Submarine Volcanism Identification with Neural Networks

Volcanic feature identification on topographic maps is an emerging area that can leverage advanced machine learning techniques. However, the field faces challenges due to the scarcity of current labeled volcanic data, class imbalances, and limitations in existing models. Current approaches often fail to effectively capture the spatial continuity and hierarchical patterns inherent in volcanic terrains, in particular for intraplate volcanism such as petit-spot volcanoes.

This study began with manually labeled data and employed Support Vector Machines as a preliminary classification tool, which revealed significant limitations in its handling of complex spatial patterns. Subsequently, efforts shifted to Convolutional Neural Networks (CNNs) with transfer learning to enhance feature extraction and classification. High-resolution topographic data from the Japanese Volcanic Islands and the Second Institute of Oceanography Ministry of Natural Resources of China were integrated. These datasets, enriched by grid-based labeling and data augmentation strategies, have provided the foundation for model training and validation. To date we have prioritized the identification of small volcanic features on Pacific-type (fast-spreading) seafloor. 

Work to date has emphasized the need for high-quality labeled datasets and innovative preprocessing techniques to have reliable machine recognition of volcanic bathymetric features. By incorporating high-precision data from Pacific expeditions, this research will also contribute to the development of future deep learning approaches, laying the groundwork for further advancements in automated volcano identification.