Spatially adaptive background delimitation strategies for species distribution models with unevenly surveyed occurrences

Resumen

The quality of species distribution models (SDM) and ecological niche models (ENM) depends critically on how the background or pseudo‐absence region is defined. Traditional background delimitation strategies include convex, concave or alpha hulls, buffers, biogeographic regions, and/or administrative units. However, in the frequent cases when species occurrence data are unevenly sampled – due to spatial variations in accessibility, survey effort or other biases –, traditional backgrounds may either include large unsurveyed areas, or exclude ecologically relevant (though insufficiently surveyed) portions of the species’ distribution range. Such mis‐specification can reduce model predictive accuracy, increase extrapolation error, and worsen biodiversity knowledge gaps. This paper presents a function in the fuzzySim R package that implements a spatially adaptive approach to delineating SDM backgrounds, in a way that reflects the spatial distribution of the available occurrence records, which is largely shaped by survey effort. The method uses data‐driven, differently-sized buffers to delimit regions that reflect local differences in occurrence data availability, with the aim of being neither overly conservative nor overly permissive. The approach is illustrated using a real-world species occurrence dataset with spatially varying degrees of survey bias. Spatially adaptive buffers yield more realistic, ecologically interpretable and justifiable background regions for modelling, which better address known biodiversity data constraints. They avoid placing pseudo-absence or unoccupied background points in unsurveyed regions, retaining isolated occurrence points without inflating the modelling area. This approach can thus help reduce the SDM uncertainty that arises from unevenly sampled biodiversity data, contributing toward the identification, visualization, and mitigation of the effects of knowledge gaps.