Habitat modeling of a rare endemic trillium species (Trillium Simile Gleason): a comparison of the methods Maxent and DOMAIN for modeling rare species-rich habitat

WCU Author/Contributor (non-WCU co-authors, if there are any, appear on document)
Ashley Mendenhall Hawk (Creator)
Institution
Western Carolina University (WCU )
Web Site: http://library.wcu.edu/
Advisor
Laura DeWald

Abstract: Many species habitat and distribution models are available that use field habitat observations to identify environmental predictor variables and quantify species-environment relationships. The relative effectiveness in terms of ease of use and accurately predicting habitat is not known for many of the models. The purpose of this study was to compare Maxent (a machine learning probability model) to DOMAIN (a simple GIS statistical profile model) in terms of their habitat prediction. Both models were used to predict habitat for the rare Trillium simile, an endemic species of the southern Appalachian Mountains found in the very rich cove environments that also provide critical habitat for other similarly sensitive species. Habitat was characterized by measuring biotic and abiotic variables at 20 sites where the species was found scattered throughout National Forests (Pisgah, Nantahala, Cherokee, and Sumter) in North Carolina and South Carolina, and the Great Smoky Mountains National Park in Tennessee. Digital environmental and climatic data for the known locations were matched to the abiotic and biotic variables measured in the field to create the models. Maxent performed with an AUC of 0.839; a DOMAIN AUC was not available because output was not automatic, and there was insufficient guidance on how to calculate the AUC. Highly suitable, suitable, and unsuitable habitat were identified for T. simile using both Maxent and DOMAIN. Model validation was performed by visiting a total of 12 highly suitable sites predicted by each model, where the original variables were collected for comparison against known occurrence sites. Model predicted data were compared to the known T. simile site data using statistical analyses and quantitative assessment. Maxent and DOMAIN models were compared using a method agreement analysis. Univariate ANOVA results, descriptive statistics, and percentages of sites withheld during model testing showed that both models successfully predicted highly suitable habitat for T. simile consistent with the characteristics of the known occurrence locations, although predictions were slightly different. Method agreement analysis resulted in a Cohen’s kappa of substantial agreement (?=0.674) between the methods Maxent and DOMAIN. Specific project objectives coupled with the complexity of understanding and evaluating model performance makes choosing a “best-fit” model a challenge for modeling rare, endemic plant species. Both models were successful at predicting suitable habitat for T. simile, and although Maxent proved to work well on a small-scale, DOMAIN was much simpler to use and is thus the recommended method. Additional experience using both models under different project circumstances and informed opinion will further assist modelers in deciding whether to use a complex model like Maxent or a more simple, less flexible model like DOMAIN when modeling habitat distribution for rare plants like T. simile.

Additional Information

Publication
Thesis
Language: English
Date: 2017
Keywords
distribution mapping, DOMAIN, Maxent, rare plant, southern Appalachians, species distribution model

Email this document to