The content of this article comes from the "Surveying and Mapping Bulletin" No. 6 in 2024, drawing review number: GS Jing (2024) No. 1024
Road extraction from multi-scale UAV remote sensing images based on deep learning
ZHANG Wei1, ZHANG Chaolong2, WANG Benlin2, CAI Anning3
1. School of Architecture, Anhui University of Science and Technology, Bengbu 233030, China; 2. College of Geographic Information and Tourism, Chuzhou University, Chuzhou 239000, China; 3. School of Tourism and Social Management, Nanjing Xiaozhuang University, Nanjing, Jiangsu, 211171
Funds: National Natural Science Foundation of China (52078237); Provincial Natural Science Research Key Project of Anhui Universities (KJ2021A1083; KJ2021A0860); Key Construction Discipline of Anhui University of Science and Technology (XK-XJJC001)
Key words: high-resolution remote sensing images, semantic segmentation, road extraction, attention mechanism
Citation format: ZHANG Wei, ZHANG Chaolong, WANG Benlin, et al. Road Extraction from Multi-scale UAV Remote Sensing Images Based on Deep Learning[J]. Bulletin of Surveying and Mapping, 2024(6): 77-81.doi: 10.13474/j.cnki.11-2246.2024.0614
Abstract:In order to solve the problems of difficulty and high cost in obtaining high-resolution remote sensing images and road image datasets in target scenes, this paper explores the optimal image resolution of the network model to perform the extraction task at different scales, and evaluates the applicability and reliability of each model in road extraction, so as to provide method reference and case reference for road identification engineering. Three classical network models in the field of image segmentation were introduced, and the public dataset was used for model training, and the images of Chuzhou City, Anhui Province taken by UAV aerial photography were used as experimental data to extract roads at different scales, so as to find out the best resolution and model applicability of each model in the new scene, and to evaluate the reliability. The experimental results show that the D-LinkNet network model has strong applicability in road extraction tasks at different scales. The reliability of the DeepLabV3+ network model is poor; The optimal resolution of the road extraction input image of the U-Net and D-LinkNet network models is 1.0 m and 0.5 m, respectively.
About author:ZHANG Wei (1983—), male, Ph.D., associate professor, mainly engaged in remote sensing application and urbanization research. E-mail:[email protected] Corresponding author: Wang Benlin. E-mail:[email protected]
First instance: Yang Ruifang review: Song Qifan
Final Judge: Jin Jun
information