Deep Learning-Based Part Labeling of Tree Components in Point Cloud Data

Point cloud data analysis plays a crucial role in forest management, remote sensing, and wildfire monitoring and mitigation, necessitating robust computer algorithms and pipelines for segmentation and labeling of tree components. This thesis presents a novel pipeline that employs deep learning models, such as the Point-Voxel Transformer (PVT), and synthetic tree point clouds for automatic tree part-segmentation. The pipeline leverages the expertise of environmental artists to enhance the quality and diversity of training data and investigates alternative subsampling methods to optimize model performance. Furthermore, we evaluate various label propagation techniques to improve the labeling of synthetic tree point clouds. By comparing different community detection methods and graph connectivity inference techniques, we demonstrate that K-NN connectivity inference and carefully selected community detection methods significantly enhance labeling accuracy, efficiency, and coverage. The proposed methods hold the potential to improve the quality of forest management and monitoring applications, enable better assessment of wildfire hazards, and facilitate advancements in remote sensing and forestry fields.

Level thesis:degree_level

Master's Degree

Year dc:date.issued

2022

Author dc:creator

• Stone, Gunner

Advisors dc:contributor.advisor

• Tavakkoli, Alireza
• Greenberg, Jonathan

Committee members dc:contributor.committeemember

• Bebis, George
• Nicolescu, Mircea
• Hand, Emily

• Clustering
• Deep Learning
• PointCloud
• Segmentation
• SpeedTree
• Synthetic

Handle dc:identifier.uri

http://hdl.handle.net/11714/9981

OAI identifier oai:identifier

oai:scholarwolf.unr.edu:11714/9981