A Machine Learning approach for Total Water storage anomaly eXtension back to 1980 (ML-TWiX)

TL;DR

This paper introduces ML-TWiX, a dataset extending global water storage data back to 1980 using machine learning, enhancing climate and hydrology studies.

Contribution

The novel contribution is a machine learning-based global TWSA dataset extending back to 1980, filling the pre-GRACE era gap.

Findings

ML-TWiX uses Random Forest, XGBoost, and Gaussian Process Regression to reconstruct TWSA from 1980 to 2012.

The dataset was validated against satellite laser ranging and sea level budget estimates.

Ensemble averaging provides a unified TWSA product with spatial uncertainty estimates.

Abstract

We present ML-TWiX, a global dataset of monthly total water storage anomalies (TWSA) reconstructed from 1980 to 2012, provided on a 0.5° × 0.5° global grid. While the GRACE and GRACE Follow-On satellite missions have provided valuable observations of global TWSA, their combined record spans just over two decades, limiting their utility for long-term climate and hydrological studies. ML-TWiX extends the GRACE-era record into the pre-GRACE period by learning from global hydrological and land surface model simulations using an ensemble of three machine learning models: Random Forest, XGBoost, and Gaussian Process Regression. The three machine learning models were independently used to reconstruct TWSA, and their outputs were subsequently combined through ensemble averaging to produce a unified product with spatially explicit uncertainty estimates. We validated ML-TWiX against multiple independent datasets, including satellite laser ranging, storage deduced from the water mass balance closure, and global mean sea level budget estimates. It provides a continuous reconstruction of global TWSA, enabling a wide range of applications in hydrology, climate science, and water resource assessment.

Figures