Joint use of machine learning and geostationary satellite data for volcanic ash cloud detection

Abstract

During an explosive eruption, a major hazard to population can be represented by the ejection in the atmosphere of gases and ash, with the consequent creation of volcanic clouds, which can compromise aviation safety. The combined use of a variety of satellite data in different spectral ranges with diverse spatial and temporal resolutions allows us to continuously monitor volcanic ash clouds in an efficient and timely manner. Specifically, the latest generation of high temporal resolution satellite sensors, such as EUMETSAT MSG Spinning Enhanced Visible and InfraRed Imager (SEVIRI) and GOES18 Advanced Baseline Imager (ABI), provide almost continuous radiometric estimates to track the entire evolution of volcanic clouds produced by worldwide volcanoes. Therefore, leveraging the strengths of these satellite sensors, which provide frequent observations at a wide variety of wavelengths, can provide critical information to help understand volcanic processes and extract eruptive products. Nevertheless, the satellite data volume is too large for ad hoc processing and analysis especially when considering daily global-scale observations. Deep learning (DL), a fastest-growing technique of artificial intelligence in remote sensing data analysis applications, has an excellent ability to learn massive, high-dimensional image features and has been widely studied and applied in classification, recognition, and detection tasks involving satellite imagery. Here, we developed a new DL model, based on Deep Convolutional Neural Networks (CNNs), which exploits a variety of spatiotemporal information mainly coming from geostationary satellite sensors. It is trained on a combination of Thermal Infrared (TIR) bands acquired by MSG-SEVIRI and GOES18-ABI. The proposed model aims to extract complex spectral and spatial patterns autonomously to recognize a volcanic ash cloud. Preliminary capabilities and limitations of this model will be presented here. Thanks to the wide area covered by these satellite sensors, it is possible to apply this model to different volcanoes. Specifically, this model has been applied to the paroxysmal events that occurred at Mt. Etna (Italy) between 2020 and 2020 and at Shishaldin (Alaska, USA) in 2023.