Performance of fish-habitat classifiers based on derived predictors from a coupled biophysical model

Abstract Fish habitat models based on remotely-sensed data may be limited by satellite coverage and availability. We compared the fit and predictive power of Random Forest habitat classifiers that were developed using predictors derived from a coupled biophysical model (i.e., modeled predictors) versus similar classifiers that used remotely-sensed satellite data for two data sets (eggs and adults) and four species that occur widely in the California Current system. When tested on independent data, classifiers of spawning habitat that used derived predictors (derived classifiers) had nearly identical accuracies (0–2% difference) to similar classifiers based on satellite data (satellite classifiers). Accuracies of derived classifiers of adult habitat were within −8% to +7% of comparable satellite classifiers. Accuracies of both types of classifiers on test data were much greater for Northern anchovy Engraulis mordax and Pacific hake Merluccius productus (0.75–0.97) than for jack mackerel Trachurus symmetricus and Pacific sardine Sardinops sagax (0.61–0.72), and generally were greater for classifiers of spawning habitat than for adult habitat. Specificity was very good for both types of classifiers, but sensitivity was poor, because classifiers identified potential habitat which was not fully occupied. Adults of all species used a broader range of habitat conditions during summer than during the spring spawning period. Derived classifiers have some advantages over satellite classifiers; they are not limited by cloud cover and they can make predictions in near real-time or the short-term future. However, there was no consistent improvement in the accuracy of derived predictors that included modeled zooplankton concentrations over comparable satellite classifiers that included reflectance/chlorophyll concentration.