Looking for an optimal hierarchical approach for ecologically meaningful niche modelling

Ecological niche models are powerful tools in ecology. Factors operating at different spatial scales are known to jointly influence species distributions, but their integration in meaningful and reliable niche models is still methodologically complex and requires further research and validation. Here, we compare six different hierarchical niche models (HNMs): two ensemble, two Bayesian, and two penalized logistic regression approaches. The six HNMs were applied to produce high-resolution (25 m) predictions for five tree species in a Biosphere Reserve in Central Spain, combining information from two spatial scales. At the regional scale (mainland Spain) climatic variables were used as predictors, and presence/absence data were derived from the Spanish Forest Inventory (76,347 plots). At the landscape scale (Biosphere Reserve) environmental variables were used as predictors while presence/absence data were derived from a local vegetation sampling (346 plots). We compared and evaluated the six HNMs using the AUC, MaxKappa, and MaxTSS statistics and the Pearson’s correlation coefficient. We obtained reliable high-resolution HNMs at the landscape scale (AUC values were greater than 0.8), although with variable performance and scope of application. Ensemble approaches delivered reliable models particularly when the sample size was not a limiting modelling factor. However, Bayesian modelling allowed considering a spatially correlated random effect that outperformed the results of all other approaches for species with a low sample size, possibly derived from a strong spatial structure in their distribution. HNMs succeeded in generating high-resolution predictions and manage to identify a significantly greater part of the climatic niche of the species than non-hierarchical models. This allows more accurate projections both in space and time, which is essential in climate change and invasive species modelling projections. The usefulness of these models for decision support in local conservation programs is therefore highlighted. Our methodological comparison is valuable to inform modellers and decision makers of the performance and implications of these approaches regarding the support they can provide for the implementation of conservation management measures at the landscape scale.