Bathymetric Surveying with Imaging Sonar Using Neural Volume Rendering

TL;DR

This work tackles high-resolution seabed mapping from forward-looking sonar by moving beyond ground-truth supervision and Lambertian assumptions. It introduces a differentiable neural rendering pipeline that represents the seafloor as a neural heightmap encoded with multi-resolution hash tables and uses a differentiable sonar renderer to model volumetric returns, jointly learning bathymetry and beam patterns in a self-supervised fashion. The approach is evaluated on both simulated and field data, where it outperforms Lambertian-based and frequency-encoded baselines and enables super-resolution mapping by fusing low-altitude FLS with high-resolution priors. The results demonstrate practical potential for detailed seabed mapping in challenging underwater environments, with future work focusing on SLAM integration and real-time capabilities.

Abstract

This research addresses the challenge of estimating bathymetry from imaging sonars where the state-of-the-art works have primarily relied on either supervised learning with ground-truth labels or surface rendering based on the Lambertian assumption. In this letter, we propose a novel, self-supervised framework based on volume rendering for reconstructing bathymetry using forward-looking sonar (FLS) data collected during standard surveys. We represent the seafloor as a neural heightmap encapsulated with a parametric multi-resolution hash encoding scheme and model the sonar measurements with a differentiable renderer using sonar volumetric rendering employed with hierarchical sampling techniques. Additionally, we model the horizontal and vertical beam patterns and estimate them jointly with the bathymetry. We evaluate the proposed method quantitatively on simulation and field data collected by remotely operated vehicles (ROVs) during low-altitude surveys. Results show that the proposed method outperforms the current state-of-the-art approaches that use imaging sonars for seabed mapping. We also demonstrate that the proposed approach can potentially be used to increase the resolution of a low-resolution prior map with FLS data from low-altitude surveys.

Figures (8)

• Figure 1: (a) ROV Ventana with the Low Altitude Survey System (LASS) mounted underneath being deployed from R/V Rachel Carson. The top right corner shows the sonar to acquire the field images. (b) An example of a sonar image captured with Gemini 720is FLS.
• Figure 2: System Overview: Given a range $r$ and azimuth angle $\theta$, we sample along the arc and for every sampled point on the arc, we sample along the acoustic ray. The spatial coordinates $(x,y)$ are encoded and then fed into the neural heightmap $\mathbf{N}$ outputting the height $h$, while at the same time the viewing angle $(\theta,\phi)$ is encoded using a spherical harmonics basis. The rendering network $\mathbf{L}$ takes spatial information, learned features $\mathbf{f}$, encoded viewing angles and surface normals $\mathbf{n}$ to learn the radiance field, which is later used together with the vertical signed distance $\Delta$ to apply sonar volumetric rendering to predict the returned intensity.
• Figure 3: (a) Each image column $\theta_i$ corresponds to a 2D fan-shape beam emitted from a transducer, recording returns between $r_\mathrm{min}$ and $r_\mathrm{max}$. (b) Geometry model of the FLS, where the blue point $(r,\theta,\phi)$ is a 3D point that is projected onto the image plane $z=0$. (c) An example of a sonar image in simulation. Each pixel at $(r,\theta)$ contains the returned intensities of all points along the elevation arc.
• Figure 4: Illustration of hierarchical sampling along the arc. Assuming the red curve is the seafloor, first $\mathbf{N_{\mathcal{A},s}}=5$ stratified samples (in green) are drawn on the arc and subsequently $\mathbf{N_{\mathcal{A},i}}=2$ importance samples (in yellow) are obtained using the S-density of the 5 green samples. The whole 7 samples are used to construct acoustic rays where we further sample along the rays to compute opacity and transmittance of the 7 samples along the arc.
• Figure 5: Reconstructed bathymetry for Freq.$+$Sampl. (left) and for the proposed method (middle) and ground truth (right) for simulation dataset, at 10 cm resolution (ROV trajectory in blue).