An Automated Georeferencing Workflow for Processing Historical Sanborn Fire Insurance Maps Using Object Detection and Optical Character Recognition

Providing georeferenced scanned map images through a geoportal facilitates their use in GIS software and encourages their utilization in research. However, using traditional manual georeferencing workflows on large map collections can be prohibitively time consuming and expensive. Here, we describe a newly developed scripted workflow for automating the georeferencing of a large collection of digitized historical Sanborn fire insurance map sheets. This scalable multi-step process involved using machine learning to train an object detection model capable of identifying street intersections, optical character recognition to locate corresponding street labels, and a scripted process for determining associated geographic coordinates. Map sheets with more than three intersections with identifiable coordinates were able to be georeferenced without any manual input by relying on these features as ground control points. Using this approach, we were able to accurately georeference 14% of processed map sheets, validating this method as a viable but currently limited option for automated georeferencing.