Automatic Extraction of Built-Up Areas for Cities in China from GF-3 Images Based on Improved Residual U-Net Network

In this paper, an automatic extraction method of multi-type built-up areas in SAR images is proposed. In order to adapt to the architectural differences in different regions, we improve the residual U -Net, one is to introduce multi -scale pyramid structure into the structure, the other is to use the diss loss function. The proposed method was verified using GF-3 SAR data in four regions of China and compared with FCN and PSPNET. Finally, the accuracy evaluation results show that the overall accuracy of the extraction results is greater than 86%, indicating that the method has a certain application potential.