High-resolution SAR image large area built-up extraction based on the improved BN U-Net – a case study of the North China Plain

Abstract:
Automatic extraction of built-up area from high resolution SAR images is still a challenging problem due to the complexity of buildings’ backscattering. In high resolution SAR images, the buildings have complicated texture and structural features under different topographic conditions. Thus, how to make fine description and accurate classification is a very worthwhile problem. To this end, we proposed an improved BN U-Net network model with migration learning strategy for built-up area extraction in large area and complex environment. The Dice Loss function was used to replace the cross entropy Loss function to solve the imbalance of the building distribution. Experimental results on GF3-FS II SAR dataset demonstrate the effectiveness of the proposed fusion scheme.