Format: ORAL

Claudio A. Bracho-Estvanez1, Salvador Arenas-Castro2, Juan P. Gonzlez-Varo1, Pablo Gonzlez-Moreno3

1Department of Biology, IVAGRO, Universidad de Cádiz, Campus Río San Pedro, Puerto Real, Cádiz, Spain. 2Area of Ecology, Department of Botany, Ecology and Plant Physiology, Faculty of Sciences, Universidad de Córdoba, Córdoba, Spain. 3Department of Forest Engineering, ERSAF RNM-360, Universidad de Córdoba, Córdoba, Spain.

Species distribution models (SDMs) are a vast collection of techniques that predict species likelihood of occurrence across spatial and temporal dimensions. These techniques barely require two inputs: presences (referred as only-presences when real absences of species are unavailable) and a set of relevant environmental variables for the niche of target species. Thus, the recent expansion of open platforms offering massive georeferenced data on biodiversity has boosted the use of SDMs. Still, the employment of such data holds a major limitation: the profusion of sampling biases. The former implies that opportunistic presences are potentially affected by heterogeneous sampling efforts across study areas and species, which demands the implementation of adequate bias corrections. As well, modelers usually take advantage of classical metrics (namely AUC and TSS) to evaluate the overall quality of model predictions. However, AUC and TSS have been widely criticized given its poor capacity to assess the biological reality of predictions, and it would be alarming if such metrics are eventually unsensitive to sampling biases. In this study, we selected 31 fleshy-fruited plants as a study case to evaluate the sensitivity of classical metrics to sampling biases, as well as to compare the effect of three distinct bias corrections on the quality of predicted distributions. We run models on a (sub-)continental scale and used both classical and alternative, spatially explicit metrics of model performance. Our results suggest that bias corrections are imperative to improve predictions, whilst model performance depends on settings as the modelling algorithm or the method to generate pseudo-absences. Likewise, classical metrics were particularly weak to discern biased predictions, and eventually contradicted alternative metrics. Altogether reveals the convenience of implementing contrasting, spatially explicit validations when assessing predictions affected by sampling biases.