Digital mapping of soil organic matter stocks using Random Forest modeling in a semi-arid steppe ecosystem

Spatial prediction of soil organic matter is a global challenge and of particular importance for regions with intensive land use and where availability of soil data is limited. This study evaluated a Digital Soil Mapping (DSM) approach to model the spatial distribution of stocks of soil organic carbon (SOC), total carbon (C_tot), total nitrogen (N_tot) and total sulphur (S_tot) for a data-sparse, semi-arid catchment in Inner Mongolia, Northern China. Random Forest (RF) was used as a new modeling tool for soil properties and Classification and Regression Trees (CART) as an additional method for the analysis of variable importance. At 120 locations soil profiles to 1 m depth were analyzed for soil texture, SOC, C_tot, N_tot, S_tot, bulk density (BD) and pH. On the basis of a digital elevation model, the catchment was divided into pixels of 90 m?×?90 m and for each cell, predictor variables were determined: land use unit, Reference Soil Group (RSG), geological unit and 12 topography-related variables. Prediction maps showed that the highest amounts of SOC, C_tot, N_tot and S_tot stocks are stored under marshland, steppes and mountain meadows. River-like structures of very high elemental stocks in valleys within the steppes are partly responsible for the high amounts of SOC for grasslands (81?C84% of total catchment stocks). Analysis of variable importance showed that land use, RSG and geology are the most important variables influencing SOC storage. Prediction accuracy of the RF modeling and the generated maps was acceptable and explained variances of 42 to 62% and 66 to 75%, respectively. A decline of up to 70% in elemental stocks was calculated after conversion of steppe to arable land confirming the risk of rapid soil degradation if steppes are cultivated. Thus their suitability for agricultural use is limited.