Improved U-Net with Global Attention Block for Hyperspectral Image Classification

Hyperspectral image (HSI) classification aims to predict the class of each pixel. The deep learning methods represented by convolutional neural network (CNN), fully convolutional network (FCN) and vision transformer (ViT) have achieved great success in HSI classification. However, CNN and FCN methods have the problem of limited receptive field, capturing global features requires the accumulation of multi layers of the network. ViT introduced self-attention mechanism to solve this problem, but the performance of ViT with limited train samples is not better than CNN. To solve the above problems, this paper introduces the global attention block (GAB) for feature extraction of HSI. GAB can be embedded in the existing CNN and FCN. In addition, this paper improves U-net for HSI classification, and combines with GAB to form GAB-UFCN. Experiment results of the proposed GAB-UFCN on three popular HSI datasets show better performance compared with the state-of-the-art methods.