Random-split validation overstates the skill of machine-learning models of global coral bleaching severity

Arav Siddharth Mestry 

Global compilations of coral bleaching surveys are increasingly used to train machine-learning models of bleaching severity, but such models are often evaluated with random train–test splits that ignore duplicated records, spatial dependence and differences among contributing survey programmes. We re-examined an earlier analysis of the Global Coral Bleaching Database (BCO-DMO dataset 773466) that reported a random-split random forest R² of 0.688. The result was reproducible, but it was obtained on 32,716 records, 34.5% of which duplicated another record exactly because Reef Check surveys are stored once per substrate category. After collapsing records to 21,527 sampling events, we compared random, site-grouped, spatial-block (1–10°), leave-one-realm-out and forward temporal validation for models using coordinates, thermal-stress metrics, other environmental variables, or indicators of the contributing data source. Random-forest R² fell from 0.60 (random) to 0.53 (new sites), 0.18 (1° blocks) and 0.02 (10° blocks), and was at most 0.06 in held-out realms. Coordinates alone predicted worse than the mean under all block designs. Indicators of the contributing source alone explained 0.35–0.37 of the variance under every cross-validation design, more than any environmental model under spatial validation. Within the single-protocol Reef Check subset, no predictor set achieved positive R² under spatial blocks. The all-predictor model retained moderate ability to rank events with at least 50% bleaching (AUC 0.73 under 5° blocks). These results indicate that the apparent skill of pooled bleaching-severity models largely reflects record duplication, interpolation between nearby surveys, and heterogeneity in how bleaching was measured, rather than transferable environment–bleaching relationships.