Zhang, Xubing ， Li, Can ， Shao, Shiwei ， Peng, Yuxin
2021-01-01 null null 12129(卷), null(期), (null页)
The automatic extraction of the roads from the remote sensing images is significant for the monitoring of the sandstorm hazards to the traffic arteries of the desert area. Since the current pixel-based road or the object-oriented extraction methods are easy to cause the noises or the adhesion phenomena, the U-Net deep learning network is applied to extract the desert roads from the Google Earth, GF-2 and JiLin-1 image datasets in this paper. Firstly, in order to improve the generalization ability of the U-Net network model, the datasets are expanded by means of rotating, mirroring, contrast stretching, and intensity dithering. Then under the constraints of the hyperparameters, the U-Net model is built and trained until the loss function value tends to be stable. Finally, the U-Net algorithm is adopted to extract the highways pass through the Takramakan desert and the Kumtg desert areas in western China. The experimental results demonstrate that the U-Net algorithm is efficient and performable.