A hybrid-attention semantic segmentation network for remote sensing interpretation in land-use surveillance

Remote sensing interpretation for surveillance of land use often needs to mark out construction disturbance on satellite imagery, such as illegal buildings or spoil area. These disturbance region annotated by a set of surveillance rules contain the corresponding image characteristics which are regarded as semantic information in computer vision. Different from the natural Landscapes interpretation, the semantic information of construciton disturbance region shows more complex to extract with lack of available training dataset and interference of the various sizes of targets. This paper proposes a hybrid attention semantic segmentation network (HAssNet) which can extract the target and its surroundings through a large receptive field for multi-scale targets. Based on the full convolutional networks (FCN), spatial attention mechanism is firstly introduced to acquire the position of segmentation target with the global correlations, so that the small targets in large scale scene are guaranteed not to be omitted in semantic features extraction. Secondly, channel attention mechanism is designed to assign higher weights to task-related channels for semantic consistency. Experimental results on an open remote sensing dataset show that HAssNet achieves average 6.7% improvement in mIoU than the state-of-the-art segmentation networks. In a land use surveillance project, HAssNet shows considerable performance compared with manual interpretation.