Adaptive Sphere-Based Feature Extraction for Efficient Classification of Bathymetric LiDAR Data

We propose a classification framework for Bathymetric LiDAR data acquired using a UAV-mounted lightweight laser scanner developed by the Fraunhofer Institute for Physical Measurement Techniques (IPM). The system captures high-resolution 3D point cloud data from both above and below the water surface, enabling detailed spatial analysis. To fully exploit the richness of this data, accurate classification of key regions, including water surfaces, submerged vegetation, seabed, and shoreline, is essential. Our approach utilizes an adaptive spherical region selection method to extract localized geometric and intensity-based features. A voxel-based representation is employed to identify high-density regions, where multiple spheres with varying radii are applied to capture fine-scale structural details. Conversely, in low-density areas, only two spheres with diameters of 1m and 2m are used to optimize computational efficiency. Extracted features include sphericity, linearity, planarity, roughness, curvature, and anisotropy, along with statistical metrics such as mean, median, and skewness for intensity values derived from green and near-infrared laser beams. The classification is performed using a Random Forest model, achieving an overall accuracy of 83%. The primary challenge lies in distinguishing submerged vegetation from the seabed and water surface due to spectral and geometric ambiguities. The proposed method effectively addresses these complexities by integrating multi-scale feature extraction with an adaptive selection strategy, ensuring both precision and computational efficiency. This approach enhances the robustness of LiDAR-based aquatic environment classification, facilitating improved analysis of submerged and terrestrial regions.