Presence-Only Methods for Predication of Eurasian Watermilfoil Habitat

Prince Czarnecki, J. M., Madsen, J. D., Shaw, D. R., & Brooks, C. P. (2010). Presence-Only Methods for Predication of Eurasian Watermilfoil Habitat. Proceedings of the Southern Weed Science Society. Little Rock, AR.

A number of modeling techniques have been developed that predict habitat suitability based on species presence (e.g., Mahalanobis distance, Maxent, etc.). While presence-only models have limitations, presence data is often the only data available for regional to large-scale research. Additionally, sampling methods can often force the use of presence-only methods because a lack of data in specific areas cannot be treated as absence of a species. In this study, Mahalanobis distance and Maximum Entropy were used to characterize and predict habitat for the invasive aquatic macrophyte, Eurasian watermilfoil (Myriophyllum spicatum). Both methods were applied within a GIS framework. The state of Minnesota was divided into a 500m grid using ArcGIS and Hawth’s Tools. Non-water areas were removed from the sample. Data for analysis were obtained from the Minnesota Department of Natural Resources and several units at the University of Minnesota. These included: Secchi depth, total alkalinity, Carlson’s Trophic State Index, lake size, distance from lake access (i.e., boat launch), distance from road, distance from reported bass habitat, M. spicatum presence, and where available, absence. Data were weighted for analysis using flow accumulation rates obtained from the National Hydrography Dataset Plus. Mahalanobis distances were calculated for each grid cell and converted to chi-square p-values with n-1 degrees of freedom (where n = the number of predictor variables). Re-classed output was compared to known values of presence and absence for validation. Validation included calculating Cohen’s kappa, specificity, and sensitivity. Low kappa (0.1) along with sensitivity of 0.55 and specificity of 0.75 suggest that the Mahalanobis distance model is a poor predictor of M. spicatum habitat. In contrast, the Maxent approach resulted in a highly predictive model. The area under the receiver operating curve for the model, which indicates the quality of the fit, is 0.968. Bass habitat (45%) followed by Carlson’s TSI (28%) explained the greatest variation in the model. Lake access contributed the least (0.6%) to the model, confirming the conclusions of previous authors with regard to anthropogenic contributions to M. spicatum presence. Results of this analysis indicate that while current M. spicatum habitat is correctly characterized by the model, the weed may not have reached all potential habitats due to some limiting factor, probably time.