Exploring the use of learning techniques for relating the site index of radiata pine stands with climate, soil and physiography

The environmental uncertainty derived from climate change suggests that some commonly used empirical indicators of forest productivity, such as site index, may not be suitable for future growth prediction in the following decades. As a consequence, the development of statistical models that relate these indicators with environmental variables may be a crucial support resource for practical forest management. In this research, we tested seven different statistical learning techniques for estimating site index of radiata pine (Pinus radiata D. Don) stands in the northwest of Spain. The predictors used for this task were a set of 43 physiographic, soil, and climatic variables obtained from available raster maps for this region, whereas the site index data came from a network of 489 plot-inventory combinations set by the Sustainable Forest Management Unit of the University of Santiago de Compostela. The proposed learning techniques produced models of easy interpretation in comparison to other machine learning approaches, and accounted for up to 50% of the response's variability. The stepwise, Elastic Net, Least Angle Regression and Infinitesimal Forward Stagewise techniques provided models with high performance but at the expense of including a large number of predictors. By contrast, the Lasso and the Partial Least Squares techniques produced more parsimonious alternatives. However, these models showed noticeable heteroscedastic residuals and a certain regression to the mean. The Multivariate Adaptive Regression Splines seemed to be the most suitable technique, as it explained 50% of the site index variability with a reduced amount of predictors, and did not show undesirable patterns in the residuals and predicted values. Besides, the growth-environment relationships represented by this model seemed to be ecologically coherent.