Urban Tree Classification from Multispectral Airborne LiDAR Using PointNet, DGCNN & RandLA-Net

Abstract. Intelligent management of urban trees is a key issue for Smart Cities, contributing to environmental sustainability and urban well-being. Geospatial technologies and artificial intelligence are increasingly being integrated into smart cites to improve resource management and urban planning. This study provides an in-depth comparison of three deep learning methods: PointNet, DGCNN RandLA-Net, applied to classification of seven urban trees species (Pine, Spruce, Birch, Maple, Aspen, Rowan, Linden) from the open-source Airborne Multispectral LiDAR dataset (MS-ALS-SPECIES). Each model was trained on a common training set, validated during training and evaluated on a separate test set, allowing a systematic evaluation of their classification performance. The comparison focuses on overall accuracy, F1-score, mean per-class accuracy and recall. The results demonstrate that PointNet achieves the best overall accuracy of test dataset of 82.07% and a mean per-class accuracy of 70.32%, with competitive performance on Pine (94.37%), Spruce (84.25%), and Maple (86.67%). DGCNN improves the capture of local structures, with 79.15% accuracy in validation, and 67.98% in testing, reflecting slight overfitting. RandLA-Net, although less accurate overall (56.03%), achieves the best inter-species homogeneity (62.32%), and high recall on minority species (Aspen: 86.36%, Linden 85.71%). These results demonstrate the potential of 3D deep learning combine with multispectral airborne lidar for automated urban tree species classification and their integration into the geospatial systems of smart cities for intelligent management of green spaces.