Abstract Soil organic carbon (C org ) maintains the fundamental properties, regimes, and functions of soils, being a key indicator of soil quality and sustainability. In recent years, complex statistical models have increasingly been used to model and map the C org content in soils. The problem of these models is their low interpretability. In this paper, we consider interpretable machine learning approaches by the example of thematic modeling of organic carbon content in soils in the south of the Bugul’minsko-Belebeevskaya Upland (Samara oblast). To interpret the results of a random forest model, several methods have been applied including permutation feature importance, partial dependence plots (PDPs), individual conditional expectation (ICE) plots, Shapley values, and interaction analysis using H-statistics. The key factors controlling the spatial variation of C org in soils have been identified. The Shapley value analysis permitted us to assess both the global and local contribution of features to the prediction results and also to find the functional form of their influence. Significant spatial variability in the contribution of individual factors emphasizes the importance of local characteristics for interpreting modeling results.
Chinilin et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: