Document: JAN-3-40-25

Predicting the distribution of vegetation communities from environmental variables in the Mojave Desert Ecoregion.

FRANKLIN, J.*

San Diego State University, San Diego, CA 92182-4493 USA ¹

Abstract:
Classification trees (CTs) have been used in ecology to predict species-habitat, vegetation-environment, and soil-environment relationships. CT modeling is one of several inductive approaches to learning generalized relationships between variables. It is a non-parametric, monothetic, divisive, clustering technique for deriving classification rules that can be used predictively by applying them to new data. I used CTs to predict vegetation type (Alliance) from 16 terrain, climate and substrate variables from a sample of ~1800 observations (field releves of vegetation composition) in the Mojave Desert. When the dataset was stratified into three groups, comprising four, 12 and 22 categories (Alliances) respectively, the resulting models predicted the alliance of each observation correctly 50-80% of the time (correct classification rate) using 10-11 variables. When it was allowed that certain classes are hard to separate, given the training sample and explanatory variables (where in limited cases a correct prediction of "either A or B" was acceptable), the correct classification rate was 75-90%. However, these accuracy estimates are based on the data used to parameterize the model. Cross-validation suggested that these models are not robust when presented with new data. There is some theoretical basis for using a CT framework to predict species distributions, when it is expected that those distributions are delimited by thresholds of tolerance, and that the environmental variables used for modeling interact hierarchically to represent environmental gradients. However, these results suggest that alternative model formulations may have theoretical and predictive advantages. For example, the presence/absence of each alliance could be modeled separately using CTs, generalized linear or generalized additive models. There is considerable statistical and ecological debate about how to validate this type of model. Should all available data be used for parameter estimation, or should new data be collected for validation? Although this sample was large, it appears it was not adequate for modeling rare alliances. The objective of this study is predictive mapping for regional biodiversity inventory, but scale differences between the training data and the map introduce additional challenges.

Keywords: predictive vegetation mapping, classification trees, species distributions

This abstract is being presented at: 8:15 AM in session:
Oral Session #59: Plant Communities: Vegetative Analysis.