Machine learning for volcano deformation: detection, interpretation and forecasting

Ground deformation is a key indicator of volcanic activity and routine acquisition by the Sentinel-1 satellite mission now allows us to monitor volcano deformation globally. However, for the data to be used in an operational way, a large amount of time-consuming processing and interpretation is needed, which is often not feasible for individual volcano observatories. We have therefore developed a system to routinely process data for volcanoes globally, and machine learning tools for detection, interpretation and forecasting, to rapidly produce useful products. Analysis of our freely-available global data set also highlights common deformation sequences operating at volcanoes, leading to deeper understanding of the underlying processes.

Our system routinely applies radar interferometry (InSAR), whenever a new Sentinel-1 image is acquired over a volcano, updates the time series, and makes them available in a portal (https://comet.nerc.ac.uk/comet-volcano-portal), which can be used directly to check activity at volcanoes of interest. However, as there are too many images to inspect routinely, we have developed an automated machine-learning approach, based on independent component analysis, to identify new deformation patterns and also changes in rate for existing patterns, both of which are key indicators of new activity. We find this approach also does better at estimating and reducing atmospheric signal than standard approaches. We then use deep learning to extract meaningful indicators of activity from the multiple independent component time series produced per volcano.

To provide quick indicators on the sources of any ground deformation we have developed a deep learning approach to localise deformation patterns and provide a first estimate of the source parameters causing the deformation, e.g. type of source, location and volume change. Our current goal is forecasting how a volcano might deform in the future, based on a time series of interferograms up to the present day. To this end, we have tested various deep-learning algorithms from the field of video prediction and are working on incorporating physical constraints, using physics-informed deep learning approaches.

Training of these networks requires a large data set of deformation time series so, in addition to processing all the available SAR data acquired over volcanoes, we also simulate data using physical models of various deformation processes that occur at volcanoes.  This has led us to new discoveries about generalisable underlying processes operating at volcanoes undergoing uplift.