Using contour loss constraining residual attention U-net on optical remote sensing interpretation.

Using deep learning in remote sensing interpretation could reduce a lot of human and material costs. Semantic segmentation is the main method for this task. It can automatically outline the objects and it has recently achieved great success in remote sensing images. However, in the appliance of remote sensing interpretation, the accuracy of contour largely determines the evaluation of remote sensing interpretation. Though the current loss functions reflect the segmentation performance, they could not guide the model to optimize itself toward a more precise contour. This paper proposed an exactly defined contour loss (CL) for remote sensing interpretation with Residual Attention U-Net (RA U-Net) as the main framework. The RA U-Net uses the residual attention module as the skip connection layer. It enhances the judgment of U-Net. In CL, image processing methods are used to extract the contours of the foreground. And elements-sum and elements-subtract operations are used to transfer the contour information to a matrix of the same size as label images. Then, these matrices would be the weights for CE. By assigning different weights for different elements in different regions, this function will guide the model to reach a balance between accurate segmentation results and precise contours. The experiment on open datasets shows a good performance. The proposed model was also trained on the Construction Disturbance Dataset collected from Jiang Xi Province, China. The dataset was labeled manually. The evaluation enhanced a lot on the Construction Disturbance Dataset and the IoU on two datasets increased 1 % to 2 % when using CL as the loss function. This paper also compared the proposed method with other state-of-the-art methods and the results showed extensive effectiveness. [ABSTRACT FROM AUTHOR]