Detection and localisation of fluid emissions in water column data using Deep Learning with acoustic and spatial information

Fluid emissions from the seafloor affect ocean chemistry and are involved in the geological processes taking place along active and passive continental margins. These emissions are linked to geological hazards, such as earthquakes, sedimentary instabilities, and extensive methane release. Detecting and locating the sources of fluid emissions is therefore of paramount importance. Hydrographic MultiBeam EchoSounders (MBES) designed for seafloor mapping can record the acoustic backscatter of the water column. Due to the impedance contrast between gas and seawater, gas bubbles form "acoustic plumes" in the echograms. Acoustic data assists in guiding the exploration of seeps and their associated geological structures. However, processing the vast amount of data generated by these sounders is a significant undertaking.

The present study is based on the data collected from two surveys, GAZCOGNE1 (Bay of Biscay, Aquitaine Basin) and GHASS2 (Black Sea) during which data were collected using a Kongsberg EM302 MBES (30 kHz transmit frequency) and a Reason Seabat 7150 MBES (24 kHz transmit frequency) respectively. These sounders have proven to be very effective in identifying fluid emissions.

Deep learning has become increasingly popular in marine science over the last few decades due to the use of Graphical Processing Units and large amounts of labelled data. This method has proven to be particularly robust in accurately analysing large datasets and identifying complex patterns. We have devised a deep-learning approach that allows us to: 1) Detect fluid-related echoes in multibeam echograms. 2) Conduct near real-time fluid detection and tracking during the acquisition surveys and provide accurate positioning of the fluid outlet beneath the seafloor based on acoustic and spatial attributes. 3) Discriminate between fluid-related echoes emanating from the primary lobe of the multibeam directivity and those originating from the side lobes, in order to accurately locate the fluid outlet. This last approach results from antenna modelling and multibeam survey simulation. The technique for echo discrimination using antenna modelling was produced with the open-source toolbox published by Urban et al 2023 (https://doi.org/10.1002/lom3.10552).

Detection on the multibeam echograms is performed by adapting the open-source You-Only-Look-Once algorithm (version 5). Training on Ifremer datasets showed that the results surpass those of a state-of-the-art method regardless of the MBES used for training and testing. Hence, this method can be applied to diverse MBES data, acoustic acquisition parameters and environmental conditions. The algorithm can detect signals throughout the entire water column, even in areas affected by acoustic artefacts such as specular side lobes and different emission sectors. We have developed methods to improve neural network learning using training sets when limited labelled MBES data are available. The method was tested during an oceanographic expedition in the summer of 2022 (MAYOBS23), demonstrating its ability to operate in near real-time with excellent performance.

The marine expedition GAZCOGNE1, part of the PAMELA project, was co-funded by TotalEnergies and IFREMER. The expedition GHASS2 was co-founded by the Agence Nationale de la Recherche for BLAck sea MEthane (BLAME) project and IFREMER. MAYOBS23 was funded by the Mayotte volcanological and seismological monitoring network REVOSIMA.