Forecasting volcanic eruptions using data fusion
PI: Christelle Wauthier, Department of Geosciences, EMS & Associate Director, ICDS
About 30 million people (about the population of Texas) live within 10 km (about 6.21 mi) of a volcano that erupted at least once in the last 10,000 years (Brown et al., 2015) and more than tens of thousands of human lives were lost from direct volcanic hazards (Brown et al., 2017). Volcanic eruptions pose a clear danger to society, with local hazards for the populations close to the volcano) but also globally (air traffic and climate disruptions) (e.g., Cassidy and Mani, 2022). Very recently, 60 people were killed by a phreatic eruption at Ontake Volcano, Japan; 22 tourists died at White Island, New Zealand in the 2019 tragedy, and over two hundred people were killed by pyroclastic flows at Fuego Volcano, Guatemala in 2018. Forecasting volcanic eruptions and their associated hazards is the Holy Grail of volcanology but remains a big challenge at this point (Poland and Anderson, 2020) as highlighted as one crucial priority in the he National Academies of Sciences, Engineering, and Medicine report (2017, 2018), Furtney et al. (2018) and Reath et al. (2018) qualitatively compared time series from various remote sensing datasets and seismic RSAM (Real-time Seismic Amplitude Measurement) over active volcanic areas. We aim to build on these efforts and develop a quantitative machine learning approach for pattern recognition in the multistream time series.
Massive Interferometric Synthetic Aperture Radar (InSAR) data sets routinely acquired by SAR satellites, such as Sentinel-1 and the upcoming NASA NISAR mission, provide geodetic measurements with a mm to cm accuracy (Hanssen, 2001; Massonnet & Feigl, 1998) in favorable conditions over broad areas globally. In particular, InSAR time series measurements have enabled the detection of subtle longer-term deformation due to tectonic motion and creep (Burgman, 2006), magmatic and hydrothermal processes (Wauthier et al., 2018), water pumping/city subsidence (Osmanoğlu et al., 2016), and slow landslides (Gonzalez Santana and Wauthier, 2021). In this project, we will first process Sentinel-1 InSAR time-series since 2014 for the three following targets affected by active volcanic and tectonic processes: Big Island, Hawaii; Pacaya Volcano, Guatemala; and Masaya Volcano, Nicaragua. We will then compile other relevant geophysical ground-based datasets such as seismic Real-time Seismic Amplitude Measurement (RSAM) which is recognized as a good first-order volcano monitoring and prediction tool (e.g., Endo and Murray, 1991), GNSS/GPS if available, as well as other remotelysensed datasets that are good indicators of volcanic unrest such as SO2 measurement made with OMI and thermal anomalies using MODIS (Wright et al, 2012). We will validate our forecasting approach in hindsight since we have a good record of recent last decade eruptions at all selected sites. Machine learning as a tool for prediction and anomaly detection has developed rapidly over the last couple of decades. Through several research studies, machine learning has been proven to be successful in identifying data anomalies (Azamathulla et al., 2010; Choubin et al., 2017; Peterson et al., 2019; Xu, 2019; Peterson et al., 2020; Bhadra et al., 2020; Hazarika et al., 2020) including anomaly detection in SAR imagery (Muzeau et al., 2022) and InSAR deformation (Brengman and Barnhart, 2021; Sun et al., 2020; Anantrasirichai et al, 2019). Although machine learning approaches are still often viewed as “black-boxes” and their use in volcanology and eruption forecasting is still in its infancy (Poland and Anderson, 2020). We will push this boundary thanks to this project.
List of specific areas of computational and/or data science expertise or skills: AI and/or data sciences using multiple streams of time series.
List of specific objectives for work supported by this call: The method should be successful at detecting “post-mortem” in hindsight volcanic eruptions at the targeted sites and will lead to at least one publication. The results will then support grant proposals to NSF, NASA, and other relevant opportunities (for example to the NSF CAIG call in February 2026).
Connection of the project to ICDS’s mission: We will develop and apply data sciences and AI methods to volcanic hazards processes and hope to improve eruption forecasting globally.