Chen et al. (2026) A novel soil moisture retrieval method via combining radiative transfer model and machine learning

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-03-20
• Authors: Yu Chen, Cheng Tong, Qixuan Sun, Yulin Shangguan, Xin Deng, Mark Crowley, Hongquan Wang, Yang Ye, Haijun Bao, Ruqi Huang
• DOI: 10.1016/j.rse.2026.115378

Research Groups

• Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China
• College of Environmental and Resource Sciences, Zhejiang University, Hangzhou, China
• School of Spatial Planning and Design, Hangzhou City University, Hangzhou, China
• Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Ontario, Canada
• Agriculture and Agri-Food Canada Lethbridge Research and Development Centre, Lethbridge, Alberta, Canada

Short Summary

This study introduces a novel, interpretable soil moisture retrieval framework by integrating a radiative transfer model (RTM) with a Kolmogorov–Arnold Network (KAN) to derive explicit mathematical expressions from satellite observations, achieving global soil moisture estimates comparable in accuracy to the SMAP Level-3 product.

Objective

• To develop an interpretable soil moisture retrieval framework that fuses the physical principles of the radiative transfer model (RTM) with the mathematical formulation capability of the Kolmogorov–Arnold Network (KAN) to provide a direct, inspectable mapping from satellite observations to soil moisture.

Study Configuration

• Spatial Scale: Global
• Temporal Scale: Daily estimates for the period 2015–2023

Methodology and Data

• Models used: Radiative Transfer Model (RTM), Kolmogorov–Arnold Network (KAN)
• Data sources: SMAP observations (brightness temperature, surface temperature, vegetation optical depth), in situ soil moisture measurements from the International Soil Moisture Network (ISMN) and the Qinghai Lake Basin dense network (QLB-NET).

Main Results

• The KAN retrieval method achieved an average correlation coefficient (R) of 0.64 and an unbiased Root Mean Square Error (ubRMSE) of 0.07 m³/m³ when validated against in situ measurements.
• These results are comparable to the SMAP Level-3 product, which showed R = 0.65 and ubRMSE = 0.06 m³/m³.
• The framework successfully captured broad spatial patterns and seasonal dynamics of soil moisture.
• The explicit mathematical formulas extracted by KAN clarify the influence of brightness temperature, surface temperature, and vegetation optical depth on soil moisture.
• The derived expressions eliminate the need for KAN retraining or iterative solving, enhancing reproducibility and operational efficiency.

Contributions

• Introduction of a novel, interpretable soil moisture retrieval framework that synergistically combines physical modeling (RTM) with data-driven machine learning (KAN).
• Generation of explicit, inspectable mathematical expressions for soil moisture retrieval from satellite observations, offering a direct mapping without complex iterative processes.
• Improvement in reproducibility and operational efficiency by providing a compact retrieval expression that does not require retraining or iterative solving, unlike traditional RTM approaches.
• Demonstration of comparable accuracy to established satellite products (SMAP Level-3) while offering enhanced interpretability.

Funding

Not explicitly stated in the provided text.

Citation

@article{Chen2026novel,
  author = {Chen, Yu and Tong, Cheng and Sun, Qixuan and Shangguan, Yulin and Deng, Xin and Crowley, Mark and Wang, Hongquan and Ye, Yang and Bao, Haijun and Huang, Ruqi},
  title = {A novel soil moisture retrieval method via combining radiative transfer model and machine learning},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115378},
  url = {https://doi.org/10.1016/j.rse.2026.115378}
}