Predicting species richness and abundance of tropical post-larval fish using machine learning

No previous studies predicted post-larval fish species richness and abundance combining molecular tools, machine learning, and past-days Remotely Sensed Oceanic Conditions (RSOCs) at different scales. Previous studies aimed at modeling species richness and abundance of marine fishes have mostly used environmental variables recorded locally during sampling. They have merely focused on juvenile and adult fishes due to the difficulty of obtaining accurate species richness estimates for post-larvae. The present work predicted post-larval species richness (identified using DNA barcoding) and abundance at two coastal sites in SW Madagascar using random forests (RF). RF models were fitted using combinations of local variables with RSOCs at a small-scale (eight days preceding fish sampling in a 50x120 km2 area), mesoscale (sixteen past-days in 100x200 km2), and large-scale (twenty-four past-days in 200x300 km2). RF models combining local and small-scale RSOC variables predicted more accurately the species richness and abundance with around 70% and 60% accuracy, respectively. We observed a small variation of RF model performance in predicting species richness and abundance among all sites, highlighting the predictive RF model consistency. Moreover, partial dependence plots showed that high species richness and abundance were predicted for sea surface temperatures <27.0°C and chlorophyll a concentrations <0.22 mg m-3. Referring to temporal changes of these variables, these thresholds were solely observed from November to December. These results suggest that, in SW Madagascar, species richness and abundance of post-larval fish may only be predicted prior to the ecological impacts of tropical storms on larval settlement success.