Classification of Imaging Artifacts in Synthetic Aperture Sonar With Bayesian Deep Learning

Synthetic aperture sonar (SAS) provides high-resolution underwater imaging but can suffer from artifacts due to environment or navigation errors. This work explores Bayesian deep learning for classifying common imaging artifacts while quantifying model reliability. We introduce a novel labeled data set with simulated imaging errors through controlled beamforming perturbations. Two Bayesian neural network variants, Monte Carlo dropout and flipout, were trained on this data to detect three artifacts induced by: sound speed errors, yaw attitude error, and additive noise. Results demonstrate these methods accurately classify artifacts in SAS imagery while producing well-calibrated uncertainty estimates. Uncertainty tends to be higher for uniform seafloor textures where artifacts are harder to perceive, and lower for richly textured environments. Analyzing uncertainty reveals regions likely to be misclassified. By discarding 20% of the most uncertain predictions, classification improves from 0.92 F<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$_{1}$</tex-math></inline-formula>-score to 0.98 F<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$_{1}$</tex-math></inline-formula>-score. Overall, the Bayesian approach enables uncertainty-aware perception, boosting model reliability—an essential capability for real-world autonomous underwater systems. This work establishes Bayesian deep learning as a robust technique for uncertainty quantification and artifact detection in SAS.