Weekly temperature variation gave an uncertain log loss gain in 4,522 Reef Check surveys from 2013–2019 under fixed logistic models

We tested whether weekly temperature variation improves later-year coral bleaching risk estimates in the Reef Check part of GCBD version 2. Two fixed logistic models used 6,706 training surveys from 1998–2012 and 4,522 test surveys from 2013–2019. The outcome was at least 10% bleaching. The base model used heat stress, depth, latitude, season and geography. Adding full-record weekly temperature anomaly standard deviation reduced test log loss from 0.283530 to 0.282724 nats per survey. The gain was 0.000805 nats, or 0.28%. Its 95% spatial cell bootstrap interval was −0.002448 to 0.004437. Its crossed space and year interval was −0.004039 to 0.006092. Both models had higher loss than a fixed training-rate risk. All 13 fixed checks retained crossed intervals that include zero. The result leaves added value uncertain for these models and surveys. The full-record temperature measure includes later years, so this is a retrospective test rather than a real-time forecast. The package preserves source data, provenance, code, all event predictions, fitted models, resampling draws, figures and reproduction checks.

coral bleaching · temperature variability · Reef Check · temporal validation · ecological prediction · log loss · open data