Model bigger to avoid unrepresentative species distribution models from truncated spatial extents, with the modulating influence of prevalence

Abstract

As background, one approach for modeling species distributions is to model samples from distributions that are truncated to administrative or ecological boundaries, with predictor variables of climate. My objective was to examine the issue of truncation, which may not generate models that represent species ranges, even inside truncated extents, due to not incorporating the amplitude (i.e., range) of climate values where species occur. For methods, I modeled distributions of 14 wide-ranging tree species in North America and 40 combinations of the tree species truncated within 10 ecological units, using three combinations of extents: both truncated presence and background absence samples (i.e., absences generated from the background extent and hereafter also generally referred to as absences), truncated presence and untruncated absence samples, and untruncated presence and absence samples. Evaluations included model representativeness or similarity to species ranges measured through visualization and the intersection over union score, predicted probabilities in the extents under current and future climate, and climate amplitude values. I also examined how extents combined with equal and unequal modeling prevalence (ratio of presence and absence classes) interacted with ten accuracy metrics, including balanced accuracy (mean of sensitivity and specificity). Presence and absence truncation to ecological units resulted in unrepresentative species distributions (visually, and with a score of 0.1 out of 1), overpredicting presence outside of species ranges and underpredicting presence inside ecological units, due to minimal differentiation in climate variables between presence and absence samples. Mean balanced accuracies for models were 0.83. Presence truncation was extremely accurate, with mean accuracies of 0.99, but models predicted unsuitable climate outside of ecological units within species ranges (score of 0.2). For untruncated samples of 14 tree species, mean accuracies were 0.96 and represented species distributions (score of 0.6). Predicted areas of presence decreased with reduced modeling prevalence, which substantially improved all accuracy metrics for the truncated presence and absence samples but not representativeness of models. In conclusion, representative distribution models, with climate variables, required untruncated presence and absence classes to achieve climate differentiation between presence and absence classes and climate amplitude from the full range of climate values.

Citation

Hanberry, Brice B. 2026. Model bigger to avoid unrepresentative species distribution models from truncated spatial extents, with the modulating influence of prevalence. Environmental Challenges. 24: 101551.