Abstract
Cross-modal road extraction from high-resolution remote sensing imagery (RSI) and auxiliary trajectory/LiDAR data remains challenging due to cross-modality inconsistency and complex background interference. In this article, we propose D2CRENet, a diffusion-driven cross-modal road extraction framework that integrates structural condition-guided learning with frequency-spatial-channel alignment to address modality inconsistency and topology degradation in complex road scenes. Unlike conventional discriminative segmentation models, D2CRENet first formulates cross-modal road extraction as a conditional generative reconstruction process within a denoising diffusion probabilistic model (DPM). By incorporating cross-modal structural guidance into latent noise prediction, the framework enables progressive road-mask reconstruction. To improve cross-modal consistency, we introduce a Cross-Frequency-Spatial-Channel Mamba (CFSCM) module that performs multidimensional alignment between RSI and GPS/LiDAR features, effectively reducing redundancy while strengthening boundary-sensitive representations. In addition, a multimodal conditional guidance (MCG) module injects fused structural guidance into the reverse diffusion trajectory, facilitating globally coherent mask reconstruction and refined high-frequency detail recovery. Extensive experiments on multiple datasets demonstrate that integrating structured cross-modal interaction with conditional diffusion modeling consistently improves both segmentation accuracy and road-network connectivity. The code will be available at https://github.com/zhouyiqingzz/D2CRENet
| Original language | English |
|---|---|
| Article number | 5646618 |
| Journal | IEEE Transactions on Geoscience and Remote Sensing |
| Volume | 64 |
| DOIs | |
| State | Published - 2026 |
Keywords
- Cross-Frequency-Spatial-Channel Mamba (CFSCM)
- cross-modal road extraction
- multimodal conditional guidance (MCG)
Fingerprint
Dive into the research topics of 'Diffusion-Driven Cross-Modal Road Extraction via Structural Guidance and Cross-Modal Alignment'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver