A sequential multi-level framework to improve habitat suitability modelling

Abstract

Context Habitat suitability models (HSM) can improve our understanding of a species' ecology and are valuable tools for informing landscape-scale decisions. We can increase HSM predictive accuracy and derive more realistic conclusions by taking a multi-scale approach. However, this process is often statistically complex and computationally intensive. Objectives We provide an easily implemented, flexible framework for sequential multi-level, multi-scale HSM and compare it to two other commonly-applied approaches: single-level, multi-scale HSM and their post-hoc combinations. Methods Our framework implements scale optimisation and model tuning at each level in turn, from the highest (population range) to the lowest (e.g. foraging habitat) level, whilst incorporating output habitat suitability indices from a higher level as a predictor. We used MaxEnt and a species of conservation concern in Britain, the lesser horseshoe bat (Rhinolophus hipposideros), to demonstrate and compare multi-scale approaches. Results Integrating models across levels, either by applying our framework, or by multiplying singlelevel model predictions, improved predictive performance over single-level models. Moreover, differences in the importance and direction of the speciesenvironment associations highlight the potential for false inferences from single-level models or their posthoc combinations. The single-level summer range model incorrectly identified a positive influence of heathland cover, whereas sequential multi-level models made biological sense and underlined this species' requirement for extensive broadleaf woodland cover, hedgerows and access to buildings for roosting in rural areas. Conclusions We conclude that multi-level HSM appear superior to single-level, multi-scale approaches; models should be sequentially integrated across levels if information on species-environment relationships is of importance.

Predicting wildlife corridors for multiple species in an East African ungulate community

“…Because the choices made during each modelling step could have repercussions for the delineation of corridors, comparing the relative effects of different model choices is a crucial step to illustrate uncertainty associated with corridor model parameterization [ 17 , 45 ]. For example, to assess species-habitat associations, a variety of approaches and algorithms are available [ 55 , 56 ]. To address this modelling uncertainty, we harnessed recent advances in species distribution modelling and used ensemble and stacked species distribution models to quantify species-habitat associations [ 57 , 58 ].…”

Characterising the relationship between suitable habitat and gene flow for Myotis bechsteinii and Eptesicus serotinus in Britain

“…The hierarchical framework has been found to improve the accuracy of HSMs by integrating predictors at their scales of effect (Bellamy et al 2020). For landscape genetics, the framework also offers insight into the models that better represent gene flow, albeit we can only make assumptions on the reasons why certain models show higher correlations (e.g.…”

“…We used the hierarchical, multi-level HSM framework developed by Bellamy et al (2020) to model habitat suitability in Britain for M. bechsteinii, a woodland specialist that performs autumnal swarming, and E. serotinus, a more generalist species that is not known to occur at swarming sites. This method nests models across three levels (population range, home range, local habitat -See below) to limit predictor collinearity and enable context dependency.…”

Landscape‐level heterogeneity of agri‐environment measures improves habitat suitability for farmland birds

“…To select the best scale for each variable and exclude highly correlated variables from the same model, we fitted univariate linear models with binomial distribution for each explanatory variable and ranked them by their Akaike information criterion corrected for small sample size (AIC c ) score. For each species, we then selected the best set of uncorrelated variables (i.e., with Spearman's correlation coefficient <0.7) with the lowest AIC c score (Bellamy et al, 2020;McGarigal et al, 2016). All calculations were carried out in R version 4.0.2 (R Core Team, 2020), using the packages sf (Pebesma, 2018), raster, and terra (Hijmans et al, 2020(Hijmans et al, , 2021, and the R codes are available in Zenodo (Roilo, 2022).…”