Wednesday, August 8, 2007
Exhibit Halls 1 and 2, San Jose McEnery Convention Center
Wenyun Zuo, Biology, University of New Mexico, Albuqureque, NM, Ni Lao, Computer Science, Carnegie Mellon University, PA, Yuying Geng, Laboratory of Quantitative Vegetation Ecology, Institute of Botany, the Chinese Academy of Sciences and Keping Ma, State Key Laboratory of Vegetation and Environmental Change, Institute of Botany, Chinese Academy of Sciences, Beijing, China
Most predictive models of species potential distributions are based on environmental variables, because they potentially important niche dimensions. Unfortunately, most predictive models suffer from the “high dimension small sample size” problem—they cannot give satisfactory results when there are only limited specimen data, and cannot handle large number of environment factors. Support Vector Machine (SVM), which is based on the Structural Risk Minimization principle, especially suitable for these kinds of data. Here, we implement a new predictive system for modeling species potential distributions based on SVM methods. To evaluate the effectiveness of the method, we perform a country-scale case study using 30 species of Rhododendron L. in China, with geographically referenced specimen data and 11 layers of 1km resolution environmental data. Here we report three results. First, using expert evaluation and Receiver Operator Characteristic (ROC) curves, we compare the SVM with the commonly used Genetic Algorithm for Rule-Set Prediction (GARP). Our results show that all SVM predictions are consistently better than GARP. Furthermore, the SVM model runs much faster than GARP. Second, we investigate the SVM performance with different numbers of environmental layers and specimen localities. We show that, even for species with few records, SVM can produce reasonable potential distribution maps. Therefore, SVM can predict potential distributions of rare species.
Finally, we use the SVM system to predict potential distribution of 400 Rhododendron species in China. We use all available Chinese specimens of Rhododendron L. and 83 layers of 1km resolution environmental data. Based on potential distribution of these species, we quantity the spatial patterns of species diversity, endemic species diversity, endangered species diversity, subgenus diversity, and different life styles diversity. These distribution maps of biodiversity not only quantity the biogeography of Rhododendron, but also provide valuable information for conservation, reintroduction, and new species discovery.
