Forest volume estimation in Eucalyptus globulus stands using LiDAR data from mainland Portugal

The rigorous characterization of forest resources is essential to support the sustainable management of forest stands and the assessment of their productive potential. In this context, airborne LiDAR data have been increasingly recognized as a valuable source of information, enabling the characterization of the three-dimensional structure of vegetation and complementing field forest inventory data. This work aimed to develop and evaluate a methodology integrating forest inventory data and open-access airborne LiDAR data provided by the Portuguese Directorate-General for Territory (DGT), acquired in 2024 with a nominal point density of 10 points m⁻², for the estimation of total stem volume with bark in eucalyptus plantations, following the area-based approach framework. The study included 461 forest inventory plots measured during the 2023–2024 inventory campaigns. Plot-level structural metrics, including canopy height metrics (e.g., height percentiles), intensity, return number, and three-dimensional shape metrics, were extracted from the normalized LiDAR point cloud. Based on their relevance and previous findings in the literature, a subset of these metrics was selected and used to fit and compare two ensemble-based machine learning models: Random Forest and XGBoost. Both models showed relevant predictive capacity. XGBoost achieved better performance on the test set, with a coefficient of determination of 0.671 and a root mean squared error of 36.46 m³ ha⁻¹, while Random Forest obtained R² = 0.616 and RMSE of 39.39 m³ ha⁻¹, showing greater stability across folds in cross-validation. These results are consistent with values reported in the literature for moderate-density airborne LiDAR data and reflect the structural variability of the studied stands. Prediction uncertainty was quantified using conformal calibration, with empirical coverage above the nominal 95% level for both models. Spatial application of the models enabled the generation of volume maps at 20 × 20 m resolution, showing spatially coherent patterns and positive spatial autocorrelation, as indicated by Moran's I. The results highlight the potential of integrating forest inventory, open-access LiDAR data and machine learning models to support spatially explicit and reproducible forest volume estimation, contributing to the incorporation of remote sensing into operational forest monitoring and decision-support systems. Nevertheless, the interpretation of the results should consider limitations related to the spatial distribution of plots, temporal mismatch between datasets, and the absence of independent spatial validation.