Deep-learning-based detection of underwater fluids in multiple multibeam echosounder data

Detecting and locating emitted fluids in the water column is necessary for studying margins, identifying natural resources, and preventing geohazards. Fluids can be detected in the water column using multibeam echosounder data. However, manually analyzing the huge volume of this data for geoscientists is a very time-consuming task. Our study investigated the use of a YOLO-based deep learning supervised approach to automate the detection of fluids emitted from cold seeps (gaseous methane) and volcanic sites (liquid carbon dioxide). Several thousand annotated echograms collected from three different seas and oceans during distinct surveys were used to train and test the deep learning model. The results demonstrate first that this method surpasses current machine learning techniques, such as Haar-Local Binary Pattern Cascade. Additionally, we thoroughly analyzed the composition of the training dataset and evaluated the detection performance based on various training configurations. 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 (fishes, dolphins). Incorporating untargeted echoes (acoustic artefacts) in the training set (through hard negative mining) along with adding images without fluid-related echoes are the most efficient way to improve the performance of the model and reduce the false positives. Our fluid detector opens the door for near-real time acquisition and post-acquisition detection with efficiency, reliability and rapidity.