Speaker
Description
Reliable forest stand delineation is a fundamental prerequisite for forest management, harvesting system allocation, and spatial planning. Conventional stand boundaries are typically derived from visual interpretation in the field or from aerial imagery, making them time-consuming, experience-dependent, and only partially reproducible. In response to increasing disturbance dynamics and the demand for consistent spatial planning units, this study develops and evaluates a reproducible, data-driven approach for automated stand classification and segmentation based on airborne 3D data.
The methodological framework integrates airborne laser scanning (ALS), UAV-based laser scanning (ULS), and image matching from aerial imagery as alternative data sources. The workflow was tested in two independent Austrian study areas where all three airborne systems were acquired within a narrow temporal window: a 130-hectare forest management unit in Styria and a larger test area exceeding 1.000 hectares in Tyrol. Structural and site-related predictors were aggregated on hexagonal grid cells (314 m²) derived from point clouds and digital terrain models. These included vegetation height quantiles, mean and median canopy height, height variability, elevation, slope, aspect, and distances to forest roads and watercourses.
Forest development stages were classified using gradient boosting models trained on independent inventory data. Model performance was assessed via k-fold cross-validation. In the Styrian test area, strict class agreement reached 56.3% (Cohen’s kappa = 0.49). Allowing a deviation of one adjacent development stage to account for structural transition zones increased overall accuracy to 82.3%. Moreover, the automated classification resulted in a 22.7% increase in within-class structural homogeneity compared to conventional stand delineation by forest experts, measured as a reduction in intra-class variance of canopy height metrics. Important predictors were the 75th percentile of vegetation height, median vegetation height, and elevation-related variables.
Pixel-level predictions were subsequently aggregated into operationally usable compartments using spatially constrained hierarchical clustering and graph-based region-growing algorithms. Small and isolated patches were reassigned to neighboring units, resulting in coherent and practically manageable stand polygons. The standardized workflow enabled direct comparison of data-source effects. ALS ensured homogeneous coverage over large areas, ULS provided higher structural detail at stand scale, and image matching represented a cost-efficient alternative but showed limitations in canopy penetration and terrain representation, influencing predictor stability and class separability.
The resulting digital stand maps provide a standardized spatial reference for deriving forest metrics such as growing stock, stem density, and diameter distributions. These compartments support harvest planning and facilitate rapid map updates following disturbance events. Overall, the integration of airborne 3D sensing and machine-learning-based segmentation provides a quantitative and scalable foundation for data-driven forest management and operational planning.
| Keywords | "LaDiWaldi"; "Remote-sensing"; "Machine-learning"; "forest-inventory" |
|---|
