A novel framework for road vectorization and classification from historical maps based on deep learning and symbol painting

Abstract

Road networks in the past are imperative for understanding evolution of transportation infrastructure, urban sprawl, and route planning, etc. Various approaches have been developed for road extraction from historical maps, among which deep learning techniques stand out as the most effective ones. However, little attention has been paid to investigating road vectorization and classification from historical maps. Moreover, road classification via machine learning methods usually requires large amounts of dedicated training data. To address these issues, this paper proposes a novel and comprehensive framework for road vectorization and classification on the basis of road segmentation from historical maps. First, deep learning is used to get pixel-wise raster road segmentation results, which are further skeletonized using morphological operations. Then, considering that each road class is represented with a certain symbol, a painting function is defined for each class able to paint the corresponding symbol. These painting functions are then used to draw road segments along the skeletons. Since the start and end points in each painting function are used to vectorise the segment, this method achieves vectorization and classification at the same time. Our method is validated on four Siegfried map sheets in Switzerland, and evaluated via both visual and quantitative assessments. The results indicate that the method is capable of classifying roads accurately. In particular, two evaluation metrics completeness and correctness achieve 90.69% and 72.71% respectively for road class 2 which accounts for the highest portion in the map. Moreover, the results of this method avoid the saw-toothed issue of vectorised road lines. This research is beneficial for creating complete vector road network datasets with class information to support decision-making in urban planning and transportation.