Comparing multi- and single-scale species distribution and abundance models built with the boosted regression tree algorithm

Abstract

Species are influenced by factors operating at multiple scales, but multi-scale species distribution and abundance models are rarely used. Though multi-scale species distribution models outperform single-scale models, when compared through model selection, multi- and single-scale models built with computer learning algorithms have not been compared. We compared the performance of models using a simple and accessible, multi-scale, machine learning, species distribution and abundance modeling framework to pseudo-optimized and unoptimized single-scale models. We characterized environmental variables at four spatial scales and used boosted regression trees to build multi-scale and single-scale distribution and abundance models for 28 bird species. For each species and across species, we compared the performance of multi-scale models to pseudo-optimized and lowest-performing unoptimized single-scale models. Multi-scale distribution models consistently performed as well or better than pseudo-optimized single-scale models and significantly better than unoptimized single-scale models. Abundance model performance showed a similar, but less pronounced pattern. Mixed-effects models, that controlled for species, provided strong evidence that multi-scale models performed better than unoptimized single-scale models. Although mean improvement in model performance across species appeared minor, for individual species, arbitrary selection of scale could result in discrepancies of up to fourteen percent for area of suitable habitat and population estimates. Scale selection should be explicitly addressed in distribution and abundance modeling. The multi-scale species distribution and abundance modeling framework presented here provides a concise and accessible alternative to standard pseudo-scale optimization while addressing the scale-dependent response of species to their environment.