Explainable data-driven spatial predictions of soil organic carbon at the national level

Accurate spatial modeling of soil organic carbon (SOC) and interpretation of the results are crucial worldwide. In this study, we performed national-scale SOC mapping in Russia with a focus on model interpretability under Shapley values technique with its geovisual analysis. Also, we compared Random Forest (RF) models with a different set of covariates and basic hyperparameter tuning. Results revealed that neither the number of covariates nor the tuning significantly affected the model performance, with the best model achieving an RMSE = 70.03 g/kg and an R² = 0.39. The generated map confirms that northern Russia holds immense SOC levels, primarily driven by the presence of organic soils and peatlands, but overestimated the lowest SOC levels in the south. Shapley value analysis revealed that both biotic and abiotic variables showed complex and non-linear relationships with SOC. Mean annual temperature and precipitation contributed positively to predictions in the north, with a sharply positive effect observed specifically within the 270-280 K (-3.15 to +6.85 °C) range and >500 mm, respectively. Furthermore, higher variability in land surface temperature and elevations below 250 m were identified as key factors promoting SOC accumulation, delineating the carbon-rich Siberian lowlands. The spatial pattern of Shapley values for the major covariates demonstrated a positive contribution to SOC prediction mainly in the northern regions. Notably, the boundary between the positive and negative contributions of some climatic variables coincided with the boundary between Dfb (warm-summer humid continental) and Dfc (subarctic) climatic zones. This study provides a critical baseline for Russia's soil carbon inventory and underscores the value of interpretable machine learning for unravelling the environmental drivers of major Earth carbon sink.