关键词: 高分辨率遥感图像, 语义分割, 道路提取, 注意力机制

Abstract: Aiming at the problems of high-resolution remote sensing images and road image datasets in the target scene in terms of difficulty in acquiring, high cost, etc., we explore the optimal image resolution of the network models to perform the extraction task at different scales, evaluate the applicability and reliability of each model on road extraction, and provide methodological reference and case study for the road recognition project. In this paper, three classical network models in the field of image segmentation are introduced, the models are trained using public datasets, and the unmanned aerial images of Chuzhou city, Anhui province are used as the test data to perform the road extraction work at different scales, to find out the optimal resolution and model applicability of each model in the new scene, and to evaluate the reliability. The experimental results show that the applicability of the D-LinkNet network model is more prominent in the road extraction task at different scales, the reliability of the DeepLabV3+ network model is poorer, and the optimal resolutions of the road extraction input images for the U-Net and D-LinkNet network models are 1.0 and 0.5 m, respectively.

Key words: high-resolution remote sensing image, semantic segmentation, road extraction, attention mechanism