Abstract High-resolution 3D forest representations are essential for remote sensing applications such as above-ground biomass estimation using radiative transfer modelling. However, existing reconstruction approaches are often time-consuming and rely heavily on manual input. A comprehensive and largely automated end-to-end workflow is presented for reconstructing realistic 3D forest representations from terrestrial laser scanning (TLS) data. The workflow includes five main steps: segmentation of individual trees, semantic classification into wood and foliage using a custom-trained PointNet++ model, reconstruction of woody structures via Quantitative Structure Models, biologically realistic foliage placement, and spatial distribution of trees. Reconstructed forest plots from Central Europe were used to simulate airborne laser scanning (ALS) data in Helios++. The results were validated against real ALS acquisitions. The simulated data showed strong agreement with real ALS data across key forest structure metrics, with correlations ranging from R² = 0.46 for height standard deviation to R² = 0.96 for mean canopy height, with corresponding nRMSE values ranging between 23.2% and 12.6%. The largest discrepancies occurred in upper canopy regions due to TLS occlusion effects, where dense vegetation blocked the scanner’s line of sight, resulting in these areas being underrepresented in the reconstructed 3D scenes. These results demonstrate that detailed 3D forest reconstructions can be achieved with minimal manual inputs, providing a robust basis for radiative transfer modelling and the generation of synthetic remote sensing datasets, which are critical for improving forest monitoring and carbon stock assessments.
Hanousek et al. (Fri,) studied this question.