Performance of modelling techniques for the prediction of forest site index: a case study for pine and cedar in the Taurus mountains, Turkey

Forest research has a long tradition in studying the relationship between stand productivity and abiotic and biotic site characteristics, such as climate, topography, soil and vegetation. Many of the early site quality modelling studies related site index (i.e. the dominant height of a forest stand at a reference age) to environmental variables that could be measured in the field relatively easily and at low cost, using basic statistical methods such as linear regression. Because most ecological variables show a typical non-linear and a non-constant variance distribution, a big source of error remained unexplained with the use of these linear models. More recently, the development of more advanced non-linear and even non-parametric and machine learning methods provided opportunities to overcome these limits. Nevertheless, these methods have also drawbacks. Due to their increasing complexity they are not only more difficult to implement and interpret, especially the ‘black box’ methods, but also more vulnerable to ‘overfitting’. The challenge is located in choosing the appropriate modelling technique for a specific situation. In this study, five different modelling techniques for predicting the site index were compared and evaluated, including multiple linear regression (MLR), classification and regression trees (CART), boosted regression trees (BRT), generalized additive models (GAM), and artificial neural networks (ANN). 167 sample plots were distributed over homogeneous stands of three important tree species of the Taurus mountains of Turkey: Pinus brutia, Pinus nigra and Cedrus libani. Soil, vegetation and topographic conditions were measured in detail and related to the site index with the five earlier mentioned modelling techniques. Except for the CART-method all methods evaluated in this study showed an improved prediction performance over the traditionally used MLR. ANN showed the overall best performance, but the complexity of the model, the training and testing effort and the interpretability taken into account, BRT and especially GAM present themselves as good alternatives.status: publishe