YOLOv5-WAL is a YOLOv5-based deep learning supervised approach to automate the detection of fluids emitted from the seafloor (e.g. methane bubbles from cold seeps and liquid carbon dioxide from volcanic sites). It concerns the detection of fluids in water column images (echograms) acquired with multibeam echosounders.  Several thousand annotated echograms from different seas and oceans acquired  during distinct surveys were used to train and test the deep learning model (Table 1). The tests were conducted on a dataset comprising hundreds of thousands of echograms i) acquired with three different multibeam echosounders (Kongsberg EM302 and EM122 and Reson Seabat 7150) and ii) characterized by variable water-column noise conditions related to sounder artefacts and the presence of biomass (e.g. fish, dolphins).

This dataset contains models trained for fluid detection issued from several multibeam echosounders (Kongsberg EM122, EM302, Reson Seabat 7150) (Table 2). This fluid detector was already used for near-real time acquisition detection during the MAYOBS23 (EM122 – 2022; Perret et al. 2023) and HAITI-TWIST (Seabat Reson 7150 - 2024) cruises.

Inference code (YOLOv5 with G3D files) is available on github repository.