Global High-Resolution Terrestrial Water Storage Anomalies through a Dynamic Soft-Constrained Deep Learning Paradigm

This dataset provides a comprehensive, high-resolution global record of monthly Terrestrial Water Storage Anomalies (TWSA) from April 2002 to December 2022. It was generated to address the spatial resolution limitations of raw satellite gravimetry observations from the GRACE and GRACE-FO missions, offering a product suitable for regional and basin-scale hydrological analysis. The dataset was simulated using a novel deep learning framework. This model spatially downscales the low-resolution (~300 km) JPL GRACE/GRACE-FO Mascon solutions (Watkins et al., 2015; Wiese et al., 2016; Landerer et al., 2020; Wiese et al., 2023) to a high resolution of 50 km (0.5-degree grid). The core of the methodology is a dynamic soft-constrained paradigm, where the model is simultaneously guided by the observational accuracy of GRACE/GRACE-FO data and the high-resolution spatial patterns from the WaterGAP Global Hydrology Model (Müller Schmied et al., 2023) and ERA5 reanalysis data (Hersbach et al., 2023). The influence of these constraints is dynamically weighted at each training step based on the evolving correlation between the model's prediction and the high-resolution inputs, ensuring an optimal simulation of observational integrity and high-resolution detail.