Hyperspectral Image Classification with Dynamic Spatial-Spectral Attention Network

Abstract

In this paper, we propose a novel dynamic spatial-spectral attention network (DSSAN) for the classification of hyperspectral images (HSIs). The input-dependent dynamic convolution is utilized to build a spatial-spectral attention module to adaptively recalibrate spatial- and channel-wise feature responses, extracting more discriminative features. We further integrate the attention module with the residual learning framework for deep feature learning. Experimental results on two popular HSI benchmark datasets verified the superiority of our DSSAN due to higher classification accuracy and fewer parameters.