Search

Satellite radar backscatter has the potential to provide useful information about the progression of volcanic eruptions when optical, ground-based, or radar phase-based measurements are limited. However, backscatter changes are complex and challenging to interpret: explosive deposits produce different signals depending on pre-existing ground cover, radar parameters and eruption characteristics. We use high temporal- and spatial-resolution backscatter imagery to examine the emplacement and alteration of pyroclastic density currents (PDCs), lahar and ash deposits from the June 2018 eruption of Volcan de Fuego, Guatemala, using observatory reports and rainfall gauge data to ground truth our observations. We use a temporally dense time series of backscatter data to reduce noise and extract deposit areas. We observe backscatter changes in six drainages, the largest deposit was 11.9-km-long that altered an area of 6.3 urn:x-wiley:21699313:media:jgrb55183:jgrb55183-math-0001 and had a thickness of 10.5 urn:x-wiley:21699313:media:jgrb55183:jgrb55183-math-00022 m in the lower sections as estimated from radar shadows. The 3 June eruption also produced backscatter signal over an area of 40 urn:x-wiley:21699313:media:jgrb55183:jgrb55183-math-0003, consistent with reported ashfall. We use transient patterns in backscatter time series to identify nine periods of high lahar activity in a single drainage system between June and October 2018. We find that the characterisation of subtle backscatter signals associated with explosive eruptions are best observed with (1) radiometric terrain calibration, (2) speckle correction, and (3) consideration of pre-existing scattering properties. Our observations demonstrate that SAR backscatter can capture the emplacement and subsequent alteration of a range of explosive deposits, allowing the progression of an explosive eruption to be monitored.

Radar satellites, such as Sentinel-1, are now able to produce time-series of ground deformation at any volcano around the world, but atmospheric effects still limit the real time detection of unrest at tropical volcanoes. Here, we test two approaches to correct atmospheric errors - phase-elevation correlations and global weather models - and assess the ability of InSAR time-series to detect deformation anomalies using either a fixed threshold or a cumulative sum control chart. We use the 2017-2018 crisis at Agung volcano as a case example because strong atmospheric signals were originally misidentified as true deformation, and obscured the subtle deformation pattern associated with magmatic activity. We assess the Receiver Operating Characteristics (ROC) of each method and found the average area under the ROC curve to be about 0.5 for the uncorrected data (corresponding to no discrimination capability), around 0.8 after combined atmospheric corrections (weather model and phase-elevation approaches), and more than 0.95 using a cumulative sum control chart (where 1 corresponds to ideal separation between classes). Our results retrospectively show that uplift could have been detected to a 95% level of confidence for both ascending and descending time series by October 2017, 15 days after the start of the seismic swarm and one month prior to the eruption. Thus our approach successfully flags anomalous behavior without relying on visual inspection or selection of an arbitrary threshold, and hence shows potential as a monitoring tool for volcano observatories globally.

Recent improvements in the frequency, type and availability of satellite images mean it is now feasible to routinely study volcanoes in remote and inaccessible regions, including those with no ground-based monitoring. In particular, Interferometric Synthetic Aperture Radar (InSAR) data can detect surface deformation, which has a strong statistical link to eruption. However, the dataset produced by the recently-launched Sentinel-1 satellite is too large to be manually analysed on a global basis. In this study, we systematically process >30,000 short-term interferograms at over 900 volcanoes and apply machine learning algorithms to automatically detect volcanic ground deformation. We use a convolutional neutral network (CNN) to classify interferometric fringes in wrapped interferograms with no atmospheric corrections. We employ a transfer learning strategy, and test a range of pretrained networks, finding that AlexNet is best suited to this task. The positive results are checked by an expert and fed back for model updating. Following training with a combination of both positive and negative examples, this method reduced the number of interferograms to ~100 which required further inspection, of which at least 39 are considered 'true positives'. We demonstrate that machine learning can efficiently detect large, rapid deformation signals in wrapped interferograms, but further development is required to detect slow or small deformation patterns which do not generate multiple fringes in short duration interferograms. This study is the first to use machine learning approaches for detecting volcanic deformation in large datasets, and demonstrates the potential of such techniques for developing alert systems based on satellite imagery.

Stress transfer between separate magma bodies is evaluated by considering how pressure changes related to magma accumulation/propagation influence the stability of a separate nearby magma body. Three-dimensional numerical models are used to evaluate the stability evolution of a magma body through the calculation of two variables: (i) the variation of the threshold pressure needed to cause failure around the magma body and (ii) the magma pressure change. A parametric study indicates that stress interactions are strongly dependent on the distance between magma bodies as well as the body's shape. Such models are then applied to evaluate stress influence of intrusive activity in 1994, 1999, and 2010 at Eyjafjallajokull volcano, which preceded two eruptions there in 2010. Two cases are considered: influence of these intrusions on (i) a magma reservoir at 20 km distance under the Katla volcano and (ii) a silicic magma body under Eyjafjallajokull. The distance between the Eyjafjallajokull intrusions and the Katla reservoir is sufficiently long to reduce the stress interaction to insignificant levels, with an amplitude of the same order as Earth tides (a few kilopascals). However, cumulative stress transfer due to the intrusions to a remnant silicic shallow body situated below the Eyjafjallajokull is much larger (0.5 2.5 MPa). This mechanical transfer could have contributed to the failure of the silicic body and promoted the chemical mixing/mingling between different magma types, which is commonly interpreted as the main cause of the 2010 explosive eruption of Eyjafjallajokull.

There are some 1,500 volcanoes with the potential to erupt, but most are not instrumentally monitored. However, routine acquisition by the Sentinel‐1 satellites now fulfils the requirements needed for interferometric synthetic aperture radar (InSAR) to progress from a retrospective analysis tool to one used for near‐real‐time monitoring globally. However, global monitoring produces vast quantities of data, and consequently, an automatic detection algorithm is therefore required that is able to identify signs of new deformation, or changes in rate, in a time series of interferograms. On the basis that much of the signal contained in a time series of interferograms can be considered as a linear mixture of several latent sources, we explore the use of blind source separation methods to address this issue. We consider principal component analysis and independent component analysis (ICA) which have previously been applied to InSAR data and nonnegative matrix factorization which has not. Our systematic analysis of the three methods shows ICA to be best suited for most applications with InSAR data. However, care must be taken in the dimension reduction step of ICA not to remove important smaller magnitude signals. We apply ICA to the 2015 Wolf Volcano eruption (Galapagos Archipelago, Ecuador) and automatically isolate three signals, which are broadly similar to those manually identified in other studies. Finally, we develop a prototype detection algorithm based on ICA to identify the onset of the eruption.