Interpretable Soil Water Retention Prediction Using Hierarchical Attention Networks With Uncertainty Quantification

Abstract Understanding which soil properties control water retention at different soil moisture states remains critical for predicting soil responses to management and environmental change. Traditional pedotransfer functions predict parameters of equations like van Genuchten's model, but parameter correlations create ill‐posed inverse problems that propagate errors. Fundamentally, parameter‐based approaches cannot reveal which properties matter most at different water potentials or how texture‐structure interactions vary across the moisture spectrum. This study presents Hierarchical Attention‐Based Pedotransfer Function (HABIT) to address these limitations through two objectives: (a) achieve superior prediction accuracy through direct soil moisture prediction that circumvents parameter correlation errors, and (b) quantify property importance and interactions across soil moisture regimes through interpretable attention mechanisms. The model employs attention architectures that dynamically weight soil properties based on water potential, enforces physical constraints including monotonicity, and accommodates incomplete soil characterization through hierarchical training. HABIT is evaluated against traditional approaches using comprehensive international soil databases spanning global textural and structural variability. Attention weight analysis reveals moisture‐dependent property interactions, including unequal bidirectional texture‐structure interactions where texture exerts stronger influence on structural interpretation than the reverse, organic carbon's dual structural and sorptive roles, and saturated hydraulic conductivity's role as a statistical surrogate encoding pore network information through correlations with unmeasured properties. These patterns generate testable predictions while explaining how direct prediction with adaptive property weighting captures soil‐specific retention behaviors that parametric forms miss. HABIT demonstrates that properly designed machine learning architectures serve as both prediction tools and hypothesis‐generating frameworks for quantitative analysis of soil property interactions.