Post provided by Jane Elith, Mike Kearney and Steven Phillips  

10th anniversary logo

To celebrate the 10th Anniversary of the launch of Methods in Ecology and Evolution, we are highlighting an article from each volume to feature in the For Volume 1, we have selected ‘The art of modelling range-shifting species’ by Elith et al. (2010).  In this post, first author, Professor Jane Elith, discusses the background and key concepts of the article, and how things have changed since the paper was published.

Illustration of the idea that model settings affect prediction.

We started work on this manuscript around 2008, prompted by increasing use of species distribution models for climate change and invasive species problems. At that stage there was growing recognition of the problems in these applications (e.g. see a recent MEE review on transferability) but relatively few tools for dealing with them. In our view, if correlative models are to be used for such purposes, the data and models require special attention.

Mike and I (Jane) first started working on the problem. Mike had been working on process-based ecophysiological models, and my focus was pattern-based correlative ones. We thought that a comparison between model outputs would give us a good basis for exploring some ideas I had about how to fit the correlative models. We had a good idea for a case study.

As we worked on this concept, it became obvious that the species distribution modelling (SDM) community lacked tools for testing the sorts of things we think are important. I mocked up what I wanted to do, and in collaboration with Steven, soon started talking about ways to implement these ideas. Steven is probably known to most of you as the computer scientist behind Maxent, so having him develop methods and code things is a big advantage!

Left to to Right: Jane, Mike, Steven.

Looking back over emails, I enjoyed seeing the to-and-fro of ideas with Steven. One particular happy day was the message “Here’s a little thing for you to play with when you return – a tool to explore model predictions.  Is this roughly what you were thinking of?” This was the start of the Explain tool, something I had wanted for ages.

So, this paper is one of those excellent collaborations where ideas emerged between the three of us and came to fruition thanks to both perseverance and our diverse skills. For more background on our approach, you can listen to the podcast we made with Graziella Iossa, who was the journal coordinator for MEE at the time of our paper publication.

Invasive Cane Toad Case Study for Model Development

To explore issues associated with fitting and predicting SDMs for range-shifting species, we used the invasive cane toad (now Rhinella marina) as a case study in Australia. This destructive species was in the process of dispersing across Australia, invading new suitable habitat every year. Mike had produced a mechanistic model for the cane toad, giving us predictions of its likely fundamental niche in Australia, both current and into future climates.

Cane toad in Ecuador. Photo credit: Nicole Kearney.

We sourced presence-absence data previously used by others to model this species, and a set of climatic predictor variables, thought relevant for both current and future times. We used four modelling methods:1) boosted regression trees (BRT); 2) generalised linear models (GLMs); 3) generalised additive models (GAMs), and 4) Maxent. We fitted models to presence-absence or presence-background records and explored options including weighting records to represent lack of equilibrium in the species distribution, choosing different geographic extents for sampling background points, and controlling the complexity of the fitted functions. We also created new predictor variables from the mechanistic model representing direct ecological processes and tested adding those as covariates.

To evaluate the models, we tested predictive success using cross-validation, assessed variable importance, fitted functions and maps, quantified extrapolation, and compared outputs with those of the mechanistic model. We found that predictions varied with modelling method and data treatment, particularly with regard to the use and treatment of absence data. Our results also showed that deliberately controlling the fit of models and integrating information from mechanistic models can enhance the reliability of correlative predictions of species in non-equilibrium and novel settings.

Images from the paper (Elith et al., 2010).

One of the important products of this work were the tools we developed to “peer into” the models. The snapshots above from the paper show how we compared pairwise correlations between predictors now and into the future (top left), and presented Multivariate Environmental Similarity Surfaces (MESS) maps to show when models need to extrapolate (top right).  The “Explain” tool, which shows why a prediction is as it is at any location by linking it to the model and the environmental conditions, is shown in the bottom left, and the “Limiting factors” analysis (identifying the variable most affecting the prediction in each map cell) shown in bottom right. These were all implemented in Maxent, with some (e.g. MESS) available from the interface, and some via the batch files.

Since Our Paper…

Here we offer a rudimentary sketch of a few changes we have noticed. Our tools have been used quite widely, particularly the MESS maps which are now also available in Robert Hijman’s R package dismo. John Baumgartner has implemented MESS and its pair, MoD maps, and also Limiting factors in R.  Others have extended the ideas of identifying novel climates, including Zurell and co-authors, Mesgaran and co-authors, Engler and Rödder, and Owens and co-authors.

Many SDM practitioners are now much more aware of the problems with transferring models to new times and places, and with modelling invasive species – there are numerous relevant papers. I have pointed to a few in an overview of SDMs for Oxford bibliographies.

Screenshot of iNaturalist records for the cane toad, May 2020.

Many modellers have turned to more process-explicit methods to model range-shifting species, and we’ve done some of that. Mike has continued working on eco-physiological models and has formally released mechanistic niche modelling software. Steven has modelled Arctic vegetation, including process understanding into predictions of future vegetation changes. I have collaborated with Natalie Briscoe and Gurutzeta Guillera-Arroita, and Damaris Zurell and colleagues (via a Germany-Australia grant), who are working to understand the potential and limitations of process-based models. And, cane toads have continued to spread across Australia – in the regions we predicted!

To read the selected Volume 1 Methods in Ecology and Evolution article in full, visit The Art of Modelling Range-Shifting Species.