SWCARE: Switchable learning and connectivity-aware refinement method for multi-city and diverse-scenario road mapping using remote sensing images

Abstract

Accurate and efficient mapping of road networks is crucial for evaluating urban development, transportation accessibility, and environmental impact. However, existing road extraction methods utilizing remote sensing images suffer from limited generalization ability and object occlusion, resulting in fragmented and discontinuous segmentation. Consequently, these limitations impede the practical applicability of these methods in multi-city and diverse-scenario road extraction applications. To address these challenges, we propose SWCARE, a road extraction method with SWitchable learning and Connectivity-Aware REfinement. We propose a flickering module with switchable learning which considers four types of auxiliary supervision information, namely road edge, road centerline, road corner, and road orientation, to improve the feature representativeness ability and enhance road extraction results. Furthermore, the proposed connectivity-aware refinement module aims to enhance the completeness and connectivity of road networks, thereby augmenting their practicality in real-world scenarios. We evaluate the performance of SWCARE on commonly used public road datasets and our constructed Large-And-Complex Road Dataset (LACRD). Our approach surpasses the state-of-the-art road extraction method in terms of both pixel-oriented and connectivity-oriented metrics, achieving a 4.41% higher Intersection over Union (IoU) and a 3.57% higher Average Path Length Similarity (APLS), respectively.

Keywords

• Deep learning
• Multi-task learning
• Iterative refinement
• Road network mapping
• Intelligent interpretation