Shapley values reveal the drivers of soil organic carbon stock prediction

Abstract. Insights into the controlling factors of soil organic carbon (SOC) stock variation are necessary both for our scientific understanding terrestrial balance and to support policies that intend promote storage in soils mitigate climate change. In recent years, complex statistical algorithmic tools from field machine learning have become popular modelling mapping SOC stocks over large areas. this paper, we report on development a method interpreting models, which implemented study variation. We fitted random forest model with 2206 measurements 0–50 cm depth interval mainland France used set environmental covariates as explanatory variables. introduce Shapley values, coalitional game theory, use them understand how influence prediction: what is functional form association between covariates, does covariate importance vary locally one location another carbon-landscape zones? Results were validated light existing well-described processes mediating regards previous studies same area. found vegetation topography overall most important drivers but varied greatly among locations zones. two spatial equivalent stocks, there was nearly an opposite pattern individual contribution yielded prediction – case variables contributed positively, whereas second negatively effect mitigated by land use. demonstrate values methodological yield useful insights space. This may provide valuable information whether empirical models predicting property interest right reasons formulate hypotheses mechanisms driving sequestration potential soil.