Abstract
Open-pit mine road extraction is of significant value for intelligent navigation, yet it faces significant challenges due to unstructured terrain, spectral homogeneity, and physical occlusions. Traditional semantic segmentation methods frequently result in fragmented road masks. This paper proposes RoadTopo-SAM, a topology-aware vectorized mine road extraction framework based on the Segment Anything Model (SAM), transforming road extraction from pixel-level classification to a graph generation task. The framework includes a geometry decoder guided by a topology-constrained loss to generate graph nodes, and a Transformer-based topology reasoning module to explicitly infer long-range connections between nodes. Additionally, a phased differential fine-tuning strategy is employed to prevent the degradation of pre-trained generic representations while effectively adapting the visual foundation model to mining scenarios. Experiments on 0.5-m resolution “Jilin-1” imagery of Tibetan copper mines and the Nanfen iron mine demonstrate that the proposed method achieves state-of-the-art performance with a TOPO F1 of 90.48% and an APLS of 45.35%. Significance testing indicates its superiority over traditional baselines. Cross-domain region-level inference further validates the model’s robustness in macroscopic topological connectivity, providing an effective solution for vectorized mine road mapping.