TABLE 1. ENVIRONMENTAL VARIABLE SUITE A SPATIAL MODEL FOR PREDICTING UNRECOGNIZED POPULATIONS OF THE OREGON SPOTTED FROG (RANA PRETIOSA) TOWARD THE SOUTHERN END OF ITS GEOGRAPHIC RANGE Klamath Basin OSF Localities B Luke Groff and Sharyn Marks; Humboldt State University, Department of Biological Sciences Marc Hayes; Washington Department of Fish and Wildlife CALIFORNIA COOPERATIVE FISH & WILDLIFE RESEARCH UNIT A C D ABSTRACT: The Oregon Spotted Frog (Rana pretiosa), endemic to the Pacific Northwest, was once considered widespread in complex warm-water wetlands. Over 70% of historic populations are thought to be extirpated with range-wide habitat loss exceeding 90%. We developed a spatial model to elucidate the probable distribution of Rana pretiosa toward the southern end of its geographic range using Maxent software and ESRI Geographic Information System technology (ArcGIS). This model was generated from two sets of spatial data, a set of occurrence points and a suite of environmental variables. The final output from this model is a predictive map that identifies areas of suitable habitat. Coupled with the National Wetlands Inventory, we intend to use this model to identify survey-deficient areas for this species, which are anticipated to occur primarily on private lands in the Klamath Basin of Oregon and the Pit River system of California. Most previous surveys of the southern range have focused on public lands. Our aim is to survey areas that the model reveals as promising during the 2010 breeding season. Any previously unrecognized populations found, particularly near the species’ range limit in California, would be important to its conservation. A B C D METHODS: We developed this model with 26 Rana pretiosa occurrence localities and a suite of 23 environmental variables. Localities include all verifiable populations within the study area and consist of 15 extant and 11 extinct populations. The variables, used to characterize habitat associated with known populations, are listed in Table 1. This model was generated with ArcGIS and Maxent, a software program used for habitat modeling. Specific model parameters are not described here. Because of our low number of occurrence localities, we implemented a novel assessment technique described in Pearson et al. (2007). This jackknife (or ‘leave-one-out’) approach requires that each locality be removed once from the dataset and a model be generated with the remaining n – 1 localities. Thus, 26 models were generated. Predictive performance was then assessed based on the ability of each model to predict the single excluded locality. We evaluated whether the success/failure of our predictions were superior to a random assignment of the excluded localities by calculating a P value, based on test criterion D. D = ∑ Xi (1 – Pi) 1. National Land Cover Data (NLCD 2001) 2. Soil Hydricity (STATSGO) 3. Elevation 4. Slope 5. Annual Mean Temperature 6. Mean Diurnal Range 7. Isothermality 8. Temperature Seasonality 9. Max Temperature of Warmest Month 10. Min Temperature of Coldest Month 11. Temperature Annual Range 13. Mean Temperature of Driest Quarter 14. Mean Temperature of Warmest Quarter 15. Mean Temperature of Coldest Quarter 16. Annual Precipitation 17. Precipitation of Wettest Month 18. Precipitation of Driest Month 19. Precipitation Seasonality 20. Precipitation of Wettest Quarter 21. Precipitation of Driest Quarter 22. Precipitation of Warmest Quarter 23. Precipitation of Coldest Quarter *Variables 5-23 obtained from WorldClim TABLE 1. ENVIRONMENTAL VARIABLE SUITE For the above formula, Xi is a success/failure variable to indicate if the ith excluded locality is included in the ith predicted area of the corresponding model. Xi is coded as 1 if the excluded locality falls in the prediction area and 0 if it does not. Pi is the proportion of the study area predicted after excluding the ith locality. RESULTS and CONCLUSION: Twenty-one of the 26 (80.8%) assessment models successfully predicted the respective excluded locality. Test results were highly significant (P < 0.001), suggesting that this model is far superior to a random assignment of the excluded localities. Landcover (47.6%), slope (23.2%), and soil hydricity (11.4%) were determined to be the most influential variables, accounting for 82.2% of the model. Importantly, the model output is continuous; each predicted cell has a value between 0 and 1 associated with it, allowing for comparisons between predicted areas. These suitability values, coupled with the National Wetlands Inventory, will help identify promising Rana pretiosa habitat and facilitate our 2010 surveys. While public lands have been the focus of most prior surveys in this region, we intend to concentrate primarily on private lands in Oregon’s Klamath Basin and California’s Pit River system. It would be very important for the conservation of this species to discover even a single unrecognized population, especially near the species’ southern range limit. ACKNOWLEDGEMENTS: We would like to thank the following individuals and organizations for their gracious assistance and support: Walter Duffy, Steven Steinberg, Kristine Preston, Paul Evangelista, Shannon Chapin, Jamie Bettaso, Blake Hossack, the California Cooperative Fish and Wildlife Research Unit, the Washington Department of Fish and Wildlife, and the U.S. Fish and Wildlife Service. REFERENCE: Pearson, R.G., Raxworthy, C.J., Nakamura, M. and Peterson, A.T. (2007). Predicting species distributions from small number of occurrence records: a test case using cryptic geckos in Madagascar. Journal of Biogeography, 34, 102-117.