Filling GRACE data gap using an innovative transformer-based deep learning approach

Abstract

The terrestrial water storage anomaly (TWSA), derived from the Gravity Recovery and Climate Experiment (GRACE) and its successor, the GRACE Follow-on (GRACE-FO) satellite, presents a remarkable opportunity for extreme weather detection and the enhancement of environmental protection. However, the practical utility of GRACE data is challenged by an 11-month data gap and several months of missing data. To address this limitation, we have developed an innovative transformer-based deep learning model for data gap-filling. This model incorporates a self-attention mechanism using causal convolution, allowing the neural network to capture the local context of GRACE time series data. It takes into account various factors such as temperature (T), precipitation (P), and evapotranspiration (ET). We trained the model using a global dataset of 10,000 time series pixels and applied it to fill all the time gaps. The validation results demonstrate its robustness, with an average root mean square error (RMSE) of 6.18 cm and Nash-Sutcliffe efficiency (NSE) of 0.906. Notably, the Transformer-based method outperforms other state-of-the-art approaches in arid regions. The incorporation of T, P, and ET has further enhanced the accuracy of gap filling, with an average RMSE decrease of 7.5 %. This study has produced a reliable gap-filling product that addresses 11-month data gaps and 24 isolated gaps, ensuring the continuity of GRACE data for various scholarly applications. Moreover, our Transformer approach holds important potential for surpassing traditional methods in predicting and filling gaps in remote sensing data and gridded observations.