KDGraph: A Keypoint Detection Method for Road Graph Extraction from Remote Sensing Images

Road graph extraction from remote sensing images is essential in navigation and urban planning. However, shadows and occlusions in these images frequently disrupt the continuity of road representations, resulting in fragmented and poorly connected road graphs extracted by existing methods. To overcome these challenges, we introduce KDGraph, a novel keypoint detection method for road graph extraction from remote sensing images. Specifically, keypoints are defined as vertices situated at the endpoints, corners, or junctions of roads, characterized by their positions and multiple directional attributes. In addition, we develop a greedy parsing algorithm to connect these keypoints and construct the road graph based on their directional information. The key innovation of our method lies in reformulating road mask extraction as a keypoint detection task. By positioning keypoints in areas less affected by shadows and occlusions, KDGraph effectively reduces their impact on road graph connectivity. Extensive experiments are conducted using the SpaceNet3 dataset and a newly constructed Shadow-Occluded Road Graph Extraction (SoR) dataset, which includes road segments from 15 cities with varying degrees of shadows and occlusions. A patch expansion strategy is introduced for large-scale inference on the SoR dataset. Results show that KDGraph outperforms all comparative methods in both generalization and the capability of handling shadows and occlusions. Code and SoR dataset available at: https://github.com/ruoyxue/KDGraph.