Spatial Modelling and Drivers of the Sum of Absorbed Bases in Agroecosystems Using Explainable Machine Learning

This is the authors' abstract. We don't add key points for this paper.

While many soil properties are routinely mapped in space and time, the sum of absorbed bases (S) remains a critically important but under‐modelled parameter. This study aims to fill this gap by providing an interpretable, machine learning (ML)‐based spatial assessment of S using a dataset comprising 4880 topsoil measurements from temperate agricultural lands. Further, we applied Shapley Additive exPlanations (SHAP) to explore the drivers of S and interpret the results of a predictive model. We tested several modelling scenarios that included co‐measured soil properties (pH, soil organic matter (SOM), thickness of the humus‐accumulative horizons, available phosphorus and exchangeable potassium), as well as environmental covariates (ECOVs) presented by key soil‐forming factors. The results revealed that the ML model using only soil properties as predictors explained 66% of the total variance, whereas adding ECOVs improved performance up to NSE of 0.74. Among them, pH, SOM and precipitation were key drivers. According to SHAP analysis, threshold values for pH (5.5–6) and SOM (5%) were identified, beyond which their further increase exerts a consistently positive influence on the predicted S values. Furthermore, precipitation variables revealed thresholds of 175 mm for the warmest quarter and 45 mm for September, beyond which higher rainfall correlates with a positive contribution to the S prediction. The spatial patterns of S values across the region showed a south‐to‐north gradient, with the highest content in the fertile Chernozems decreasing toward Phaeozems and acidic Albeluvisols zones. Our findings contribute insights into the spatial modelling of S and exploring its important drivers using an interpretable ML framework.