Nowcasting convective cores using deep learning

Abstract Extreme precipitation events in the Tropics are often linked to convective storms. Storm nowcasting systems for early warning have been recognized as a necessary adaptation strategy to increase public resilience in a rapidly evolving climate. Ongoing advances in deep‐learning techniques now allow better exploitation of satellite data, including where ground‐based radar networks and computational resources are limited. This study presents an easily reproducible model (Network for Nowcasting Convective Cores, NetNCC) for nowcasting convective cores in near real time at a fine spatial resolution (). NetNCC uses Meteosat Second Generation thermal infrared brightness temperature images at three successive time steps to predict the probability of convective cores for lead times between 1 and 6 hr. A spatially enhanced loss function (fractions skill score) captures the atmospheric conditions within the neighbourhood of each pixel. Skilful probabilistic predictions are achieved at fine spatial resolutions () for short lead times of and at coarser resolution (180–270 km) for lead times . Nowcast accuracy shows a clear diurnal cycle, with higher accuracy during the daytime when convection frequency is at a maximum in eastern and southern Africa. NetNCC performs better for short‐lived convective cores that develop over eastern Africa than for cores embedded in typically long‐lived West African convection. Compared with persistence, NetNCC nowcasts are considerably better particularly for fine spatial scales. Further input information on atmospheric and land‐surface states alongside thermal infrared imagery may be required to improve longer (4 hr) lead‐time predictions. Nevertheless, our results indicate considerable potential of NetNCC for nowcasting convective activity, and hence intense rainfall, in at‐risk regions and illustrate the utility to forecasters for operational applications.