Road Topology Extraction Based on Point of Interest Guidance and Graph Convolutional Neural Network From High-Resolution Remote Sensing Images

Abstract

Road topology networks play a crucial role in expressing road information, as they serve as the fundamental representation of road systems. Unfortunately, in high-resolution remote sensing images, roads are often obscured by buildings, tree trunks, and shadows, resulting in poor connectivity and extraction of topology. To address this challenge, this paper proposes a multilevel extraction method for road topology based on a graph structure. The main contributions of this work are as follows. First, a point of interest (POI) extraction model based on the improved D-LinkNet network is constructed. This model captures relevant information about POIs, such as road intersections and large curvature points. Second, the extracted POIs and the feature maps from the POI model are combined to form triplet information. This information is then fed into a binary classifier, which identifies reliable edges with high confidence levels. These edges contribute to the formation of a subgraph representing the topological structure. Third, a graph convolutional neural network model is employed to predict and supplement the aforementioned subgraphs, resulting in the final road topology. This approach effectively addresses the problem of road interruption caused by occlusion from other ground objects in deep learning-based road topology extraction. The proposed method is supported by both data and experimental results, demonstrating its effectiveness in road topology extraction.