North America arbovirus vectors Aedes aegypti and A. albopictus have increased their northern range in recent years. Management and surveillance initiatives commonly apply species distribution models (SDMs) to advise resource allocation. However, many different SDM algorithms exist which provide different predictions, posing problems for SDM selection. Further, North America occurrence records of Aedes species are predominately limited to management region level (e.g., County, Health Region), without reliable absence data. The generalization to a region’s occurrence may hinder SDM use for preventative management. Applying the virtual ecologist approach, we designed an artificial Aedes species based on shared characteristics of A. aegypti and A. albopictus. We provide an extensive analysis of 11 popular SDMs ability to predict across North America trained with regional presence data and pseudo-absences. SDMs projections were evaluated against an independent presence-absence occurrence set at a resolution of 1 km.
Results/Conclusions
We demonstrate disagreement between traditional threshold-independent measures and predictions against known truth. The resulting predictions of high threshold accuracy SDMs had greater deviation from the true distribution compared to those with lower threshold accuracy. This problem highlights the importance of selecting appropriate measures of performance. Assessment of model performance with a known truth is vital to determine the agreeability of threshold and prediction accuracy to aid in SDM selection of generalized presence data of Aedes.