Virtual Laser Scanning using HELIOS++ - Applications in Machine Learning and Forestry

Virtual Laser Scanning (VLS) provides a remote sensing method to generate 3D point clouds, which can, in certain cases, replace real data acquisition. A prerequisite is a suitable substitute of reality for modelling the 3D scene, the scanning system, the platform, the laser beam transmission, the beam-scene interaction, and the echo detection. The suitability of simulated laser scanning data largely depends on the application, and simulations that are more realistic come with stricter requirements on input data quality and higher computational costs. It is therefore important to have a good capability for corresponding trade-offs in the simulation software.

With the scientific software HELIOS++ [1], we provide an open source solution to acquire VLS data, where this trade-off can be tuned easily. HELIOS++ is implemented in C++ for optimized runtimes, and provides bindings in Python to allow integration into scripting environments (e.g. GIS plugins, Jupyter Notebooks).

The HELIOS++ VLS concept is based on a modular design. This allows the user to quickly exchange single simulation components, such as the scanner or the 3D scene. The simulation of diverging laser beams and the recording of full waveforms is supported via a subray tracing approach: depending on the desired physical realism and accuracy, a user-defined number of concentric circles approximate a single laser beam. On each circle, individual subrays are cast into the scene, which can then intersect with a single object and produce a hit. The returned waveforms are subsequently added together. This allows the simulator to detect multiple echoes for each pulse. The waveforms can be exported for further analysis.

In this contribution, we present main applications of HELIOS++ as a general-purpose LiDAR simulator. The first application is forestry, where green vegetation can be represented by different 3D model types. As the simulation of individual leaves as 3D mesh models requires high computational power, voxel-based methods have recently been proposed. HELIOS++ also supports simulation of semitransparent voxels, where a subray has a certain probability of creating a return when traversing. This probability depends on the incidence angle and the leaf angle distribution (e.g., planophile, erectophile …), the traversal length through the voxel, and the leaf area density of the voxel, which can, e.g., be derived from a terrestrial laser scanning point cloud. Tuning of the subray-parameters allows recreating vertical point density profiles of real surveys.

A second use case is the generation of training data for machine learning algorithms. Recently, several methods for Deep Learning on point clouds have been presented. However, such methods require immense amounts of training data to achieve acceptable performance. We present how VLS can be used to generate training data in machine learning classifiers, and how different sensor settings influence the classification results.

This contribution provides an introduction to VLS, possible use cases, pitfalls and best practices for successful application of laser scanning simulation.

[1] Winiwarter et al., 2021, DOI: 10.1016/j.rse.2021.112772