Interpretable machine learning for digital soil pH mapping using an optimized AdaBoost algorithm

Abstract

Soil pH is a fundamental parameter determining nutrient availability, microbial activity, and crop productivity. Unlike previous studies that often prioritize prediction accuracy over explainability, this study proposes an interpretable machine-learning framework integrating hyperparameter optimized adaptive boosting (AdaBoost) with Shapley Additive exPlanations (SHAP) to unravel the spatial drivers of soil pH. A systematic workflow was implemented to evaluate a diverse set of algorithms, followed by Bayesian optimization to fine-tune the best-performing models. The results demonstrated that the optimized AdaBoost model yielded the largest performance improvement (~7.5%), achieving excellent accuracy on independent test data with a coefficient of determination (R²) of 0.817 and a mean absolute error (MAE) of 0.293. Furthermore, SHAP analysis identified iron (Fe) and calcium carbonate (CaCO₃) as the most influential predictors, revealing that Fe exhibits a strong inverse relationship with pH, while CaCO₃ shows a positive association. This framework successfully balances high predictive accuracy with pedological interpretability, offering a robust tool for digital soil mapping and precision agriculture.