A Multi-Sensor Machine Learning Framework for Global Volcanic Cloud Monitoring

Volcanic eruptions represent major natural sources of atmospheric pollution, injecting large amounts of silicate ash and volcanic gases, particularly sulfur dioxide (SO₂), into the atmosphere. These emissions generate complex volcanic clouds that pose immediate hazards to aviation safety and public health, while also producing significant long-term environmental effects. When SO₂ reaches the stratosphere, it undergoes chemical transformation into sulfate aerosols, which scatter incoming solar radiation, modify the Earth’s radiative balance, and can influence global climate patterns.

Continuous and systematic monitoring of volcanic activity is therefore crucial, both for rapid hazard mitigation and for advancing long-term climate research. Achieving comprehensive global surveillance of volcanic events requires the capabilities of geostationary Earth orbit (GEO) satellites. Unlike polar-orbiting systems, GEO platforms provide high temporal resolution, enabling near real-time observation of eruption onset and volcanic cloud dispersion. Geostationary satellites, like the Advanced Baseline Imager (ABI) aboard GOES, the Spinning Enhanced Visible and Infrared Imager (SEVIRI) and the Flexible Combined Imager (FCI) onboard Meteosat, and the Advanced Himawari Imager (AHI) on Himawari, continuously generate large volumes of multi-spectral data, capturing dynamic atmospheric processes at high temporal resolution. The spectral signatures of volcanic ash and SO₂ are often subtle and can overlap with those of meteorological clouds, desert dust, or other atmospheric constituents, making traditional threshold-based or rule-based approaches insufficiently robust.

This work aims to develop a global and near real-time monitoring framework for volcanic clouds by synergistically integrating data from major geostationary satellite systems, including GOES, Meteosat, and Himawari. The objective is to leverage their complementary spatial coverage and high temporal resolution to achieve continuous observation of volcanic emissions worldwide. By applying machine learning (ML) techniques to multi-spectral satellite imagery, this study seeks to automatically detect, classify, and track volcanic ash and SO₂ clouds, improving discrimination from meteorological clouds and other atmospheric phenomena. A hybrid deep learning (DL) model combining VGG16 and U-Net was previously trained and validated on MSG-SEVIRI imagery to segment volcanic clouds from explosive eruptive events at Mt. Etna. This study extends that framework to other geostationary satellite platforms with the goal of enabling global, real-time volcanic monitoring. The architecture processes Ash RGB false-color composites, which explicitly highlight volcanic ash and SO₂ clouds resulting from explosive eruptive activity.

By integrating multi-sensor geostationary data with advanced ML methodologies, this approach provides global coverage, rapid response capability, and enhanced accuracy in tracking volcanic cloud evolution worldwide.