Development of a Machine Learning Classifier to retrieve Time-Series of Ash Componentry at Tungurahua Volcano, Ecuador, 1999-2016

Petrological monitoring of active volcanoes is an often underutilised tool for eruption forecasting due to the high cost and long lead times of petrological analyses, even though these analyses can provide vital context to interpret geophysical monitoring signals¹. Ash componentry specifically is the process of classifying ash particles by grain type (e.g. juvenile, accidental, etc.), and is extremely useful for understanding the state and driving processes of the volcanic conduit and shallow hydrothermal system², potentially helping to anticipate transitions into intensified explosive activity³. However, componentry analysis is time-consuming, and classification of particles can be subject to various classification schemes and interpretations depending on the observer. To overcome these problems, we adopted a machine learning (ML) approach to classify particles in an automatic and consistent manner.

In this work, we describe the development of a ML model, coupled with a tailored classification scheme, to classify ash from a collection of 30 samples between 1999–2016 from Tungurahua volcano, Ecuador. We analysed 180 grains in-depth to develop a systematic classification scheme for optical images of grains based on diagnostic optical features (e.g., colour, lustre, degree of alteration, and edge-angularity) and tested the robustness of the classification using evidence of the grains’ petrogenesis acquired via high-resolution surface imaging on a Quanta-SEM and automated minerology maps on a TIMA. We then imaged ~10,000 grains across the samples using a HIROX HRX-01 digital microscope at the Natural History Museum, London. The images were segmented using FastSAM⁴ and labelled according to our classification scheme. To set up our model, we split the dataset into training and test sets, and we followed the steps described in Benet et al.⁵ to obtain the Volcanic Ash Database (VolcAshDB) classifier. The model classifies relatively accurately, and performance should improve as we collect more particle images and re-train the model. We find that the obtained component proportions as time-series are instrumental to interpret the evolution of the volcanic conduit and shallow storage system throughout the studied period by linking these proportions to concurrent monitoring data such as seismicity or SO₂ flux. This work is carried out in collaboration with the Ecuadorian monitoring authority (IG-EPN), and we aim to create a model that can be operational for near-real-time petrological monitoring of any future activity at Tungurahua volcano, as well as to set out a methodology that can be used to re-train the model for other volcanic systems.

 

1. Re et al. 2021, JVGR, https://doi.org/10.1016/j.jvolgeores.2021.107365.

2. Gaunt et al. 2016, JVGR, https://doi.org/10.1016/j.jvolgeores.2016.10.013.

3. Cashman & Hoblitt. 2004, Geology 32, https://doi.org/10.1130/G20078.1.

4. Zhao et al. 2023, Preprint, arXiv. https://doi.org/10.48550/arXiv.2306.12156.

5. Benet et al. 2024, GGG, https://doi.org/10.1029/2023GC011224.