Deep Learning-Based Bathymetric Reconstruction of Forward-Looking Sonar Intensity Data/Imagery using Multi Beam Echo Sounder Ground Truth
Document Type
Presentation
Date of Original Version
3-27-2026
Abstract
Forward-looking sonar (FLS) provides real-time imaging capability for autonomous underwater vehicles, but extracting 3D bathymetry from FLS images is difficult because the sensor's wide vertical aperture maps multiple seafloor points onto the same pixel. This thesis addresses the elevation ambiguity problem using a three-stage neural network trained on multi-beam echo sounder (MBES) measurements as ground truth. The three stages handle out-of-range detection, valid/partial return classification, and elevation angle regression sequentially, with the hierarchical design motivated by the severe class imbalance in FLS data where over 96\% of pixels carry no valid return. The network is trained and evaluated on simulated Stonefish data from a 35-minute water tank mission. Results show 94.3\% classification accuracy and an elevation angle RMSE of $0.018$\,rad on valid pixels, corresponding to mean reconstruction errors below $0.1$\,m on flat and moderately curved terrain.
Recommended Citation
Naderi, Farhang, "Deep Learning-Based Bathymetric Reconstruction of Forward-Looking Sonar Intensity Data/Imagery using Multi Beam Echo Sounder Ground Truth" (2026). Oral Presentations. Paper 27.
https://digitalcommons.uri.edu/gradcon2026-presentations/27