Objective Airborne LiDAR has become a core technology for large-scale estimation of tree- and stand-level attributes due to its high spatial resolution. Incorporating LiDAR-derived variables into mixed-effects models can further enhance the accuracy of individual-tree diameter at breast height (DBH) estimation.

Method The study was conducted across seven counties/cities in Guangdong Province—Lechang, Yingde, Heping, Lianshan, Longmen, Shixing and Yunan—where 133 pure Chinese fir (Cunninghamia lanceolata) plantation plots (30 m × 30 m) were established. Field-measured tree measurements were integrated with synchronously acquired airborne LiDAR point clouds to extract crown projection area (CA), height (LH) and stand canopy density (SCD) as predictors. A basic linear model using CA alone was first constructed, followed by a two-level nested mixed-effects model with region and plot treated as random effects. Model validation was performed via leave-one-plot-out cross-validation, and the impacts of four sampling strategies (random, median, largest and smallest tree) on prediction accuracy were compared.

Result Among crown-shape variables, CA was highly correlated with crown diameter (CD) and crown volume (CV). Nevertheless, the linear model using CA alone provided the best fit and was selected as the base model. Incorporating random effects for region and plot markedly improved DBH estimation. The optimal model specification included random effects on parameters a, b and c, resulting in a 13.15% reduction in root-mean-square error (RMSE) compared with the base model. Cross-validation results indicated that both RMSE and total relative error (TRE) decreased as the number of sample trees used for local calibration increased. The random sampling strategy yielded the best overall performance.

Conclusion The developed two-level nested mixed-effects model provides and effective approach for estimating individual-tree DBH in Chinese fir plantations across southern China. Considering the trade-off between survey cost and prediction accuracy, randomly selecting three sample trees per plot is recommended for estimating random effects in operational inventories.