Long-term reconstruction or prediction of spatial and temporal information on land or terrestrial water storage (TWS) dynamics globally is critical and highly challenging. A hybrid approach was proposed to combine the strengths of physically-based modeling and deep learning for estimating global TWS anomalies (TWSA). Specifically, we developed a spatiotemporal attention-based deep learning model (STAU-Net), integrating the U-Net architecture with ConvLSTM layer and convolutional block attention module (CBAM) to learn the spatiotemporal patterns of TWSA observed by GRACE, driven under different predictor combinations, e.g., meteorological forcings, soil properties, and modeled TWSA. Once trained and validated, the model can estimate long-term global TWS dynamics without requiring GRACE TWSA as inputs. The evaluation results suggest that the hybrid approach can provide improved predictions of global TWSA compared to others. This study demonstrates the unique ability of the hybrid approach in global freshwater availability monitoring and prediction.
|