Machine Learning-Based Automated Detection of Seafloor Gas Seeps

Recognizing the importance of seeps and the significant potential for further seep discovery, a research team has been developing an efficient and cost-effective machine learning-based software system to automatically detect seafloor gas seeps (i.e., cold seeps) in mapping sonar water column data. Seeps are hotspots of deep-sea biodiversity, an important part of the global carbon cycle, indicators of potential marine geohazards, and an energy production resource. However, the deep ocean is largely unexplored, and most seeps are discovered through manual visual analysis of water column sonar imagery, which is time consuming, costly, and inconsistent. This project is designed to automate that process, accelerating the speed, accuracy, and consistency of seep discovery, and resulting in improved understanding of seep abundance and distribution.

Training the Model

The core element of the new system is a machine learning model, which needed to be trained to detect seeps. To train and validate the model, the team relied on multibeam sonar water column data collected by NOAA Ocean Exploration during expeditions along the U.S. Atlantic margin on NOAA Ship Okeanos Explorer. They visually reviewed over 428,396 multibeam sonar water column images — searching for gas bubble plumes — and digitally labeled them, noting the presence or absence of seep targets. Seep targets were present in 6,583 (1.54%) of these images. Where seeps were present, the researchers drew bounding boxes around the seep targets to indicate their locations, which resulted in 8,324 bounding boxes.