A fast and effective deep learning approach for road extraction from historical maps by automatically generating training data with symbol reconstruction

Historical road data are often needed for different purposes, such as tracking the evolution of road networks, spatial data integration, and urban sprawl investigation. However, road extraction from historical maps is challenging due to their dissatisfying quality, the difficulty in distinguishing road symbols from those of other features (e.g., isolines, streams), etc. Recently, although deep learning, especially deep convolutional neural networks (CNNs), have been successfully applied to extract roads from remote sensing images, road extraction from historical maps with deep learning is rarely seen in existing studies. Apart from this, it is time-consuming and laborious to manually label large amounts of training data. To bridge these gaps, this paper proposes a novel and efficient methodology to automatically generate training data through symbol reconstruction for road extraction. The proposed methodology is validated by implementing and comparing four training scenarios using the Swiss Siegfried map. The experiments show that imitation maps generated by symbol reconstruction are especially useful in two cases. First, if little manually labelled training data are available, models trained on imitation maps alone can already provide satisfactory road extraction results. Second, when training data from imitation maps are mixed with real training data, the resulting models even outperform the models trained on real data alone for some metrics, thus indicating that imitation maps can be a highly valuable addition. This research provides a new insight for fast and effective road extraction from historical maps using deep learning.