Monitoring Above Ground Biomass of Managed Forests via Semantic and Instance Segmentation of UAV-LiDAR Point Clouds Using Deep Learning Models.
Abstract
With rising demand for managed forests with high-yielding biomass production and high capacity to absorb CO2, accurate quantification of their above ground biomass (AGB) is indispensable to document sustainable resource management. Unmanned Aerial Vehicles (UAVs) and airborne Light and Detection Ranging (LiDAR) scanners can constrain the uncertainty inherent to estimate AGB by producing point clouds (PCs) that represent the 3D structure of forests. While the processing of PCs has proven to be challenging due to the non-uniform sampling density of the data, the evolution of Deep Learning models (DLs) that can handle this irregularity paves the way for assigning semantic information and recognizing 3D structures in PCs describing forests. In this study, we evaluate the effectiveness of DLs to: i) segment plantation forests (into terrain, low and high vegetation) scanned by a UAV-LiDAR system and ii) detect individual conifer or broadleaf trees to retrieve their geometric characteristics for estimating the respective AGB. The aerial campaign was conducted in six experimental sites in Denmark where ground truth data of trees positions and dimensions are used for validating the accuracy of the DLs. The forest types consisted of: Grand fir/ Norway spruce (homogeneous or admixed with other conifers), beech-oak and beech-Douglas fir forests. The instance segmentation for extracting individual trees is assessed using a data-driven point cloud up-sampling network that learns multilevel features per point from the sparse geometry of the training data, and generates denser PCs that may describe the underlying geometry of a tree. For the semantic segmentation, we explore the performance of hierarchical feature learning operating directly on PCs and of the O-CNN and Octree Generating networks that convert PCs to volumetric octree grids. The DLs are trained using manually labelled PCs, publicly available labels, and evaluated based on the resulting distribution of local uniformity. A comparison between the in situ observations of AGB among different forest types and the respective predictions using DLs highlights the limitations and the potential of DLs to be utilized as a scientific tool for monitoring AGB without the need of regressing field-based measured AGB with LiDAR-derived height metrics.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.B51B..03T