How Machine Learning and Satellite Data Enhance Near-Real Time Detection of Volcanic Activity Worldwide

Abstract

Nowadays, several satellite missions provide thermal infrared data at various spatial resolutions and revisit time, enabling nearly continuous monitoring of thermal volcanic activity worldwide. In addition, computer vision techniques, based on Machine Learning (ML) and Deep Learning (DL) models, offer unique advantages in automatically extracting valuable information from large datasets. Here, we propose a Machine Learning approach that leverages the capabilities of such models, combined with nearly continuous high spatial resolution images (20 m) acquired from the Sentinel-2 MultiSpectral Instrument (S2-MSI), to detect high-temperature volcanic features and to quantify volcanic thermal emissions. The impact of the volcanic activity is assessed based on the spatial distribution of the erupted products. Spatially characterizing the detected thermal anomalies allows us to both highlight significant pre-eruptive thermal changes and estimate the areal coverage, length of the erupted products, and the lowest altitude reached by them. We utilize the entire archive of high spatial resolution Sentinel-2 data, which comprises more than 6000 S2-MSI scenes from ten different volcanoes around the world. By employing a “top-down” cascading architecture that integrates two distinct Machine Learning models, a scene classifier (SqueezeNet) and a pixel-based segmentation model (Random Forest), we achieve very high accuracy, specifically 95%. This result comes from overcoming the limitations of the scene-level DL classification model, which compresses the entire spatial and spectral information into one unique label, by relying on the pixel-level Random Forest (RF) model. The use of multiple models allows to create more robust and powerful predictors making this tool suitable for near real time volcanic activity detection. These results demonstrate that the cascading system processes any available S2-MSI image in near-real time, providing a significant contribution to the monitoring, mapping, and characterization of volcanic thermal features worldwide.