VolcAshDB open services for visualization and semi-automated classification of volcanic ash particles

Damià Benet, Kévin Migadel, Fidel Costa

Volcanic ash particles contain critical pieces of information about their origin and the processes driving the eruptive activity. However, the classification of particles into different types (juvenile, lithic, free crystal, altered material) is not standardized and varies from observer to observer. As a result, datasets produced by different research groups are often difficult or not possible to compare, which limits the intercomparison and thus generalization between eruptive episodes within and between volcanoes. To address this, we developed the Volcanic Ash DataBase (VolcAshDB), an open repository of ash-particle data accessible through a web platform (volcashdb.ipgp.fr), and used it to build machine-learning classifiers for more objective and reproducible classification. Here, we present three new web services for particle visualization and semi-automatic classification. We show how users can browse and download different datasets of particles using advanced filters, generate visual summaries, and compare particle populations across volcanoes and eruptive styles. We also provide a service for semi-automatic ash particle classification that requires small prior user knowledge based on a Vision Transformer (ViT) trained on the VolcAshDB dataset in January 2024. The service includes (i) optional image processing to help users process images before upload, (ii) a pre-classification check that evaluates whether uploaded images are similar to the training set using an Out-of-Distribution (OOD) method, and (iii) an automatic classification into the main particle types using probabilities. We evaluated the classifier on 600 new multi-focused binocular particle images from five volcanoes. The model yielded an average accuracy of 0.86. The pre-classification check adds a reliability layer by identifying uploads that fall outside the range covered by the training images, helping users interpret predictions. Together with the relatively good performance on new data, the classification service should be useful for componentry studies. As VolcAshDB grows to include additional samples and imaging setups, we expect performance to further improve, supporting a more robust and generalizable classifier for global comparisons of ash componentry and, ultimately, near real-time petrologic monitoring of volcanic eruptions.