Creation of a Volcanic Ash DataBase (VolcAshDB) and exploitation with machine learning

Volcanic ash provides unique pieces of information that can help understand the likely evolution of volcanic activity during early stages of a crisis and possible transitions towards different eruptive styles. Ash contains different types of particles that are indicative of ascent-related processes. For instance, identification of the so-called juvenile particles can be associated with the movement of magma at shallow depths. However, classifying ash particles is not straightforward. Diagnostic observations may vary from one sample to another, and a standardized classification methodology hasn’t emerged yet, hampering reproducibility and robustness of classification. To address this problem, we created a new Volcanic Ash DataBase (VolcAshDB), a web-based platform aimed at hosting a curated dataset of classified particle images. Using the contents of about 6,300 particles, we have trained various machine learning (ML) models for particle classification. In my talk, I will provide an overview of the methodology for creating VolcAshDB, and discuss our exploration of different ML models for classification. This includes an XGBoost model with high interpretability, and a Vision Transformer, which achieved a classification accuracy of 93%. ML models could be potentially used by volcano observatories to monitor volcanic ash componentry during activity, aiding more informed decisions for hazard mitigation. To attend online (Zoom meeting): https://u-paris.zoom.us/j/81632721094?pwd=Mi9FQmVCNjFCaStpSEt2NFFjaHJyZz09 Meeting ID : 816 3272 1094 Password : 642426