Conference Agenda

Satellites systematically acquire imagery that can be used for volcano monitoring, characterising magmatic systems and potentially forecasting eruptions on a global scale. However, exploiting the large dataset is limited by the need for manual processing and inspection, meaning timely dissemination of information is challenging. Many new machine learning algorithms and architectures have been proposed for detecting and locating volcano deformation. Here we focus on deep learning methods developed by the COMET group and their application to the COMET-LICSAR database. We use a 3-stage approach to analyse the large dataset of satellite imagery available from Sentinel-1. The first step is the automatic generation of InSAR images using the COMET-LICSAR automated processing system (Lazecky et al, 2020,2021). The next step is the automatic analysis of the processed images, which requires a machine learning approach (Anantrasirichai et al, 2018,2019a,2019b). For volcano monitoring, false negatives are far more problematic than false positive, so we use a conservative thresholding approach. Finally, expert reviews are needed to identify the true positives and characterise the signal at each volcano making use of any external information available (Biggs et al, 2023). In this talk, we will provide an overview of the latest developments in deep learning from the COMET group, covering both method development and the real-time application to large datasets. The expanding dataset of systematically acquired, processed and flagged images enables the quantitative analysis of volcanic monitoring signals on an unprecedented scale, but tailored processing is still needed for routine monitoring applications.

Our first approach uses a Convolutional Neural Network to distinguish between deformation and atmospheric artefacts in individual interferograms. We use a transfer-learning strategy to fine-tune the AlexNet architecture using a combination of real and synthetic data (Anantrasirichai et al, 2018, 2019a,2019b). This method is now applied in real-time to flag volcano deformation in automatically processed Sentinel-1 imagery through the COMET volcano deformation portal: https://comet.nerc.ac.uk/comet-volcano-portal/. It can also be applied retrospectively to create a catalogue of past events and we summarise the results from a dataset of ~600,000 automatically processed interferograms covering >1000 volcanoes from 2015-2020 (Biggs et al, 2023). Of the 16 volcanoes flagged most often, 5 experienced eruptions, 6 showed slow deformation, 2 had non-volcanic deformation and 3 had atmospheric artefacts. Finally, we summarise the latest proof-of-concept studies including unsupervised and semi-supervised deep learning methods.

Ground deformation is a key indicator of volcanic activity and routine acquisition by the Sentinel-1 mission now provides the possibility to monitor volcano deformation globally, with at least one acquisition every twelve days. Building on the low-resolution COMET processing chain, we have developed a system to routinely apply radar interferometry (InSAR) at higher resolution, whenever a new Sentinel-1 image is acquired, and extract the ground deformation. As there are too many images to inspect individually, we have developed an automated machine learning approach, based on independent component analysis, to identify any new deformation patterns and also any changes in the rate of existing deformation patterns, both of which are key indicators of changes in activity. In addition, we have developed a deep-learning based algorithm to automatically classify the potential source of the deformation.

Our key current goal is to forecast how a volcano might deform in the future, based on a time series of interferograms up to the present day. As this is analogous to a video prediction problem, we are testing various deep-learning algorithms from this field and will report on the results here. Training of these networks requires a large data set of deformation time series, and we are therefore processing all available SAR data acquired over volcanoes. This will still leave us far short of sufficient examples and we have therefore also developed a deformation simulator, based on physical models of various deformation processes that occur at volcanoes.

We aim for our forecasts to be a useful tool for volcano observatories, and we are working with pilot observatories to achieve this. In addition, we expect the resulting forecasts to highlight common deformation sequences operating at volcanoes, leading to deeper understanding of the underlying processes. Already, characterising how deformation develops in time, which feeds into the building of our deformation simulator, has led us to new discoveries about generalisable underlying processes operating at volcanoes undergoing uplift.

The Earth’s subaerial volcanoes pose a variety of threats, yet the vast majority remain unmonitored. However, with the advent of the latest synthetic aperture radar (SAR) satellites, interferometric SAR has evolved into a tool that can be used to monitor the majority of these volcanoes. Whilst challenges such as the automatic and timely creation of interferograms have been addressed, further developments are required to construct a comprehensive monitoring algorithm that is able to automate the interpretation of these data.

To monitor volcanoes using SAR data, we have previously developed an algorithm that uses independent component analysis to first separate deformation signals, topographically correlated atmospheric phase screens, and turbulent atmospheric phase screens in time series of InSAR data. Using these separated signals, the algorithm is able to detect both when an existing signal changes in rate (e.g. accelerating deformation), and when a new signal enters a time series (e.g. new deformation). These two detection metrics make our algorithm especially useful for routine global monitoring as we detect changes in time series that indicate a volcano has entered a period of unrest, rather than simply detecting deformation.

We present results from a new algorithm that incorporates both our original algorithm and enhancements required for routine global monitoring. We term this algorithm LiCSAlert, and it addresses multiple challenges that our previous algorithm identified, such as working with deformation signals that are of low rate or those that are correlated with topography, providing visualisation tools to allow the status of all the ~1300 monitored volcanoes to be assessed easily, and providing web-based visualisation of results for use by other organisations (e.g. volcano observatories).

To detect deformation signals that are either low-rate, or display subtle changes in rate, we present results of how we updated our algorithm to maximise the variance of deformation signals, and to then monitor these signals for changes in rate. To detect deformation signals that are correlated with topography, we present results of how we updated the LiCSAlert algorithm to use tICA to recover temporally (rather than spatially) independent signals. To verify the results that our new approach yields, we apply the LiCSAlert algorithm to volcanoes at which there are independent (GPS derived) displacement measurements. At Campi Flegrei, our algorithm is able to isolate deformation and to detect subtle changes in its rate, and at Vesuvius our algorithm is able to isolate low-rate deformation that is coincident with a topographically correlated atmospheric screen.

For global monitoring, the ~1300 volcanoes that we monitor in both ascending and descending time series produce ~2600 LiCSAlert results, which update every ~12 days (and will increase in frequency after the launch of Sentinel-1C). Our novel 2D visualisation tool allows these ~2600 results to be easily interpreted, and functions by assigning a probability of unrest due to a new signal entering a time series in one dimension, and to unrest due to the change of an existing signal in the second dimension. We show how this representation varies through time from 2018 as new data are acquired by Sentinel-1, and present examples of interesting unrest episodes that we detect.

To disseminate our results, we have made our results available to view and explore via an online tool. We demonstrate how this can be used in conjunction with interferograms from LiCSAR and time series from LiCSBAS by other parties to utilise our volcano monitoring results.

Even though they are situated on the Pacific Ring of Fire, the threat of volcanic eruption is often underestimated in British Columbia and the Yukon Territory in Canada. Within this zone of active tectonism are dozens of potentially active volcanoes, many of which erupted during the Quaternary period. Canada is home to 348 known vents that are Pleistocene in age or younger, and 54 of these vents are known to have been active in the Holocene. The annual probability of an eruption in Canada has been estimated at 1/200 for any eruption, and 1/3333 for a major explosive eruption. Notable recent events are the ~220 BP eruption at Tseax cone which reportedly resulted in ~2000 fatalities to the Nisga’a First Nation as well as the ~2360 BP eruption at Mt. Meager, a major Plinian eruption which had a Volcanic Explosivity Index (VEI) of 4 and dispersed pumiceous tephra over thousands of kilometers throughout western Canada. However, with no eruptions in living memory and no systematic monitoring in place, the false perception that Canada's volcanoes are extinct persists.

In 2021 the Geological Survey of Canada undertook a systematic inventory and relative threat ranking of the volcanic hazard as it exists in Canada. This study used a well-established method developed by the United States Geological Survey (USGS) as part of a National Volcano Early Warning System (NVEWS). Known volcanic vents were lumped into 28 volcanic fields and complexes and were assigned a threat score based on geology/eruptive history and exposure factors. Each volcano grouping was assigned both an “overall threat score” and an “aviation threat score”. Of the 28 sites considered, Mt. Meager and Mt. Garibaldi ranked “very high threat” while Mt. Cayley, Mt. Price and Mt. Edziza ranked “high threat”. This relative threat ranking has major implications for developing an optimal monitoring strategy using finite resources.

In this study, we present an overview of the Government of Canada's first operational volcano monitoring system. We describe how the highest priority monitoring sites were determined and how the RADARSAT Constellation Mission, Canada’s newest generation of Earth observation satellites, as well as other SAR platforms are leveraged to provide an efficient and cost-effective monitoring system. We describe the cloud-based infrastructure of our fully automated InSAR monitoring system and provide details on how the system ingests, processes, stores and disseminates InSAR deformation results for interpretation. We present what is, to our knowledge, the highest revisit frequency (4-day) satellite based InSAR measurements routinely observed over volcanic hazards. We discuss the implications of low period InSAR on temporal decorrelation and maximum observable displacement rates and discuss our geological framework for discriminating between slope instabilities, glacial movement and magmatic deformation. We discuss a use case of this temporally dense dataset in constructing a training dataset for automated deformation detection. Finally, we discuss possible implementations of time-series for use in monitoring, e.g. ground deformation or probability-of-unrest as a function of time.