Soil property variation mapping through data mining of soil category maps

Spatial information on soil properties is an important input to hydrological models. In current hydrological modelling practices, soil property information is often derived from soil category maps by the linking method in which a representative soil property value is linked to each soil polygon. Limited by the area-class nature of soil category maps, the derived soil property variation is discontinuous and less detailed than high resolution digital terrain or remote sensing data. This research proposed dmSoil, a data-mining-based approach to derive continuous and spatially detailed soil property information from soil category maps. First, the soil-environment relationships are extracted through data mining of a soil map. The similarity of the soil at each location to different soil types in the soil map is then estimated using the mined relationships. Prediction of soil property values at each location is made by combining the similarities of the soil at that location to different soil types and the representative soil property values of these soil types. The new approach was applied in the Raffelson Watershed and Pleasant Valley in the Driftless Area of Wisconsin, United States to map soil A horizon texture (in both areas) and depth to soil C horizon (in Pleasant Valley). The property maps from the dmSoil approach capture the spatial gradation and details of soil properties better than those from the linking method. The new approach also shows consistent accuracy improvement at validation points. In addition to the improved performances, the inputs for the dmSoil approach are easy to prepare, and the approach itself is simple to deploy. It provides an effective way to derive better soil property information from soil category maps for hydrological modelling. Copyright (c) 2014 John Wiley & Sons, Ltd