Extracting Human Settlement Footprint from Historical Topographic Map Series Using Context-Based Machine Learning

Abstract

Information extraction from historical maps represents a persistent challenge due to inferior graphical quality and large data volume in digital map archives, which can hold thousands of digitized map sheets. In this paper, we describe an approach to extract human settlement symbols in United States Geological Survey (USGS) historical topographic maps using contemporary building data as the contextual spatial layer. The presence of a building in the contemporary layer indicates a high probability that the same building can be found at that location on the historical map. We describe the design of an automatic sampling approach using these contemporary data to collect thousands of graphical examples for the symbol of interest. These graphical examples are then used for robust learning to then carry out feature extraction in the entire map. We employ a Convolutional Neural Network (LeNet) for the recognition task. Results are promising and will guide the next steps in this research to provide an unsupervised approach to extracting features from historical maps.