Zhang et al. (2025) Multi-layer grid-scale soil moisture estimation using spatiotemporal deep learning methods with physical constraints

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Authors: Tuantuan Zhang, Zhongmin Liang, Jianhong Zhou, Quanxi Shao, Ranjan Sarukkalige, Haishen Lü, Jiangjiang Zhang, Chenglin Bi, Jun Wang, Yiming Hu, Binquan Li
• DOI: 10.1016/j.jhydrol.2025.133086

Research Groups

• College of Hydrology and Water Resources, Hohai University, Nanjing, China.
• School of Earth System Science, Tianjin University, Tianjin, China.
• CSIRO Data61, Australian Resources Research Centre, Kensington, WA, Australia.
• Department of Civil Engineering, Curtin University, Bentley, WA, Australia.
• The National Key Laboratory of Water Disaster Prevention, Hohai University, Nanjing, China.
• Yangtze Institute for Conservation and Development, Hohai University, Nanjing, China.
• Northwest Engineering Corporation Limited, Xi’an, China.

Short Summary

This study develops a physics-guided deep learning (PGDL) model that integrates the Richardson-Richards equation into a CNN-LSTM architecture to estimate multi-layer soil moisture. The approach significantly improves the accuracy and physical consistency of soil moisture predictions across depths of 10 cm to 50 cm compared to standard machine learning methods.

Objective

• To improve the accuracy, interpretability, and physical consistency of multi-layer grid-scale soil moisture (SM) estimation by constraining a spatiotemporal deep learning model with the Richardson-Richards equation.

Study Configuration

• Spatial Scale: Grid-scale estimation at a 10 km resolution.
• Temporal Scale: Spatiotemporal dynamics focusing on multi-layer depths (10 cm, 20 cm, 30 cm, 40 cm, and 50 cm).

Methodology and Data

• Models used: PGDL-CNN-LSTM (Physics-Guided Deep Learning combining Convolutional Neural Networks and Long Short-Term Memory). Comparison models included standard CNN-LSTM, Extreme Gradient Boosting (XGBoost), and Random Forest (RF).
• Physical Constraints: The Richardson-Richards equation was integrated into the CNN-LSTM loss function. Additionally, the LSTM activation function was constrained to ensure estimated values remained within the residual and saturated SM range.
• Data sources: Remote sensing data (surface SM), auxiliary topographic data, and meteorological data.

Main Results

• Performance Improvement: The PGDL-CNN-LSTM model outperformed all baseline machine learning methods across all soil layers (10–50 cm).
• Error Reduction: For the 10 cm soil layer, the proposed method achieved a 9.6% reduction in root mean square error (RMSE) compared to the standard CNN-LSTM.
• Physical Consistency: The integration of physical constraints led to a 92.0% reduction in physical inconsistency.
• Data Efficiency: The model maintained high effectiveness and accuracy even when training data was reduced by 50%.
• Root-Zone Estimation: The method successfully extended surface SM data to estimate root-zone SM at a 10 km resolution, capturing dynamic temporal changes effectively.

Contributions

• Methodological Innovation: Successfully integrated the Richardson-Richards equation as a physical constraint within a hybrid CNN-LSTM architecture for hydrological applications.
• Enhanced Interpretability: Addressed the "black box" nature of deep learning by guiding the model to adhere to known physical laws of moisture transport.
• Robustness: Demonstrated that physics-guided models are more resilient to data scarcity than purely data-driven approaches.

Funding

• Not explicitly detailed in the provided text.

Citation

@article{Zhang2025Multilayer,
  author = {Zhang, Tuantuan and Liang, Zhongmin and Zhou, Jianhong and Shao, Quanxi and Sarukkalige, Ranjan and Lü, Haishen and Zhang, Jiangjiang and Bi, Chenglin and Wang, Jun and Hu, Yiming and Li, Binquan},
  title = {Multi-layer grid-scale soil moisture estimation using spatiotemporal deep learning methods with physical constraints},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.133086},
  url = {https://doi.org/10.1016/j.jhydrol.2025.133086}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)