Flexible species distribution modelling methods perform well on spatially separated testing data

Aim To assess whether flexible species distribution models that perform well at nearby testing locations still strongly when evaluated on spatially separated data. Location Australian Wet Tropics (AWT), Ontario, Canada (CAN), north-east New South Wales, Australia (NSW), Zealand (NZ), five countries of America (SA), and Switzerland (SWI). Time period Most data were collected between 1950 2000. Major taxa studied Birds, mammals, plants reptiles. Methods We compared 10 modelling methods with varying flexibility in terms the allowed complexity their fitted functions [boosted regression trees (BRT), generalized additive model (GAM), multivariate adaptive splines (MARS), maximum entropy (MaxEnt), support vector machine (SVM), variants linear (GLM) random forest (RF), an Ensemble model]. used established practices for selection to avoid overfitting, including parameter tuning learning methods. Models trained presence–background 171 tested presence–absence Training using both spatial partitioning, latter based 75-km blocks. calculated average performance mean rank (focussing area under receiver operating characteristic precision-recall gain curves, correlation) assessed statistical significance differences them. Results The ranking did not change strongest predictive nonparametric known be flexible. An ensemble formed by averaging predictions pre-selected was best followed MaxEnt a variant forest. Main conclusions Whilst some modellers expect limited simple smooth predict better data, we found no evidence blocks 75 km. conclude are tuned enough overfitting effective predicting distinct areas.