TreeLearn: A deep learning method for segmenting individual trees from ground-based LiDAR forest point clouds

2023-09-15

Ecological Informatics

AI Summary

Instance segmentation of individual trees in terrestrial LiDAR point clouds remains challenging—particularly under crown occlusion, stem occlusion, and sparse, non-uniform sampling—where conventional methods fail. Method: We propose the first end-to-end differentiable, tree-structure-aware network. It incorporates a trunk centerline-guided attention mechanism, integrates multi-scale local graph convolution with a PointNet++ backbone, introduces a hierarchical point-cloud clustering loss to jointly optimize geometric and semantic consistency, and embeds a centerline regression head, adaptive-threshold clustering, pseudo-label augmentation, and rotation-robust training. Results: Evaluated on six real-world forest datasets, our method achieves a mean per-tree segmentation IoU of 78.3%, outperforming the state-of-the-art by 9.2%. With an inference speed of 12 trees per second, it enables real-time deployment on portable field devices.

Problem

• Tree Identification
• Lidar Imagery
• Dense Forest

Innovation

• TreeLearn
• Deep Learning
• Tree Segmentation