Modeling the distribution of four vegetation alliances using generalized linear models and classification trees with spatial dependence

Generalized linear models (GLMs) and classification trees were developed to predict the presence of four vegetation alliances in a section of the Mojave Desert in California. Generalized additive models were used to provide response shapes for parameterizing GLMs. Environmental variables used to model the distribution of the alliances included temperature, precipitation, elevation, elevation-derived terrain variables (slope, transformed aspect, topographic moisture index, solar radiation, and landscape position), and categorical landform/surface composition variables. Vegetation distributions exhibit spatial dependence and therefore we used indicator kriging to derive neighborhood values of "presence" also used as predictors in the models. The models were developed using 2859 observations coded present or absent for each of the four alliances, and assessed using 960 observations. In general, all of the models were improved with the addition of the kriged dependence term. However, models that relied heavily on the kriged dependence term were less generalizable for predictive purposes. Classification tree models had higher classification accuracy with the training data, but were less robust when used for predictions with the test data. Each of the models was used to generate a map of predictions for each alliance and the results were often quite different. The predicted maps with the kriged dependence terms looked unrealistically smooth, particularly in the classification tree models where they were often selected as the most important variables, and therefore heavily influenced the spatial pattern of the resulting map predictions.