COS 66-5
Generating maps of uncertainty for ecological niche models using a single-algorithm ensemble approach

Wednesday, August 13, 2014: 9:20 AM
Carmel AB, Hyatt Regency Hotel
Robert A. Boria, Biology, City College (CUNY), NY, NY
Robert P. Anderson, Biology, City College of New York, City University of New York, New York, NY
Background/Question/Methods

This study aims to create a consensus ecological niche model (ENM) using a single-algorithm approach, while adjusting model parameters to maximize performance. Generally, a single model from one algorithm or an ensemble of different algorithms is used to generate a prediction. Additionally, several recent studies have shown the need tune model settings for a single algorithm. Nevertheless, uncertainty cannot be measured with a single model, and multiple settings may lead to essentially co-optimal models. To address this issue, we used Maxent, 19 bioclimatic variables, and occurrence records of a Malagasy tenrec, Microgale gracilis. To make higher-quality models, we reduced the effects of sampling biases (via spatial filtering) and used a custom study region. We calibrated and evaluated preliminary models using a jackknifing approach, tuning two model settings (feature classes and the regularization multiplier) to estimate optimal model complexity. Based on omission rates and AUC, we chose the top 10 performing preliminary models, and then generated a consensus prediction by averaging the values for each grid cell. Furthermore, we calculated the standard deviation to obtain a map showing variation in geography. 

Results/Conclusions

The top-performing models fell into two groups, those made with linear and quadratic feature classes, and those with the hinge feature class. These models were 56–95% similar in geographic space (D-statistic), again falling into two clusters; models made with the same feature classes were similar, whereas cross-cluster comparisons showed much lower similarity). In conclusion, similarly performing models can have high variation in their geographic predictions. A consensus model allows researchers to use several high-performing models, which may vary in geography, to generate a useful prediction and detect areas of discrepancy.