Continuous monitoring of volcanoes is essential for advancing scientific understanding and issuing alerts to at-risk populations. Many volcanoes are equipped with ground-based RGB cameras, whose recordings are analyzed by experts or semi-automatic models. Automatically distinguishing clouds from volcanic columns and other emissions remains a major challenge. To date, no open-source labeled database of volcanic images exists, and no study has attempted to localize multiple emissions beyond ash columns. We introduce VIGIA-PlumeNet, capable of operating across diverse environments a,nd sky conditions. It performs multi-class segmentation, identifying the volcano and distinguishing plumes, gases, and lava, with 88 % accuracy on the 58 test images when excluding the background class. This is achieved by using the versatile DINO-v2 AI model as a general visual processor, paired with a specialized component that we specifically trained to pinpoint and outline volcanic features; and constructing VIGIA-PlumeData, a dataset of 250 manually annotated images from over 60 volcanoes worldwide, collected through community effort. VIGIA-PlumeData and VIGIA-PlumeNet are released publicly, establishing the first benchmark dataset for volcanic image segmentation.
Add the full text or supplementary notes for the publication here using Markdown formatting.