Abstract

Historical map documents are unique witnesses of landscapes in the past and contain valuable information about the spatiotemporal evolution of geographic phenomena such as forest coverage, transportation networks, and human settlement patterns. However, this information needs to be extracted (from map documents) and converted into machine-readable data to be used for quantitative analysis. The extraction of information from historical maps is a persistent challenge due to the low graphical quality of the scanned documents and the massive data volume of digital map archives, which can hold hundreds of thousands of scanned map sheets. Recently, several digital map archives have been made available to the public including the United States Geological Survey (USGS) historical topographic maps (Fishburn et al., 2017) and the Sanborn Fire insurance map collection (Library of Congress 2018). For example, the USGS has systematically scanned and georeferenced approximately 200,000 historical topographic map sheets at scales of up to 1: 24,000, produced between 1884 and 2006. This map archive is publicly available and represents a unique data source for various applications and studies requiring retrospective geographical data.
However, information extraction from such large amounts of data covering considerable spatial and temporal extents requires systematic, robust, and automated approaches lending from the fields of computer vision, image processing, and machine learning. Convolutional neural networks (CNNs) and other deep learning methods have recently shown promising performance in image recognition …

Date

September 21, 2025

Authors

J Uhl, S Leyk, Y Chiang, W Duan, C Knoblock

Journal

Proceedings 22nd International Research Symposium on Computer-based Cartography and GIScience (Auto-Carto 2018), Madison, WI. https://www. ucgis. org/assets/docs/AutoCarto-2018Proceedings. pdf