Pointwise deep learning for leaf-wood segmentation of tropical tree point clouds from terrestrial laser scanning

Abstract

Terrestrial laser scanning (TLS) is increasingly used in forest monitoring, providing detailed 3D measurements of vegetation structure. However, the semantic understanding of tropical tree point clouds, particularly the segmentation of leaf and wood components, remains a challenge. Deep learning (DL) on point clouds has been gaining traction as a valuable tool for automated leaf-wood segmentation, but its widespread adoption is impeded by data availability, a lack of open-source trained models, and knowledge on its influence on subsequent woody volume reconstruction. To address these issues, this paper makes three key contributions. First, it introduces a new dataset consisting of 148 tropical tree TLS point clouds from north-eastern Australia with manual leaf-wood annotations. Second, it uses this dataset to compare several state-of-the-art point-wise DL networks and benchmark these against traditional approaches, using a common training and inference pipeline to allow for a fair model comparison. We conduct an ablation study to examine the effects of various hyperparameters and modelling choices, focusing solely on point coordinates as input to develop a model adaptable to different forest types, platforms, and point cloud qualities. Third, we assess the impact of point-wise segmentation quality on tropical tree volume estimation using quantitative structure model (QSM) reconstruction on the extracted woody component. Results show that our newly trained DL models significantly outperform traditional benchmarks for leaf-wood segmentation of tropical tree point clouds from TLS, with PointTransformer achieving the highest performance (mIoU = 92.2 %). Quantitative and qualitative analyses reveal that DL methods excel in distinguishing woody points, crucial for woody volume estimation via QSMs, but may suffer from connectivity issues due to lack of physical constraints. Volumes of trees segmented using PointTransformer closely match those of manually segmented trees (MAE = 7.1 %), highlighting its suitability for automated woody volume estimation. Although this study demonstrates the effectiveness of state-of-the-art neural architectures for tropical tree point cloud processing, advocating for their integration into forest structure analysis pipelines, future work should focus on enhancing quantity, quality and variety of training data, to increase model robustness and generalisability. We make the dataset, code and trained DL models publicly available.¹