Deep learning algorithms to image the tree interiors via scans on a straight trajectory

• Main Author: Bui, Quang Huy
• Other Authors: Abdulkadir C. Yucel
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Earth and Environmental Sciences; Engineering; Deep learning; Straight trajectory scans; Tree interiors imaging; Image-to-image translation; Model optimization; Dataset development
• Online Access: https://hdl.handle.net/10356/177081
• Institution: Nanyang Technological University

This project introduces a deep learning approach to reconstruct images of tree interiors via a straight trajectory scan. The radar captures signals reflected from the tree trunk while it moves along a linear path at a distance from the trunk's surface. These signals, represented as B-scans, are processed, and employed as the datasets for a modified U-Net model, with the ground truth as tree cross sectional image. The goal of the project is to reconstruct the image of the tree cross sectional image to be as close as possible to the ground truth, so that it can be applied to detect trees that is on the verge of falling in real life, therefore avoiding accidents. After experimenting with different model training approaches, loss functions and hyperparameter combinations, a very high accuracy has been achieved in imaging the tree interiors. The project hence holds significant practical potential in the future. By providing accurate and efficient tree assessment method, the project can assist the authorities in creating effective forest management and conservation strategies, hence minimizing the overall number of tree fall accidents.