Exploring environmental coverages of species: a new variable contribution estimation methodology for rulesets from the genetic algorithm for rule-set prediction

Variable contribution estimation for, and determination of variable importance within, ecological niche models (ENMs) remain an important area of research with continuing challenges. Most ENM algorithms provide normally exhaustive searches through variable space; however, selecting variables to include in models is a first challenge. The estimation of the explanatory power of variables and the selection of the most appropriate variable set within models can be a second challenge. Although some ENMs incorporate the variable selection rubric inside the algorithms, there is no integrated rubric to evaluate the variable importance in the Genetic Algorithm for Ruleset Production (GARP). Here, we designed a novel variable selection methodology based on the rulesets generated from a GARP experiment. The importance of the variables in a GARP experiment can be estimated based on the consideration of the prevalence of each environmental variable in the dominant presence rules of the best subset of models and its coverage. We tested the performance of this variable selection method based on simulated species with both weak and strong responses to simulated environmental covariates. The variable selection method generally performed well during the simulations with over 2/3 of the trials correctly identifying most covariates. We then predict the distribution of Toxostoma rufum (a bird with a cosmopolitan distribution) in the continental United States (US) and apply our variable selection procedure as a real-world example. We found that the distribution of T. rufum could be accurately modeled with 13 or 10 of 21 variables, using an UI cutoff of 0.5 or 0.25, respectively, arriving at parsimonious environmental coverages with good model accuracy. We also provide tools to simulate species distributions for testing ENM approaches using R.