Autor:innen:
Dr. Felix Stumpf | Kompetenzzentrum Boden, Schweiz | Switzerland
Dr. Thorsten Behrens | Kompetenzzentrum Boden, Schweiz
Dr. Armin Keller | Kompetenzzentrum Boden, Schweiz
Soil maps are indispensable for an adequate evaluation of soil quality and soil related ecosystem services and thus to allow sustainable soil management. In Switzerland, existing soil maps are limited with respect to spatial coverage and scale, while attempts to update legacy soil data using recent Digital Soil Mapping (DSM) techniques are rare. There is a clear need for detailed and area covering soil property maps. This study is framed by three dimensional (3D) DSM approaches to predict the spatial and vertical distribution of soil texture (sand, silt, clay) and pH at national scale. The focus of the study lies on investigating DSM models with respect to performance and uncertainty for three depths intervals (0-30 cm, 30-60 cm, 60-120cm) and for crop-, grass- and woodland, as well as with respect to covariate importance. The DSM models build on Quantile Regression Forests, a fivefold data split approach for independent evaluation and a legacy soil data set from various regional soil surveys across 35 years. The soil covariate set is extensive, including a spatially multiscale terrain analysis, a temporarily multiscale land use and vegetation analysis, a spectral bare soil analysis, and a spatio-temporal climate analysis. Main data sources of the soil covariate set are an elevation model based on airborne laserscan data, as well as spectral raster timeseries based on satellite imagery from Landsat and Sentinel missions. Soil texture models showed R2 values between 0.55 and 0.8 across all depth intervals and landcovers, while soil pH models showed R2 values between 0.61 and 0.86. The highest R2 values were found in the depth interval 30-60cm for soil texture and pH. With respect to landcover, soil texture models performed best on cropland, while soil pH models showed an increased R2 on woodland. Soil depth was the most important covariate for all soil texture and pH models, followed by climate and the equally ranked covariates for bare soil spectral reflectance, land use and vegetation, as well as terrain.