AQUA 3D: Topo-bathymetric LiDAR data Classification

LiDAR bathymetry has emerged as a powerful remote sensing technique for mapping underwater and nearshore environments with high precision. The technology utilizes laser-based scanning to measure topographic and bathymetric features by emitting laser pulses and analyzing their reflections. However, accurately classifying LiDAR bathymetry data remains a significant challenge due to the complex interactions between laser beams and different surface types, such as water, vegetation, and submerged terrain.

In this study, we have enhanced a classification algorithm tailored for high-resolution LiDAR bathymetry data captured by a novel laser scanner developed by our research team at the Fraunhofer Institute for Physical Measurement Techniques (IPM). This scanner integrates both green and near-infrared (NIR) laser channels, optimizing data acquisition for both terrestrial and aquatic environments. The NIR laser primarily aids in detecting water surface areas, as it is strongly absorbed by water, effectively delineating the boundary between land and water. In contrast, the green laser penetrates the water column, enabling precise measurement of submerged features, including underwater vegetation, the seabed, and other submerged objects.

The developed system captures ultra-high-resolution data, covering regions both above and below the water surface with exceptional detail. The primary classification categories include water surface, submerged vegetation, seabed, soil ground and terrestrial vegetation such as trees surrounding the lake. Since the captured data by newly designed system is super dense, its processing requires expensive computational complexity. Therefore, in this study we have focused on improving an algorithm to overcome this problem and can efficiently handle processing of this data.

The proposed classification algorithm is designed to handle the intricate variations in LiDAR return signals caused by factors such as water turbidity, surface reflectivity, and vegetation density. By leveraging both spectral and spatial features of the captured data, our approach ensures precise delineation of different classes. The enhanced algorithm contributes to improved ecological monitoring, hydrological modeling, and environmental management, making it a valuable tool for scientific research and practical applications in hydrography and environmental science.

Nevertheless, the inherent complexity of data in overlapping regions particularly at the interface of submerged vegetation, the water surface, and the seabed pose significant challenges. A considerable amount of classification errors is associated with vegetation close the seabed or the water surface, where the geometric and spectral distinctions become less pronounced. A promising approach to meet these challenges is the sphere-based feature extraction, since it analyzes the data in multilevel steps including both local and global regions; however, spherical region selection with multiple radii can result in redundant feature generation. To address this, we employ adaptive spherical region selection to minimize unnecessary computations in areas with low point density. In this study, a voxel-based representation of point clouds is used to identify regions of higher point density. These high-density regions are prioritized for considering multiple spheres with varying radii to extract features at different levels of detail. Conversely, low-density regions are less critical for fine-grain spherical region selection, and only two spheres are selected to analyze these areas efficiently. For each selected sphere, a variety of geometrical features are extracted. Features at different sphere scales capture the distribution shape of the point clouds, including measures of sphericity, linearity, and planarity. Additionally, roughness, curvature, and anisotropy of the point clouds are assessed. Statistical features such as mean, mode, median, standard deviation, range, and skewness are also computed for the intensity values of point clouds embedded by different sphere scales. Note that, intensity values from both green and near-infrared signals are considered. This approach enables a comprehensive analysis of point cloud structures while optimizing computational efficiency. Extracted features are fed into a Random Forest classifier to determine different classes.