Machine Learning Assessment of Soil Carbon Sequestration Potential: Integrating Land Use, Pedology, and Machine Learning in Croatia

Spatially quantifying the soil carbon sequestration potential (SCSP) is crucial for targeting climate change mitigation strategies like carbon farming. However, static mapping approaches often fail by assuming that the drivers of soil organic carbon (SOC) are stationary. We hypothesized that the hierarchy of SOC controllers is fundamentally non-stationary, shifting from intrinsic stabilization capacity (pedology) in stable ecosystems to extrinsic flux kinetics (climate) in dynamic systems. We tested this by developing a land-use-specific (LULC; Cropland, Forest land, Grassland) ensemble machine learning (ML) framework to quantify the soil carbon saturation deficit (SCSD) across Croatia’s pedologically diverse landscape on 622 soil samples. The LULC-stratified ensemble models (SVM, RF, CUB) achieved moderate to good predictive accuracy under cross-validation (R2 = 0.41–0.60). Crucially, the feature importance analysis (permutation MSE loss) proved our hypothesis: in Forest land, SOC was superiorly controlled by intrinsic capacity (Soil CEC, Soil pH), defining the mineralogical C-saturation “ceiling”; in Grasslands, control shifted to extrinsic C-input kinetics (Precipitation: Bio19, Bio12), which “fuel” the microbial carbon pump (MCP) via root exudation; and in Croplands, the model revealed a hybrid control, limited by remaining intrinsic capacity (CEC, Clay) but strongly influenced by C-loss kinetics (Temperature: Bio08), which regulates microbial carbon use efficiency (CUE). This study demonstrates that LULC-specific dynamic modeling is a prerequisite for accurately mapping SCSP. By identifying soils with both high intrinsic capacity (high CEC/Clay) and high degradation (high SCSD), our data-driven assessment provides a critical tool for spatially targeting carbon farming interventions for maximum climate mitigation return on investment (ROI).