Bridging the spatiotemporal ice sheet mass change data gap between GRACE and GRACE-FO in Greenland using machine learning method

Abstract

The Gravity Recovery and Climate Experiment (GRACE) mission terminated its operation in October 2017, and its successor, the GRACE Follow‐On mission, started operation in May 2018, leading to a 11-months data gap. As a prospective way to evaluate the Greenland ice sheet (GrIS) mass change, it would be vital to bridge the data gap. Despite efforts to bridge the gap with machine learning (ML) techniques, there are still two flaws in terms of the lack of spatial distribution and selecting forcing variables. Firstly, we applied the partial least squares regression (PLSR) model to assess the most explanatory variables of GRACE-derived ice mass change, and thus selected satellite altimetry elevation, and runoff, sublimation, rainfall, and snowfall from MAR v3.12 (Modèle Atmosphérique Régionale) model as predictors for predictive models. SVM (Support Vector Machine), BPNN (Back Propagation Neural Network), and MLR (Multiple Linear Regression) are utilized to reconstruct the GRACE-like gridded data. The SVM outperforms the other two predictive models, with 25.6 % of grid cells having Nash-Sutcliffe efficiency (NSE) coefficient above 0.75 and Scaled Root Mean Square Error (RSR) above 0.50. The average RMSE values for all testing periods are 3.86, 4.37, and 5.57 cm, for SVM, BPNN, and MLR, respectively. The reconstructed GRACE-like grid data demonstrated accurate spatiotemporal information with less noise, and exhibited a higher mean correlation of 0.93 across six sub-regions of GrIS, indicating a stronger agreement with original GRACE data compared to other methods. Furthermore, we investigated the changes in precipitation, runoff, and temperature pattern driven by the North Atlantic Oscillation (NAO), Greenland Blocking Index (GBI) and Atmospheric Rivers (ARs) to corroborate the shifts in GrIS mass loss feature. It was found that there was a correlation between precipitation, runoff, and ARs frequency, with mean correlation of 0.48 and 0.51 for wintertime and summertime, respectively. The moderate mass loss rate during the data gap can be attributed to the positive NAO and higher ARs frequency in GrIS. In general, the reconstructed data presented in this study provided a better comprehension of ice mass change and valuable information on how the ice mass responds to climate variations.