Combining Satellite Precipitation Products and Deep Learning to Increase Lead Times in Real-Time Riverine Flood Risk Forecasting

This research addresses the difficulties to increase lead time on predictions of riverine flooding risks represented by the rainfall in extreme or long-duration weather systems originated from ungauged areas. The methodology explores the opportunity to enhance rainfall data coverage and simplify forecasting tasks by combining the global reach and high resolution of IMERG V07 satellite precipitation products (SPPs) with the ability of deep learning models to capture complex spatiotemporal relationships in time series data. In a real-world case study, the method applies a long short-term memory (LSTM) model to capture patterns in the historical relationship between IMERG rainfall estimates from selected areas of the Atlantic Ocean and stream level variations in three UK catchments (C1, C2, and C3). The model then utilizes near-real-time (NRT) data from the IMERG early run product to make real-time predictions. Lead times are determined by considering three key factors: the latency of the NRT data, the distance between the catchment and the IMERG data collection point, and the forward speed of the weather system carrying rainfall toward the catchment. The method was applied to predict stream level variations during two extreme rainfall events and results compared to those obtained from a similar LSTM model using local rain gauge data. Through this comparison, across all catchments the proposed methodology demonstrated significantly smaller prediction errors for lead times exceeding 1.5 hours on both events. For example, with NRT IMERG data, 6.5-hour lead time predictions for C1, C2, and C3 had RMSE values of 19 mm, 21 mm, and 26 mm, respectively, for the 2022 event, and 16 mm, 29 mm, and 45 mm for the 2023 event. In contrast, predictions with the same lead time using rain gauge data resulted in RMSE values of 77 mm, 64 mm, and 59 mm for the 2022 event, and 165 mm, 44 mm, and 112 mm for the 2023 event. More importantly, considering that during the rainfall events water level rose about 600mm in C1, up to 700 mm in C2 and up to 1000mm in C3, the errors with the proposed methodology remained below 10% of the total water level rise in each catchment on predictions with up to 9 hours lead time. While these are excellent results for real-time applications of flooding forecasts, the 4-hour latency of NRT IMERG data limits the method's applicability for predictions with less than 4 hours lead time and for floods triggered by localized or short-duration rainfall events.