Skip to content
Open access

DGNet: A Direction-Aware and Discrepancy-Guided Reconstruction Network for Road Extraction From High-Resolution Remote Sensing Images

2026 · IEEE Geoscience and Remote Sensing Letters · Vol 23, pp. 2507605-2507605 · 0 citations · 17 references

Abstract

Road extraction from high-resolution remote sensing images is essential for urban planning, traffic management, and disaster assessment. However, occlusions caused by buildings, trees, and shadows often lead to broken road segments, while curved roads and intersections introduce complex directional variations, making connectivity-preserving extraction challenging. Existing U-Net-like encoder–decoder methods mainly rely on direct skip fusion but often ignore the degradation of road evidence across skip transmission, bottleneck inference, and decoder reconstruction. To address this issue, we propose DGNet, a direction-aware and discrepancy-guided reconstruction network that formulates broken-road recovery as progressive structural reconstruction rather than simple multistage feature aggregation. Specifically, the cross-layer feature enhancement module (CFEM) refines skip features by aligning cross-layer information and enhancing road-oriented directional cues. The dual-context collaboration module (DCCM) promotes semantic–structural contextual interaction at the bottleneck for occluded-road inference. The discrepancy-guided directional decoder (DGD-Decoder) further exploits skip–decoder response differences and jointly activated responses for adaptive directional reconstruction. Experiments on the DeepGlobe and Massachusetts road datasets show that DGNet achieves superior $F1$ -score, IoU, and APLS. Compared with CGCNet, DGNet improves them by 1.14%, 1.49%, and 1.19% on DeepGlobe and by 0.95%, 1.29%, and 1.76% on Massachusetts, respectively.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.