Optimizing ensembles of small models for predicting the distribution of species with few occurrences

Abstract

Ensembles of Small Models (ESM) represent a novel strategy for species distribution modelling with few observations. ESMs are built by calibrating many small models and then averaging them into an ensemble model where the small models are weighted by their cross‐validated scores of predictive performance. In a previous paper (Breiner, Guisan, Bergamini, & Nobis, Methods in Ecology and Evolution, 6, 1210–1218, 2015), we reported two major findings. First, ESMs proved largely superior to standard models in terms of model performance and transferability. Second, ESMs including different modelling techniques did not clearly improve model performance compared to single‐technique ESMs. However, ESMs often require a large computation effort, which can become problematic when modelling large numbers of species. Given the appealing new perspectives offered by ESMs, it is especially important to investigate if some techniques yield increased performance while saving computation time and thus could be predominantly used for building ESMs. Here, we present results from a reanalysis of a subset of the data used in Breiner et al. (2015). More specifically, we ran ESMs: (1) fitted with 10 modelling techniques separately (in Breiner et al., 2015 we used only three techniques); and (2) using various parameter options for each modelling technique (i.e., model tuning). We show that ESMs vary in model performance and computation time across techniques, and some techniques are advantageous in terms of optimizing model performance and computation time (i.e., GLM, CTA and ANN). Including one of these modelling techniques could thus optimize computation time compared to using more computing‐intensive techniques like GBM. Next, we show that parameter tuning can improve performance and transferability of ESMs, but often at the cost of computation time. Parameter tuning could therefore be used when computing resources are not a limiting factor. These findings help improve the applicability and performance of ESMs when applied to large numbers of species.

Data quantity is more important than its spatial bias for predictive species distribution modelling

“…In real applications, it is unlikely that the functional form of the model will exactly match the form of the true species responses. Indeed, the species distribution modelling literature has many examples of different modelling methods performing best in different studies, suggesting that no modelling method consistently outperforms others (Bahn & McGill, 2007;Breiner, Nobis, Bergamini, & Guisan, 2018;Cutler et al, 2007;Elith et al, 2006;Elith & Graham 2009).…”

Unraveling the habitat preferences of two closely related bumble bee species in Eastern Europe

“…Araújo & New, 2007; Elith & Graham, 2009; Marmion, Luoto, Heikkinen, & Thuiller, 2009). Because the number of known locations of species presence was limited, we employed the ensemble of small models approach implemented in the “ecospat” R package (Breiner, Guisan, Bergamini, Nobis, & Anderson, 2015; Breiner, Nobis, Bergamini, Guisan, & Isaac, 2018; Di Cola et al, 2017). Ecospat fits bivariate models of presence/(pseudo‐)absence with two predictor variables at a time, creating an ensemble of “small” models weighted by each bivariate model's performance.…”

“…We used ecospat v.3.1 and Biomod2 v.3.3‐19 to run Maxent models (specifically the MAXENT.Phillips models, as implemented by Phillips, Anderson, & Schapire, 2006), generalized linear models (GLM), classification tree analysis (CTA, also known as classification and regression trees (CART); Breiman, Friedman, Olshen, & Stone, 1984), and artificial neural networks (ANN; Ripley, 1996). In a recent study comparing 10 different modeling approaches implemented in ecospat and Biomod2, these were shown to be the top performing ones, while keeping computation times manageable (Breiner et al, 2018). We used model tuning to optimize the parameter settings for each model.…”

Modelling of rare flood meadow species distribution by a combined habitat surface water–groundwater model

“…However, in the past, it was considered a challenge to model the distribution of rare and endangered species (Elith et al, 2006;Guisan et al, 2006); rare species datasets are mostly characterized by low occurrences, resulting in potentially overfitted models when multiple predictors are included (necessary for describing the species' niches, in particular, for rare specialist species with complex habitat requirements; Lomba et al, 2010). A promising step forward to overcome this obstacle was made by considering model ensembles of small models (ESM) to improve the reliability of habitat models (Breiner, Guisan, Bergamini, & Nobis, 2015;Breiner, Nobis, Bergamini, & Guisan, 2018). Few studies have tested the novel ESM approach (Breiner et al, 2018;Di Febbraro et al, 2017), but hitherto, ESM studies on rare flood meadow species cannot be found in literature.…”

“…A promising step forward to overcome this obstacle was made by considering model ensembles of small models (ESM) to improve the reliability of habitat models (Breiner, Guisan, Bergamini, & Nobis, 2015;Breiner, Nobis, Bergamini, & Guisan, 2018). Few studies have tested the novel ESM approach (Breiner et al, 2018;Di Febbraro et al, 2017), but hitherto, ESM studies on rare flood meadow species cannot be found in literature.…”