Improved grid-based semi-centennial reconstruction and diagnostic assessment of global water storage leveraging ensemble machine learning

Abstract

Gravity Recovery and Climate Experiment (GRACE)-based terrestrial water storage (TWS)—an essential climate variable—has revolutionized our understanding of global climate systems, terrestrial water cycles, glacier mass balance, and sea level changes over the past two decades. However, the data brevity of GRACE hinders a holistic understanding of the hydrological cycle and water fluxes, and the available reconstructed data have several limitations. Here, we leverage a suite of machine learning models, namely, Artificial Neural Network, Random Forest (RF), and Long Short-Term Memory (LSTM), to reconstruct the grid-based global TWS from 1980 to 2024 and validate with GRACE-based observations. Further, we examine the secular trends and variability of the 45-year long-term TWSA data using a recently proposed metric, namely the trend to variability ratio (TVR). The results show that three models performed well in over 88% of the study area, with RF as the best-performing model covering 48.5% of the area, followed by Long Short-Term Memory (LSTM) (27.9%) and Artificial Neural Network (ANN) (23.6%). Global TWS experienced five multiyear depletion events, with an exceptional water loss of 34.41 Gt during April 2011–January 2013, and has been continually below-average since 2012, primarily concentrated in regions with high aridity, heavy ice mass loss, extreme climatic shifts, and extensive human activities. TVR identifies significant water losses and gains in the majority of hotspot regions, which would have been otherwise overlooked in conventional linear trend analysis. The study offers crucial insights into identifying long-term water depletion areas, enabling proactive planning, prioritizing interventions, and sustainable water resource allocation for decision-makers.