Physics-informed reconstruction of terrestrial water storage anomalies over the Tibetan Plateau (1982–2024)

We develop a physics-informed deep learning framework (PI-CNN-LSTM) that integrates water-cycle mass conservation into a CNN–LSTM architecture by constraining precipitation, evapotranspiration, and runoff processes. The model substantially outperforms conventional deep-learning approaches and enables reconstruction of monthly TWS variability across the TP at 0.25° resolution from 1982 to 2024.