Machine Learning Based Automated Detection of Seafloor Gas Seeps

Generating a comprehensive understanding of the locations and distribution of seafloor gas seeps is a task that has become increasingly of interest to marine research organizations and industries. Currently, the feasible scale of mapping efforts is limited by a dependence on manual inspection of water column sonar images by trained experts to detect the presence of seafloor seeps, a time-consuming, costly, and inconsistent procedure. Here, we explore the mitigation of this constraint by creating a sonar dataset labeled for seep presence and using it to develop a machine learning-based method for automatic seafloor seep detection. The presented convolutional neural network model achieves an average accuracy of more than 95% in 5-fold cross-validation.