【摘要】 针对光学遥感影像道路提取过程中连通性弱、细微分支遗漏以及预测路网与真实路网拓扑结构不一致等问题，本文提出了一种方向引导与拓扑感知的道路提取网络。首先，设计多路径方向引导模块关注多方向连接关系，分离并独立学习不同方向的连通特征，增强道路分支间的联系，提高分割连续性；其次通过全粒度互补特征融合模块融合细、粗粒度特征的互补信息，强化道路细节与语义表达，提高网络对细微分支的捕捉能力；最后设计了拓扑感知损失函数，旨在从拓扑视角探索几何结构差异，增强模型对拓扑变化的敏感度，约束预测路网与真实路网的拓扑一致性。所提模型在DeepGlobe与Massachusetts数据集上的F1值分别为81.95%和79.98%，对比现有最佳方法提升了0.73%和1.5%,IoU为69.35%和66.38%，提升了0.98%和0.66%。实验结果表明，RDTA-Net提取出的道路精确度和完整度优于其他主流方法。此外，在遮挡、噪声以及光照等复杂场景下，仍具有优越的提取效果。更多还原

【Abstract】 To address the challenges of weak connectivity, subtle branch omission, and topological inconsistency between predicted and real road networks in optical remote sensing image road extraction, this paper proposed a road extraction network with directional guidance and topological awareness. First, the multi-path directional guidance module was designed to model multi-directional connectivity relationships. By decoupling and independently learning connectivity features across distinct directions, this module enhanced inter-branch linkages and improved segmentation continuity. Second, the full granular complementary feature guidance module integrated fine-grained and coarse-grained features, reinforcing both road details and semantic representations to strengthen the network’s capability in capturing subtle branches. Finally, a topological awareness function was introduced to quantify geometric structural discrepancies from a topological perspective, thereby constraining the topological consistency between predicted and real road networks. The proposed model achieves F1 scores of 81. 95% and 79. 98% on the DeepGlobe and Massachusetts datasets, outperforming the state-of-the-art methods by 0. 73% and 1. 5%, respectively. The IoU metrics reach 69. 35% and 66. 38%, with improvements of 0. 98% and 0. 66% over existing benchmarks. Experimental results demonstrate that RDTA-Net significantly surpasses mainstream methods in both road extraction accuracy and completeness. Furthermore, it exhibits robust performance in complex scenarios involving occlusions, noise, and illumination variations.更多还原