Bio-optical signatures of in situ photosymbionts predict bleaching severity prior to thermal stress in the Caribbean coral species Acropora palmata

The identification of bleaching tolerant traits among individual corals is a major focus for many restoration and conservation initiatives but often relies on large scale or high-throughput experimental manipulations which may not be accessible to many front-line restoration practitioners. Here, we evaluate a machine learning technique to generate a predictive model which estimates bleaching severity using non-destructive chlorophyll-a fluorescence photo-physiological metrics measured with a low-cost and open access bio-optical tool. First, a 4-week long thermal bleaching experiment was performed on 156 genotypes of Acropora palmata at a land-based restoration facility. Resulting bleaching responses (percent change in Fv/Fm or Absorbance) significantly differed across the four distinct light-response phenotypes (clusters) generated via a photo-physiology-based dendrogram, indicating strong concordance between fluorescence-based photo-physiological metrics and future bleaching severity. The proportion of thermally tolerant Clade D symbionts also differed significantly across photo-physiology-based dendrogram clusters, linking light-response phenotypes and bleaching response with underlying symbiont species. Next, these correlations were used to train and then test a Random Forest algorithm-based model using a bootstrap resampling technique. Correlation between predicted and actual bleaching responses in test corals was significant (p < 0.0001) and increased with the number of corals used in model training (Peak average R² values of 0.45 and 0.35 for Fv/Fm and absorbance, respectively). Strong concordance between photo-physiology-based phenotypes and future bleaching severity may provide a highly scalable means for assessing reef corals.