Individual Tree-Scale Aboveground Biomass Estimation of Woody Vegetation in a Semi-Arid Savanna Using 3D Data

Abstract

Allometric equations are the most common way of assessing Aboveground biomass (AGB) but few exist for savanna ecosystems. The need for the accurate estimation of AGB has triggered an increase in the amount of research towards the 3D quantification of tree architecture through Terrestrial Laser Scanning (TLS). Quantitative Structure Models (QSMs) of trees have been described as the most accurate way. However, the accuracy of using QSMs has yet to be established for the savanna. We implemented a non-destructive method based on TLS and QSMs. Leaf-off multi scan TLS point clouds were acquired in 2015 in Kruger National Park, South Africa using a Riegl VZ1000. The 3D data covered 80.8 ha with an average point density of 315.3 points/m2. Individual tree segmentation was applied using the comparative shortest-path algorithm, resulting in 1000 trees. As 31 trees failed to be reconstructed, we reconstructed optimized QSMs for 969 trees and the computed tree volume was converted to AGB using a wood density of 0.9. The TLS-derived AGB was compared with AGB from three allometric equations. The best modelling results had an RMSE of 348.75 kg (mean = 416.4 kg) and a Concordance Correlation Coefficient (CCC) of 0.91. Optimized QSMs and model repetition gave robust estimates as given by the low coefficient of variation (CoV = 19.9% to 27.5%). The limitations of allometric equations can be addressed by the application of QSMs on high-density TLS data. Our study shows that the AGB of savanna vegetation can be modelled using QSMs and TLS point clouds. The results of this study are key in understanding savanna ecology, given its complex and dynamic nature.

PointMM: Point Cloud Semantic Segmentation CNN under Multi-Spatial Feature Encoding and Multi-Head Attention Pooling

“…Subsequently, the features η i are input into the multi-head attention pooling module, which integrates neighborhood features through pooling operations to generate a larger receptive field and more global feature vector MP(F i ). It is worth noting that we set up four downsampling layers, so the number of attention heads for each layer is 2 n (n ∈ [1,4]). The initial input to the downsampling layer in this paper is a point cloud of dimensions N × K × D, and the number of sampled points in each subsequent layer is multiplied by 4 −n (n ∈ [1,4]), where n represents the downsampling layer.…”

“…It is worth noting that we set up four downsampling layers, so the number of attention heads for each layer is 2 n (n ∈ [1,4]). The initial input to the downsampling layer in this paper is a point cloud of dimensions N × K × D, and the number of sampled points in each subsequent layer is multiplied by 4 −n (n ∈ [1,4]), where n represents the downsampling layer. Additionally, the output of the downsampling layer is feature maps of dimensions N/4 × 64, N/16 × 128, N/64 × 256, and N/256 × 512.…”

“…Compared to 2D images, three-dimensional point clouds obtained using 3D scanners and depth sensors (such as LiDAR and RGB-D cameras) can more comprehensively and intuitively express the spatial relationships between various targets in the scene. They have been widely utilized in various industries, including 3D modeling [1], autonomous driving [2], and metaverse [3], and natural resource surveys [4]. Point cloud semantic segmentation is a crucial supporting technology for understanding and analyzing 3D scenes [5].…”