Evaluation of modelling techniques for forest site productivity prediction in contrasting ecoregions using stochastic multicriteria acceptability analysis (SMAA)

Accurate estimation of site productivity is crucial for sustainable forest resource management. In recent years, a variety of modelling approaches have been developed and applied to predict site index from a wide range of environmental variables, with varying success. The selection, application and comparison of suitable modelling techniques remains therefore a meticulous task, subject to ongoing research and debate. In this study, the performance of five modelling techniques was compared for the prediction of forest site index in two contrasting ecoregions: the temperate lowland of Flanders, Belgium, and the Mediterranean mountains in SW Turkey. The modelling techniques include statistical (multiple linear regression - MLR, classification and regression trees - CART, generalized additive models - GAM), as well as machine-learning (artificial neural networks - ANN) and hybrid techniques (boosted regression trees - BRT). Although the selected predictor variables differed largely, with mainly topographic predictor variables in the mountain area versus soil and humus variables in the lowland area, the techniques performed comparatively similar in both ecoregions. Stochastic multicriteria acceptability analysis (SMAA) was found a well-suited multicriteria evaluation method to evaluate the performance of the modelling techniques. It has been applied on the individual species models of Flanders, as well as a species-independent evaluation, combining all developed models from the two contrasting ecoregions. We came to the conclusion that non-parametric models are better suited for predicting site index than traditional MLR. GAM and BRT are the preferred alternatives for a wide range of weight preferences. CART is preferred when very high weight is given to user-friendliness, whereas ANN is recommended when most weight is given to pure predictive performance. (C) 2011 Elsevier Ltd. All rights reserved.