Long short-term memory exploitation of satellite gravimetry to infer floods

• LSTM successfully model the TWS spatio-temporal behavior. • LSTM have been able to clearly detect the amount of TWS leap in the month before the flood. • LSTM increases spatial resolution and reveals the pre-flood signal. Flood forecasting is a vital segment of disaster risk management in that it contributes to the prediction of the magnitude, occurrence, duration and timing of floods. Owing to the nonlinear nature of atmospheric phenomena, however, forecasting becomes a challenging task that requires a multifaceted approach involving various sensors. Indeed, there exist compounding evidence that flood processes would benefit from use of various sensors. One such sensor is the Gravity Recovery and Climate Experiment (GRACE) and its follow-on (GRACE-FO), which provides Total Water Storage (TWS) products that could potentially be useful for flood monitoring and forecasting. However, GRACE/GRACE-FO’s coarse spatial resolution of 300 km remains a bottleneck to the full exploitation of its products for flood studies and management. Herein, a deep learning Long Short-Term Memory (LSTM) method with high learning capability that optimizes the hyperparameters is proposed to downscale the coarse GRACE/GRACE-FO TWS products (from 300 km to 55 km). Its spatial and temporal learning is subjected to three different training scenarios (i.e., 60 %, 70 % and 85 %), where the one with least root-mean-square-errors (RMSE) is selected as the best-case scenario. The proposed LSTM deep learning approach is tested based on the 2019 Lorestan flood in Iran, where the results show that it successfully models the spatio-temporal behavior of TWS changes with its long-term and short-term memory capabilities. In March and April 2019, heavy precipitation caused a significant increase in TWS changes, approximately 40 ± 2 cm. This is captured by the LSTM-downscaled products but not the coarse GRACE/GRACE-FO TWS changes. Furthermore, the LSTM downscaled GRACE-FO TWS for the period after 2018 shows a strong and statistically significant mean correlation (above 0.70 at the 95 % confidence level) with both river discharge and precipitation. The original GRACE-FO on the other hand shows a correlation of 0.40, indicating the superiority of the LSTM-derived GRACE-FO’s TWS changes. The coarse resolution of the GRACE satellite is a major cause of low correlation, which improves after downscaling. LSTM thus has the potential of downscaling GRACE products, providing data that are useful for flood process, management and studies.