Training Deep Learning Models for Geographic Feature Recognition from Historical Maps

Abstract

Historical map scans contain valuable information (e.g., historical locations of roads, buildings) enabling the analyses that require long-term historical data of the natural and built environment. Many online archives now provide public access to a large number of historical map scans, such as the historical USGS (United States Geological Survey) topographic archive and the historical Ordnance Survey maps in the United Kingdom. Efficiently extracting information from these map scans remains a challenging task, which is typically achieved by manually digitizing the map content. In computer vision, the process of detecting and extracting the precise locations of objects from images is called semantic segmentation. Semantic segmentation processes take an image as input and classify each pixel of the image to an object class of interest. Machine learning models for semantic segmentation have been progressing rapidly with the emergence of Deep Convolutional Neural Networks (DCNNs or CNNs). A key factor for the success of CNNs is the wide availability of large amounts of (labeled) training data, but these training data are mostly for daily images not for historical (or any) maps. Today, generating training data needs a significant amount of manual labor that is often impractical for the application of historical map processing. One solution to the problem of training data scarcity is by transferring knowledge learned from a domain with a sufficient amount of labeled data to another domain lacking labeled data (i.e., transfer learning). This chapter presents an overview of deep-learning semantic segmentation models and discusses their strengths and weaknesses concerning geographic feature recognition from historical map scans. The chapter also examines a number of transfer learning strategies that can reuse the state-of-the-art CNN models trained from the publicly available training datasets for the task of recognizing geographic features from historical maps. Finally, this chapter presents a comprehensive experiment for extracting railroad features from USGS historical topographic maps as a case study.