BIRDS AND BIAS: Evaluating the effects of Spatial Autocorrelation in Bird Distribution Models

By

Abstract
Globally, bird populations are declining dramatically due to various reasons. Species distribution models have proven to be very useful to monitor bird populations. However, it is known that species distribution models are affected by spatial autocorrelation, which, if not accounted for, may lead to biased model results, resulting in an increased likelihood of Type-1 errors and incorrect variable selection.

Sovon (the Dutch organisation for field Ornithology) frequently uses species distribution models to model the distribution of birds throughout the Netherlands. By assessing several modelling factors, 120 binomial model scenarios were created and assessed for two bird species (i.e. Common Blackbird Turdus Merula & Golden Oriole Oriolus oriolus) based on Atlas data from Sovon.

Literature has shown that spatial autocorrelation is common in ecology; therefore, SDM models may be vulnerable to the effects of spatial autocorrelation. Furthermore, spatial models often have an improved fit over non-spatial models. Solutions to account for spatial autocorrelation in SDM models are spatial filtering, systematic sampling, using independent data, adjusting spatial resolution, and including distance and spatial autocorrelation in SDM models. Given that different bird species can differ greatly in distribution and habitat size, considering an appropriate scale is found to be important when modelling bird distribution and is species dependent. In addition, the ecological relevance of a species distribution model seemed to be one of the most important factors to successfully explain the distribution of a species, which requires customisation for each species independently.

Results from the modelling indicated that the different SDM models (GLM, RF, BRT & MaxEnt) performed comparably well. Moreover, this study confirms that specialist species are easier to predict than generalist species. Sampling strategy and class balancing both did not seem to have a big effect on the amount of spatial autocorrelation in the model residuals. Spatial resolution had a noticeable effect on both the amount of spatial autocorrelation and model performance. The spatial congruence analysis confirmed the finding that an appropriate scale is bird dependent, which was also found in the literature.

Based on the outcomes of this study, future implications require continuing model assessments and comparison studies to further strengthen our understanding of spatial autocorrelation and how it affects species distribution models, thereby enhancing the conservation of bird populations and bird habitats within the Netherlands and beyond.

Keywords: Spatial autocorrelation; Niche; Species distribution; Spatial bias