New AI model can assist with early warning for coral bleaching risk

(Press-News.org) MIAMI — Scientists have created an AI model that forecasts moderate heat stress — a major precursor to coral bleaching — at sites along Florida’s Coral Reef up to six weeks ahead, with predictions generally accurate within one week.

The study presents a site-specific, explainable machine-learning framework to support coral scientists and restoration practitioners with local reef management and emergency response planning.

“This model gives coral scientists and resource managers advance notice of whether heat stress is likely to occur in a season — and, more importantly, the week it is most likely to begin,” said lead author Marybeth Arcodia of the University of Miami Rosenstiel School of Marine, Atmospheric, and Earth Science. Arcodia holds a dual appointment within the Department of Atmospheric Sciences and the Frost Institute for Data Science and Computing. “Through explainable AI, we can also identify the environmental factors driving those predictions at each reef site.”

“Our model identifies potential factors that influence thermal stress at a given reef site,” said Richard Karp, a co-author and post-doctoral research associate at the Rosenstiel School’s Cooperative Institute for Marine and Atmospheric Studies, “This information gives managers an opportunity to identify trigger points in emergency action plans, which can support planning and response decisions.”

The research team combined atmospheric science, coral ecology, and data science to build a prediction tool tailored to Florida’s Coral Reef.

Localized AI predictions on actionable timescales Using an XGBoost machine-learning model, the team forecast the onset of moderate coral heat stress at three reef sites using environmental data from 1985–2024. Inputs included accumulated and instantaneous heat-stress metrics, sea-surface temperature anomalies, air temperature, winds, solar radiation, and indicators of Loop Current and El Niño conditions, drawing on NOAA Coral Reef Watch and other public datasets.

“Our prediction system produced skillful forecasts up to six weeks in advance, and in most cases it was accurate to within about one week of when heat stress actually began,” said Karp. “It also outperformed two benchmark approaches — a multiple logistic regression model and a frequency-based method — both in predicting whether heat stress would occur and in identifying when it would start.”

The researchers also applied explainable AI techniques using SHAP — a method that shows which environmental factors most strongly influence each prediction — to understand how the drivers of heat stress differ by reef site and by forecast lead time.

Surface air temperature consistently ranked among the most important predictors, while other key environmental factors varied by site and lead time, underscoring the value of localized prediction.

“These insights are delivered on timescales when management actions are still possible,” Karp added. “They help prioritize monitoring, inform when to initiate emergency actions, and guide where resources are most effectively targeted.”

Actionable forecasts to support proactive reef conservation Florida and Caribbean reefs are experiencing increasingly frequent and severe heat-stress events and bleaching — including the record-breaking 2023 marine heatwave — heightening the need for site-level early-warning tools.

The authors emphasize that the new AI framework is intended to complement — not replace — existing operational systems such as NOAA Coral Reef Watch, by adding a localized, season-by-season timing signal for heat-stress onset.

The study, “An explainable machine learning prediction system for early warning of heat stress on Florida’s Coral Reef,”was published open access in Environmental Research Communications on December 16, 2025.

The study was completed over a two-year period from conceptualization to publication. Funding was provided by the Regional and Global Model Analysis program area of the U.S. Department of Energy’s Office of Biological and Environmental Research as part of PCMDI, and by NOAA grants #NA19OAR4590151 and #NA24OARX431C0022, and by NOAA Coral Reef Conservation Program grants #31476 and #31640.

Authors include Marybeth C. Arcodia from the University of Miami Rosenstiel School of Marine, Atmospheric, and Earth Science and the University of Miami Frost Institute for Data Science, Richard Karp, from the Rosenstiel School’s Cooperative Institute for Marine and Atmospheric Studies, and Elizabeth A. Barnes from the Department of Atmospheric Science, University of Colorado.

 

END