Semantic Segmentation of High Resolution Remote Sensing Images with Extra Context Attention Mechanism

High Resolution Remote Sensing Images (HRRSIs) usually have a larger size compared with natural images. Because of the limitation of GPU memory, it is not possible to train semantic segmentation models on HRRSIs directly. Commonly used methodologies perform training and prediction on cropped sub-images. Thus they fail to model potential dependencies between pixels beyond sub-images. To solve this problem, we firstly propose extra context attention to capture global information from larger receptive fields and discriminative information from surrounding pixels beyond sub-images. Secondly, we apply feature map refinement module to better fuse extra context information and primary semantic information. Finally, we apply channel attention module to improve the performance of the decoder so that features from different levels can be better integrated. Experimental results on ISPRS Potsdam dataset demonstrate the effectiveness of our proposed network for semantic segmentation in HRRSIs.