COS 16-4
Optimizing, and comparing multiple hypotheses with, Maxent habitat suitability models with AIC: Camas (Camassia spp.) as case study

Monday, August 11, 2014: 2:30 PM
315, Sacramento Convention Center
Braden T. Elliott, Environmental Studies, Dartmouth College, Hanover, NH

The increasingly common use of information-theoretic measures such as Akaike Information Criterion (AIC) in the analysis of spatially-explicit maximum entropy (Maxent) habitat suitability models presents an opportunity to measure the relative evidence for multiple working hypotheses explaining species distribution. This research applies AIC to Maxent in addressing two kinds of questions: appropriate model complexity, and incorporation of human cultural geography. The organism of interest is camas (Camassia spp.), a perennial angiosperm in the Pacific Northwest and an important ancestral food source for many indigenous peoples of the region. The region of analysis is the state of Oregon. Occurrence data come from the Oregon Flora Project Atlas. Environmental data sources include 19 BIOCLIM layers, 3 PRISM layers, and 3 NED-derived layers. Cultural data sources include an aggregation of 20th-century culture-area maps of the Pacific Northwest, and cost distance from archaeologically-documented earth ovens from the Oregon State Historic Preservation Office. The ecological model set incorporates only information from the environmental data sources, and the socioecological model set incorporates information from both the environmental and cultural data sources.


Dropping covarying or weak environmental predictors improved model fit (lower AIC, or smaller ΔAIC). Altering the default parameters of the Maxent package by increasing the regularization parameter decreased AIC down to an optimal level. The best model (smallest ΔAIC) in the ecological model set was pared down to two non-covarying strong predictors and optimized through regularization. The best model (smallest ΔAIC) overall was a socioecological model pared down to four non-covarying strong predictors and optimized through regularization. These results demonstrate the capacity of information-theoretic measures to weigh the evidence for different mechanisms at play in species distribution, and lend quantitative support to the well-documented qualitative relationship between indigenous peoples and traditional food plants.