Narrow road extraction from high-resolution remote sensing images: SWGE-Net and MSIF-Net
Zhebin Zhao Wu Chen San Jiang Yaxin Li Jingxian Wang a Department of Land Surveying and Geo-Informatics,The Hong Kong Polytechnic University,Hong Kong,Chinab School of Computer Science,China University of Geosciences,Wuhan,Chinac Algorithm Development Department,Micro Dimension Technology Limited,Hong Kong,ChinaZhebin Zhao received his B.S. degrees from Arizona State University,and Hainan University,in 2021,and his M.S. degree from The Hong Kong Polytechnic University,in 2023. He is currently a research assistant in the Department of Land Surveying and Geo-Informatics at The Hong Kong Polytechnic University. His research interests include visual perception,localization and mapping.Wu Chen received the Ph.D. degree from Newcastle University,in 1992. He is currently a Professor with the Department of Land Surveying and Geo-Informatics,The Hong Kong Polytechnic University. His current research interests include the global navigation satellite system (GNSS) positioning quality evaluation,system integrity,various GNSS applications,seamless positioning,and indoor positioning.San Jiang received the B.S. degree in remote sensing science and technology and the M.Sc. and Ph.D. degrees in photogrammetry and remote sensing,all from Wuhan University,in 2010,2012 and 2018,respectively. From 2012 to 2014,he worked as an Assistant Engineer with the Tianjin Institute of Surveying and Mapping. From 2014 to 2015,he was with the State Key Laboratory of Information Engineering in Surveying,Mapping,and Remote Sensing (LIESMARS),Wuhan University,as a Research Assistant. He is currently an Associate Professor with the School of Computer Science,China University of GeoSciences. His research interests include image matching,structure-from-motion-based aerial triangulation,and 3-D reconstruction.Yaxin Li received the B.S. degree from Wuhan University,in 2013,and the M.S. and Ph.D. degrees from The Hong Kong Polytechnic University,in 2015 and 2020,respectively. His current research interests include indoor positioning and navigation,simultaneous localization and mapping (SLAM),indoor 3-D modeling,semantic segmentation,and auto building information modeling (BIM) generation.Jingxian Wang received his B.S. degree and M.S. degree from Nanjing University of Aeronautics and Astronautics,in 2015 and 2018. He is currently pursuing the Ph.D. degree with the Department of Land Surveying and Geo-Informatics,The Hong Kong Polytechnic University. His current research interests include multi-sensors fusion,indoor position,pedestrian and vehicle navigation.
DOI: https://doi.org/10.1080/10095020.2024.2405017
IF: 4.278
2024-09-22
Geo-spatial Information Science
Abstract:Accurate and complete road network extraction plays a critical role in urban planning, street navigation, and emergency response. At present, narrow roads are a main feature in most public road datasets. However, the continuity and boundary completeness of the extraction results for these narrow roads are relatively poor, due to their varied shapes, uneven spatial distribution, and the presence of various interfering elements. To address these issues, this study introduces a novel network, the Self-weighted Global Context Road Extraction Network (SWGE-Net), which integrates a dilate block and an improved coordinate attention mechanism to effectively capture the complex details and spatial information of narrow roads. Furthermore, most public road training datasets often lack labels for very narrow roads, this omission leads to poor extraction results for these roads in test datasets. In order to further improve the extraction capability for unlabeled, extremely narrow roads, this study introduces another network called the Multi-scale Information Fusion Road Extraction Network (MSIF-Net), which uses the same encoders as SWGE-Net and has a special module for merging information at different scales. This module, with a dilate block and pyramid pooling-based decoder, makes the network better at recognizing and combining features of different sizes. Experimental results indicate that SWGE-Net outperforms the baseline network with road IoU scores of 71.57% and 60.67% on the DeepGlobe and CHN6-CUG road datasets, respectively an improvement of 18.51% and 5.40%. Meanwhile, MSIF-Net not only exceeds the baseline in road IoU scores for both datasets, but also achieves the best performance in extracting unlabeled, extremely narrow roads in qualitative experiments.
remote sensing