Folha highlights Pi Center research on rain forecasting with AI

The article “Precipitation nowcasting with geostationary satellite imagery and physics-constrained diffusion models”, developed by Antônio Catão, Melvin Poveda Quimiz and Leonardo Voltarelli, doctoral students at IMPA, under the guidance of researcher Paulo Orenstein, was featured in an article published by the Folha de S.Paulo last Sunday (31). The paper presents Tupann, an artificial intelligence model capable of predicting rainfall from satellite images with high spatial resolution and updated at ten-minute intervals.

In the article, Folha highlights the technology’s potential to expand meteorological monitoring in regions with limited radar coverage, as well as contributing to issuing alerts and responding to extreme weather events. The system produces forecasts in less than three minutes after receiving satellite images, offering an alternative for areas where weather observation infrastructure is limited.

“The Brazilian AI works on immediate rain forecasting – ‘nowcasting’. The task is complex because storms change direction and intensity in a matter of minutes. Orenstein compares the challenge to watching the first few seconds of a video and trying to guess the next frames,” said the report.

The research was carried out at the Pi Center, an IMPA initiative aimed at applying artificial intelligence to problems of public interest. The work originated in a project carried out in partnership with Rio de Janeiro City Hall to improve rainfall forecasts in the city. This collaboration gave rise to the chuvas.impa.br platform, used as a tool to support meteorological monitoring.

The Folha article also contextualizes the relevance of the research in light of the increase in extreme weather events around the world and the growing adoption of artificial intelligence techniques in meteorology.

In addition to the press coverage, the research was also recognized by the international scientific community. The article won the Best Student Paper award at the workshop on machine learning for remote sensing at the International Conference on Learning Representations (ICLR), one of the most important artificial intelligence conferences, held this year in Rio de Janeiro. Read the full article here.