Extracting Wetlands from Swiss Historical Maps with ConvolutionalNeural Networks

Abstract

Historical maps can serve as valuable resources for various kinds of researches such as ecology, land reclamation (Ngo et al, 2015), toponymy, history, etc. (Chiang, 2016). Specifically, extracting wetlands from historical maps facilitates researchers to investigate the spatio-temporal dynamics of the hydrological and ecological situation. Deep learning methods, especially Fully Convolutional Neural Networks (FCNN), provide an efficient and effective way to extract features from raster maps. To extract wetland areas from the Swiss Siegfried maps, we trained a U-Net based architecture and applied the learnt model to 573 map sheets across Switzerland. The pixel-wise prediction results were converted to polygons through a vectorization and generalization step. The vector wetland layers will be used for studies on land-cover change.