A Comparison of Modeling Methods for Predicting Forest Attributes Using Lidar Metrics

Abstract

Recent advancements in laser scanning technology have demonstrated great potential for the precise characterization of forests. However, a major challenge in utilizing metrics derived from lidar data for the forest attribute prediction is the high degree of correlation between these metrics, leading to multicollinearity issues when developing multivariate linear regression models. To address this challenge, this study compared the performance of four different modeling methods for predicting various forest attributes using aerial lidar data: (1) Least Squares Regression (LSR), (2) Adaptive Least Absolute Shrinkage and Selection Operator (ALASSO), (3) Random Forest (RF), and (4) Generalized Additive Modeling Selection (GAMSEL). The study used three primary plot-level forest attributes (volume, basal area, and dominant height) as response variables and thirty-nine plot-level lidar metrics as explanatory variables. A k-fold cross-validation approach was used, with consistent folds to assess the performance of each method. Our results revealed that no single method demonstrated a significant advantage over the others. Nonetheless, the highest R2 values of 0.88, 0.83, and 0.87 for volume, basal area, and dominant height, respectively, were achieved using the ALASSO method. This method was also found to be less biased, followed by GAMSEL and LSR.