Evaluating digital soil mapping approaches for mapping GlobalSoilMap soil properties from legacy data in Languedoc-Roussillon (France)

International audience

English. Digital Soil Mapping is becoming increasingly operational because of shared approaches, clear specifications (e.g., GlobalSoilMap) and more “practical” applications across the planet. In a 27,236 km2 French region located in the Northern Mediterranean area, this study evaluated four well-known Digital Soil Mapping approaches that were possibly applicable in the French context for inferring the GlobalSoilMap (GSM) grid using freely available data from the French spatial data infrastructure. These approaches used for soil input either a legacy 1:250,000 scale soil map—an area-weighted mean of soil mapping units and spatial disaggregation of soil mapping units (simulated)—or a set of 2024 legacy-measured soil profiles (1 profile/13.5 km2) associated with a set of soil covariates using random forest and random forest plus kriging. These methods provided estimates of 29 soil properties over a 100 m by 100 m grid, i.e., clay, sand, silt, coarse fragment, organic carbon contents, pH and CEC, at the four upper GSM-specified depth intervals, plus soil depth. This experiment showed that the performances in mapping soil properties were highly variable (from R2 = 0 to R2 = 0.79). They were strongly correlated with the amount of spatially-structured variance of the soil properties captured by the available soil data. For example better performances were observed for organic carbon that is driven by large-scale climate variations than for texture and soil depth that are driven by short scale variations of parent materials and erosion–deposition ratio along a slope. Furthermore, the profile-based DSM models generally outperformed the soil map-based DSM models, and kriging did not improve the results of random forest. Finally improving the quality of soil data input was considered as the best way to improve the current mapping performances. In addition to providing a GlobalSoilMap proof-of-concept for Northern Mediterranean areas, this paper illustrates how a prior analysis of the available soil data may help in the anticipated estimation of the mapping performances and in the prior selection of the most promising DSM models to reach these performances.