Answers are grounded in this paper and cite their sources.
Accurate maps of chemical industrial park pipelines are important for infrastructure inventory, safety supervision, emergency response, and maintenance planning, yet pipeline networks are difficult to delineate from high-resolution optical images because they are narrow, low-texture, partially occluded, and embedded in visually complex industrial backgrounds. This study develops CDBNet, a detail-preserving and structure-aware deep network for pipeline mapping from Earth observation images. A frozen DINOv3 ViT-L backbone pretrained on satellite imagery is used as a transferable feature extractor. A cross-layer detail-preserving foundation pyramid adapter converts multi-layer transformer tokens into high-resolution pyramid features, a direction-aware aggregation module enhances elongated structures in horizontal, vertical, oblique, and curved layouts, and a boundary-skeleton-guided refinement module improves contour localization and centerline continuity. We also construct PipelineRS, a pixel-level annotated dataset from GF-2 and GF-7 images covering four chemical industrial parks in China. On PipelineRS, CDBNet achieves 0.9499 precision, 0.9365 recall, 0.9432 F1-score, 0.8925 foreground IoU, 0.9460 mean IoU, and 0.9029 average boundary localization accuracy. Cross-dataset tests on Massachusetts and DeepGlobe road datasets, robustness experiments, and independent cross-regional evaluations further show that the proposed method can produce stable geospatial information for slender industrial infrastructure mapping.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: