Is standard Maxent model over-fitted? A test based on simulation
Understanding the distribution of species in relation to environment and human disturbance is critical for basic ecology and biodiversity conservation. Maxent is commonly used to model species distribution based on presence-only species data and to estimate local species richness. The default value of beta, an key parameter for controlling model complexity is set at 1. It is a general concern whether the default value may lead to model over-fitting. Recent studies have recommended using AICc to select a simpler and better model at a higher beta value. I modeled the distributions of 72 fish species in Illinois stream based on sampling data and landscape-level environmental variables, and took them as the truth, against which the predictions of the default and selected models were assessed. The sensitivity of the selected models increased, but the specificity decreased. Consequently, the overall performance measured with True Statistical Skill changed little. It is true of the similarity between the predicted and true species composition. I concluded that the default Maxent modeling is highly resistant to over-fitting, however, high beta values may be needed to minimize omission errors in modeling endangered species.
Yong Cao (Primary Presenter/Author), Illinois Natural History Survey, University of illinois, firstname.lastname@example.org;