Exploring hybrid consensus models to assess roadkill

Abstract

Accurate information provided by reliable models is essential for identifying hotspots and mitigating roadkill. However, existing methods, such as kernel density estimation (KDE) and maximum entropy modeling (ME) may individually identify only a subset of the suitable locations for mitigation, because KDE cannot detect hotspots once local abundances are depressed, and ME may only partially identify current hotspots due to imperfect discrimination skill. Here, we propose a hybrid consensus modeling (HCM) approach that leverages the strengths of both KDE and ME by using their consensus to identify the core subset of hotspots. We collected herpetofauna (amphibians and reptiles) roadkill data (N = 839) along four roads in Taiwan (R.O.C.) to evaluate the statistical performance and theoretical mitigation efficiency of HCM, KDE and ME, and to compare the allocation among roads, spatial clustering, and environmental conditions in the identified hotspots. HCM was applied on the herpetofauna dataset as well as separately on amphibians and reptiles. Although the discrimination skill of KDE and ME models for both target clades together was good to excellent (AUC_KDE = 0.944, AUC_ME = 0.822), the highest theoretical mitigation efficiency, was displayed by HCM Consensus (2.89), followed by KDE (2.58), and ME (1.91). Furthermore, we show that theoretical mitigation efficiency increases with decreasing spatial clustering (Moran's I). Given pervasive budget constraints, we recommend to limit permanent mitigation measures such as fenced culverts to HCM Consensus hotspots, temporary measures to KDE hotspots, and to target additional monitoring at ME hotspots.

Keywords: Consensus; Hotspot; Kernel density; Maximum entropy; Mitigation; Roadkill.