Automated classification of tree species using graph structure data and neural networks

Abstract

The classification of tree species in urban contexts is pivotal in assessing ecosystem services and fostering sustainable urban development. This paper explores using graph neural networks (GNNs) on graph structure data derived from quantitative structure models (QSMs) and tree structural measurement for appropriate species classification. The study addresses gaps in existing methods by integrating relationships between tree components, such as branches and cylinders, and considering the entire tree structure in a novel graph data format. The results demonstrate the efficacy of GNNs, particularly the Gated Graph Convolutional Network (GatedGCN), in appropriately classifying urban tree species. It gained an overall classification accuracy and weighted F1 score of 0.84. An analysis of confusion matrices revealed similarities in visual characteristics among several species, including A. platanoides and T. cordata, which pose significant challenges in accurately distinguishing between them. However, certain species, such as A. hippocastanum and P. nigra var. italica, have proved easier to classify than others. Furthermore, the results highlight the importance of relationships between different tree components in species recognition, such as the ratio between branch radius and parent branch radius, the factors often overlooked by previous methods. This underscores the novelty and effectiveness of the proposed approach in this study. Future research could explore integrating additional data sources, such as Leaf Area Density (LAD) calculated from LiDAR and hyperspectral data, to enhance classification accuracy. In conclusion, the evaluation results of the GatedGCN model demonstrated its effectiveness in classifying tree species using a novel data structure format derived from QSM tree characteristics. Advancing urban tree species classification through such methods can enhance future urban tree management using automated AI and robotics solutions.

Keywords

• Species classification
• Graph structure data
• Graph neural networks
• QSM
• LiDAR
• Urban tree