Sidescan Only Neural Bathymetry from Large-Scale Survey

Abstract

Sidescan sonar is a small and low-cost sensor that can be mounted on most unmanned underwater vehicles (UUVs) and unmanned surface vehicles (USVs). It has the advantages of high resolution and wide coverage, which could be valuable in providing an efficient and cost-effective solution for obtaining the bathymetry when bathymetric data are unavailable. This work proposes a method of reconstructing bathymetry using only sidescan data from large-scale surveys by formulating the problem as a global optimization, where a Sinusoidal Representation Network (SIREN) is used to represent the bathymetry and the albedo and the beam profile are jointly estimated based on a Lambertian scattering model. The assessment of the proposed method is conducted by comparing the reconstructed bathymetry with the bathymetric data collected with a high-resolution multi-beam echo sounder (MBES). An error of 20 cm on the bathymetry is achieved from a large-scale survey. The proposed method proved to be an effective way to reconstruct bathymetry from sidescan sonar data when high-accuracy positioning is available. This could be of great use for applications such as surface vehicles with Global Navigation Satellite System (GNSS) to obtain high-quality bathymetry in shallow water or small autonomous underwater vehicles (AUVs) if simultaneous localization and mapping (SLAM) can be applied to correct the navigation estimate.

Towards Differentiable Rendering for Sidescan Sonar Imagery

“…However, the "ground truth" bathymetry is needed for creating a training set, which is not always practical underwater. Neural rendering [7], [8] methods that leverage the continuity and differentiablity of implicit neural representations have been recently proposed to fit many sidescan lines into a self-consistent bathymetry with a global optimization. Specifically, a multi-layer perceptron (MLP) with sine activation functions, known as SIREN [27] was used to represent the bathymetry where the gradients of the bathymetry were constrained by SSS intensities through a Lambertian model.…”

“…Specifically, a multi-layer perceptron (MLP) with sine activation functions, known as SIREN [27] was used to represent the bathymetry where the gradients of the bathymetry were constrained by SSS intensities through a Lambertian model. Extended from [7], a nadir model was proposed in [8] to model the nadir region in SSS waterfall images so that the optimization can converge without any external bathymetric data, e.g., altimeter readings. However, acoustic shadows cannot be explained by the Lambertian model in [8].…”

“…Nevertheless, one can address such a chicken-and-egg problem by iteratively improving the estimates of the bathymetry and the pose estimation. Following this principle, we present a framework † , NeuRSS, that leverages the advances of neural rendering to estimate bathymetry from SSS [7], [8], which provides an elevation prior addressing the degeneracy in SSS SLAM problems and showcase that one can significantly improve the AUV's DR estimate and subsequently produce high-quality bathymetry using SSS data from a standard survey.…”

“…The contribution of this work is to extend our previous work [7], [8], [21] and combine them into a new framework. For the sonar scattering model in [8], we extend it to be able to model shadows.…”

“…The contribution of this work is to extend our previous work [7], [8], [21] and combine them into a new framework. For the sonar scattering model in [8], we extend it to be able to model shadows. For the back-end of the SSS SLAM framework in [21], we extend it to a submap-based optimization with a prior map to better address the elevation degeneracy.…”