Résumé
Identifying indicator species is critical for understanding ecosystem health and dynamics. In this study, we evaluate the use of an interpretable machine learning approach, named ``Predomics,'' to identify key coral reef fish indicator species surveyed in different Lagoon habitats by baited remote underwater video stations (BRUVS). We compared its performance with the Indicator Species method (IndVal), traditionally used in ecological science. Despite the small number of features (here fish species), Predomics models maintain comparable generalization performance to those built with IndVal identified species and, in some cases, outperform them. This shows that a smaller set of indicator species can effectively capture ecological patterns. In particular, the predictive importance of the indicator species derived from Predomics shows significant positive correlations with the indicator value of IndVal, indicating its biological pertinence. We evaluated the performance of these methods on both quantitative abundance (MaxN) and presence/absence profiles of fish species. Moreover, we examined how feature prevalence influences the robustness of selected species in generalization to unseen samples during model training. We observed that removing fish species with low prevalence enhanced the generalization robustness of Predomics models, particularly the bininter model, which outperformed others in presence/absence scenarios. By identifying a streamlined set of indicator species with high generalization performance, Predomics offers a biologically relevant and efficient alternative to standard ecological methods like IndVal. This simplifies ecological assessments and improves the accuracy of biodiversity monitoring in marine environments.