Residual Attention Mechanism for Construction Disturbance Detection from Satellite Image

Semantic segmentation could not distinguish the spot's contour, which has both the construction disturbance region and original physiognomy. This paper proposed a semantic segmentation network named Residual Attention U-Net (RA U-Net) in the appliance of construction disturbance interpretation. With Inception-v3 as the backbone, the proposed model used the residual attention module replaced the skip connection in U-Net and Conditional Random Field (CRF) as the post-processing. Regarding natural landform as noise, the residual attention module could retain the natural landform. Then, CRF was used to fine-grained outline. The experiment on Standford Background Dataset proved the capability of the proposed model in semantic segmentation. It shows a good performance in the construction disturbance interpretation dataset labelled by ourselves. Moreover, it has been used for the Soil and Water Conservation project held by the local government.