PARENT SESSION
Contributed Oral Session 93: Modeling: Movement, Populations, and Communities
Wednesday, August 10, 1:30 PM - 5:00 PM, Meeting Room 513 E, Level 5, Palais des congrès de Montréal
Comparing species distribution models using artificial species in a real landscape.
Meynard, Christine*,1, Quinn, James1, 1 University of California, Davis, Davis, CA95616
ABSTRACT- Species distributions can be modeled by establishing a relationship between occurrence locations and the environmental or habitat characteristics of those locations. Different strategies can be used to establish these relationships. These range in complexity from describing minimum environmental envelopes encompassing all locations to statistical analyses including several regression-like and neural network approaches. Among the most frequently used techniques are General Linear Models (GLMs), General Additive Models (GAMs), Classification Trees (CT) and Genetic Algorithm for Rule Set Production (GARP). Each of these statistical models work in different ways and make different assumptions. There are also important differences in terms of their availability in statistical packages and their compatibility with Geographic Information Systems (GIS). Here we compare these models in terms of their prediction ability using artificial species. California environmental information was gathered on a GIS and several model species were created. For each one we mathematically populated the California landscape by assuming a different "fundamental niche" determined by a few climatic and topographic variables and interactions between them. The landscape was then randomly sampled for species presence/absence data and for all environmental information available from a standardized GIS library of the state. Predictive ability was assessed by comparing the area under the receiver operator curve, as well as by looking at the explanatory variables selected by each analysis compared with variables initially used to create deterministic distributions. Models differ in their performance depending on the ecology of the model species. GLMs and GARP generally performed well, but GLMs provide a more straightforward analysis than the other approaches. On the other hand, classification trees did not perform as well, and tended to over emphasize categorical variables. It is important to consider these differences when applying these kinds of strategies to conservation and biogeographic studies.
Key words: species distribution, modeling, artificial species, niche