PDAGCN: Attention-Enhanced Dual Graph Network With Progressive Dilated Convolutions for High-Precision Urban Road Mapping

Road networks, inherently characterized by complex graph structures, play a critical role in urban planning, traffic management, and geographic information systems. Traditional convolutional neural networks (CNNs) have advanced automatic road extraction from remote-sensing images. However, they often fail to capture global topology or to guarantee topological connectivity, especially in large-scale scenes. This limitation arises because most CNNs treat road extraction as a pixel-level segmentation task. Graph Convolutional Networks (GCNs) can effectively model long-range dependencies and preserve road connectivity. However, their reliance on local adjacency matrices restricts the receptive field, limits global-context modeling capacity, and can be computationally inefficient. To address these challenges, we propose Progressive Dilation Attention Graph Convolutional Network (PDAGCN), a novel model that synergistically integrates channel and spatial attention mechanisms, progressive multi-grid dilated convolution, and GCNs to better exploit both local details and global topological relationships. This approach is particularly effective for extracting continuous and complete road networks in high-resolution imagery, be it sparse, irregular layouts in mountainous terrain or dense urban patterns. Experimental results demonstrate that PDAGCN achieves state-of-the-art performance on the publicly available and challenging ZY-3 sparse mountain-road dataset, with a mean Intersection-over-Union (mIoU) of 61.37% and an F1 score of 41.95%, surpassing the performance of other tested methods. On a 3-m very-high-resolution GF-2 aerial imagery dataset, PDAGCN maintains high performance, reaching an mIoU of 81.75% and an F1 score of 79.17%. Qualitative results corroborate that PDAGCN effectively maintains road network connectivity and extracts fine-grained structures in both sparse and dense scenarios. These results highlight the synergy between progressive dilated convolution and graph attention, which together enhance both the continuity and boundary precision of the extracted roads.