Attributing terrestrial water storage changes on the Tibetan Plateau to climate and human drivers using a hybrid deep learning approach

The Tibetan Plateau (TP), referred to as “Water Tower” for Asia, is highly sensitive to both climate change and anthropogenic activities. However, assessing the spatiotemporal dynamics of terrestrial water storage (TWS) and attributing its changes across the TP remains challenging. This is primarily due to discontinuous satellite observations and the complex interactions between climatic drivers and potentially human-related influences. To address these limitations, we develop a VMD‑CNN‑LSTM‑based TWS Prediction and Hierarchical Attribution (VCLHA) framework. VCLHA can quantify the independent contributions of climatic factors (precipitation, evapotranspiration, runoff, and temperature) and a residual component potentially associated with human activities to TWSC. The evaluation results demonstrate that the framework has robust performance, achieving a Nash-Sutcliffe Efficiency (NSE) value of 0.983. Robustness is further supported by intercomparison with independent GRACE-like reconstructions (e.g., GRACE-REC, CSRM-Bayesian, and BNML_TWSA) during the GRACE/GRACE-FO observational gap (July 2017–May 2018), showing consistent plateau-mean evolution. The prediction results reveal pronounced spatial heterogeneity, with water increases of 1.87 mm yr −1 in northern TP and severe water decreases of 8.17mm yr −1 in the south. The precipitation is identified as the dominant climatic driver, explaining 43.2% of the variance in TWS change. However, while the residual component potentially associated with human activities shows no clear long-term trend (0.05 mm yr −1 ), it exhibits strong seasonality, characterized by positive effects in spring and negative effects in summer (−2.81 mm) and autumn (−1.62 mm). This seasonal variability amplifies water-storage deficits during summer–autumn (July–October), particularly in southern and outflow basins. These findings provide quantitative evidence for the relative roles of climate variability and human-related influence in regional water storage variability, thereby supporting adaptive and region-specific water-resource management strategies on the TP. Plain Language Summary. The Tibetan Plateau (TP) supplies freshwater to many of Asia’s major rivers and supports hundreds of millions of people downstream. However, it is challenging to determine how much recent changes in water availability are related to climate change and human-related influences. In this study, we develop a hybrid deep learning framework to reconstruct satellite gravity observations and to assess the relative roles of natural processes and a residual component potentially associated with human activities in water storage change. From 2003 to 2023, climate-related processes have led to a steady long-term decline in water storage (−0.10 mm per year). While the residual component potentially associated with human activities did not exhibit a consistent long-term trend, it had a pronounced seasonal impact, contributing to water gains in spring but intensifying losses in summer (−2.81 mm) and autumn (−1.62 mm). These seasonal deficits were most severe in southern, downstream basins, whereas northern and interior regions showed partial recovery in spring. Precipitation was found to be the most important climate factor influencing water storage. These insights can help improve drought preparedness, irrigation planning, and reservoir management as climate change continues to reshape the regional water cycle.