Aarhus University Seal / Aarhus Universitets segl

No. 389: Automated production of spatial datasets for land categories from historical maps

Levin, G., Groom, G.B., Svenningsen, S.R. & Perner, M.L. 2020. Automated production of spatial datasets for land categories from historical maps. Method development and results for a pilot study of Danish late-1800s topographical maps. Aarhus University, DCE – Danish Centre for Environment and Energy, 121 pp. Scientific Report No. 389 http://dce2.au.dk/pub/SR389.pdf

Summary

Historical maps contain valuable, geographic information about past land use and land cover (LULC). However, historical maps are merely graphic raw material comprising colours, symbols, lines, and text labels that represent different aspects of LULC. In order to make these maps useful for assessment and quantification of LULC and LULC changes (LULCC) with modern geographical information systems (GIS), the LULC categories represented in the historical maps must be transformed into machine-readable spatial datasets. Traditionally, this production of LULC category digital geo-data from historical maps has been done by time and resource demanding visual interpretation of the maps and manual digitisation. Consequently, analyses of historical maps have typically been elaborated only for rather small areas. Recent developments within digital image processing have led to unprecedented possibilities for less time-consuming automated extraction of LULC categories from historical maps.

In 2019, Aarhus University and the Royal Danish Library conducted a pilot study to develop, test, and evaluate automated production of land category digital geo-data from Danish 1:20,000 topographical maps from the late 1800s (the Høje Målebordsblade map series). The study was undertaken for two areas: Hirtshals, in northern Jutland and Hobro in central Jutland, covering around 300 km² and a large variety of different Danish landscapes. Produced digital geo-data were for the land categories heath, dune sand, wetland, forest and water bodies. The automated geo-data production comprised a combination of object based image analysis, vector GIS, colour segmentation and machine learning processes. An accuracy assessment, based on visual interpretation of the applied map sheets for around 27,500 control points, was conducted. For most of the target categories, results indicated producers accuracies of around 95 % or higher and users accuracies of around 90 % or higher. The pilot study revealed that colour variations between map sheets is a significant factor in determining the level of method sophistication required to achieve satisfactory results. A comparison with contemporary maps revealed LULC changes, which are generally characterised by a decrease in open habitat types, such as heath, dunes and wetland due to cultivation and afforestation. We conclude, that automated production of LULC category digital geo-data from historical maps offers a less time consuming and consequently more resource efficient alternative to traditional manual vectorisation.