Rule-based vs parametric approaches for developing climate-sensitive site index models: a case study for Scots pine stands in northwestern Spain

Abstract

Parametric indirect models derived from stem analysis of dominant trees were more robust than rule-based machine learning techniques for predicting Site Index of Scots pine stands as a function of climate. The uncertainties derived from climate change make it necessary to develop new methods for representing the relationships between site conditions and forest growth. To compare parametric vs nonparametric approaches for modeling site index (SI) of Scots pine stands using bioclimatic variables. We used Random Forest, Boosted Trees, and Cubist techniques for directly predicting the SI of 41 research plots of Scots pine stands, and six parametric models for indirectly predicting SI using stem analysis data. As predictors, we used raster maps of 19 bioclimatic variables. The fitted models explained up to $$\sim$$ 80% of the SI variability, using from five to nine bioclimatic predictors. Though the apparent performance of the parametric models was lower than the rule-based, their bootstrap validation statistics were noticeably higher. Parametric indirect models seemed to be the most robust modeling alternative.