Testing the sensitivity of species distribution models

Background/Question/Methods

Finding reliable methods to evaluate and predict species distributions under past and future climate scenarios continues to be a challenge in evolution and ecology.  Species distribution models (SDMs) use population locations and environmental variables to determine a species’ habitat suitability and forecast species’ historical or future ranges under specified climate scenarios. These predictive tools have many modifiable parameters and there are a multitude of options for formatting spatial data. To date, little is known about how some of these parameters may change the relative accuracy of SDM models when considering species with different sizes and patterns of distribution. We examined how applying different settings and data formats in the commonly used SDM program MaxEnt affected  model outcomes, and we assessed which methods may be more suitable for certain types of species or data. We performed our analyses for two closely related plant species with contrasting distributions in Europe – Silene vulgaris, a widespread generalist, and S. uniflora, a habitat specialist restricted to coastal areas – and used SDM to predict suitable habitat in the last glacial maximum (LGM). This species comparison allowed us to test whether species with different habitat preferences differ in their sensitivity to changes in SDM program parameters.

Results/Conclusions

We explored the differences in the model output by adjusting seven different parameters. Point thinning, replicated run type, test percentage, and raster resolution did not greatly affect LGM habitat suitability predictions for either S. vulgaris or S. uniflora.  However, choice of LGM climate scenario (based on CCSM or MIROC global circulation models), selection of feature types, and removal of correlated variables affected the outcome of the models for both species. We only saw a marked difference in model sensitivity between the species when changing the LGM data for the different climate scenarios. Based on our results, it is important for SDM users to test sensitivity to different climate scenarios, feature types, and make sure to remove correlated variables. Due to the many nuances of this type of modeling and the lack of literature comparing the effects of adjusting parameters, similar tests should be run on many different types of data sets. It will be important to compare the sensitivity of models predicting the distribution of both animal and plant taxa with different habitat requirements and contrasting dispersal mechanisms at varying spatial scales.