Predictive performance of presence?only species distribution models: a benchmark study with reproducible code

Species distribution modeling (SDM) is widely used in ecology and conservation. Currently, the most available data for SDM are species presence-only records (available through digital databases). There have been many studies comparing performance of alternative algorithms data. Among these, a 2006 paper from Elith colleagues has particularly influential field, partly because they several novel methods (at time) on global set that included independent presence–absence model evaluation. Since its publication, some further developed new ones emerged. In this paper, we explore patterns predictive across methods, by reanalyzing same (225 six different regions) using updated knowledge practices. We apply well-established such as generalized additive models MaxEnt, alongside others received attention more recently, including regularized regressions, point-process weighted random forests, XGBoost, support vector machines, ensemble framework biomod. All use include background samples (a sample environments landscape) fitting. impacts weights presence points introduce ways evaluating fitted to these data, area under precision-recall gain curve, focusing rank results. find way matters. The top method was an tuned individual models. contrast, ensembles built biomod with default parameters performed no better than single moderate performing Similarly, second forest parameterized deal (contrasted relatively few records), which substantially outperformed other implementations. that, general, nonparametric techniques capability controlling complexity traditional regression MaxEnt boosted trees still among code working examples provided make study fully reproducible.