Research map / Graph learning / Knowledge and temporal graphs
Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Collects a global-scale satellite road graph dataset (Global-Scale) and proposes SAM-Road++, which uses node-guided resampling and an extended-line connectivity cue on top of SAM-Road to extract road graphs.
Problem
Road graph extraction from satellite images is limited by scarce, geographically narrow labeled data (mostly urban) and by a train/inference mismatch in two-stage global graph methods such as SAM-Road, plus occlusion of road connectivity.
Approach
Resample training nodes from the predicted road mask (highest-probability points near ground-truth nodes) so the connectivity classifier sees inference-like nodes, and add an extended-line feature sampling the mask along extensions beyond two nodes to handle occlusion.
Key results
- On Global-Scale in-domain, SAM-Road++ reaches 62.33 F1 and 62.19 APLS versus 59.80 F1 and 59.08 APLS for SAM-Road.
- On SpaceNet, SAM-Road++ attains 81.57 F1 and 73.44 APLS, improving over SAM-Road's 80.52 F1 and 71.64 APLS.
- On the out-of-domain Global-Scale split, SAM-Road++ achieves 43.17 APLS versus 40.51 for SAM-Road.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · City-Scale | F1: 77.23 -> 80.01 | Helped |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · City-Scale | APLS: 68.37 -> 68.34 | Hurt |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · SpaceNet | F1: 80.52 -> 81.57 | Helped |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · SpaceNet | APLS: 71.64 -> 73.44 | Helped |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · Global-Scale (In-Domain) | F1: 59.80 -> 62.33 | Helped |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · Global-Scale (In-Domain) | APLS: 59.08 -> 62.19 | Helped |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · Global-Scale (Out-of-Domain) | F1: 46.64 -> 48.34 | Helped |
| Use SAM-Road++ instead of SAM-Road | Road graph extraction · Global-Scale (Out-of-Domain) | APLS: 40.51 -> 43.17 | Helped |
| Add node-guided resampling (ablation) | SAM-Road++ · SpaceNet | F1: 80.89 -> 81.57 | Helped |
| Add node-guided resampling (ablation) | SAM-Road++ · SpaceNet | APLS: 73.22 -> 73.44 | Helped |
| Add extended-line strategy (ablation) | SAM-Road++ · SpaceNet | APLS: 71.90 -> 73.44 | Helped |
| Add extended-line strategy (ablation) | SAM-Road++ · SpaceNet | F1: 81.77 -> 81.57 | Hurt |
Surprising findings
- Precision of SAM-Road++ on City-Scale (88.39) is slightly lower than SAM-Road (90.47), attributed to node-guided resampling making the model accept nodes deviating from ground truth.
Limitations
- Out-of-domain performance is noticeably lower than in-domain performance on Global-Scale.
- City-Scale test set (27 images) makes APLS results unreliable according to the authors.
- Precision lags SAM-Road on City-Scale.
Experimental setup
Datasets: Global-Scale City-Scale SpaceNet
Benchmarks: City-Scale SpaceNet Global-Scale
Metrics: TOPO F1 TOPO precision TOPO recall APLS
Baselines: Sat2Graph RNGDet RNGDet++ SAM-Road SegFormer
Builds on: SAM-Road SAM Sat2Graph RNGDet++
Closest papers by meaning
- TopDiG: Class-Agnostic Topological Directional Graph Extraction From Remote Sensing Images (CVPR 2023 · similarity 0.85)
- Patched Line Segment Learning for Vector Road Mapping (AAAI 2024 · similarity 0.84)
- OpenSatMap: A Fine-grained High-resolution Satellite Dataset for Large-scale Map Construction (NeurIPS 2024 · similarity 0.82)
- Bezier Everywhere All at Once: Learning Drivable Lanes as Bezier Graphs (CVPR 2024 · similarity 0.80)
- Structure-Adaptive Multi-View Graph Clustering for Remote Sensing Data (AAAI 2025 · similarity 0.79)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.