RoadCorrector: A Structure-Aware Road Extraction Method for Road Connectivity and Topology Correction

Road extraction from high-resolution remote sensing images has been an important research problem for decades. Despite the breakthrough progress of road extraction studies in recent years due to the rapid advancement of deep learning techniques in a remote sensing domain, the vast majority of methods pay little attention to the basic structure of roads. Especially in complex scenes such as a tree or shadow occlusion and stacking of multiple roads, existing road extraction methods still suffer from generating broken road surfaces, inaccurate topology, and connections. In this work, we propose a novel structure-aware road extraction method, named RoadCorrector, which solves the above limitations of existing methods by adding structure-related assistance branches and two correction modules. Specifically, RoadCorrector encompasses three main stages: road segmentation, connectivity refinement, and topology correction. First, we design a multibranch road extraction network (MBRE-Net) that combines intersection and frequency domain information to efficiently extract road masks. Subsequently, a connectivity refinement strategy based on energy function is introduced to deal with the discontinuity of roads in the occluded and intersection regions. Finally, the topology correction module aims at constructing vectorized road networks with more accurate connection relations. Experimental results on several public datasets show that our RoadCorrector achieves remarkable improvements compared with state-of-the-art methods, with the $F1$ -score and intersection over union (IoU) improved by 3.3%-4.5% and 2.0%-5.1%, respectively. Moreover, the road network extraction results of RoadCorrector have more accurate topological properties, demonstrating its great potential in actual application scenes. The code and dataset of RoadCorrector will be released at https://github.com/Lijp411/RoadCorrector.