A comparative study of regression methods to predict forest structure and canopy fuel variables from LiDAR full-waveform data

Abstract

Regression methods are widely employed in forestry to predict and map structure and canopy fuel variables. We present a study where several regression models (linear, non-linear, regression trees and ensemble) were assessed. Independent variables were calculated using metrics extracted from full-waveform LiDAR data, while the reference data used to generate the dependent variables for the prediction models were obtained from fieldwork in 78 plots of 16 m radius. Transformations of dependent and independent variables with feature selection were carried out to assess their influence in the prediction of response variables. In order to evaluate significant differences and rank regression models we used the non-parametric tests Wilcoxon and Friedman, and post-hoc analysis or post-hoc pairwise multiple comparison tests, such as Nemenyi, for Friedman test. Regressions using transformation of the dependent variable, like square-root or logarithmic, or the independent variable, increased R2 up to 6% with respect to linear regression using unprocessed response variables. CART (Classification and Regression Tree) method provided poor results, but it may be interesting for categorisation purposes. Square-root transformation of the dependent variable is the method having the best overall results, except for stand volume. However, not always has a significant improvement with respect to other regression methods.