Improving Global Surface Soil Moisture Prediction Through Physics‐Guided Deep Learning and Cluster‐Based Regionalization

Abstract Surface soil moisture (SSM) is essential to the hydrological cycle and land–atmosphere interactions, and its accurate simulation is crucial for climate prediction and resource management. This study developed an innovative modeling framework for global SSM prediction by integrating physics‐guided deep learning (PGDL) and clustering‐based regionalization. The PGDL model combines the physical knowledge from the Terrestrial Ecosystem Model (TEM) and the temporal learning capacity of long short‐term memory (LSTM) networks. By introducing a clustering strategy based on multi‐source features, the global land was divided into subregions with consistent characteristics. Within this framework, cluster‐specific models were trained using in situ observations and evaluated at the global scale against independent satellite observations. This clustering approach enhanced model generalization across diverse climatic and geographic conditions, yielding more robust predictions based on environmentally consistent samples. Results show that the PGDL model (RMSE: 0.081, : 0.55) outperformed both the process‐based (PB) model (RMSE: 0.167, : 0.43) and the purely deep learning (DL) model (RMSE: 0.085, : 0.40) at the global scale, while also exhibiting stronger physical consistency with water balance diagnostics. After excluding regions with high uncertainty in SSM observations, the performance of all models improved, with PGDL maintaining the best performance (masked RMSE: 0.064, masked : 0.58). Overall, this study demonstrates the superiority of the PGDL model and highlights the importance of clustering strategies in model construction and evaluation for achieving more accurate and robust SSM predictions across heterogeneous environments.