Modeling pine forest growing stock volume in subtropical regions of China using airborne Lidar data

Pine forests, particularly Masson pine (Pinus massoniana), are widely distributed across the subtropical regions of China. Understanding the spatial distribution of pine forest growing stock volume (GSV) is essential for effective management and planning of forest resources. However, accurately estimating pine forest GSV over large areas using airborne Lidar is challenging due to the complex stand structure and the influence of environmental conditions. This research employed airborne Lidar data and sample plots from 11 typical sites across the northern, central, and southern subtropical regions of China to explore modeling approaches for pine forest GSV estimation. Ordinary Linear Regression (OLR), Geographically Weighted Regression (GWR), and Hierarchical Bayesian Approach (HBA) were employed to model GSV through comparative analysis of using various sample sizes. The results indicate that: (1) HBA(Site), which models different pine forest types (i.e. pure pine forest (PPF) and mixed pine forest (MXF)) separately, with typical site as a stratification factor, provided the best estimation results with coefficient of determination (R2) of 0.80 and 0.74, root mean square error (RMSE) of 25.15 m3/ha and 23.86 m3/ha for PPF and MXF, respectively. When combining PPF and MXF together to develop GSV models, HBA(Site/Type) with double stratification factors (i.e. typical site and forest type) showed similar performance, with only a minor reduction in RMSE by 0.45 m3/ha for PPF and 0.15 m3/ha for MXF. This modeling approach effectively alleviated the overfitting problem caused by relatively small sample sizes. (2) Compared to OLR which only used global variables, both GWR which incorporated spatial information and HBA which utilized stratification significantly improved modeling performance. (3) The accuracy of OLR and GWR models remained relatively stable with various sample sizes, indicating their low sensitivity to sample sizes, whereas HBA exhibited high sensitivity due to the influence of stratification factors. (4) For a single tree species, a stratification-based modeling method was valuable for GSV estimation. This study provided a quantitative analysis and accurate estimation of pine forest GSV over large areas with different environmental conditions and offered new insights for GSV estimation of other forest types. More research is needed to quantitatively examine different contribution of sample sizes, modeling algorithms, variables from different sources, and stratification factors on modeling results, so that we can design an optimal procedure for GSV modeling using airborne Lidar data.