Remote Sensing Scene Classification with Dual Attention-Aware Network

Remote sensing scene classification is of great importance to remote sensing image analysis. Most existing methods based on Convolutional Neural Network (CNN) fail to discriminate the crucial information from the complex scene content due to the intraclass diversity. In this paper, we propose a dual attention-aware network for remote sensing scene classification. Specifically, we use two kinds of attention modules (i.e. channel and spatial attentions) to explore the contextual dependencies from the channel and spatial dimensions respectively. The channel attention module intends to capture the channel-wise feature dependencies and further exploit the significant semantic attention. On the other hand, the spatial attention module aims to concentrate the attentive spatial locations and thus discover the discriminative parts inside the scene. The outputs of two attention modules are finally integrated as the attention-aware feature representation for improving classification performance. Experimental results on RSSCN7 and AID benchmark datasets show the effectiveness and superiority of the proposed methods for scene classification in remote sensing imagery.