Résumé
Coral reef monitoring often relies on indicator species to reflect ecological conditions across habitats, yet existing identification methods such as IndVal and TWINSPAN vary widely in the number and identity of the taxa they select. Here, we compare these two traditional approaches with an interpretable machine learning method, Predomics, to identify parsimonious sets of indicator species using Baited Remote Underwater Video Stations (BRUVS) in a tropical lagoon of New Caledonia. TWINSPAN and Predomics consistently identified far fewer indicator species than IndVal, yet all methods achieved equivalent predictive accuracy in habitat classification. Notably, all approaches converged on the same core taxa for distinguishing inshore from offshore habitats, lending strong cross-method support to the ecological relevance of these species. Results were robust across both abundance and presence/absence data, and cross-validation confirmed the generalizability of the selected indicators to unseen samples. The identified indicator species, primarily wrasses, goatfishes, and threadfin breams, align with established habitat preferences. These results show that parsimonious, interpretable machine learning methods can match or complement classical approaches while delivering simpler, more actionable indicator sets for efficient and scalable reef health assessments and conservation planning.