3D forest semantic segmentation using multispectral LiDAR and 3D deep learning

Regular forest inventory is essential for conservation and management. Over the past decades, laser scanning has emerged as a remote and non-destructive solution to streamline this laborious process. Advanced multispectral (MS) laser scanning systems simultaneously acquire 3D spatial and spectral information across multiple wavelengths, enabling estimation of forest biophysical and biochemical traits. This study investigates the potential of airborne MS laser scanning for fine-grained forest semantic segmentation into six components (ground, low vegetation, trunk, branches, foliage, and woody debris), thereby supporting forest inventory and analysis. We evaluate three state-of-the-art 3D deep learning models (kernel point convolution (KPConv), superpoint transformer (SPT), and point transformer V3 (PTv3)) and random forest model. Our analysis reveals the superiority of PTv3, outperforming the other models by 21.8 percentage points (pp) with the mean intersection over union (mIoU) of 69.1%. Additionally, our rigorous spectral ablation study demonstrates that MS laser scanning data substantially improves the segmentation results, increasing the IoU of woody debris, branches, and trunks by 12.7 pp, 4.5 pp, and 2.5 pp, respectively. This study highlights the strong potential of MS laser scanning to enable automated and accurate forest inventory through prior fine-grained forest semantic segmentation.