Machine learning based soil maps for a wide range of soil properties for the forested area of Switzerland

• Published in: Geoderma Regional Vol. 27; p. e00437
• Main Authors: Baltensweiler, Andri, Walthert, Lorenz, Hanewinkel, Marc, Zimmermann, Stephan, Nussbaum, Madlene
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2021
• Subjects: biodiversity; carbon sequestration; clay; data collection; Digital soil mapping; ecosystems; Forest soils; forests; geostatistics; gravel; kriging; landscapes; Machine learning; Model averaging; prediction; Quantile regression forest; regression analysis; sand; soil density; soil depth; soil organic carbon; Switzerland; uncertainty; Uncertainty maps; water purification; Switzerland; Quantile regression forest; Uncertainty maps; Digital soil mapping; Model averaging; Machine learning; Forest soils
• DOI: 10.1016/j.geodrs.2021.e00437
• ISSN: 2352-0094

Spatial soil information in forests is crucial to assess ecosystem services such as carbon storage, water purification or biodiversity. However, spatially continuous information on soil properties at adequate resolution is rare in forested areas, especially in mountain regions. Therefore, we aimed to build high-resolution soil property maps for pH, soil organic carbon, clay, sand, gravel and soil density for six depth intervals as well as for soil thickness for the entire forested area of Switzerland. We used legacy data from 2071 soil profiles and evaluated six different modelling approaches of digital soil mapping, namely lasso, robust external-drift kriging, geoadditive modelling, quantile regression forest (QRF), cubist and support vector machines. Moreover, we combined the predictions of the individual models by applying a weighted model averaging approach. All models were built from a large set of potential covariates which included e.g. multi-scale terrain attributes and remote sensing data characterizing vegetation cover. Model performances, evaluated against an independent dataset were similar for all methods. However, QRF achieved the best prediction performance in most cases (18 out of 37 models), while model averaging outperformed the individual models in five cases. For the final soil property maps we therefore used the QRF predictions. Prediction performance showed large differences for the individual soil properties. While for fine earth density the R2 of QRF varied between 0.51 and 0.64 across all depth intervals, soil organic carbon content was more difficult to predict (R2 = 0.19–0.32). Since QRF was used for map prediction, we assessed the 90% prediction intervals from which we derived uncertainty maps. The latter are valuable to better interpret the predictions and provide guidance for future mapping campaigns to improve the soil maps. •Legacy soil data were used to map soil properties for the entire Swiss forest area.•Seven chemical and physical soil properties for six depth intervals were predicted.•Performances of six modelling approaches and model averaging were compared.•Quantile regression forest performed best and was used for prediction.•Accompanying uncertainty maps provide guidelines for future mapping campaigns.