Machine Learning Inference of the Ocean Environment from Acoustic Data

The use of machine learning models to determine physical quantities of a complex underwater environment via acoustic data is relatively underdeveloped. Traditionally, research in underwater acoustics (the study of sound wave generation, propagation, scattering and reception in water) has mainly been applied for the use of sound wave navigation and ranging (SONAR) systems for communication, long-range sensing, target detection, marine wildlife monitoring and exploration.

However, recent breakthroughs in data mining and analysis, supported by powerful super-computing capabilities, promise exciting new possibilities for the implementation of machine and deep learning approaches to develop methods to extract and predict important properties of the ocean that are relevant to underwater acoustics. The main properties of interest have been the sound speed profile and sediment geo-acoustics for which there are established model-based inversion schemes. It is expected that these and many other acoustically-relevant properties, across multiple scales, from internal waves and eddies to ocean spice, together with their spatial and temporal variability could now be derived from acoustic data using new machine learning inference methods. The result would be a much richer and truer description of the ocean environment which would improve the prediction accuracy of acoustic models and the associated performance of sonar systems.

UK maritime forces are deployed 365 days a year across the world to provide a forward national presence that projects influence to safeguard our interests. This pattern of operational deployment depends on our ability to use and exploit the ocean environment. Consequently, the operational use of sonar systems depends on having a description of the ocean environment that includes all acoustically-relevant properties of the environment - the more complete this description is, the greater our ability to operate successfully.

Unfortunately, direct measurement of such properties is difficult and expensive and the quality of the output from acoustic models that need this description are typically limited by the information that is available. The availability of ocean environment information is equally important for our understanding of complex ocean processes and sustainable use of the oceans. Healthy oceans are vital for life on Earth: they buffer the globe from the effects of anthropogenic CO2, support biodiversity and marine life, provide critical resources and a means of global transportation.

The process of extracting information indirectly from measured acoustic data is called inversion. This acoustic measurement results from an acoustic signal that has propagated through the environment (changes to the signal result from different acoustically-relevant properties of the environment) before being received by a hydrophone or similar system. The acoustic measurement therefore contains information about the ocean environment that can be derived using suitable models and methods. The aim of this project is to use data-driven machine learning models to improve this resulting description of the ocean environment, to represent a wider range of ocean properties that are relevant to underwater acoustics, and to support our understanding, use and exploitation of the ocean environment.

The focus of the project will be on machine learning models that can be use acoustic data collected by in-situ sensors and remote sensors, modelled data, historical data, and data from other sources, to infer acoustically relevant properties of the ocean environment, from which to build an up-to-date and accurate representation of the acoustic environment for any sonar deployment.