Bridging the data gap between GRACE and GRACE-FO using artificial neural network in Greenland

Abstract

The Gravity Recovery and Climate Experiment (GRACE) provided an entirely new way to measure mass changes on the Earth at unprecedented accuracy and resolution. However, the delayed launch of the GRACE Follow On (FO) mission led to an approximately 1-year gap between GRACE and GRACE-FO data, breaking the continuity of observation and hampering data analysis. Efforts have been made to bridge this gap for the major river basins, but little has been done on ice sheets. To address this limitation, we evaluated multiple linear regression (MLR), a back propagation neural network (BPNN), and a deep belief network (DBN) to fill the data gap in Greenland and its six sub-regions. We employed these methods to establish the relationships between precipitation, runoff, evapotranspiration, and ice discharge and GRACE-estimated ice mass changes. A sliding window testing method shows that the BPNN outperformed the two other methods with a root mean square (RMS) error of 1.5 cm in the metrics for the equivalent water height for the whole of Greenland and 1.5–3.6 cm (2.5 cm in average) for the six sub-regions. This accuracy is rather high given an uncertainty of ∼ 2 cm in GRACE estimates, but higher for the whole of Greenland than in the sub-regions. This difference was caused by mass leakage error between adjacent sub-regions in the GRACE observations, and can be mitigated by merging these sub-regions. The BPNN predicted ice mass changes agree with the GRACE estimates for long-term and seasonal signals, but are less accurate and precise for interannual signals. Based on the GRACE observations and the BPNN predicted values, we re-estimated the ice mass change in Greenland finding that ice loss on Greenland decelerated after mid-2012, primarily due to deceleration in SW, SE, and NE regions.