Automatic Detection of Volcanic Unrest Using Interferometric Synthetic Aperture Radar

A diverse set of hazards are posed by the world's 1500 subaerial volcanoes, yet the majority of them remain unmonitored. Measurements of deformation provide a way to monitor volcanoes, and synthetic aperture RaDAR (SAR) provides a powerful tool to measure deformation at the majority of the world's subaerial volcanoes. This is due to recent changes in how regularly SAR data are acquired, how they are distributed to the scientific community, and how quickly they can be processed to create time series of interferograms. However, for interferometric SAR (InSAR) to be used to monitor the world's volcanoes, an algorithm is required to automatically detect signs of deformation-generating volcanic unrest in a time series of interferograms, as the volume of new interferograms produced each week precludes this task being achieved by human interpreters. In this thesis, I introduce two complementary methods that can be used to detect signs of volcanic unrest. The first method centres on the use of blind signal separation (BSS) methods to isolate signals of geophysical interest from nuisance signals, such as those due to changes in the refractive index of the atmosphere between two SAR acquisitions. This is achieved through first comparing which of non-negative matrix factorisation (NMF), principal component analysis (PCA), and independent component analysis (ICA) are best suited for solving BSS problems involving time series of InSAR data, and how InSAR data should best be arranged for its use with these methods. I find that NMF can be used with InSAR data, providing the time series is formatted in a novel way that reduces the likelihood of any pixels having negative values. However, when NMF, PCA, and ICA are applied to a set of synthetic data, I find that the most accurate recovery of signals of interest is achieved when ICA is set to recover spatially independent sources (termed sICA). I find that the best results are produced by sICA when interferograms are ordered as a simple ``daisy chain'' of short temporal baselines, and when sICA is set to recover around 1-3 more sources than were thought to have contributed to the time series. However, I also show that in cases such as deformation centred under a stratovolcano, the overlapping nature of a topographically correlated atmospheric phase screen (APS) signal and a deformation signal produces a pair of signals that are no longer spatially statistically independent, and so cannot be recovered accurately by sICA. To validate these results, I apply sICA to a time series of Sentinel-1 interferograms that span the 2015 eruption of Wolf volcano (Galapagos archipelago, Ecuador) and automatically isolate three signals of geophysical interest, which I validate by comparing with the results of other studies. I also apply the sICA algorithm to a time series of interferograms that image Mt Etna, and through isolating signals that are likely to be due to instability of the east flank of the volcano, show that the method can be applied to stratovolcanoes to recover useful signals. Utilising the ability of sICA to isolate signals of interest, I introduce a prototype detection algorithm that tracks changes in the behaviour of a subaerial volcano, and show that it could have been used to detect the onset of the 2015 eruption of Wolf. However, for use in an detection algorithm that is to be applied globally, the signals recovered by sICA cannot be manually validated through comparison with other studies. Therefore, I seek to incorporate a module into my detection algorithm that is able to quantify the significance of the sources recovered by sICA. I achieve this through extensively modernising the ICASO algorithm to create a new algorithm, ICASAR, that is optimised for use with InSAR time series. This algorithm allows me to assess the significance of signals recovered by sICA at a given volcano, and to then prioritise the tracking of any changes they exhibit when they are used in my detection algorithm. To further develop the detection algorithm, I create two synthetic time series that characterise the different types of unrest that could occur at a volcanic centre. The first features the introduction of a new signal, and my algorithm is able to detect when this signal enters the time series by tracking how well the baseline sources are able to fit new interferograms. The second features the change in rate of a signal that was present during the baseline stage, and my algorithm is able to detect when this change in rate occurs by tracking how sources recovered from the baseline data are used through time. To further test the algorithm, I extended the Sentinel-1 time series I used to study the 2015 eruption of Wolf to include the 2018 eruption of Sierra Negra, and I find that my algorithm is able to detect the increase in inflation that precedes the eruption, and the eruption itself. I also perform a small study into the pre-eruptive inflation seen at Sierra Negra using the deformation signal and its time history that are outputted by ICASAR. A Bayesian inversion is performed using the GBIS software package, in which the inflation signal is modelled as a horizontal rectangular dislocation with variable opening and uniform overpressure. Coupled with the time history of the inflation signal provided by ICASAR, this allows me to determine the temporal evolution of the pre-eruptive overpressure since the beginning of the Sentinel-1 time series in 2014. To extend this back to the end of the previous eruption in 2005, I use GPS data that spans the entire interruptive period. I find that the total interruptive pressure change is ~13.5 MPa, which is significantly larger than the values required for tensile failure of an elastic medium overlying an inflating body. I conclude that it is likely that one or more processes occurred to reduce the overpressure within the sill, and that the change in rate of inflation prior to the final failure of the sill is unlikely to be coincidental. The second method I develop to detect volcanic deformation in a time series of interferograms uses a convolutional neural network (CNN) to classify and locate deformation signals as each new interferogram is added to the time series. I achieve this through building a model that uses the five convolutional blocks of a previously state-of-the-art classification and localisation model, VGG16, but incorporates a classification output/head, and a localisation output/head. In order to train the model, I perform transfer learning and utilise the weights made freely available for the convolutional blocks of a version of VGG16 that was trained to classify natural images. I then synthesise a set of training data, but find that better performance is achieved on a testing set of Sentinel-1 interferograms when the model is trained with a mixture of both synthetic and real data. I conclude that CNNs can be built that are able to differentiate between different styles of volcanic deformation, and that they can perform localisation by globally reasoning with a 224 x 224 pixel interferogram without the need for a sliding window approach. The results I present in this thesis show that many machine learning methods can be applied to both time series of interferograms, and individual interferograms. sICA provides a powerful tool to separate some geophysical signals from atmospheric ones, and the ICASAR algorithm that I develop allows a user to evaluate the significance of the results provided by sICA. I incorporate these methods into an deformation detection algorithm, and show that this could be used to detect several types of volcanic unrest using data produced by the latest generation of SAR satellites. Additionally, the CNN I develop is able to differentiate between deformation signals in a single interferogram, and provides a complementary way to monitor volcanoes using InSAR.