A deep learning time-series approach for liana and tree classification from terrestrial LiDAR point clouds

Lianas are woody climber plants with a relatively thicker stem that use trees as structural support to reach the forest canopy. They compete with trees for above and below ground resources, and they also occupy gaps and illuminated areas on the upper part of the canopy more efficiently. Therefore, their increases can suppress tree generation, promote tree mortality, and decrease tree growth, thereby affecting the whole forest carbon sequestration. Terrestrial Laser Scanning (TLS), also named terrestrial Light Detection and Ranging (LiDAR) is an active remote sensing method than can generate clouds with million points to describe three-dimensional structure of vegetations in high accuracy. Also, recent advances in deep learning algorithms provides us an excellent oppotunity to understand forest structure, since the performance of deep learning algorithm is always better than traditional machine learning algorithm, when it comes to data-intensive programs. This project aims to explore the utilization of deep learning for liana and tree classification based on TLS data. We want to develop an automatic deep learning algorithm to separate liana from its host tree, with high accuracy. The presented approach can facilitate more studies to investigate the impact of lianas on the structure and dynamics of forests.

Faculty Supervisor:

Gerardo Arturo Sanchez-Azofeifa

Student:

Partner:

Ghent University

Discipline:

Earth science

Sector:

Education

University:

University of Alberta

Program: