Automated Detection of Seafloor Gas Seeps in Multibeam Echosounder Data With an Attention-Guided Convolutional Neural Network

Seafloor gas seeps are a globally distributed feature that hold substantial implications for ocean carbon cycling, chemosynthetic ecology, energy production, and geohazards. Detection of seafloor gas seeps currently requires human visual interpretation of sonar water column imagery by a trained individual, which is time consuming, costly, and inconsistent. In this study, we introduce a novel dataset composed of 428 396 high-resolution multibeam echosounder (MBES) sonar water column images annotated for the presence and position of gas bubble plume targets generated by seafloor seeps. Leveraging this dataset for training and validation, we present a machine learning model based on an attention-guided convolutional neural network framework for automated gas seep detection in seafloor environments. In addition, we present an analysis of a range of processing approaches for MBES sonar water column data, evaluating their impact on the seep detection performance of the machine learning model. Finally, our proposed ML framework underwent a rigorous evaluation using a location-independent validation strategy, achieving an overall accuracy of 87%, with 78% precision and 72% recall rate ensuring robustness and reliability. This research represents a substantial advancement in automated gas seep detection methodology.