Virtual Laser Scanning of Dynamic Scenes (VLS-4D): A Novel Opportunity for Advancing 3D Forest Monitoring

Virtual laser scanning (VLS) [1] has been used intensively for method development and machine learning in forestry, e.g., for quantification of leaf angle distribution [2], aboveground biomass prediction [3], or leaf-wood segmentation [4]. So far, these applications have been limited to mono-temporal VLS acquisitions where scenes were simplified to being completely static. However, forests are inherently dynamic environments with processes occurring at different timescales and rhythms, such as wind-induced movement, response to varying water potential, seasonal changes, stress-induced changes or growth.

Given the increasing availability of multi- and hyper-temporal point cloud data [5] as well as the potential of cloud-to-cloud data fusion [6], we propose virtual laser scanning of dynamic scenes (VLS-4D) [7] to develop methods for monitoring vegetation movement, tree health, and forest growth. Unlike real-world data collection, which is limited by time or equipment to one or a few scenarios, VLS-4D allows the creation of many different scenarios. This is achieved by combining different scene compositions and dynamics, acquisition modes and sensor settings. Furthermore, VLS-4D includes perfect reference data of the underlying scene, including semantic labels, geometry and changes (e.g., as deformation/movement values or change labels). Such data is usually difficult, time-consuming or impossible to obtain when working with real point clouds, or it is associated with considerable errors or unknown ambiguities. The scenario building capabilities, together with the availability of reference data, make VLS-4D a promising data generation tool for the ever-growing pool of deep learning methods for the analysis of forest point clouds and point cloud time series.

We distinguish three concepts of how dynamic scenes can be implemented in LiDAR simulation [7]:

a) Few static representations of the forest scene at different epochs, e.g., one before and one after a forest disturbance event.
b) Many static snapshots sampled from an animated scene, e.g., daily snapshots to simulate a permanent laser scanning setup.
c) Animation within the scene, e.g., vegetation moving in the wind during a single terrestrial laser scan.

We will present simulation workflows for each of these concepts using the open-source software HELIOS++ [8], from animated 3D scene generation in Blender to final simulated point clouds and point cloud time series. With these simulation examples, we illustrate the research gaps that can be filled by such virtual experiments, address strategies and challenges in implementing VLS-4D, and discuss future directions. We expect VLS-4D data to play an essential role in the development of innovative methods for forest monitoring, complementing the still limited and typically unlabelled real-world multitemporal datasets.

References:

[1] Winiwarter, L., et al. (2022): DOI: https://doi.org/10.1016/j.rse.2021.112772

[2] Liu, J. et al. (2019): DOI: https://doi.org/10.1016/j.isprsjprs.2019.01.005

[3] Schäfer, J. et al. (2023): DOI: https://doi.org/10.1093/forestry/cpad061

[4] Esmorís, A. et al. (2024): DOI: https://doi.org/10.1016/j.isprsjprs.2024.06.018

[5] Eitel, J.U.H. et al. (2016): DOI: https://doi.org/10.1016/j.rse.2016.08.018

[6] Balestra, M. et al. (2024): DOI: https://doi.org/10.1007/s40725-024-00223-7

[7] Weiser, H. & Höfle, B. (2024): DOI: https://doi.org/10.31223/X51Q5V

[8] HELIOS++: https://github.com/3dgeo-heidelberg/helios