Enhanced rainfall nowcasting of tropical cyclone by an interpretable deep learning model and its application in real-time flood forecasting

Reliable Tropical Cyclone (TC) rainfall and flood forecasts play an important role in disaster prevention and mitigation. Numerous studies have demonstrated the promising performance of deep learning in hydrometeorological forecasts. However, few studies have investigated the potential enhancement of advanced TC track forecasts in predicting rainfall and induced flood. In this study, a novel rainfall nowcasting model (TCRainNet) is developed by fusing TC track characteristics with antecedent rainfall in a Convolution LSTM to predict hourly rainfall with a lead time of 6 h. The nowcasts are subsequently used to drive an event-based Xin’anjiang hydrological model for real-time flood forecasting. The model performance is interpretated by the occlusion sensitivity approach, and the propagation of errors from TC track forecasts to flood forecasts is quantified. The results underscore the superiority of TC track characteristics as input features for rainfall nowcasts, as indicated by a Mutual Information value of up to 0.51. The generated nowcasts are found to have averaged Probability of Detection (POD) and Critical Success Index (CSI) greater than 0.27 and 0.2 respectively. The Mean Absolute Error (MAE) of the nowcasts falls below 2.6 mm, which is only 46 % of the ECMWF operational high-resolution forecasts. The rainfall-driven flood forecasts have NSE greater than 0.7 and PBIAS smaller than 20 % with lead time up to + 4 h. It is shown that the position error of $0.{45}^{°}$ and intensity error of 10 hPa&7.8 m/s in TC track forecasts generally result in 0.9 mm degradation in rainfall forecasts and 10% decline in the accuracy of rainfall-driven flood forecasts. The effectiveness of our method presents favorable applicability in advancing disaster mitigation efforts.