‘Give me a point and I will place the map’: automated georeferencing historical maps using spatially explicit AI

Georeferencing provides a foundation for transforming historical maps into usable data for geographic analysis. Existing methods remain limited in their ability to integrate spatial semantics with geometric feature extraction, accommodate complex street geometries and stay robust under sparse or uneven ground control points (GCPs). These limitations restrict the scalability and reliability of automated georeferencing for historical map collections. Building on advances in vision–language models (VLMs), this study reconceptualizes georeferencing as a spatially contextual inference problem rather than a geometric fitting task. We propose a framework that integrates geometric feature extraction, semantic interpretation and spatial dependence. Specifically, the framework (1) employs VLMs for topology-aware semantic enrichment of GCPs, (2) supports complex morphologies through convolutional neural network–based spatial anchor extraction with graph-based reasoning and (3) stabilizes georeferencing under sparse and uneven GCP conditions using a projection-similarity prior. Evaluation of Sanborn maps from Chicago showed that 84.3% of map sheets were successfully georeferenced. The method achieved an 80% success rate with fewer than three GCPs and a 74.3% success rate under uneven GCP distributions, demonstrating robustness and scalability for historical map georeferencing. This study provides a proof of concept that can be generalized to a wider class of historical cartographic materials.