Resurrecting Lost Landscapes: Automated Sanborn Map Digitization and 3D Reconstruction through Machine Learning and Open-Source Tools

Abstract

Sanborn fire insurance maps contain vital information for creating historic reconstructions. However, manual digitization of the map sheets is time consuming and thorough feature extraction is constrained. Applying machine learning to this problem offers an efficient solution to extract comprehensive datasets of these features, enhancing the ability to create more accurate and authentic 3D historic cityscapes. Denver’s former Auraria neighborhood was a landscape lost in the wake of early twentieth century zoning practices and federal urban renewal initiatives of the 1950s. It provides a useful case study for testing and improving workflows of automated Sanborn map digitization and 3D reconstruction. To digitize Auraria Sanborn maps, I apply R-based geocoding and automated georeferencing tools. A support vector machine (SVM) algorithm from the scikit-learn Python library performs pixel-based classification to detect building materials and footprints, and vectorization is conducted in QGIS. State-of-the-art deep neural networks for text detection extract building utilization, names, and detailed descriptions. Machine learning models are run with free computing resources from the Google Colaboratory. 3D reconstruction of the neighborhood is created with Blender’s open-source 3D modeling software. This approach efficiently performs these tasks and uses automation to reduce manual input and open-source alternatives to minimize costs, increasing the number of digitized historic landscapes available and allowing broader participation in their production. Ultimately, additional highly detailed 3D reconstructions will reveal deeper insights into patterns of disinvestment, displacement from zoning, and neighborhood change across twentieth century American urban landscapes.