Region resampling did not show a clear mean gain from basin calibration of Reef Check bleaching risk in 2002–2019

We tested whether a local risk adjustment improves a coral bleaching model fitted in other ocean basins. We used 10,542 Reef Check surveys at 3,711 sites in the Atlantic, Indian and Pacific Oceans from 2002 to 2019. The outcome was at least 10% bleaching. The input was log(1 + CoRTAD heat accumulation at the survey date). Each ecological region stayed out of training in turn. The adjusted model kept the external heat slope and fitted an intercept adjustment on other regions in the target basin. Sites had equal weight within each basin; the three basins had equal weight in the mean score. Mean squared probability error fell from 0.08941 to 0.08261, a 7.6% relative reduction. The gain was 0.00680, with a 95% interval of -0.00688 to +0.01179 from 499 region bootstrap draws with full model refits. With predictions fixed, the interval was +0.00324 to +0.00975. Thus the mean gain was not clear after training uncertainty was included. Atlantic and Pacific gains had positive nominal intervals; the Indian gain was negative and imprecise. In the planned 2014–2017 subset, adjustment increased mean error by 4.0% in the point estimate, with an interval that included no change. The result applies to one survey programme and the tested historical model. It does not establish a common ecological response or future alert skill. The package preserves source data, provenance, all held-out predictions, code, scores, bootstrap draws and figures.

coral bleaching · model transfer · calibration · spatial validation · Reef Check · CoRTAD · training uncertainty