Keywords: Hierarchical Clustering, Bayesian Structural Time Series, Factor selection,Predictive modeling,MCMC,Red Sea Surface Temperature
Prediction of future spatio-temporal patterns in coral reef bleaching conditions with an estimate of uncertainty, is a key role in enhancing de- cision making for environment management over the Red Sea. Sea surface temperature (SST) variability is at first level the factor impacting coral reefs through increasing thermal bleaching events. Therefore, detecting and understanding the drivers of SST variability is critical for making meaningful projections of coral bleaching risk. In this work, we applied a practical statistical framework based on the Bayesian structural time series model and we showed its efficiency to predict and analyze bleaching risk alerts by (1) predicting SST for the next season with a very high accuracy and by (2) detecting the main predictive factors of its variabil-ity. SST predictions were obtained over three main regions of the Red Sea, clustered hierarchically based on long term variability dissimilarities. Thanks to this approach, we were able to isolate predictors of SST in an efficient and systematic way considering a large number of global climate indices. Efficiency in the prediction scheme is ensured by avoiding redundancy thanks to the combination of (1) prediction over the clustered regions and (2) the use of Bayesian paradigm and the MCMC algorithm with a structural time series model to perform variable selection at the same time as model training. The new insight gained from the application of this approach is that large-scale spatial patterns of ENSO and MJO are important predictive factors over the southern part of the Red sea only. On the other hand, AO and NAO are predictors of monthly SSTA over the Northern part. The connection between ENSO, AO and NAO climate indices and coral time series over the Red Sea is supported by northern Red Sea coral records during recent centuries.