Forecasting volcanic activity using deep learning

DEEPVOLC will radically advance the way future activity is forecast at volcanoes by applying advances in artificial intelligence to transformative new geodetic datasets. 200 million people live within 30 km of a volcano. Accurate forecasting of volcanic eruptions is problematic because 1) it relies on human interpretation at individual volcanoes, 2) a volcano can behave in unexpected ways not previously seen at that location, and 3) most volcanoes are not instrumented. DEEPVOLC will address this by i) applying artificial intelligence, ii) using data for all volcanoes worldwide, and iii) exploiting advances in satellite monitoring. A key indicator of potential volcanic activity is deformation of a volcano's surface due to magma migrating beneath. Surface movements as small as a few millimetres can now be measured from space, using satellite-borne radar. A recently-launched European satellite mission, Sentinel-1, has transformed our ability to measure surface deformation at all of the world's volcanoes, acquiring data at least twice every twelve days. However, forecasting how deforming volcanoes will behave in the future remains challenging. In this project I will apply recently developed deep learning approaches to the satellite data. This is an entirely new approach to forecasting volcanic activity, which currently relies on the individual expertise available at each observatory, and which is only now made possible due to the launch of Sentinel-1 and advances in deep learning algorithms. DEEPVOLC will combine knowledge from all volcanoes that have been active in the era of satellite deformation observations, and will continue to improve as it ingests data from new activity. The main deliverable will be a system for volcano observatories that uses knowledge of how volcanoes behave globally to automatically identify deformation at volcanoes locally, and forecast how the deformation will evolve, indicating the probability of eruption.