A Hybrid Physics‐Guided Deep Learning Modeling Framework for Predicting Surface Soil Moisture

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Xi, Xuan, Zhuang, Qianlai, Liu, Xinyu
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2025JH000682

Accurate prediction of surface soil moisture (SSM) is vital for understanding the complex interactions between terrestrial and atmospheric processes with significant implications for weather forecasting, agriculture, and water management. In this study, we introduce an innovative physics‐guided deep learning (PGDL) model by integrating the process‐based insights of the terrestrial ecosystem model (TEM) with the dynamic predictive capabilities of long short‐term memory (LSTM) networks to improve SSM prediction. The PGDL model leverages the complementary strengths of the deterministic framework of TEM and the data‐driven prowess of LSTM, providing predictions that are deeply rooted in physical processes while capturing complex patterns in data. We evaluated the PGDL model against traditional process‐based (PB) models and pure deep learning (DL) approaches in both single‐site and multisite simulations. For single‐site simulations, we performed time‐based partitioning on 7 sites with different vegetation types, whereas for multisite simulation, data from 13 sites within the same vegetation type were used for model training and evaluation. In single‐site simulations, PGDL achieved significantly lower RMSE (0.04) and higher R 2 (0.66) compared to PB (RMSE: 0.13, R 2 : −2.32) and DL (RMSE: 0.06, R 2 : 0.36). In multisite simulations, PGDL (RMSE: 0.06, R 2 : 0.63) also outperformed PB (RMSE: 0.09, R 2 : 0.28) and DL (RMSE: 0.08, R 2 : 0.38). Our results show that the PGDL modeling framework improves the predictive accuracy of DL models and the physical interpretability of PB models, which can serve as a robust tool to predict SSM dynamics. In our study, we developed a new model to predict how wet the soil is at the surface, which is important for understanding how land interacts with the atmosphere. This information is crucial for better weather forecasting, farming, and managing water resources. Our new method combines traditional process‐based models with advanced deep‐learning techniques. By blending these approaches, our model not only relies on established physical principles but also learns from patterns in data that process‐based methods might miss. We evaluated our approach through single‐site and multisite simulations. For single‐site simulations, we independently analyzed data from 7 sites each with a different vegetation type. For multisite simulation, we used data from 13 sites within the same vegetation type. Compared to traditional process‐based and pure deep‐learning models, our new model achieves higher accuracy and more physically consistent predictions. The success of our model highlights the advantage of combining the interpretability of traditional scientific methods with the flexibility of big data analytics. This could lead to better ways of predicting soil moisture, which is becoming increasingly important as the climate changes. A new deep learning modeling framework guided by physical processes was developed to predict surface soil moisture The physics‐guided deep learning (PGDL) model outperforms the traditional process‐based model and pure deep learning model in prediction accuracy Hydrological process consideration increases the interpretability and reliability of the PGDL model