Nowcasting with Artificial Intelligence for African Rainfall: NAIAR

Lead Research Organisation: UK CENTRE FOR ECOLOGY & HYDROLOGY
Department Name: Hydro-climate Risks


This project aims to use new digital solutions to create 0 to 6 hour predictions - nowcasting - for tropical storms using satellite data. The methods will be developed and rolled-out for Africa, where people urgently need information about storm hazards, through our existing online platforms and smartphone apps. In this way the results of the research will be used to deliver information on storm hazards to users within minutes. The project very closely addresses the NERC Digital Strategy.

Tropical storms are very unpredictable, changing very rapidly - explosively - over timescales of an hour or so. For this reason, predictions are naturally very uncertain. Very often, the most important information people need regarding a storm hazard is what is happening now, and some information about how the storm likely to move and develop in the next couple of hours. This process is called "nowcasting" and in the USA, nowcasting of tornados saves many lives every year. The lack of weather radars in most African countries means that nowcasting is almost completely absent, but we have recently shown that satellite methods can provide useful nowcasting of storms too. The new Meteosat Third Generation (MTG) satellite will provide even better data coverage, from about 2024, at higher frequency and finer spatial scale. There is a tremendous opportunity to innovate in the creation of new nowcasting methods and communicate them to weather services, organisations and the public across Africa.

While existing satellite nowcasting methods have some skill, they also have major shortcomings. They work by extrapolating observed patterns forward in time, but are not constrained to obey the laws of physics, and unphysical predictions commonly occur. The most challenging problem in storm nowcasting is to predict the initiation and subsequent development of new storms in future: there is no accepted way to do this, and our considerable knowledge of the physics of initiation is not being exploited. It takes about 30 minutes to generate these nowcasts, and when their accuracy is degrading after an hour or two, their use becomes limited. We aim to create useful 6-hour nowcasts.

Nowcasting is an obvious application where new data-science methods, in particular machine-learning (ML), have the potential to make a massive impact, and a number of groups have begun to propose practical solutions. We need fundamental research to understand and improve the performance of these data-driven solutions, on the basis of the underlying physics and fluid-dynamics of storms. For instance, existing methods can extrapolate an image of a storm forward in time using ML to predict its future movement or growth, but the result may grow and be distorted in shape in a way which is incompatible with the laws of physics. These unrealistic predictions are obvious to an experienced forecaster but ordinary users of the data will be vulnerable to the consequences of inaccurate nowcasts. When nowcasts are used to predict hazards such as floods, unphysical solutions could lead to bad decisions.

In this project, we aim to combine machine-learning, theoretical fluid dynamics, operational prediction and meteorology, to create innovative approaches to nowcasting of tropical storms. We will develop ML methods which are fast, and which obey physical laws, like the weather prediction models. Our solutions will include statistical forecasts of rainfall probabilities, as well as ensembles of forecast realisations, and an automated evaluation system will be created. Recent advances in physical understanding and the new data offered by MTG, will be used to create statistical nowcasts of storm initiation and its subsequent evolution. We will apply these methods through our existing web-based and mobile-phone communication portals delivering information to Africa, and support colleagues in Africa to exploit the methods locally.


10 25 50