Sanborn Fire Insurance maps are some of the most complete records of the historical built environment that researchers have access to, with building-level data for over 12,000 North American cities contained in large atlases that date back more than a century. These maps have become invaluable resources for digital humanities research, helping to visualize, disseminate, and interpret the history of urban environments (Ross, 1971; Krafft, 1993). Figure 1 shows an example of one such map. A key challenge in working with these maps, however, is ensuring the preservation and accessibility of their information through digitization, which has traditionally been largely contained to tedious handtracing methodologies. While there is existing research to extract building information through machine learning, the methodology is costly in both time and processing power due to the need to train large datasets, which can limit its scalability to broader areas (Tollefson et al., 2021; Lin et al., 2023).
Otto et al. (Mon,) studied this question.