Abstract:

Terrestrial water storage (TWS) is a key indicator of the water cycle and climate change, and accurate monitoring of its variations is essential for understanding changes in the water cycle under climate warming. To address the inherent limitations of GRACE/GRACE-FO satellite gravity products in spatial resolution and temporal continuity, as well as the spatial discreteness and complex noise of GNSS vertical crustal displacement observations, this study proposes a TWS inversion framework based on physics-informed neural networks (PINN). The proposed method deeply integrates GNSS imaging results with fused GRACE products, enabling a paradigm shift from discrete station observations to continuous spatial field reconstruction. First, GNSS station vertical displacement time series are transformed into a regular-grid deformation field using spatial imaging techniques to enhance spatial continuity and hydrological signal characterization. On this basis, a PINN model is constructed with time and GNSS vertical crustal displacement as inputs and GRACE TWS as the output, and is compared with a backpropagation neural network under consistent parameter settings. The results show that, after introducing physical information constraints, the model significantly outperforms the purely data-driven model, and the reconstructed results show higher consistency with GRACE and GLDAS products in amplitude, phase, and interannual variability. In addition, the dynamic weighting strategy enables the model to effectively balance data fitting and physical constraints during training, thereby improving convergence stability and overall inversion accuracy. This study demonstrates that PINN provides a robust and transferable technical pathway for regional high-resolution TWS inversion constrained by geodetic observations.

Key words: GNSS, GRACE/GRACE-FO, physical information, machine learning, terrestrial water storage