For more than 10 years, the University of Bretagne Occidental has maintained a network of 7 to 9 hydrophones moored in the Indian Ocean. These hydrophones have collected an enormous data set of continuous waveform data that needs to be analyzed. Given the nature and size of this enormous data set, machine learning is a natural tool to building better understanding of the various phenomena that occurs in the ocean.
This master’s thesis will be focused on using U-net neural networks to automatically process the waveform data. The goal of this project is to detect seismic events related to seafloor spreading. The network of hydrophones in the southern Indian Ocean encompass three spreading ridges, which create new seafloor and continuously generate earthquakes. These earthquakes can be used to better understand this process (e.g. through their geographic and time distribution). Due to the size and uniqueness of the data set, machine learning, specifically U-net based neural networks can detect events in continuous waveform data. The student will develop a novel neural network to detect and catalog events. This information will then be used to understand the physics of seafloor spreading. Additionally, the student may also detect whale songs.
This project includes a potential research stay to Laboratoire Geosciences Ocean, given available funding.
More details on the project data
In the Indian Ocean, the OHASISBIO network comprises 7 to 9 distant hydrophones continuously recording low-frequency sounds (0-120Hz) since 2010. Its objective is to monitor the seismic activity of mid-ocean ridges, but also the presence and migration patterns of large whales, and the oceanic ambient noise in general. Indeed, mid-oceanic spreading centers generate a large number of earthquakes and thus acoustic waves, indicative of the intervening seafloor spreading processes. Moreover, large baleen whales produce many loud and distinctive calls and songs, which provides clues as to when and where species dwell and migrate. Other sounds of interest are cryogenic sounds produced by icebergs or man-made noises (ship traffic, seismic exploration).
Over the years, passive acoustic monitoring of the ocean results in very large data sets (e.g. 25G/yr/instrument x 10 instr. x 10 years). The preliminary but indispensable, and time consuming step in the data analysis consists in identifying the different types of acoustic events. To achieve a more complete and efficient analysis, we wish to develop a deep learning application for event detection and signal discrimination in our acoustic database.