MB-UDF: A self-supervised model for continuous representation of seafloor topography using multibeam echo sounder data

Conventional methodologies for processing Multibeam Echo Sounder (MBES) data have primarily relied on the generation of Digital Elevation Models (DEMs) from bathymetric point clouds. Nevertheless, these approaches demonstrate inherent limitations in their capacity to faithfully represent the continuous and intricate morphology of the seafloor, a critical requirement for comprehensive geophysical analyses. This study introduces MB-UDF, a novel self-supervised learning framework that trains neural networks to represent Unsigned Distance Functions (UDF) for high-fidelity reconstruction of continuous 3D seafloor surfaces from MBES data. The primary contributions of MB-UDF encompass a specialized point cloud sampling mechanism and an efficient self-supervised learning strategy, both meticulously designed to address the inherent characteristics of MBES data. Our PyTorch implementation is open-sourced and available at https://github.com/Parallelopiped/MB-UDF . To rigorously evaluate the performance of MB-UDF, we established a comprehensive MBES dataset incorporating diverse bathymetric terrains. Experimental results demonstrate that our method significantly outperforms existing 3D reconstruction techniques in terms of normal consistency and surface continuity, exhibiting enhanced robustness and superior precision compared to conventional DEM approaches. The proposed MB-UDF framework provides innovative methodological tools for advancing research in marine geology, seafloor mapping, and related domains. • INR-based algorithm effectively reconstructs 3D seafloor from multibeam data. • Novel point cloud sampler helps model learn UDF from bathymetric data. • Self-supervised learning strategy shows excellent convergence performance.