Abstract

This study investigates the integration of UAV LiDAR-derived individual tree crown (ITC) metrics and distance-dependent competition indices (CI) as input to machine learning (ML) models—random forest (RF) and support vector machines (SVM)—for predicting the individual tree level yield of Pinus taeda (L.) plantations after 4-years. We hypothesized that RF and SVM would outperform multiple linear regression (MLR) in prediction accuracy and that certain ITC and CI variables would disproportionately impact yield predictions. Analysis of variance (ANOVA) shows a significant impact of planting density across the models. SVM achieved the highest individual tree-level yield accuracy (normalized RMSE (nRMSE): 9.59 %, R²: 0.59), followed by RF (nRMSE: 10.86 %, R²: 0.48). When the individual tree-level yields were aggregated at the stand-level, SVM underpredicted total stem volume by −1.50 %, while RF overpredicted by 1.53 %. Using the top seven predictors, the reduced ML models maintained similar accuracy (RFreduced–nRMSE: 11.88 %, R2: 0.37; SVMreduced–nRMSE: 9.14 %, R2: 0.55). The SVMreduced model underpredicted stand volume by 0.90 %, while the RFreduced overpredicted by 0.71 % at the stand level. The study also evaluated a best subset MLR model (nRMSE: 12.04 %, R2: 0.44), however, the model failed to meet the assumptions of homoscedasticity, primarily due to the presence of trees that died over the four-year prediction interval. These findings show LiDAR's potential for accurate tree- and stand-level yield predictions, with broader uses in yield, carbon, and biomass modeling. © 2025 Elsevier B.V.