Self-Constructing Graph Convolutional Networks for Semantic Segmentation of Historical Maps

Historical maps precisely capture the spatial reality of the past. Hence, they are a fundamental source of geographic information (Chiang et al., 2020), e.g. as a prerequisite for backtracking environmental changes including land-use change, urbanization, epidemiology and landscape ecology (Schlegel, 2019). To avoid a deterioration of this information and consequently preserve it and make it widely available, often a digitization is performed, which mostly corresponds to the process of scanning a historical document and storing the resulting raster graphic in a database. However, in this case the scanned historical documents, i.e. maps, neither have searchable metadata nor are semantically enriched (Chiang et al., 2020). Hence, invaluable spatial information remain unreachable (Schlegel, 2021). Thus, the ultimate goal is to digitize historical maps in a way that their information and content is actually readable, searchable and analyzable by machines. In combination with a database this enables an effective and intuitive analysis and comparison with current map features (Schlegel, 2019, 2021).