The making of a Volcanic Ash DataBase (VolcAshDB) and its exploitation with machine learning

The study of volcanic ash and its different components provides key information that can help understand the likely evolution of volcanic activity during early stages of a crisis and possible transitions towards different eruptive styles. However, classifying ash particles into components such as juvenile or lithic is not straightforward. Diagnostic observations may vary depending on the style of eruption, and there is no standardized methodology, which may lead to ambiguities in assigning a given particle to a given class. To address this problem, we created the web-based Volcanic Ash DataBase (VolcAshDB) which is made of > 6,300 multi-focused binocular images of particles from a range of magma compositions and types of volcanic activity (https://volcash.wovodat.org/). For each particle image, we quantitatively extracted 33 features of shape, texture, and color, and visually classified each particle into one of the four main components: free crystal, altered material, lithic, and juvenile. We used the data in VolcAshDB to setup a variety of machine learning-based models aimed at improving ash particle classification. We identified the features that are discriminant of a given particle type through explanatory AI and the Shapley values from the predictions made by an XGBoost model. We have also developed an accurate Vision Transformer model (93% accuracy) that could be potentially used by volcano observatories to obtain a relatively rapid and objective score on a particle-by-particle basis. Such models could be used for petrologic monitoring in a reproducible and systematic manner aiding in making more informed decisions for hazard mitigation.