Evaluation of spatial predictions of site index obtained by parametric and nonparametric methods—A case study of lodgepole pine productivity

Abstract

We demonstrate the potential of using least-squares regression, generalized additive model, tree-based model, and neural network model on layers of environmental data grids for mapping site index in a case study. Grids of numerical environmental variables represented layered data, and a sparse site index plot network was located in the grids. Site index data were based on stem analysis (observed height at the index age of 50 years) of 431 lodgepole pine trees in 88 sample plots. The plots were established in a 17,460 km 2 boreal mixedwood forest of Alberta, Canada dominated by mature and over-mature stands. The generalized additive model presented a better fit and better adaptability to extreme data (i.e., mature stands) than the least squares nonlinear and other nonparametric techniques, such as the tree-based model and neural network model. Among the four models tested, nonlinear regression is of the data modeling culture, which assumes a stochastic data to relate productivity to environmental variables, and such models are optimized for estimation. Other three models belong to the algorithm modeling culture, which treat the relationship between productivity and independent variables as an unknown black box and try to find a function between them; therefore, these models are more suitable for prediction purpose. Implications for biophysical site index modelling with extreme data are discussed.