A new study published in Science of Remote Sensing presents a substantial improvement in reconstructing missing data from NASA's Gravity Recovery and Climate Experiment (GRACE) and its successor, GRACE Follow-On (GFO).

The research, led by The Ohio State University's Byrd Center Postdoctoral Scholar Jaydeo Dharpure, utilizes machine learning (ML) and deep learning (DL) techniques alongside ERA5-Land reanalysis data at an enhanced, more precise 0.5-degree grid cell resolution to estimate terrestrial water storage anomalies (TWSA),  TWSA measures the temporal variability in the total water stored on land relative to its long-term average.

Other co-authors of the study are Byrd Center principal investigators, Earth Sciences Professor Ian Howat from the Glacier Dynamics Research Group, Geography Professor Bryan Mark from the Glacier Environmental Change Research Group, and former Postdoctoral Scholar Saurabh Kaushik.

The GRACE and GFO satellite missions have provided crucial data for tracking global water storage changes. Yet, interruptions in their observations, particularly between July 2017 and May 2018, have posed challenges for long-term hydrological and climate studies.

Unlike past studies that relied on a single prediction method, this study employed a multi-model approach to improve accuracy across different regions by selecting the best-performing model for each 0.5° grid cell. Researchers trained five machine learning (ML) and deep learning (DL) models—Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGB), Deep Neural Network (DNN), and Stacked Long-Short Term Memory (SLSTM)—using eight predictor variables from April 2002 to August 2022. These variables included six hydroclimatic factors (temperature, precipitation, runoff, evapotranspiration, ERA5-Land-derived TWSA, and cumulative water storage change), along with a vegetation index and timing variables. 

The models were tested across five climate zones and twenty six major river basins, particularly during extreme events like floods, droughts, and sea-level rise. Results showed that SVM was the most effective ML model, while SLSTM proved to be the most accurate overall, especially in regions with strong seasonal water variations, such as the Amazon Basin in South America.

Interestingly, no single model was optimal everywhere. For this reason, the researchers used a "Leader" model approach, selecting the best-performing model for each grid cell rather than applying the same model worldwide. This strategy reduced errors and improved accuracy across different environments. Compared to previous studies, this multi-model approach showed superior performance and produced a continuous, reliable datasets for tracking water resources.

The resulting high-resolution global TWSA datasets generated through the robust machine learning and deep learning framework, shed light on global water storage dynamics. By enabling improved assessments of extreme climate events, this innovative approach enhances our capacity to monitor Earth's hydrological changes, benefiting researchers and offering insights for global water management.

To learn more, visit Combining machine learning algorithms for bridging gaps in GRACE and GRACE Follow-On missions using ERA5-Land reanalysis .