Nowcasting for urban flash floods in Africa: a machine-learning and satellite-observation based model

On the world’s fastest urbanizing continent, Africa, urban floods are a real and growing problem. Early warning is the first important step in flood risk management. This requires continuous and reliable precipitation measurements and forecasts, which are not always available in African cities.

In this study a nowcasting model based on Convolutional Neural Network (TrajGRU) was developed for short-term, 0-2 hours, precipitation forecast in Ghana, West Africa. The nowcasting model is trained on historical rainfall estimates derived from the MSG-SEVIRI instrument by the Nighttime Infrared Precipitation Estimation (NIPE) model. Input for the model are real-time NIPE MSG-SEVIRI estimates.

The Meteosat Second Generation (MSG) SEVIRI instrument provides high-resolution and short-latency data, covering Europa and Africa. Especially in areas without radar observations, MSG offers unique and relevant information for early warning with respect to fast occurring events such as urban flash floods. Its resolution allows for the retrieval of convective rainfall, often a cause of flash floods in tropical areas.

To assess the performance of the model, we compare our method to current state-of-the-art Lagrangian nowcasting system from the pySTEPS library applied to the NIPE-MSG-SEVIRI data.

The result is an operationally running model for nowcasting two hours ahead with 15 minutes temporal and approximately three kilometer in Ghana (rainsat.net). The method is readily applicable in other regions in Africa.