Reducing wind effects in multi-scan terrestrial laser scanning point clouds of trees combining LiDAR simulation and deep learning [data and source code]
Multi-scan terrestrial laser scanning (TLS) is the state-of-the-art-technique for precisely characterizing the 3D structure of trees. In the acquired point clouds, wind effects such as branch duplication can significantly impact data quality and can lead to errors in derived estimates such as wood v...
Enregistré dans:
| Auteurs principaux: | , , |
|---|---|
| Format: | Base de données Research Data |
| Langue: | anglais |
| Publié: |
Heidelberg
Universität
2026-08-10
|
| DOI: | 10.11588/DATA/JQNDRP |
| Sujets: | |
| Accès en ligne: | Verlag, kostenfrei, Volltext: https://doi.org/10.11588/DATA/JQNDRP Verlag, kostenfrei, Volltext: https://heidata.uni-heidelberg.de/dataset.xhtml?persistentId=doi:10.11588/DATA/JQNDRP 
|
| Notes sur l'auteur: | Hannah Weiser, Ronald Tabernig, Bernhard Höfle |
| Résumé: | Multi-scan terrestrial laser scanning (TLS) is the state-of-the-art-technique for precisely characterizing the 3D structure of trees. In the acquired point clouds, wind effects such as branch duplication can significantly impact data quality and can lead to errors in derived estimates such as wood volume. As it is difficult to fully avoid wind during data acquisition, those effects are omnipresent in TLS data, widely known but still an unsolved issue. A method that can reliably correct wind effects would be of enormous value to the scientific community. We present a wind correction approach that is based on local non-rigid registration of scans. For this, we compare one deep learning method with two geometric methods. Training and evaluation of these methods requires pointwise motion vectors and windless representations of the trees. Since this reference can hardly be obtained for real-world datasets, we employ virtual laser scanning of swaying trees to generate simulated multi-scan TLS data and error-free reference data. Our comprehensive assessment using motion vectors, point cloud distances and visualisations shows that wind correction successfully improves point cloud representation. Trained from scratch on domain-specific tree data, the deep learning method outperforms non-learning baselines in terms of non-rigid registration accuracy on simulated data and better preserves points in occluded areas. Applied to a small real-world tree dataset, both the deep learning method and the geometric methods achieve an increase in scan overlap ratio from 60% to over 79% and a decrease in chamfer distance between scans from 2.86 cm to under 1.72 cm, matching the values of a low-wind reference dataset. We believe that our approach presents a valuable pre-processing step to reduce errors in downstream applications such as quantitative structure modelling or change analysis without the need to discard wind-affected scans or to adapt algorithm settings. |
|---|---|
| Description: | Gefördert durch: Deutsche Forschungsgemeinschaft (DFG): 496418931; Deutsche Forschungsgemeinschaft (DFG): 528521476 Gesehen am 10.08.2026 |
| Description matérielle: | Online Resource |
| DOI: | 10.11588/DATA/JQNDRP |