Local-scale GIS Modeling of Eurasian Watermilfoil
Prince Czarnecki, J. M., Shaw, D. R., Madsen, J. D., Harvill, J. L., Ervin, G. N., Samson, S., & Martin, James L. (2009). Local-scale GIS Modeling of Eurasian Watermilfoil. Proceedings Southern Weed Science Society /Weed Science Society of America Abstracts. Orlando, FL.
A spatial analysis using generalized linear mixed models was conducted to estimate the predictive probability for the presence of Eurasian watermilfoil (Myriophyllum spicatum L.) in Pend Oreille Lake and outflowing river (Idaho). Predictor variables included water depth, fetch length, and distance from nearest M. spicatum population. Water depth was interpolated from NOAA sounding data using simple kriging with ArcGIS Geostatistical Analyst. Fetch length was determined using methods outlined in the Shoreline Protection Manual. These methods were automated using python scripts obtained from USGS. Input wind speed and direction data were taken from a U.S. Navy buoy on Pend Oreille Lake. Distance from nearest population was a 0/1 variable which identified if the point was within a reasonable distance to expect asexual spread from the nearest population. Presence/absence data were obtained by field surveys conducted in summer 2007. Data were interpolated and re-sampled to a 250-meter point grid for analysis in SAS. Re-sampling and grid generation were done with Hawth’s tools in ArcGIS. Only points where water depth was less than 20 m were considered for model use, representing 2 times the maximum preferred depth of growth for M. spicatum reported in literature.
Model parameters estimates were obtained using Proc GLIMMIX in SAS. Using data on Pend Oreille Lake and separately on data from the river flowing out of the Lake, five statistical models were considered for estimating the predictive probability of the presence of M. spicatum in terms of the three predictors. The first was a traditional logistic regression model. This model did not include a spatial autocorrelation structure, but did include the Boolean distance variable. The remaining four models incorporated spatial autocorrelation via a spherical spatial covariance function, and so did not require the Boolean distance variable. Specifically the four spatial models considered in this study were a (1) binomial regression model with over-dispersion, (2) random effects model, (3) conditional spatial generalized linear mixed model, and (4) marginal spatial GLM. Depth and fetch were highly significant in every model considered. All four of the spatial models had a lower predictive probability error variance than the logistic regression model. However, the additional complexity of spatial models requires advanced computing algorithms for parameter estimation. For this study, this additional complexity resulted in a lack of convergence in some cases. When it was possible to estimate all the parameters in the model, the four spatial models performed equally well. Depending upon the desired interpretation of the model and covariance parameters, and the ability to estimate those parameters, one of these four models may be preferred over the other three. Based on this study, in terms of predictive probability error variance, all spatial models are preferred over the traditional logistic regression model.
Model outputs indicate predictive probabilities for the presence of M. spicatum at each point in the lake and river. A spatial view of these probabilities created in ArcGIS illustrates areas where M. spicatum is likely to occur based on existing depth and fetch.