Speaker
Description
Bark stripping is a major disturbance factor in Central European forests and leads to reduced growth, increased susceptibility to fungal infection, and elevated mortality risk. In operational forest management, severely damaged trees often need to be removed in order to avoid economic losses and secondary infections. Reliable and objective detection of bark stripping is therefore highly relevant for precision forestry and technology supported harvesting decisions. This study investigates whether bark stripping can be detected automatically using Light Detection and Ranging (LiDAR) data and Convolutional Neural Networks (CNN) and evaluates the potential integration of such an approach into LiDAR based forest inventory workflows.
Data acquisition was carried out in Styria in Austria on 37 sample plots with and without visible bark stripping. In total, 71 trees with bark damage were recorded across these plots, containing 172 individual bark stripping wounds. Three laser scanning systems were used, including a terrestrial laser scanner (TLS) Riegl VZ600i, a mobile laser scanner XGrid Lixel L2 Pro (MLS), and a consumer grade device Apple iPad Pro. The TLS point clouds provided intensity, reflectance, amplitude, and RGB attributes for each measured point, whereas the MLS point clouds delivered RGB and intensity attributes, and the tablet based system RGB attributes only.
Tree stems were unwrapped from the point clouds to generate two dimensional stem representations. Bark wounds were manually delineated and served as reference data. From these labeled stem surfaces, 100 × 100 pixel image patches with a spatial resolution of 2.5 mm per pixel were extracted to generate training and validation datasets. Initial model development focused exclusively on TLS data, as this dataset provided the most complete geometric and radiometric information. A pre trained ResNet 18 architecture was fine tuned for binary classification of damaged versus undamaged stem patches.
First experiments using grayscale patches derived from TLS point clouds achieved an overall classification accuracy of 76 percent on the independent test dataset. Precision exceeded recall, indicating a conservative model behavior with a tendency to classify uncertain patches as undamaged. The resulting F1 score reflected this imbalance between omission and commission errors. Ongoing work investigates multi channel input configurations combining RGB and LiDAR derived features as well as larger spatial contexts in order to increase detection sensitivity.
The presented workflow demonstrates the feasibility of automated bark stripping detection from LiDAR data and provides a foundation for integrating damage detection into digital forest inventory and decision support systems. Such an approach can support automated timber harvesting in identifying trees that require priority removal and contributes to the advancement of precision technology and remote sensing in forest operations.
| Keywords | LaDiWaldi; LiDAR; Inventory; CNN |
|---|
