Cross-Island Transferability of Multi-Horizon Probabilistic Coastal Sargassum Occurrence Forecasts: A Frozen-Model Evaluation from Puerto Rico to the Dominican Republic

Satellite observations reveal the offshore distribution of pelagic Sargassum, but predicting its occurrence at individual Caribbean coastlines remains challenging where local observational records are limited. This study tests whether forecasts developed on one Caribbean island can transfer to another without retraining or recalibration. Satellite-derived predictors from the 0–10, 10–50, and 50–100 km offshore bands were evaluated at one-, two-, and four-week horizons. Rolling-origin validation in Puerto Rico selected Random Forest models at one and two weeks and a one-dimensional convolutional neural network (CNN1D) ensemble with Platt calibration at four weeks. Frozen pipelines were applied unchanged at nine Dominican Republic sites and evaluated against 283 station-week targets. Puerto Rico 2026 provided independent testing. At one and two weeks, area under the receiver operating characteristic curve (ROC AUC) values were 0.745 and 0.757, area under the precision-recall curve (PR AUC) values were 0.689 and 0.687, and Brier Skill Score (BSS) values were +0.110 and +0.059. After transfer, ROC AUC values were 0.686 and 0.663, PR AUC values were 0.548 and 0.509, and BSS values were +0.098 and +0.056. Site removal did not eliminate short-horizon skill, and BSS remained positive by year. Transfer was clearest at one and two weeks, suggesting potential value for targeted coastal Sargassum monitoring, field verification, and response planning.