Comparing voxelisation methods of 3D terrestrial laser scanning with Radiative Transfer simulation to assess vegetation density

Terrestrial Laser Scanning (TLS) measurements are increasingly used to characterize forest structure, because of their high potential to provide accurate information on some key structural features that are hard to measure in the field, e.g. tree height, crown dimensions, stem shape or the 3D distribution of vegetation material. The later is of great interest as it plays a major role in many ecological processes and influences both productivity and biodiversity. Used within radiative transfer models, this feature is also useful to improve our understanding on how the signal interacts with vegetation in remote sensing studies, thereby enhancing our ability to analyze remote sensing data for forest ecosystem monitoring purposes. However to be used either in ecological models or in remote sensing studies, very dense point clouds acquired using TLS must be processed and synthesized in order to become usable by foresters or in models. To that aim many approaches have been developed. Some of them seek to identify and characterize the several parts of the trees, e.g. the stems, the main branches, the crowns... Others, like voxel-based approaches, provide information for spatial units irrespective of the trees, e.g. voxels or plots. This presentation focus on voxel-based approaches used to estimate the distribution of vegetation density material in a 3D grid (also referred as voxelization ) from single or multi echo TLS data. Different methods have been proposed (Hosoi and Omasa, 2006; Beland et al., 2014; Durrieu et al., 2008; Beland et al., 2011), but further studies are needed to validate these approaches and better characterize their limits and their sensitivity to instrument settings, vegetation characteristics and voxel geometric features (size, geometry) or other methodological choices made to compute vegetation density. Using a modeling approach to that aim is highly beneficial because it allows testing many configurations and does not require, at least for a theoretical validation, to acquire reference data on the actual 3D distribution of the vegetation from field surveys, which is highly challenging.\nThe objective of this study was threefold: (1) to develop a simulation framework to simulate TLS data based on DART (Discrete Anisotropic Radiative Transfer) model (Gastellu-Etchegorry et al.,2004), whose capabilities make it highly suitable for the purpose of this study, (2) to propose an improved voxelisation approach suitable for processing multi-echoes TLS data sets, and (3) to validate this approach and propose a series of guidelines to retrieve, from TLS data of a forest stand, 3D vegetation density and LAI into a voxelized space. To achieve this last objective a sensitivity analysis was performed to evaluate the impact of several parameters on voxelisation results. Firstly, three voxel attributes were analysed, namely their shape (cubic versus spherical), dimensions and sampling rate. Secondly, instrumental parameters were analysed to determine suitable scanner angular resolution and the benefit of processing multiple TLS returns. Thirdly, the influence of two vegetation attributes, size and angular distribution of leaves, were evaluated. Lastly, the resulting guidelines were applied to a more practical case involving real trees to provide a reliable assessment of the accuracy of vegetation densities that can be achieved from voxelisation approaches. Results show a general good agreement (r2 > 0.8 for realistic trees) with multi echo management, while the use of single returns only leads to an overestimation of leaf density. Theoretical cases show that cubic voxels gives best results overall, with RMSE increase from 0.05 to 0.3 with incerasing leaves size or voxel dimensions when there is no clumping effect inside the voxels, and a good voxel sampling.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.