Score-Based Multibeam Point Cloud Denoising

TL;DR

The paper tackles outliers and noise in multibeam echo-sounder bathymetry by adapting a score-based denoising framework from 3D point clouds to MBES data. It trains a Score Estimation Network to learn the local score S_i(r) that guides gradient-based denoising via the ensemble score E(r) and applies an IQR-based outlier detector. A ground-truth MBES dataset is created from real AUV surveys with manual cleaning and mesh draping to enable evaluation on 32-ping patches, and the model is benchmarked against traditional CUBE-like and simple statistical baselines. Results show consistent improvements in outlier rejection and, with a mean-interpolation extension and a second denoising pass, competitive denoising performance, with open-source code and pretrained models provided for reproducibility.

Abstract

Multibeam echo-sounder (MBES) is the de-facto sensor for bathymetry mapping. In recent years, cheaper MBES sensors and global mapping initiatives have led to exponential growth of available data. However, raw MBES data contains 1-25% of noise that requires semi-automatic filtering using tools such as Combined Uncertainty and Bathymetric Estimator (CUBE). In this work, we draw inspirations from the 3D point cloud community and adapted a score-based point cloud denoising network for MBES outlier detection and denoising. We trained and evaluated this network on real MBES survey data. The proposed method was found to outperform classical methods, and can be readily integrated into existing MBES standard workflow. To facilitate future research, the code and pretrained model are available online.

Figures (7)

• Figure 1: Left: Illustration of the score estimation network applied to an MBES patch. Right: An intuitive description of the score function used to denoise MBES points. In this case, the score essentially represents the 1D vector from the z value of the noisy point to the z value the corresponding clean point.
• Figure 2: Visualization of MBES dataset used in this paper. The AUV trajectory is overlaid on the cleaned bathymetry in gray. The test set is enlarged and highlighted on the top right corner. Note that the depth range in the test set (85-95m) is significantly different than that of the training set (30-100m).
• Figure 3: Distribution of $z$ value differences between the manually identified outliers in the raw MBES data and clean draping results in training, test and validation set.
• Figure 4: Visual outlier detection results of a patch where the proposed score net achieves better results than baseline methods (radius-based method is visualized here). Typically, these patches have very small range in $z$ compared to the rest of the test set. As a result, the baselines with globally tuned parameters performs poorly. The score-based methods, on the other hand, learn the local score around points and are less sensitive to the exact numeric range within a patch.
• Figure 5: Visual outlier detection results of a patch where Score (64) performs worse than the baselines (radius-based method is visualized here). Score (256) tends to outperform Score (64). Typically, these patches have large range of $z$, and the outliers are significantly removed from the inliers. In the patch shown, the outliers range between 19.25-25.6m, whilst inliers reside within 19.3-19.7m. Due to the large amount of outliers closed to each other, the estimated scores will be inflated, leading to worse performance of score-based methods. This inflation can partly be mitigated by using larger neighborhood for score ensemble (see differences between Score (64) in (c) and Score (256) in (d)).