Development of Detection Method for CO₂ Seepage in the Ocean Using Machine Learning

CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> seepage detection from the seabed in carbon storage projects is crucial, but conventional methods based on the relationship between dissolved oxygen saturation and CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> partial pressure have resulted in false detections. This study aimed to develop an accurate seepage detection method using machine learning models. We collected 66 seepage data points from natural seepage in Tachibana Bay and artificial seepage off Tateyama, along with 502 non-seepage data points from the Tomakomai demonstration project and other sources. Three machine learning models-Local Outlier Factor, One Class Support Vector Machine, and Random Forest-were evaluated for performance. Random Forest demonstrated superior accuracy. While optimizing input feature combinations enabled more accurate seepage detection compared to conventional indicators, the imbalanced dataset (11.6 % seepage, 88.4 % non-seepage) remains a challenge requiring further data collection.