Modeling soil organic carbon with Quantile Regression : dissecting predictors' effects on carbon stocks

Soil organic carbon (SOC) estimation is crucial to manage natural and anthropic ecosystems. Many modeling procedures have been tested in the literature, however, most of them do not provide information on predictors' behavior at specific sub-domains of the SOC stock. Here, we implement Quantile Regression (QR) to spatially predict the SOC stock and gain insight on the role of predictors (topographic and remotely sensed) at varying SOC stock (0\u201330cm depth) in the agricultural areas of an extremely variable semi-arid region (Sicily, Italy, around 25,000km2). QR produces robust performances (maximum quantile loss = 0.49) and allows to recognize dominant effects among the predictors at varying quantiles. In particular, clay mostly contributes to maintain SOC stock at lower quantiles whereas rainfall and temperature influences are constantly positive and negative, respectively. This information, currently lacking, confirms that QR can discern predictor influences on SOC stock at specific SOC sub-domains. The QR map generated at the median shows a Mean Absolute Error of 17 t SOC ha-1 with respect to the data collected at sampling locations. Such MAE is lower than those of the Joint Research Centre at Global (18 t SOC ha-1) and at European (24 t SOC ha-1) scales and of the International Soil Reference and Information Centre (23 t SOC ha-1) while higher than the MAE reached in Schillaci et al. (2017b) (Geoderma, 2017, issue 286, page 35\u201345) using the same dataset (15 t SOC ha-1). The results suggest the use of QR as a comprehensive method to map SOC stock using legacy data in agro-ecosystems and to investigate SOC and inherited uncertainty with respect to specific subdomains. The R code scripted in this study for QR is included