Abstract

The widespread adoption of artificial intelligence in digital soil mapping has achieved remarkable predictive accuracy, but the "black box" nature of complex models limits their acceptance among soil scientists, agronomists, and land managers who require understanding of prediction rationale for decision-making. This study presents a comprehensive evaluation of Explainable AI (XAI) techniques for enhancing interpretability in soil property prediction models while maintaining predictive performance. We implemented and compared five XAI approaches: SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), Integrated Gradients, Attention Mechanisms, and a novel Soil-Specific Explanation Framework (SSEF) across 18,934 soil samples from diverse agricultural landscapes in North America and Europe. The models predicted soil organic carbon (SOC), pH, clay content, and available nitrogen using 68 environmental covariates including climate, topography, remote sensing, and geological variables. Model architectures included Random Forest, XGBoost, Neural Networks, and Transformer-based approaches to evaluate XAI effectiveness across different algorithm types. The SSEF approach achieved superior explanation quality with fidelity scores of 0.94 for SOC, 0.91 for pH, 0.89 for clay content, and 0.87 for nitrogen predictions. Feature importance rankings showed high consistency (Spearman correlation >0.85) across different XAI methods, with climate variables (precipitation, temperature) and topographic indices (elevation, slope) emerging as primary predictors. SHAP analysis revealed non-linear relationships and interaction effects previously undetected through traditional statistical approaches, including threshold effects of precipitation on organic carbon accumulation and complex terrain-climate interactions affecting soil pH. Local explanations through LIME successfully identified region-specific prediction drivers, with coastal areas prioritizing salinity-related variables and mountainous regions emphasizing elevation and slope factors. Attention mechanism visualization in transformer models revealed spatial patterns in feature importance that aligned with known soil formation processes. User evaluation studies with 47 soil scientists and agronomists demonstrated significant improvements in model trust (78% increase), decision confidence (65% increase), and practical adoption intentions (82% increase) when XAI explanations were provided. Computational overhead for explanation generation averaged 12-18% of prediction time, making real-time interpretability feasible for operational applications. The study establishes practical frameworks for implementing explainable AI in soil science applications, bridging the gap between predictive accuracy and scientific understanding while supporting evidence-based agricultural decision-making.