Combining classification and regression trees and the neuro-fuzzy inference system for environmental data modeling

A procedure is presented for dynamic statistical modeling of predictands with nonlinear predictand-predictor relationships when there are many potential predictors. Classification and regression trees (CART) are used for predictor selection and data stratification. The CART output is suitable for piecewise-continuous predictands. Using predictors selected by CART, a neuro-fuzzy inference system (NFIS) algorithm produces an output model for continuous predictands. An application to modeling ground-level ozone is discussed.