Creating 3D Models of Historic Urban Neighborhoods Using Machine Learning

By Ashutosh RoyReviewed by Susha Cheriyedath, M.Sc.Jul 5 2023

Cities in the United States have experienced significant transformations throughout the 20th century. The advent of tramway and the individual automobile significantly impacted the long-standing patterns of urban development that were limited by pedestrian transportation as the predominant mode of travel.

The Sanborn Fire Insurance maps possess a substantial amount of detailed information at the building level pertaining to cities in the United States, with records dating as far back as the late 19th century. Extracting building-level information from Sanborn maps poses a significant challenge due to the considerable quantity of map entities and the absence of suitable computational methods for detecting these entities in an effective and efficient manner.

A research paper published in the journal PLOS ONE presents a novel approach to developing a scalable workflow that leverages machine learning techniques for the purpose of identifying building footprints and their corresponding properties on Sanborn maps. The aforementioned information has the potential to be efficiently utilized in the development of three-dimensional visual representations of historic urban neighborhoods, as well as in the provision of insights for urban transformations.

What are Sanborn fire insurance maps, and what is their use?

The field of geo-humanities has gained prominence recently as an interdisciplinary research area that combines humanities and geography. Scholars in this field are increasingly focused on developing methodologies that can improve the research, dissemination, and interpretation of urban history. Historical maps are a significant asset for geo-humanities research due to their tendency to encompass retrospective geographic data that may be challenging to locate through other means. Sanborn fire insurance maps are a collection of historical maps that offer extensive and meticulous urban information at the building level. These maps cover more than 12,000 towns and cities across the United States from the 19th century.

The process of extracting data from the Sanborn maps presents challenges due to the inherent design of the maps, which do not prioritize machine readability and pose difficulties in organizing the information within an organized database set-up. The Sanborn maps were produced through the lithographic printing process and were meticulously hand-colored using waxed paper stencils. One commonly employed approach to derive geographic data from Sanborn maps involves the manual process of georeferencing and annotating map features utilizing software for geographic information systems.

The manual approach is constrained due to the extensive information in each map that necessitates processing. In recent years, there has been significant progress in machine learning, which has contributed significantly to developing semi-automated and fully automated processes for extracting geographical data from historical maps.  As such, developing effectual systems for mining building-level data from Sanborn maps remains a challenging task.

What does this study involve?

This study aims to discuss the constraints of previous research by introducing a scalable workflow for extracting geographical data from Sanborn maps. It primarily centers on the analysis of building footprints and their corresponding attributes, such as construction materials, utilization, and the number of stories.

Buildings serve as the fundamental infrastructure of urban areas, and the utilization of three-dimensional historical building data can facilitate the quantification, examination, and comprehension of the transformations in neighborhood environmental, social, and health circumstances, as well as lifestyles, within cities over time. This can also contribute to the advancement of high-fidelity, three-dimensional visualizations and virtual reality experiences of historic neighborhoods. Such improvements can facilitate education, outreach, and engagement regarding urban history, encompassing the various impacts and disturbances caused by urban policies, developments, and infrastructure projects during the 20th century. The capacity to produce such data on a large scale for cities in the United States can facilitate extensive research and outreach efforts on a city-wide level, allowing for longitudinal analysis and comparisons between cities on a national scale.

The methods employed in this study involve the utilization of Sanborn maps to analyze two specific neighborhoods in Columbus, Ohio, USA. These neighborhoods were subjected to the division caused by the construction of highways during the 1960s.

Major findings

The analysis of the results of this study, both quantitatively and visually, indicates a significant level of accuracy in the extracted information at the building level. The F-1 score for building footprints and construction materials is 0.9, while for building utilization and number of stories, it exceeds 0.7.

In essence, this study has the following significant contributions:

• It introduces a scalable computational workflow that facilitates the automated geographic information extraction from historical maps. The purpose of this workflow is to provide insights into urban changes.
• The experimental findings demonstrate the efficacy of workflow in accurately extracting information and generating authentic 3D digital representations of historic urban neighborhoods.
• This research constitutes a crucial milestone in the investigation and validation of computational techniques for urban studies.
• The suggested workflow exhibits the potential for application in various geographical regions and historical periods, contingent upon the accessibility of Sanborn maps, software, and computational resources.  

Conclusion

The Sanborn Fire Insurance maps possess a substantial amount of detailed information about individual buildings within American cities, with a historical record dating back to the late 1800s. Urban environments can be effectively studied by utilizing them as valuable resources, particularly in examining the impact of urban highway construction and urban renewal on their development and transformation over time.

This research paper presents a novel approach to developing a scalable workflow that leverages machine-learning techniques to identify building footprints and their associated properties on Sanborn maps. This information has the potential to be effectively utilized in the creation of 3D visualizations of historic urban neighborhoods and to inform urban changes.