BeschreibungForest stratification is a powerful tool to determine the vitality of forest stands with implicit forecasting support. It enables the extraction of vital information such as biomass, forest physiology and rejuvenation capacity [1-4]. Given their importance, in situ stratification studies are performed despite their complex nature and high intrinsic costs (time, people and equipment). These studies entail per-tree manual parameter acquisition, such as crown height and width, tree height and trunk radius ; and statistical models . In contrast, airborne laser scans (ALS) and unmanned aerial vehicle laser scans (ULS) can generate high-density wide-area scans but lack annotations crucial to derive information and train neural networks. To contribute towards automating the classification of stratification, we: 1) propose a generative forestry scene pipeline that merges real and synthetic tree objects into ALS/ULS-like point clouds and provides their per-point annotation; 2) study this data usage for training a point cloud segmentation neural network for stratification [7,8]; 3) measure the model performance on real data to assess its viability as a stratification tool. The pipeline exploits synthetic trees and vegetation descriptor values like height and radius by randomly permuting them within defined limits, subsequently generating an arbitrary number of parametrised mesh objects. Besides, hand annotated real ULS and terrestrial laser scans are captured for training and validation. Synthetic and real objects are then randomly placed in a scene to simulate ULS data acquisition . The resulting point cloud is annotated using the known relation between the object parameters and its point cloud representation.
Although the pipeline can create arbitrary trees, this contribution focuses on alpine forest stands with various species, such as larch, spruce and beech. Results verification is conducted using real ULS and terrestrial laser scans collected in Ebensee, Upper Austria. Moreover, we explore potential benefits of mixing real with synthetic annotations for training.
|6 Sept. 2023
|London, Großbritannien/Vereinigtes Königreich
- Assistive and Autonomous Systems