and propose a cross-layer multi-attention guided classification network (CLMA-Net) for HSIs. First, a backbone network, which is a two-branch convolutional neural network, is developed to extract spectral and spatial features. Then, cross-layer multi-attention modules, which integrate attention information of multiple convolutional layers, are embedded into two branches. As a result, spectral and spatial features are optimized by making the network attend to interested parts. Finally, spectral and spatial features are concatenated and used to predict class label by a fully connected layer. Experimental results demonstrate the effectiveness of the proposed method. The code will be available at https://github.com/mengkai-liu/CLMA-Net.