Physics‐Informed Neural Networks to Develop Site‐Specific Pedotransfer Functions

Abstract Pedotransfer functions (PTFs) are widely used to estimate soil hydraulic parameters based on easily accessible soil information, playing an important role in the parameterization of earth surface models. However, conventional PTFs, developed using measurements from small volume soil samples, often exhibit significant deviations from field observations and substantial variability when applied to field‐scale hydrological models. Here, we introduce new Site‐Specific Pedotransfer Functions (SPTFs) that combine deep learning with physics‐based modeling of soil hydrological processes. SPTFs differ from conventional PTFs in two aspects: they utilize time‐series data as input and they directly optimize simulated soil water content by the 1‐D Richardson–Richards equation with observations, ensuring improved applicability to field conditions. We trained and tested the model using two years of soil moisture observations from 1,181 sites in the International Soil Moisture Network. Evaluation using field data demonstrates that SPTFs achieve a Nash‐Sutcliffe Efficiency of 0.65 and root mean squared error of 0.072 cm 3 cm −3 in simulating soil water content at the depth of 0.05 m on the test set ( n = 179), which is close to the values predicted by the inverse modeling method, while maintaining the computational efficiency of PTFs. This study highlights the promise of SPTFs as a robust parameterization framework for localized field applications.