University of British Columbia
Enhancing aboveground biomass component estimation using airborne laser scanning and deep learning across Canadian forests
Abstract
dc:descriptionForest biomass accounts for a large proportion of Earth’s total biomass and comprises 50 - 65% of terrestrial carbon. Accurate estimation of forest aboveground biomass (AGB) and its components (wood, bark, branch, and foliage) is essential for forest management and carbon accounting. Currently, one of the most accurate approaches to map forest AGB is using airborne laser scanning (ALS) as it provides a detailed 3D point cloud representation of forest structure. However, current statistical and machine learning (ML) methods used to predict forest AGB rely on summarizing point cloud data as statistical metrics across a regular grid, thereby losing a significant amount of horizontal and vertical spatial context. This thesis investigates the application of deep neural networks (DNNs) to extract information directly from ALS point cloud data to enhance total and component forest AGB estimation. Four research gaps regarding total and component AGB estimation are explored: 1) comparing DNN architectures with conventional ML; 2) developing a data fusion DNN to synthesize ALS, satellite multispectral and topographic data; 3) employing self-supervised learning (SSL) to boost DNN performance in limited data situations; and 4) evaluating transfer learning (TL) across different forest ecosystems. Results demonstrate that DNNs can provide improved performance compared to random forest for total and component AGB estimation, but this varies by DNN architecture and biomass component, with the octree convolutional neural network performing best (R2 = 0.78). The data fusion DNN combining ALS with multispectral and topographic data yielded minor improvements for total and component AGB estimation (1 – 2% improvement in root mean squared error; RMSE). The use of SSL provided benefits for DNNs trained on smaller datasets (2 – 6% RMSE improvement). DNNs trained using TL generalized effectively to small datasets with different forest ecosystems and ALS characteristics (up to 29% RMSE improvement). This dissertation concludes that point cloud DNNs can enhance predictive performance for total and component AGB compared to conventional ML, and demonstrates that when appropriate training strategies such as SSL and TL are implemented, larger improvements are achievable. Overall, forest inventory improvements offered by DNNs must be weighed against trade-offs such as model explainability and computational demands.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Forestry
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Seely, Harrison
- Contributors dc:contributor
-
- Coops, Nicholas C.
Rights
dc:rights- Statement dc:rights
-
- Attribution-ShareAlike 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/94938