Usage
  • 118 views
  • 174 downloads

The Estimation of Buttresses Volume and Classification of Leaf, Wood, and Lianas from Terrestrial Laser Scanning data

  • Author / Creator
    Han, Tao
  • Forests account for one third of land area, and two third of global photosynthesis. The better we know about forests, the better decisions we can make on forest management and carbon cycle modeling. During the last decades, we see the development in remote sensing techniques for forest monitoring. For example, satellite images are efficient methods to monitor forest changes. However, those satellite images present forests in two dimensions, while they do not support rapid and robust assessment of accurate ground reference data. By contrast, Terrestrial Laser Scanning (TLS) can retrieve the three-dimensional vegetation structure with millimeter accuracy.

    The main objectives of this PhD thesis are to: (1) present a new machine learning model to separate lianas and trees; and (2) develop a deep learning model to classify leaf and woody components; and (3) develop a non-destructive method to estimate buttress volume. Here, we use 3D point clouds collected by terrestrial laser scanning (TLS) to reach above objectives, where TLS can describe the 3D forest structure in millimeter-level details. This PhD research contributes to filling important knowledge gaps in contemporary scientific fields.

    Chapter 2 describes the utilization of a new machine learning model, based on Random Forest (RF) and eXtreme Gradient Boosting (XGBoosting), to separate lianas and trees using TLS point clouds. In this chapter, we find that the XGBoosting algorithm achieves an overall accuracy of 0.88, with a recall of 0.66, higher than the RF algorithm (accuracy of 0.85 and recall of 0.56). In addition, we find the optimal radius search method is as accurate as the multiple radius search method, with F1 scores of 0.49 and 0.48, respectively. We also find that the RF algorithm shows a recall of 0.88 on the independent data. We conclude that the model in chapter 2 provides a new flexible approach to extract lianas from 3D point clouds, enabling more studies to evaluate lianas impact on tree and forest structure.

    Chapter 3 explores the utilization of a deep learning time-series approach to classify leaf and woody components from TLS point clouds. We found that the multivariable time series (MTS) method (accuracy of 0.96) outperformed the univariable time series (UTS) method (accuracy of 0.67 to 0.88) to classify leaf and woody components. Meanwhile, Residual Network (ResNet) spent much more time than Fully Convolutional Neural Network (FCN) and Long Short-Term Memory Fully Convolutional Neural Network (LSTM-FCN) in model development, while those three networks demonstrated similar performance on an independent dataset. Furthermore, we found that the Class Activation Map (CAM) of the proposed model can explain the black-box effect of deep learning. We demonstrated that deep learning algorithms coupled with TLS point clouds could accurately separate leaf and woody components, providing a good start point for future research to estimate forest structure parameters such as Leaf Area Index (LAI) and Wood Area Index (WAI).

    A non-destructive method is developed to estimate buttress volume using 3D point clouds in chapter 4. In this chapter, we found that the alpha shape algorithm (ASA) and slice triangulation (ST) performed better than allometric models for buttress volume estimation. Moreover, the ASA tended to work better than ST when the trees presented more and shallower horizontal buttresses. Concerning the allometric models, Darea130 was the most accurate predictor to estimate buttress volume, with a lower Akaike information criterion (AIC, -66.25) than DAB (-59.55) and Dconvex130 (30.56). At the same time, the DAB (RRMSE of 0.23) and Darea130 (RRMSE of 0.21) showed similar performance on validation data. Our results indicate that the ASA can help to correct the bias in the present and past estimates of volume and biomass of large trees, which are keystone components to understanding biomass allocation and dynamics in tropical forests.

  • Subjects / Keywords
  • Graduation date
    Spring 2023
  • Type of Item
    Thesis
  • Degree
    Doctor of Philosophy
  • DOI
    https://doi.org/10.7939/r3-atwn-5h82
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.