Publikation
Adapting Diffusion-Based LiDAR Scene Completion to Underwater SONAR Bathymetry
Varchaswi Madamanchi; Ahmed H. Elsayed; Benjamin Lehmann; Frederic Theodor Stahl
In: Max Bramer; Frederic Theodor Stahl (Hrsg.). Artificial Intelligence XLIII - SGAI 2026. SGAI International Conference on Artificial Intelligence (AI-2026), 46th SGAI International Conference on Artificial Intelligence, December 14-18, Cambridge, United Kingdom, Springer Nature Switzerland AG, 11/2026.
Zusammenfassung
Multibeam SONAR point clouds are important for bathymetric reconstruction, aquatic vegetation monitoring, and robotic lake maintenance. However, they differ strongly from terrestrial LiDAR data because they are sparse, noisy, unevenly sampled, affected by acoustic artefacts, and dominated by surface-like bathymetric geometry rather than object-rich 3D structure. This paper investigates whether LiDiff, a diffusion-based scene completion model originally developed for terrestrial LiDAR data, can be adapted to underwater SONAR bathymetric point clouds. We use multibeam SONAR data from Masch Lake, Hanover, and the external AURORA underwater bathymetry dataset. Since the available SONAR data consists of aggregated bathymetric maps rather than sequential partial scans, a SONAR-to-KITTI-style preprocessing pipeline is used to create local incomplete-to-complete training samples. Ordinary kriging is included as a geometric 2.5D bathymetric surface reconstruction method, while LiDiff is evaluated as a learning-based 3D point cloud completion method. The LiDiff experiments analyse training on AURORA, fine-tuning on Masch Lake, and synthetic diagnostic scenarios. On Masch Lake strip-masked tiles, ordinary kriging reconstructs the bathymetric surface accurately, achieving a Chamfer Distance (CD) of approximately 0.171 m and an F1-score of 0.982. LiDiff learns to fill strip-like gaps on AURORA and improves numerically on Masch Lake after fine-tuning. However, the LiDiff metrics must be interpreted together with qualitative inspection because the model uses fixed point sampling, while SONAR bathymetry has uneven point density. This can lead to denser outputs and over-completion of elevated vegetation-like structures. The results show that diffusion-based completion is promising for SONAR point cloud completion, but reliable bathymetric reconstruction requires surface-aware representations and better separation between lakebed, vegetation, and acoustic artefacts.