Abstract:
This thesis is part of the Collaborative Research Centre (SFB) 299 – Land Use Options for Peripheral Regions founded by the German Research Foundation (DFG). The general aim was to develop methods and concepts for representative sampling and high-resolution digital soil mapping in large-scale landscapes (>1000km²). The main focus was on the development of a multi-stage sampling scheme in combination with data mining techniques in order to regionalize soil properties and soil classes. The main study areas included the Nidda catchment in Hesse, Germany, the Pfälzer Wald in Rhineland-Palatine, Germany, and one region in the Republic of Niger.
As a first step of the sampling scheme (see manuscript 1) the Nidda catchment was divided in homogeneous, non-fragmented segments. The 1:50.000 soil map (HLUG), which served as a basis for the segmentation, was evaluated by a moving-window frequency distribution analysis and subsequently classified by a spatial k-means cluster analysis. The number of clusters as well as the size of the moving window were determined semi-automatically. Based on this approach the soilscapes Vogelsberg, Lower Vogelsberg, Forest of Büdingen, North-East Wetterau, South-West Wetterau and Taunus could be statistically separated. Thus homogeneous soilscapes for the additional use of digital soil mapping approaches were generated. Based on a patch sampling method developed in the SFB 299 representative patches were derived. The spatial extent of each patch was 3x3 km.
The second step covers the development of methods for a representative generation of transects for geophysical surveys (see manuscript 2). The so called Singleline approach includes all spatial units along a single transect in a representative subspace. The Multiline method designates several transects along defined transition zones between different soil types in a representative subspace. Statistical analyses and model calculations showed that the Multiline approach generates longer transects. With regard to the frequency distribution of characteristic terrain attributes such as slope, aspect and local elevation, however, it describes the feature space more thoroughly.
Further, this thesis focused on an objective and reproducible improvement of the prediction quality of soil information. Traditional soil maps have been compiled with different standards and mapping techniques and, as expert-based values, they are generally affected by subjective influences. The resulting uncertainties and generalizations have a direct effect on the boundaries. Therefore, a raster-based approach was developed in order to adjust the boundaries (see manuscript 3) on the basis of higher resolution terrain attributes. This method produced a plausible adjustment of the boundary regions for a soil map of the Nidda catchment (1: 50.000) and a geological map of the Republic of Niger (1:1.000.000). As a result, errors and artefacts in predictive models can be reduced.
The last part of this work focused on the combination of sampling schemes and predictive models and their application in order to handle redundant, uncertain (noisy) information in large datasets, as they tend to occur when existing, rasterized soil maps are used as a training base (see manuscript 4). By applying different randomized sampling procedures focussing on relevant information and improving prediction quality up to 12% was enabled.
Based on the modular structure of the methods introduced, it is possible to obtain a reproducible and objective segmentation of the landscape, a valid identification of representative subsets and transects and an improved quality of the geodata basis used for soil-landscape-modelling. Therefore, comprehensive and valid high-resolution soil scientific information can be provided even for large scale landscapes.