An explainable machine learning prediction system for early warning of heat stress on Florida’s Coral Reef

Abstract Coral reefs are facing increasing threats from rising ocean temperatures, necessitating timely and localized prediction tools to inform reef management and conservation. This study introduces a machine learning framework capable of forecasting the onset of moderate coral heat stress at site-specific resolution on Florida’s Coral Reef. Leveraging the XGBoost algorithm, the data-driven prediction system forecasts whether heat stress will occur in a given season and, if so, the week in which moderate stress will begin. The prediction system achieves skillful forecasts up to six weeks in advance with a mean absolute error of approximately ±1 week. Two baselines are defined to compare performance– a multiple logistic regression model and a frequency-based model that predicts onset using the most common onset week, with the number of predicted onsets matched to the historical onset rate through random sampling. At the three reef sites analyzed, the machine learning model outperforms both baseline approaches in overall performance, including accurate stress onset timing. We systematically disentangle predictor importance across reef sites, lead times, and onset occurrence using SHAP—a novel multi-dimensional analysis that reveals how heat stress drivers vary by location and lead time. Surface air temperature consistently ranked as a top predictor, while other key variables varied by location and lead time, underscoring the importance of localized analysis for drivers of heat stress onset. This framework provides an explainable prediction tool on actionable timescales for anticipatory conservation with insight into stress onset at specific reefs, potentially allowing managers to develop reef-specific monitoring for emergency actions.