Reliable estimation of tree-level diameter at breast height (DBH) and stem volume from remote sensing data remains challenging across structurally heterogeneous plantation forests due to cross-site domain shift. This study proposes a structurally site-aware modelling framework designed to mitigate site-induced errors by prioritising training samples structurally proximate to the target site in predictor space. Using unmanned aerial vehicle-based laser scanning (ULS)-derived metrics from 20 geographically independent radiata pine plantation sites in New Zealand, we compared standard pooled workflows with site-aware implementations across multiple feature selection and regression combinations under leave-one-site-out (LOSO) validation. For DBH, the optimal site-aware Elastic Net configuration achieved a mean rRMSE of 16.0% and coefficient of determination (R2) of 0.607, reducing relative error by up to 23.7% compared with the corresponding standard workflow. Gains were more pronounced for stem volume, where the site-aware model achieved a mean rRMSE of 34.5% and R2 of 0.648, substantially reducing cross-site errors observed under standard parametric formulations by 85.2% and outperforming a previously published high-dimensional Random Forest benchmark built on the same dataset (mean rRMSE of 35.6% and R2 of 0.631). Feature selection patterns revealed that standard workflows converged on a narrow set of universally dominant structural predictors, whereas the site-aware approach redistributed predictor importance across sites, reflecting adaptive alignment to local structural variations. These findings demonstrate that correcting structural domain misalignment can enhance model transferability while maintaining parsimony, offering a scalable solution for operational multi-site forest inventory modelling.
Felix et al. (Mon,) studied this question.