Global-Local Channel Attention for Hyperspectral Image Classification

Hyperspectral image classification (HSIC) assigns a pixel-wise semantic label leveraging the rich information in the broad spectral band. However, most of the existent HSIC algorithm fail to take advantage of varying importance of different channels, which hinder the further improve of the performance. To this end, we originally propose a Global-Local Channel Attention (GLCA) module to assist the process of selecting useful channels while suppressing others. Working in a plug-and-play fashion, GLCA precisely re-calibrate channel-wise feature responses in a pixel-wise manner, flexible enough to be applied to any existing depth-based HSIC model with little additional computational cost. Rich experiments prove the effectiveness of our algorithm, and to the best of our knowledge, we establish new state-of-the-arts on multiple HSIC datasets.