Spatially explicit information describing reef fish communities is critical for effective management of coral reef ecosystems. However, geographic coverage of in situ data is often limited. To overcome this information gap, statistical modeling can be used to make predictions across space by relating sample data to predictor variables describing the associated environment. As part of a marine biogeographic assessment to inform the Bureau of Ocean Energy Management’s renewable energy policy decisions in Hawaii, spatial predictions of several reef fish community metrics were generated from visual survey data compiled by University of Hawaii’s Fisheries Ecology Research Lab and environmental predictors representing seafloor topography, benthic habitats, geography, and oceanography. Boosted regression trees, an ensemble approach combining machine learning with tree-based statistical models, were fit to the data in R using the ‘dismo’ package. Model parameters were tuned by fitting models for a range of learning rate, tree complexity, and bag fraction values, and identifying for each combination of values the number of boosting trees that minimized predictive deviance. Predictors that contributed least to model performance were eliminated using a recursive model simplification procedure. Non-parametric bootstrapping, in which the survey data were randomly resampled and a model was fit on each sample, was then used to create an ensemble of spatial predictions across the study area. The coefficient of variation was calculated to visualize the spatial precision of model predictions. Model performance was evaluated by calculating the percent deviance explained by the model when evaluated on data withheld from model fitting.
