Improving species distribution model predictive accuracy using species abundance: Application with boosted regression trees

Auxiliary information in the form of species abundance is frequently available as part of data collected for ecological investigations, yet when modeling distributions of species over large regions, species presence (and sometimes absence) are typically used. Incorporating abundances into species distribution models may greatly improve model predictive accuracy in practice. Boosted regression trees (BRT) models have been widely used in species distribution modeling, however no ecological study has been conducted to date that has assessed the predictive accuracy of BRT models that incorporates species abundance weights. We compared traditional, unweighted BRTs with species abundance-weighted BRTs for 55 fluvial fish species native to the Northeastern U.S. Overall model deviance explained and six diagnostic measures of predictive performance were compared between traditional BRTs and weighted BRTs. These comparisons indicated that unweighted BRTs performed better for fluvial fish species considered common, including those with greater numbers of presences and higher prevalence. Conversely, weighted BRTs were better suited for modeling distributions of species that had fewer presences, lower prevalence, and higher rarity, indicating the potential of species abundance-weighted distribution modeling to improve results for species of high conservation importance. Last, we offer insights into the applicability of using weighted approaches with other commonly used species distribution modeling methods.