AI for Rapid Intensification of Hurricanes: research possible due to NOAA Cooperative Institutes
By Theresa Barosh | April 2025
About 80% of major hurricanes experience rapid intensification, an especially concerning turn of events when occurring shortly before landfall because it strains emergency managers undertaking preparedness activities. Major hurricanes start with at least 111 miles per hour winds, and rapid intensification occurs when maximum wind speeds increase by at least 35 miles per hour within 24 hours.
CIRA researchers have developed AI to forecast rapid intensification of hurricanes. The research could support hurricane forecasts in the future to better protect homes and lives in the Southeastern United States. An agreement between NOAA and Colorado State University, CIRA provides researchers with unique access to weather satellite resources and technology.
“One of the nice things about working for these Cooperative Institutes is that you can really explore – you can think big, and you can do a lot of experiments,” said CIRA researcher Ryan Lagerquist. “This is just freedom that you typically don’t have in the private sector because you have to meet very short-term goals.”
Based out of NOAA’s Global Systems Laboratory in Boulder, Colorado, Lagerquist led the research, leaning on his thirteen years of experience with machine learning and weather. Being a university researcher working out of a NOAA office supports Lagerquist in building research collaborations. His team of NOAA and CSU researchers developed successful convolutional neural networks, a form of deep learning algorithm, for hurricane rapid intensification prediction. Lagerquist’s AI model outperforms existing operational rapid intensification models in multiple ways – using only a single satellite image and a few parameters representing the surrounding environment as input.
Why AI?
There are many factors that play a role in whether a hurricane intensifies. The problem of putting all these data sources together and making a good prediction is challenging. This is where machine learning excels because it is good at harnessing lots of data to make predictions.
Lagerquist’s approach for predicting rapid intensification of hurricanes is to use satellite data, CIRA’s specialty. He starts with the current state of the hurricane and uses properties of the near storm environment, like sea surface temperature and ocean heat content. Lagerquist feeds all those values into a machine learning algorithm that he trained on historical data of the past several decades of hurricanes. His AI uses the environmental data and current hurricane state to predict which hurricanes are going to rapidly intensify.
Predictions are fast, made in seconds. In contrast, using a traditional numerical weather prediction – which simulates the entire atmosphere, including the hurricane itself – is very computationally expensive. Numerical weather predictions take hours to run on a big supercomputer.
Lagerquist’s AI model provides skillful rapid intensification predictions that are not strongly correlated with those of existing operational hurricane models. Lagerquist said this result suggests his AI would be a useful addition to an operational ensemble because it brings something unique. Forecasters use an ensemble, or collection, of resources to make decisions about predictions and warnings that go to the public. Each piece of the ensemble provides a unique perspective or set of information for forecasters to consider. For example, a forecaster may reference both an AI model and a numerical weather prediction using nuanced knowledge about their location to determine which information is most relevant.
What is special about a NOAA Cooperative Institute?
With previous experience at Google and the National Center for Atmospheric Research, Lagerquist reflected on the benefits of working at a NOAA Cooperative Institute versus working in the private sector.
“You have freedom to spend more time figuring out what works and what doesn’t,” said Lagerquist. “And that’s often where we get the most bang for our buck out of science.”
What is needed: Strong support for NOAA Cooperative Institutes so they can continue to facilitate the large and meaningful advances in science that cannot be replicated in the private sector.
Case study for Hurricane Laura. The forecast time is 0200 Eastern on Aug 25 2020; all models are forecasting whether rapid intensification (RI) will occur in the next 24 hours. Panels a-c show input data used by our neural networks. Panel a shows the satellite image, with the green arrow marking the direction of the hurricane’s motion. Panel b shows variables summarizing the satellite image, which are part of the widely used Statistical Hurricane Intensity Prediction Scheme (SHIPS) developmental dataset. Panel c shows variables describing the environment surrounding the hurricane, also part of the SHIPS dataset. The variables in panels b-c are shown in normalized units, where blue means lower than usual and red means higher than usual. Panel d shows RI forecasts from several models: three pre-existing models and our three neural networks. The correct answer is “yes”: Laura did undergo RI between 0200 ET Aug 25 and 0200 ET Aug 26. Each pre-existing model produces one RI probability, shown with a green dot: the Deterministic-to-probabilistic Statistical Model, the SHIPS Rapid Intensification Index (RII), and the SHIPS consensus. Each neural network provides an ensemble of RI probabilities (many different answers), shown with an orange violin plot; the average probability is shown with a black line. One neural network uses only the SHIPS variables in panels b-c; one neural network uses only the satellite image in panel a; and one neural network uses all data. Review the full paper.