Scalable Deep Learning for Individual Tree Species Classification from Cross-Platform LiDAR Point Clouds

Abstract

Accurate individual tree species classification from point cloud data are essential tasks with significant implications for forest inventory, biomass estimation, and carbon monitoring. Recent advancements in Light Detection and Ranging (LiDAR) technologies, such as airborne LiDAR, Unmanned Aerial Vehicle (UAV)-based LiDAR, and handheld mobile LiDAR, provide rich data sources for these applications. Additionally, the rapid rise of deep learning techniques has demonstrated considerable potential for enhancing the accuracy and efficiency of data interpretation tasks. However, effectively leveraging deep learning to utilize diverse LiDAR datasets for individual tree segmentation and species classification remains challenging. However, deep learning methods typically require extensive annotated data, posing a critical challenge in forestry applications, where labelling individual trees in point clouds is particularly difficult, unlike the abundant large-scale datasets available in image-based domains. This issue leads to three primary research questions: Firstly, how can individual tree segmentation be achieved more efficiently and reliably in complex forest environments characterized by crown overlap, occlusion, and varying point cloud densities? Secondly, can deep learning models effectively classify individual tree species using low-density LiDAR point clouds? Lastly, how can we enhance cross-platform generalization and enable efficient adaptation to newly introduced tree species with limited labelled samples?

To address these questions, this thesis presents three main contributions. First, it develops an interactive deep learning pipeline for individual tree segmentation from cross platform laser-scanning data. By integrating user-guided prompts, such as bounding boxes or point clicks, with point cloud-based instance segmentation, the pipeline produces accurate and flexible tree delineations while reducing manual delineation time, thereby supporting the efficient preparation of tree-level training samples. Second, it proposes the Attribute-Aware Cross-Branch Transformer, tailored for tree species classification from low density LiDAR data. The model jointly exploits geometric and radiometric attributes and is designed to learn discriminative, species-specific features under sparse, and uneven point cloud conditions. Third, it investigates a transfer learning framework to enhance the generalization of tree species classification models across heterogeneous LiDAR platforms and to support rapid fine tuning for unseen species. By pretraining on multi platform datasets and adapting to new domains with limited labels, the framework improves scalability and data efficiency. These contributions provide a unified and scalable framework that integrates interactive segmentation, tree species classification, and transfer learning for forestry LiDAR applications, enabling automated, data-efficient forest monitoring in support of operational inventory and carbon-related assessment.