Friday, August 7, 2009 - 9:00 AM

COS 120-4: Predicting hotspots of plant invasion from small sample sizes

Sunil Kumar, Colorado State University and Thomas J. Stohlgren, USGS Fort Collins Science Center.

Background/Question/Methods

Accurate and reliable spatial prediction of habitats that are likely to be invaded by nonnative plant species is urgently required to guide field crews and policy makers, identify priority areas for early detection and rapid response, and evaluate the cost-effectiveness of monitoring and control programs. Most early invaders have small number of occurrence locations making commonly used habitat modeling approaches problematic. Our objectives in this study were to: (1) predict “hotspots” of plant invasions using small sample sizes; and (2) test the predictions using a newly collected data from the field.

Native and nonnative plant species occurrence data were collected from 180 0.1-ha modified-Whittaker plots from Rocky Mountain National Park, Colorado, USA. Nonnative plant species were found at 115 plot locations out of 180. For modeling, we selected nonnative plant species that occurred on at least four plots; total 21 species with sample sizes (n) ranging from 4 to 74. We used maximum entropy distribution modeling or ‘Maxent’ method for predicting these 21 dominant and co-dominant nonnative species. Model performance for species with n > 20 (six species) was tested using area under the receiver operating characteristic curve (AUC). However, for species with n < 20 (15 species) we used a “jackknife or leave-one-out” procedure of model performance. Predictions for 21 species were combined together to generate a map of “plant invasion hotspots.”

Results/Conclusions

Predictive models for all the species performed better than a random prediction with the average test AUC values ranging from 0.84 to 0.96 for the six species with relatively large sample sizes (n from 26 to 74). Predictive models for the remaining 15 species had high success rates (i.e., low omission rate) ranging from 50-90% and were statistically significant. For example, model for Lolium pretense (n = 4) had 50% success rate (P = 0.008), while the model for Agrostis gigantea (redtop, n=10) had a 90% success rate (P < 0.0001). The field test showed that the combined “hotspots” map was statistically significant with an R2 value of 0.44 (P < 0.0001). Our results show that presence-only or niche modeling methods such as Maxent can be used to predict “hotspots” of plant invasion, even when using small sample sizes.