Study areas
Our study areas reflect two different Danish landscape types. The Hirtshals study area in northern Jutland (Fig. 1b) covers 170 km² and is characterised by aeolian sand along the coastline and late- and post- glacial marine sand and moraine clay inland. In the late 1800s the area was dominated by vast areas of dune sand and heath along the coast and a mixture of agricultural land and wetland inland. The Hobro study area in mid-Jutland (Fig. 1c) covers 425 km² and is characterised by mixture of moraine sand and sandy clay and organic soils in river valleys. In the late 1800s, this area included extensive areas of heath, wetland, forest, and agricultural land.
Applied maps
The first large-scale survey of the Danish landscape was done by the military in the second half of the 19th century, resulting in topographic maps in scale 1:20,000 published between 1870 and 1899. Although the landscape was surveyed from a military perspective (Svenningsen 2016), these so called Høje Målebordsblade (HMB) maps are the most important source for detailed and spatial explicit information about the Danish landscape before the onset of industrial intensification of agriculture in the 20th century. The HMB include detailed information about a range of LULC categories as well as point and linear features. However, only a sample map with the range of different symbols was published not containing definitions of the different LULC categories and features, indicating that such information was deemed self-evident by the military surveyors at the time. Therefore, an archive-based historical study was conducted tracing the development of LULC categories back to the original survey instructions from the first half of the 19th century. This led to the definition of a set of overarching LULC categories (Svenningsen et al. 2022).
The HMB comprise approximately 1,200 sheets and were published in two different versions. A black and white print, which was later hand-coloured and a three-colour version (black, brown, and blue). While the three-coloured version relies on symbols for representing areal LULC categories, the hand-coloured version also utilizes colours. The sheets of the hand-coloured version have been scanned and geo-rectified to a seamless raster dataset (SDFI 2023). This dataset still needs to undergo map processing to become available as machine-readable GIS layers. Since the original hand-coloured or three-colour paper maps are of higher quality in terms of the fidelity of line work and colour consistency, the required sheets were acquired from the Map Collection at the Royal Danish Library, scanned at a resolution of 600 dpi, which is appropriate for 1:20,000 scale source material (Tobler 1988) and georeferenced using tie-points to the other set of, already georeferenced, HMB raster data, applying a spline model and nearest-neighbour resampling to 1 m cells (Royal Danish Library 2024).
Contemporary LULC information was derived from Basemap03, a nationwide LULC map for the year 2018, created by superimposing and spatially aligning current geographical layers into a raster map with a resolution of 10x10 meters (Levin 2019). We grouped the LULC categories from Basemap03 and from the HMB into ten major categories (Table 1). Although the HMB do not include categories for dry grassland and agriculture and built/infrastructure features were not extracted in the context of this project we incorporated these categories from Basemap03 as they can yield valuable insights into landscape dynamics.
Table 1
Description of employed LULC categories from applied historical and contemporary maps.
| Høje Målebordsblade (HMB) (1880) | Basemap03 (2018) |
---|
Forest | Areas with tree cover. | Areas, dominated by woody vegetation. Includes temporary un-stocked areas. Excludes Christmas trees, energy forest and orchards. |
Wetland | Areas with soft soil (often organic), and with potential for seasonal flooding. Vegetation cover includes permanent grass and natural vegetation. | Uncultivated areas, dominated by grass and herb vegetation on permanently or seasonally waterlogged soils. |
Heath | Areas covered with heather vegetation (Heather or similar dwarf shrub). | Uncultivated areas, dominated by vegetation of heather or similar dwarf shrub. |
Dry grassland | Was not mapped as a category in the HMB | Grass and herb vegetation on dry and/or calcareous soils. Includes dunes covered with grass and herb vegetation. |
Dune sand | Areas characterised with loose soil in form of sand. | Uncultivated areas, dominated by dune sand and coastal sand without or with sparse vegetation. |
Freshwater | Permanent open water bodies of freshwater. Includes water bodies covered by reed vegetation. | Permanent water bodies, including lakes, ponds and streams. Includes water bodies covered by reed vegetation. |
Sea | Permanent open water bodies of open sea and fjords. | Permanent water bodies of open sea and fjords. |
Agriculture | Was not mapped as a category in the HMB | Cropland and grassland for agricultural production, except grassland, which is also mapped as heath or dry grassland. |
Built-up | Was not extracted from the HMB in this study | Built-up land, including buildings and transport infrastructure and associated land, such as gardens, parks, concrete surfaces and road and rail verges. |
Other | Other LULC, including built-up land and transport infrastructure, which, in this study was not extracted and agricultural land, which was not mapped as a category in the HMB | Other LULC, which is not included in any of the other LULC categories. |
Automated production of machine-readable layers of LULC categories
Production of machine readable geodata from the scanned historical maps (hereafter referred to as “the mapping”) was done as a set of modular sequences of raster, vector and object processing procedures that included image enhancement, CIS, OBIA and ML methods (Groom et al. 2020, 2021; Levin et al. 2020). Modules and module sequences were customised for each of five target categories and were developed to operate fully and without additional user input for all parts of both study areas. Figure 2 presents an overview of the general workflow. We selected five target categories: Heath, dune sand, wetland, forest and water bodies as these are of relevance to issues of 100 + year LULCC and are extensively represented within the study areas. The mapping of each target category was made as an independent module and was also done independently for the extent of each HMB map sheet, cut into quarters (with 1 km E-W and N-S overlaps) on account of computer RAM considerations. In general, it was the Royal Danish Library sourced raster data there were used in the mapping, on account of the superior linework and colour quality; for some sheets however colours in that material had faded, requiring instead use of raster data map sheet extracts from the seamless raster dataset. Initial inspection was made of each sheet to check for the consistency of target category representations to the general HMB patterns and for any systematic HMB map quality issues, such as localised fading. If quality issues were noted decisions were made related to (a) which of the two alternative HMB rasters to work with, (b) the need for use of a more customised, interactive setting of segmentation thresholds, (c) the need for independent work upon other sub-parts of a map sheet.
For the target categories heath, wetland, forest, and water bodies occurrences in the HMB are characterised by use of colour. For these four the applied mapping method was in general the same, consisting of determining which of a set of image enhancement rasters best isolated the target category, segmenting on a threshold in that raster and application of spatial object processes to improve upon the initial segmentation (Levin et al. 2020). Image enhancements to provide a set of derived rasters included Hue, Saturation and Intensity (HSI) transformations, and redness (R/(G + B)), greenness (G/(R + B)), blueness (B/(R + G)) transformations, applied upon raw, histogram normalisation and inverted forms of the image data. Figure 3 shows examples of these derived rasters, including how they enable isolation of the target categories. For smaller water bodies and for forest, the colour of the target category areas showed sufficient consistency across the map sheets to allow for application of a single (raster + threshold) rule for each. The colour of both wetland and heath in the applied HMB map sheets was far more variable, requiring a different approach (Groom et al. 2020), drawing upon the property that wetland and heath are each also associated with use of distinct respective symbols. In this, a ML algorithm was trained and then applied to detect occurrences of each symbol. From accurate ML detection of even just some of the symbols it was possible to determine in an automated way, for each map sheet, the appropriate derived (raster + threshold) rule to apply as part of the general method.
The dune sand category is distinct in that its representation is not associated by any use of areal colour hand-painting, but only by use of a black stippling (dots). The dot density is relatively constant, so having isolated all dune sand dots as black parts with an object size of just a few pixels, object-based distance and density methods and criteria were applied to form image object polygons of the extents of the dune sand (Groom et al. 2021).
As assists to the mapping of each of the target categories masks were produced from the HMB and applied in the general method. A mask of all the black parts was generally applied for the post-segmentation processes, such as for filling parts of the target category’s extent where there was map text or linework (e.g. height contours). A mask was also formed of any large water bodies, both inland and off-coast; this was prescribed since these geographic features in the HMB often have highly variable colourations (e.g., large lakes are often hand coloured in blue only close to their shores). Furthermore, there is often marked bi-modality in the overall size range of “blue” water body extents (i.e. some very large plus many much smaller water bodies), which presented challenges for mapping of all water bodies with the general method. The mapping of water bodies in this work has not included mapping of rivers except where these widen to form lake-like features; the coincidence of many rivers in the HMB with administrative boundaries, which are presented as broad features in various colours led to the decision that it was not possible to include rivers in the mapping of the water bodies target category in a consistent manner. The solid black lines that mark the limit of water surfaces were applied, via the mask of black parts, as an assist in the mapping of the water body target category, since there were cases were the blue hand-colouring was spread over parts outside the apparent HMB feature.
The HMB include delineation of the forest and heath categories on steeply sloping ground. Steeply sloping ground is itself designated by use of lines crossing between the height contours. That results in the representations of forest and heath on steep ground being broken into extents comprising just a few image pixels, which would not be detected by the general methods for forest or heathland detection. The applied solution was to form a mask of steeply sloping ground from the mask of the black parts. The general methods for forest and heath were then applied within those mask extents with a far smaller minimum size threshold.
For improvement upon the initial target category object mappings, size thresholds were applied for filling small holes in target category extents and for removal of tiny cases of commission errors. Pixel-based object extent re-sizing methods were also applied including with control via surface tension values. The final mappings of target categories for the entireties of the two study areas was done by a second OBIA workflow. A first part of this addressed inconsistencies, such as arising from the use of object size criteria, within the quarter-sheet overlap zones. Target category disjoints across map sheet borders, occasionally occurring locally where mappings from the two sets of source HMB material met, were addressed using a set of neighbourhood analysis based decision rules.
A key characteristic of the applied workflow for the automated production of machine readable geodata is that it represents a toolbox of a set of tried-and-tested techniques that were then applied, in various combinations, for each of the five target categories. Whilst the mapping itself was done by running for each target category of a single workflow, visual inspections of each HMB raster data set and of the product of each mapping are considered as key steps of the overall workflow.
Accuracy assessment
A validation layer was created as a regular point grid with an equidistance of 100 meters and a total of 27,498 validation points. The distance of 100 meters was small enough to capture the variation of different LULC categories while being manageable to interpret within the resources of this project. Each point was classified visually based on the colours and signatures present up to 50 meters from the point. A point was classified either as a pure category, such as heath or wetland or as a combination of several categories, such as heath and wetland or as heath and forest. The extracted polygon features were overlaid with the validation layer, and for each target category, proportions of false positive and false negative were calculated.
Assessing land use and land cover dynamics
To map LULC dynamics from 1880 to 2018, we integrated LULC category layers derived from HMB map sheets with contemporary LULC data from Basemap03. The produced vector layers of HMB LULC categories were converted to 10x10 meter raster corresponding to the resolution of Basemap03. To reduce the bias from small discrepancies between the delineations of LULC categories in the two maps, we spatially aligned the rasters of HMB categories to Basemap03 if they were located within 40 meters from the same categories in Basemap03.