Plan2Map benchmark lets AI reconstruct geospatial boundaries from planning documents
New system reads planning records to draw accurate property boundaries from text and maps
A team led by Fabian Degen and Oishi Deb from the University of Oxford has released Plan2Map, a multimodal benchmark designed to test AI systems on the task of reconstructing geospatial boundaries from text-heavy planning records. The dataset comprises 208 real UK planning documents, each containing notice text, schedules, map plates, labels, and boundary annotations. Systems must produce a valid GeoJSON boundary without seeing the reference—a challenging exercise in document-grounded spatial reasoning.
To tackle this, the authors propose GeoPlanAgent, a modular agent that decomposes the task into six steps: evidence extraction, localization, map registration, boundary segmentation, projection, and verification. The agent uses a tool-in-the-loop approach, calling geospatial libraries and vision models as needed. Tested on Plan2Map, GeoPlanAgent achieves a mean Intersection over Union (IoU) of 0.736 and a median of 0.904, with nearly 68% of predictions scoring above 0.8 IoU. This far exceeds direct VLM-to-GeoJSON baselines, which remain unreliable. Error analysis shows that most failures stem from poor localization and map registration, while supervised segmentation substantially improves pixel-level accuracy. Plan2Map offers a concrete testbed for automating boundary extraction in urban planning, real estate, and land management.
- Plan2Map benchmark includes 208 UK planning records with text, maps, and annotations for boundary reconstruction
- GeoPlanAgent achieves 0.904 median IoU and 67.8% of predictions above 0.8 IoU
- Remaining errors are concentrated in localization and map registration, not segmentation
Why It Matters
Automating boundary extraction from planning documents could save hours of manual work for urban planners and surveyors.