Abstract: Monitoring methane bubble plumes seeping from seafloor cold seeps is crucial for marine research and resource exploration. Current cold seep research primarily utilizes remotely operated vehicles and autonomous underwater vehicles equipped with acoustic and optical devices for detection. However, accuracy issues remain in the subsequent data analysis and image recognition stages, necessitating the use of deep learning and intelligent discrimination. Therefore, this study introduces convolutional neural network (CNN) technology for the intelligent identification and classification of seafloor cold seep eruption states. By integrating high-quality video data obtained from marine surveys and manual identification, the bubble coverage index (BCI) is calculated. These data are then used to establish a calibration dataset for cold seep eruption images based on the ResNet model. In the application within the Haima cold seep research area of the Qiongdongnan Basin, the model achieved a recognition accuracy of up to 98.9% through comparative analysis with calibration results. This provides a new method for data analysis and image processing, as well as reliable support for on-site intelligent decision-making and in-situ sample collection of seafloor cold seeps. This novel approach enhances the efficiency of automated cold seep eruption image processing, addresses the challenge of recognizing cold seep eruption states, and offers a foundational tool for environmental assessment and energy exploration of cold seeps.