SCANet: Spatial-Channel Attention Feature Extraction Network for Remote Sensing Images

This study focuses on adding channel and spatial attention modules to a residual block for more effectively interpreting high-resolution remote sensing images. Three feature extraction networks are constructed through parallel and sequential combinations to select a network for high-resolution remote sensing images. Furthermore, these three networks are combined with classification, semantic segmentation, and object detection networks. The results of ablation and comparative experiments on different remote sensing data sets—NWPU-RESISC45, DeepGlobe, and DOTA—reveal that the three constructed methods have achieved good results in semantic segmentation and classification tasks. The construction method of sequential combination (spatial–channel) is the most improved. The effectiveness and general applicability of the proposed feature extraction networks are verified. These feature extraction networks can be used in classification, semantic segmentation, and object detection tasks for remote sensing images and can effectively improve the accuracy of the network.