Parametric and non-parametric statistical methods for predicting plotwise variables based on Landsat ETM+: a comparison in an Araucaria araucana forest in Chile

Abstract

The Araucaria araucana forests have a high level of both ecological and scientific importance, because they are long-lived and endemic. Although there have been several ecological studies conducted concerning A. araucana forests, none has produced quantitative models. We compared parametric and non-parametric statistical methods for predicting stand variables from Landsat ETM+ derived variables from two A. araucana stands in south-central Chile. The assessed parametric methods were multiple linear regressions (MLR), generalized least squares with a non-null correlation structure (GLS), linear mixed-effects models (LME), and partial least squares (PLS); while the non-parametric methods were: k-nearest neighbor (k-NN) and most similar neighbor (MSN). In descending order, number of trees per ha (N), stand gross volume (V), stand basal area (G), and dominant height (H_dom) were the most difficult variables to be modeled by all the methods. LME with known random effects (i.e., LME1) performed best, achieving a root mean square showing differences (RMSD) for N and V of 18.31 and 4.08 % versus 33.06 and 33.05 % for the second-best method, respectively. However, within the parametric methods, LME1 cannot be used for predicting new observations with no data. After LME1, GLS performed the best; also accounting for the spatial correlation of the data. Parametric methods achieved lower errors. Furthermore, differences were greater among non-parametric than those among parametric methods, with a difference of 10-15 % between k-NN and MSN. Although, given our results, we favor parametric methods; we point out that non-parametric methods are also useful, and the choice between parametric and non-parametric methods depends on the ultimate objective of the study.