Semantic segmentation of sparse irregular point clouds for leaf/wood discrimination

TL;DR

This work tackles leaf/wood discrimination in sparse, irregular ULS forest point clouds, a prerequisite for accurate leaf-area estimation. It introduces SOUL, a geometry-centric semantic segmentation pipeline that uses coordinates-only input, multi-scale geometric features, geodesic voxelization decomposition (GVD) for data partitioning, and a rebalanced loss to address extreme class imbalance. The approach achieves state-of-the-art performance on tropical ULS data and shows promising generalization to other forest datasets and LiDAR devices, aided by the GVD pre-partitioning and ablation-driven feature design. By removing reliance on intensity data and focusing on intrinsic geometric structure, SOUL offers a robust, transfer-friendly tool for forest monitoring and leaf-area research, with potential extensions to denser canopies and broader LiDAR configurations.

Abstract

LiDAR (Light Detection and Ranging) has become an essential part of the remote sensing toolbox used for biosphere monitoring. In particular, LiDAR provides the opportunity to map forest leaf area with unprecedented accuracy, while leaf area has remained an important source of uncertainty affecting models of gas exchanges between the vegetation and the atmosphere. Unmanned Aerial Vehicles (UAV) are easy to mobilize and therefore allow frequent revisits to track the response of vegetation to climate change. However, miniature sensors embarked on UAVs usually provide point clouds of limited density, which are further affected by a strong decrease in density from top to bottom of the canopy due to progressively stronger occlusion. In such a context, discriminating leaf points from wood points presents a significant challenge due in particular to strong class imbalance and spatially irregular sampling intensity. Here we introduce a neural network model based on the Pointnet ++ architecture which makes use of point geometry only (excluding any spectral information). To cope with local data sparsity, we propose an innovative sampling scheme which strives to preserve local important geometric information. We also propose a loss function adapted to the severe class imbalance. We show that our model outperforms state-of-the-art alternatives on UAV point clouds. We discuss future possible improvements, particularly regarding much denser point clouds acquired from below the canopy.

Figures (21)

• Figure 1: Point clouds produced by three scanning modes on the same area ($20\meter \times 20\meter \times 42\meter$), illustrate how much the visibility of the understory differs. The colors in the figure correspond to different labels assigned to the points, where red and green indicate leaves and wood, respectively. Blue points are unprocessed, so labeled as unknown. TLS data were initially subjected to semantic segmentation using LeWos LeWos and subsequently manually corrected. Next, the K-nearest neighbors (KNN) algorithm was used to assign labels to ALS and ULS data based on the majority label among their five nearest neighbors in TLS data.
• Figure 2: Overview of SOUL. (a) We use only the coordinates of raw LiDAR data as input. (b) Four geometric features linearity, sphericity, verticality, and PCA1 are calculated at three scales using eigenvalues, then standardized. (c) GVD proposes partitioned components and performs data normalization within these components. (d) Training deep neural network. (e) Finally, output are point-wise predictions.
• Figure 3: (a) displays the components found by the GVD algorithm. (b) displays the geodesic distance within the corresponding component. The figures are illustrative, actual ULS data lacks such depicted complete tree trunks.
• Figure 4: The labeled ULS data set in French Guiana, where only 4.4% of the points correspond to wood. This significant class imbalance of leaf & wood presents a considerable challenge in discriminating wood points from ULS forest point cloud.
• Figure 5: Qualitative results on ULS test data. Because of the class imbalance issue, methods such as FSCTFSCT, LeWosLeWos, and other existing approaches developed for dense forest point clouds, like TLS or MLS, (cf. Section \ref{['related_work']}) are ineffective.