The construction of causal networks to estimate coral bleaching intensity

Current metrics for predicting bleaching episodes, e.g. NOAA's Coral Reef Watch Program, do not seem to apply well to Brazil's marginal reefs located in Bahia state and alternative predictive approaches must be sought for effective long term management. Bleaching occurrences at Abrolhos have been observed since the 1990s but with a much lower frequency/extent than for other reef systems worldwide. We constructed a Bayesian Belief Network (BN) to back-predict the intensity of bleaching events and learn how local and regional scale forcing factors interact to enhance or alleviate coral bleaching specific to Abrolhos. Bleaching intensity data were collected for several reef sites across Bahia state coast (∼12°–20°S; 37°–40°W) during the austral summer 1994–2005 and compared to environmental data: sea surface temperature (SST), diffuse light attenuation coefficient at 490 nm (K490), rain precipitation, wind velocities, and El Niño Southern Oscillation (ENSO) proxies. Conditional independence tests were calculated to produce four specialized BNs, each with specific factors that likely regulate bleaching intensity. All specialized BNs identified that a five-day accumulated SST proxy (SSTAc5d) was the exclusive parent node for coral bleaching producing a total predictive rate of 88% based on SSTAc5d state. When SSTAc5d was simulated as unknown, the Thermal-Eolic Resultant BN kept the total predictive rate of 88%. Our approach has produced initial means to predict beaching intensity at Abrolhos. However, the robustness of the model required for management purposes must be further (and regularly) operationally tested with new in situ and remote sensing data.

► Analysis of environmental conditions modulating coral bleaching intensity using satellite and in situ data.
► Bayesian network is used to learn how multi-scale forcing factors interact to enhance bleaching.
► Persistence of high sea surface temperature exerts the strongest influence.
► Models achieved a total predictive rate of 88%.