Machine learning in ocean acoustic thermometry

Project description
Ocean acoustic thermometry builds on the concept that acoustic signals travel faster in warm water than in cold water. The sound speed is obtained by measuring acoustic travel time between a source and receiver. To obtain the accurate travel times, high accuracy clocks are used in the instruments and mooring motions are monitored by a long baseline transponder network.

The acoustic signals at low frequencies are propagating over large distances. The more dynamic the ocean is, the more difficult it is to understand the acoustic propagation. How the acoustic signal propagates is determined by the vertical stratification and horizontal variability e.g., fronts and eddies and the bathymetry between source and receiver. Acoustic models such as ray tracing or full wave models are used to understand how the acoustic signals propagate in the ocean.

At the receiver several arrivals will be observed because the sound follows different paths. Part of the received signal has been refracted internally in the ocean. Other parts have interacted with the surface e.g., sea surface, sea ice, as well as with the sea floor causing losses due to reflection and scattering.

The acoustic signals used in ocean acoustic thermometry can be frequency modulated sweeps or M-sequences (phase coded signals). The traditional approach is to correlate the known signal with the acoustic recording to detect the signals in the recordings (Pulse compression). Despite the pulse compression the signal can be smeared out and hidden in natural background noise.

In this project the acoustic recordings from the UNDER-ICE project will be revisited and analysed. The UNDER-ICE experiment was carried out in the Fram Strait and consisted of five moorings that were deployed between 2014 and 2016. Each mooring was equipped with 10 hydrophone modules equidistantly spaced by 9 m, creating an array with an aperture of 81 m. These recordings have been pre-processed e.g., pulse compressed, corrected for clock drift and mooring motion. Ray based inversion routines depend on detection of peaks in the pulse compressed result, and to identify corresponding ray paths. In the UNDER-ICE there are clear peak arrivals in some periods, while in other periods it is very difficult to detect any signals.

Machine learning can be used to detect and utilise features in acoustic data if enough test data is available. To our knowledge machine learning has not yet been used to detect tomographic signals in recordings. This master project will investigate the usefulness of ML in detecting the acoustic signal in the acoustic recordings obtained in UNDER-ICE. The objective is to compare traditional and machine learning based detection and travel time estimation for acoustic thermometry.

The following work steps are planned for the project:
-Get an overview of methods used in traditional detection of tomographic signals.
-Analyse the acoustic recordings and the pre-processed acoustic arrivals.
-Familiarise with ML techniques used in acoustics and how to build test data.
-Review and test different ML packages to analyse the tomographic data.

Proposed course plan during the master's degree (60 ECTS)
GEOV217 (10 sp.)
GEOV276 (10 sp.)
GEOV277 (10 sp.)
GEOV316 (10 sp.)
GEOV302 (10 sp.)
GEOV375 (10 sp.)

Prerequisites
Bachelor i geofysikk

Field- lab- and analysis work
The following work steps are planned for the project:

• Get an overview of methods used in traditional detection of tomographic signals.
• Analyse the acoustic recordings and the pre-processed acoustic arrivals.
• Familiarise with ML techniques used in acoustics and how to build test data.
• Review and test different ML packages to analyse the tomographic data.