A comparative study of voxel-based leaf area density estimation from quantitative structure models of trees

This study explores the potential of using Quantitative Structure Models (QSM) to predict trees' voxel-based Leaf Area Density (LAD) to reduce the workload and data redundancy in studying deciduous trees. For this purpose, leaf-on and leaf-off Terrestrial Laser Scanning (TLS) of 16 Platanus x hispanica trees on streets were utilized. QSMs were extracted and interpreted into QSM indexes corresponding to voxels, a novel approach introduced in this study. Twelve standard regression models were tested to predict the LAD value for each voxel using its QSM indexes. The Hist Gradient Boosting Regressor (HGBR) model demonstrated the best performance, with an R-squared score of 0.56 and a mean absolute error of 0.0187 m²/m³ (16.33 %) in the LAD prediction. This deviation mainly happened at the crown center, where branches were dense while leaves were few. The trained model was also applied to another set of 13 young plane trees of different tree sizes at a nursery. Their predicted Leaf Area Index (LAI) was compared to the LAI measured indirectly by hemispherical photography, showing a deviation of 0.12 m²/m² (8.6 %) for the 3 largest trees with the closest Diameter at Breast Height (DBH) to the street trees. The deviations are larger for young nursery trees with smaller DBHs. Therefore, further experiments are needed to optimize the voxel size and adapt the model to different species with varying crown sizes.