28/04/2025

IMPA presents AI for global-scale rainfall forecasting.

After developing artificial intelligence models to improve rainfall forecasting in the city of Rio de Janeiro, IMPA presented a similar project with potential for global reach this Monday (28), at the Climate Informatics 2025 conference, at the Getúlio Vargas Foundation (FGV), in Rio.

Led by the Pi Center (IMPA's Center for Projects and Innovation), the new model will be presented in the lecture "Precipitation nowcasting of satellite data using physically conditioned neural networks". The work uses exclusively satellite data to estimate rainfall volume over short periods of time. Its advantage is benefiting regions lacking coverage from weather radars.

“This new version, focused solely on satellite data, significantly expands the reach of the technology. For example, it can be applied in regions of the Global South where there is no radar infrastructure, in parts of Africa, Oceania, and remote areas of Brazil,” explains Leonardo Voltarelli, a doctoral student at IMPA who is part of the research conducted by the Pi Center.

The project is a continuation of the one carried out in partnership with the Rio de Janeiro City Hall – which last year received a set of machine learning-based models trained with local data. The tool allows for predicting the evolution of radar and satellite images, offering a valuable tool for managing urban climate risks. In this case, the models allow for predictions every 10 minutes for areas of 5km x 5km, up to three hours in advance.

IMPA's participation in Climate Informatics highlights the leading role of Brazilian science in strategic themes such as climate, sustainability, and technological innovation. "Being present at an event of this magnitude, with a project conceived in Rio de Janeiro and with the potential for global impact, demonstrates the importance and potential of integrating cutting-edge science with the demands of society," concluded Voltarelli.

IMPA researcher Paulo Orenstein also participated in the event. In the panel on Regional Collaborations, he addressed work in the area of very short-term forecasting – 30 minutes to 3 hours, and medium-term forecasting – 2 to 6 weeks, both focusing on regions with limited data that could benefit from better forecasts in areas such as agriculture, energy, water resource allocation, and preparedness for natural disasters.