Forecasting volcanic activity using deep learning

Información del proyecto

DEEPVOLC

Identificador del acuerdo de subvención: 866085

DOI

10.3030/866085

Fecha de la firma de la CE 19 Diciembre 2019

Fecha de inicio 1 Junio 2020

Fecha de finalización 31 Julio 2026

Financiado con arreglo a

• EXCELLENT SCIENCE - European Research Council (ERC)

Coste total

€ 1 999 495,00

Aportación de la UE

€ 1 999 495,00

Coordinado por

UNIVERSITY OF LEEDS
United Kingdom

Periodic Reporting for period 3 - DEEPVOLC (Forecasting volcanic activity using deep learning)

Período documentado: 2024-02-01 hasta 2025-07-31

Resumen del contexto y de los objetivos generales del proyecto

Some 200 million people live within 30 km of a volcano, but accurate forecasting of volcanic activity is difficult. It relies on human expertise at individual volcanoes, but volcanoes can behave in unexpected ways not previously observed at that location. An additional complication is that most volcanoes are not instrumented. A key indicator of volcanic activity is deformation of a volcano's surface due to magma migrating beneath. Recent advances in satellite monitoring now allow us to monitor this deformation at volcanoes worldwide. The overall aim of DEEPVOLC is to apply the latest deep learning algorithms to the satellite data to combine knowledge from all volcanoes that have been active in the satellite-monitoring era. This will enable us to use knowledge of how volcanoes behave globally to identify deformation at volcanoes locally, and forecast how it will evolve. Through working with volcano observatories throughout the project we will deliver tools that can be used to aid in the forecasting of volcanic activity.

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

There are four main pillars towards achieving the goals of DEEPVOLC. The first involves extracting examples of deformation at volcanoes from all satellite data that have been acquired to date. We have made significant progress in achieving this for data from Sentinel-1, which is the only satellite constellation to acquire data routinely over all volcanoes. We have developed algorithms to extract clean time series of deformation from the data and they are now running routinely on the 250 most active volcanoes, and selectively on other volcanoes. They include the development of a deep learning algorithm to count up displacement contours in the deformation maps, which is a critical step in extracting the deformation time series that sometimes fails using existing approaches. The cleaned deformation time series are being placed in a public repository to allow others to also use them to train their machine learning algorithms.

The second pillar involves simulating deformation data to represent all physical processes that can cause deformation at volcanoes. We have analysed all examples of uplift at volcanoes in the literature and found a common pattern in how the uplift evolves in time, which informs our understanding of the physical processes leading to the uplift. We have developed numerical models to simulate the deformation from these processes. We have also carried out an extensive review of what is known about these processes from fields other than volcano deformation, which we use to constrain the range of values we pick for the controlling processes.

The third pillar concerns using both the real and simulated deformation data to train machine-learning algorithms to identify, classify and forecast deformation. We have developed an algorithm that detects new deformation patterns and also detects changes in rate for existing deformation patterns. This now runs routinely on the 250 most active volcanoes. We have also developed prototype deep-learning algorithms to forecast deformation caused by the propagation of magma filled cracks beneath the surface, which are trained on simulated data.

The final pillar involves working with volcano observatories to deliver tools that can be useful to them. So far, we have worked closely with the Icelandic volcano observatory, and also sought input from US volcano observatories.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

The work described above all represents progress beyond the state of the art. This includes the technological developments required to process the data, and the scientific progress in terms of understanding the physical processes in operation at volcanic plumbing systems. Our machine-learning approaches use existing algorithms developed in the field of computer science, but their adaptation and application to deformation time series also represents progress beyond the state of the art.

In the second half of the project, we will continue progress in all four pillars. We will extend our Sentinal-1 deformation time series to include all ~1400 potentially active volcanoes and extend our processing to include data from a new satellite to be launched in 2024 (NISAR), which will also acquire data routinely over volcanoes globally. We will extend our simulated data to include the full range of deformation signals expected at volcanoes, including landslides, tectonics and hydrothermal systems. We will further develop our prototype deep-learning forecasting algorithm, which currently only forecasts magma-filled crack propagation, into an algorithm that can forecast multiple deformation-causing processes. We will visit five representative volcano observatories, covering a range in terms of resources and volcanic settings, and towards the end of the project will run a workshop to which we will invite staff from all observatories globally.