news and update ISSN 1948‐6596 update Choosing the right path for species distribution modeling Species distribution models (SDM) have been widely used to address the lack of knowledge about species’ distributions (i.e. the so‐called Wal‐ lacean shortfall; Lomolino 2004). However, SDM show great variability in their predictions due to the uncertainties that accumulate during the modeling process (Barry & Elith 2006). These un‐ certainties are generally separated in three major classes: data, model and predictions. Uncertainty in the data refers to variation arising from incom‐ plete knowledge about the patterns and proc‐ esses studied, bias in the sampling procedure, and quality and choice of covariates; it applies to data used to build and fit the models, as well as to as‐ sess their success. Model uncertainty is due to discrepancies in the assumptions and algorithms used to fit the data. Finally, given that the true distribution is unknown and the relationship be‐ tween species’ presence and the environment may differ in studied and unstudied geographic regions, we are uncertain whether a prediction is a perfect fit and even if it were, it may not be so at a different time or place. Currently, SDM applica‐ tions lack approaches to manage such uncertainty, which could result in using them without consider‐ ing the reliability of their outputs with the subse‐ quent cost in the quality of the research based in these analyses. If, on the contrary, such uncer‐ tainty is acknowledged but is thought to be un‐ manageable, SDM would be flagged as unreliable for conservation planning or basic research de‐ spite their potential to provide good quality bio‐ geographical data. Beale and Lennon (2012) review the main sources of uncertainty associated with SDM and highlight research directions to improve SDM pre‐ dictions. These authors categorize SDM along an axis ranging from those that are purely statistical and try to identify process from pattern (i.e. niche ‐based models) to those that identify directly the processes and mechanisms to then generate the distributional pattern (i.e. process‐based models). For practical reasons the authors classified SDM into three types: niche‐based distribution models (which estimate the niche from the species’ geo‐ graphic distribution and re‐project it on a geo‐ graphic space), demographic models (which corre‐ late demographic parameters of the species with climate or weather to characterize its distribution) and process‐based models (which identify the physiological responses of the species and use them to determine the geographical distribution). The authors argue that, for all model types, the most critical sources of uncertainty are model uncertainty and prediction uncertainty. While the type of model is clearly relevant to niche‐based models for its influence in niche identification, the type of model also is important for demographic models to identify the actual links between popu‐ lation growth and weather, and for process‐based models to correctly estimate parameters. The un‐ certainties associated with predictions also affect all model types because of current deficiencies in the measures of the model fit. Beale and Lennon (2012) state that the performance of models can be improved by the incorporation of uncertainty in environmental covariates and by development of measures of model fit that take into account model complexity but are not affected by preva‐ lence and spatial autocorrelation. Specifically for niche‐based SDMs, Beale and Lennon (2012) identify that the quality of dis‐ tribution data presents particular challenges, since these data are not only used for model fit but also for model building. They argue, however, that suitable tools exist to assess the uncertainty at all steps of the modeling process of niche‐based models (data quality, choice of covariates, model‐ ing technique and evaluation). Once these uncer‐ tainties are estimated, it is possible to incorporate spatial error terms into the model building. This allows evaluation of the effects of uncertainties in the predictions, or even obtaining SDM results corrected by the underlying uncertainty. Process‐based models, however, require detailed species‐specific information which often is unavailable, and present the additional difficulty of identifying the interaction between species and frontiers of biogeography 4.3, 2012 — © 2012 the authors; journal compilation © 2012 The International Biogeography Society
Read full abstract
Jul 17, 2013
David Bush + 3
Open Access
