Incorporating spatial dependence in predictive vegetation models

Predictive vegetation modeling can be defined as predicting the distribution of vegetation across a landscape based on the relationship between the spatial distribution of vegetation and certain environmental variables. Often these predictive models are developed without considering the spatial pattern that exists in biogeographical data. When explicitly included in the model, this spatial dependence can increase the predictive ability significantly. In this study, presence/absence models of vegetation alliances in a portion of the Mojave Desert (California, USA) are developed using classification trees and generalized linear models and two methods of incorporating spatial dependence in the models are explored. The first method of incorporating spatial dependence involves interpolation and simulation techniques to “fill in the blanks” of the sample data to obtain an additional variable of neighborhood presence/absence. The second method considers that the model residuals are a direct indication of spatial dependence, typically in the form of an unmeasured yet important environmental variable. The model residuals are interpolated to a continuous map and added to the model predictions. In general, incorporating spatial dependence resulted in improved model accuracy for a majority of the eleven vegetation alliances studied here. However, incorporating spatial dependence did decrease the accuracy for some alliances, typically the rarer alliances. Simulation, while more computationally intensive than interpolation, provided more realistic looking predictions. When focusing on the spatial dependence in the model residuals, more robust model predictions resulted, as the alliances which are predicted well by environmental variables were “left alone”.