Abstract

The integration of artificial intelligence (AI) in soil property prediction has revolutionized digital soil mapping, yet the "black box" nature of complex machine learning models limits their adoption in agricultural decision-making and policy formulation. This study presents a comprehensive evaluation of Explainable AI (XAI) techniques applied to soil property prediction models, enhancing interpretability without compromising predictive accuracy. We implemented and compared four machine learning algorithms (Random Forest, XGBoost, Support Vector Machine, and Neural Networks) with three XAI methods (SHAP, LIME, and Permutation Feature Importance) for predicting soil organic carbon (SOC), pH, and available nitrogen across 3,247 sampling points in diverse agricultural landscapes. The Random Forest model achieved the highest accuracy (R² = 0.89 for SOC, 0.82 for pH, 0.78 for nitrogen) while maintaining superior interpretability through SHAP analysis. Key findings revealed that elevation, precipitation, and normalized difference vegetation index (NDVI) were the most influential predictors across all soil properties. SHAP waterfall plots successfully explained individual predictions, showing how each feature contributed to model decisions. The XAI framework identified non-linear relationships and feature interactions that traditional statistical methods failed to capture, including threshold effects of temperature on soil organic carbon and complex interactions between topographic variables. Model explanations demonstrated high consistency across different XAI methods, with correlation coefficients >0.85 between SHAP and LIME importance rankings. The developed XAI framework provides transparent, trustworthy soil property predictions, enabling informed agricultural management decisions and supporting sustainable farming practices. This research establishes a foundation for implementing explainable machine learning in precision agriculture applications.