Qi et al. (2026) Physics‐Informed Neural Networks to Develop Site‐Specific Pedotransfer Functions

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-01-01
• Authors: PengFei Qi, Yunquan Wang, Rui Ma, Jieliang Zhou, Harry Vereecken, Budiman Minasny, Ziyong Sun, Gaofeng Zhu, Kun Zhang
• DOI: 10.1029/2025wr041265

Research Groups

Not explicitly stated in the abstract.

Short Summary

This study introduces new Site-Specific Pedotransfer Functions (SPTFs) that combine deep learning with physics-based modeling to improve the estimation of soil hydraulic parameters. SPTFs achieve high accuracy (Nash-Sutcliffe Efficiency of 0.65, RMSE of 0.072 m³ m⁻³) in simulating soil water content at field scale, comparable to inverse modeling but with the computational efficiency of conventional PTFs.

Objective

• To develop and evaluate new Site-Specific Pedotransfer Functions (SPTFs) that integrate deep learning with physics-based modeling of soil hydrological processes to accurately estimate soil hydraulic parameters for field-scale applications.

Study Configuration

• Spatial Scale: Distributed across 1,181 sites globally (International Soil Moisture Network).
• Temporal Scale: Two years of soil moisture observations.

Methodology and Data

• Models used: Deep learning, 1-D Richardson-Richards equation.
• Data sources: Soil moisture observations from the International Soil Moisture Network (ISMN).

Main Results

• SPTFs achieved a Nash-Sutcliffe Efficiency (NSE) of 0.65 in simulating soil water content on the test set (n = 179).
• The root mean squared error (RMSE) for simulated soil water content was 0.072 m³ m⁻³ at a depth of 0.05 m.
• The performance of SPTFs is comparable to values predicted by the inverse modeling method.
• SPTFs maintain the computational efficiency of conventional pedotransfer functions (PTFs).

Contributions

• Introduction of novel Site-Specific Pedotransfer Functions (SPTFs) that combine deep learning with physics-based hydrological modeling.
• Utilization of time-series data as input for PTFs, departing from conventional static inputs.
• Direct optimization of simulated soil water content against observations using the 1-D Richardson-Richards equation, enhancing applicability to field conditions.
• Demonstrated improved accuracy and applicability for field-scale hydrological models compared to conventional PTFs, while retaining computational efficiency.
• Provides a robust parameterization framework for localized field applications.

Funding

Not explicitly stated in the abstract.

Citation

@article{Qi2026PhysicsInformed,
  author = {Qi, PengFei and Wang, Yunquan and Ma, Rui and Zhou, Jieliang and Vereecken, Harry and Minasny, Budiman and Sun, Ziyong and Zhu, Gaofeng and Zhang, Kun},
  title = {Physics‐Informed Neural Networks to Develop Site‐Specific Pedotransfer Functions},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041265},
  url = {https://doi.org/10.1029/2025wr041265}
}