Machine Learning for classifying the underwater environment from Transmission Loss data

This paper describes the development and testing of probabilistic Machine Learning (ML) algorithms capable of classifying the underwater environment using only acoustic Transmission Loss (TL) data. The context of the work is to support the autonomous repositioning of vehicles to optimise sonar operations without the aid of external mission planning.

Methods of classifying the underwater environment from TL data were developed and demonstrated on simulated data generated using a variant of the RAM Parabolic Equation model over a 200-Hz band centred at 1 kHz. Training and independent test TL datasets were generated for nine classes of shallow water environment (all combinations of upward refracting, downward refracting and stratified sound speed profiles and low, medium and high loss seabeds). Three classifier algorithms were investigated; the main Probabilistic Principal Component Analysis (PPCA) algorithm was fully developed and tested including simulations of TL data collected over short spatial intervals representing a receiver mounted on an autonomous underwater vehicle. The PPCA model gave a classification accuracy of >93% when test data was provided over the maximum spatial interval (50 km). For 5-km intervals, classification accuracy varied with starting range but remained high, and confusion occurred only between environments with similar TL trends, resulting in minimal operational impact. An alternative Latent Variable Gaussian Process (LVGP) model also showed strong performance, achieving a classification accuracy of 96%.

This paper details the mathematical basis of the algorithms and measures of performance, the simulated data and testing methodology, and the results and conclusions of the work.