Abstract
Historical topographical maps contain valuable, spatially and thematically detailed information about past landscapes. Yet, for analyses of landscape dynamics through geographical information systems, it is necessary to "unlock" this information via map processing. For two study areas in northern and central Jutland, Denmark, we apply object-based image analysis, vector GIS, colour image segmentation, and machine learning processes to produce machine-readable layers for the land use and land cover categories forest, wetland, heath, dune sand, and water bodies from topographic maps from the late nineteenth century. Obtained overall accuracy was 92.3%. A comparison with a contemporary map revealed spatially explicit landscape dynamics dominated by transitions from heath and wetland to agriculture and forest and from heath and dune sand to forest. However, dune sand was also characterised by more complex transitions to heath and dry grassland, which can be related to active prevention of sand drift but that can also be biased by different categorisations of dune sand between the historical and contemporary data. We conclude that automated production of machine-readable layers of land use and land cover categories from historical topographical maps offers a resource-efficient alternative to manual vectorisation and is particularly useful for spatially explicit assessments of long-term landscape dynamics. Our results also underline that an understanding of mapped categories in both historical and contemporary maps is critical to the interpretation of landscape dynamics.