Towards Improved Forecasting of Volcanic Hazards Using Machine Learning Applied to InSAR Data

There are ~ 1400 subaerial volcanoes with the potential to erupt, but less than 10% are instrumentally monitored. Routine acquisition by the European Sentinel-1 radar mission now offers the potential to monitor most of them with at least two acquisitions every twelve days. We have developed a system to routinely apply radar interferometry (InSAR) whenever a new image is acquired by Sentinel-1 over a volcano. Displacement of the ground between images shows up in the resulting interferograms, but there are too many of these to inspect individually. We have therefore developed an automated approach to identify signs of both new deformation patterns and changes in rate for existing deformation patterns, using independent component analysis. We first use a set of training interfero-grams to identify components associated with background deformation and common atmospheric patterns. We then analyse new interferograms in the context of those components, and flag changes in rate above the background variation, and significant unexplained new signals. We have found this approach to be successful even without having to first explicitly estimate the atmospheric signal in individual interferograms. We have also developed an alternative algorithm using a deep-learning approach, with the aim of extending this approach to forecasting of volcanic activity. This is implemented with a “double-headed” convolutional neural network, which is able to both locate deformation and classify the type of deformation, in single interferograms. We achieve this through creating a large dataset of synthetic interfero-grams, which features labels of both the type and location of any deformation, and is used for initial training of the network. We then fine-tune the network's performance through the inclusion of a smaller amount of real data.