(left) Philipp collecting leaves of Taraxacum palustre during a trait sampling campaign in 2019. Photo taken close to Zurich, Switzerland. (right) Adonis vernalis – a rare beauty that was part of our plant trait sampling campaign in 2019. Photo taken close to Martigny, Switzerland.
Institution: Swiss Federal Institute for Forest, Snow, and Avalanche Research
Current academic life stage: Postdoc
Research interests: I am a macroecologist with a broad interest in questions related to species and trait distributions and biodiversity.
Current study system: After working with plankton biogeography during my PhD, in my postdoc I now focus on terrestrial plants in Europe, in particular on the communities of the European Alps. What I particularly like about these plant communities is (i) that I can develop and validate ideas about them while being active outdoors and (ii) that the wealth of observational, trait, and phylogenetic data combined with environmental and remote sensing data provide countless possibilities for creative analyses to deepen our understanding.
Recent paper in Journal of Biogeography: Brun P, Thuiller W, Chauvier Y, Pellissier L, Wüest RO, Wang Z, Zimmermann NE. 2020. Model complexity affects species distribution projections under climate change. Journal of Biogeography 47: 130– 142. DOI: 10.1111/jbi.13734. FULL ACCESS FOR 2020 & 2021.
Motivation for the paper: Since the paper by Merow et al. (2014), the species distribution modelling (SDM) community is generally aware that decisions about model complexity can have important and sometimes problematic implications for study results, but it had never been thoroughly assessed what these implications are for the important SDM application of projecting future range changes. We comprehensively studied how three aspects related to model complexity (parameterization complexity, number of predictor variables, and multicollinearity) affect analyses, using dominant European tree species.
(left) Major European forests (Data Source: https://land.copernicus.eu/). (right) Strong environmental gradients in the European Alps. While the growing season is well on its way close to the valley bottom, the higher altitudes are still deeply covered with snow. Photo taken close to Mount Titlis in the Swiss Alps.
Key methodologies: We made an effort to comprehensively investigate the implications of model complexity in our analysis. We randomly subsampled 300 sets of predictors from a substantial pool of climate, soil, and terrain variables. This gave us the possibility to study the effects of the number of variables considered and their multicollinearity, independent of the actual predictors used. We also compared different levels of parameterization complexity, restricting the algorithms to fit very coarse occurrence-environment relationships at one extreme and allowing them to closely follow the data and identify relationships with very complex shapes at the other. In addition, we varied factors that are often permuted in projection ensembles, i.e., SDM algorithms, emission scenarios, and climate models. All in all, we made almost a million future projections. Given that our observational and environmental data were of high quality (regular, rich sampling design, confirmed absences), our results provide a robust and reliable assessment of the implications of decisions on model complexity relative to aspects of SDM projection design.
Unexpected outcomes: I was surprised to see that, apart from the most common species, high multicollinearity did not notably decrease model performance, even when assessed under environmental block cross-validation. Yet, multicollinearity systematically increased range loss projections, indicating that this violation of model assumptions has a distortive effect on projections that may pass below the radar of common model evaluation.
Major result and contribution to the field: Parameterization complexity should be varied along with SDM algorithms in ensemble projections. The range of suitable options depends on the dataset at hand and may be identified by decent model performance. The number of predictors included should be balanced between providing sufficient information for well-performing models and avoiding too much noise, which deteriorates performance and introduces disagreement between projections. We found 10 predictors to be ideal, but the number may be smaller for less well-designed survey data or flatter environmental gradients. Multicollinearity should be constrained by maximum absolute Pearson correlation coefficients of 0.7, in order to avoid distorted projections.
What are the next steps? I see this work as a contribution to the call of Araújo et al. (2019), to develop standards for methods and data in SDM-based studies. Ultimately, the goal is to reliably predict the impacts of global change on biodiversity.
If you could study any organism on Earth, what would it be and why? Studying the ecology of individual plants at scale using remote sensing data is something I would love to do. Otherwise, I would be keen to know more about the fine-scale distribution and fruiting behaviour of the morel mushroom (Morchella esculenta).