Towards the Automatic Detection of Volcanic Unrest using Sentinel-1 InSAR data and Machine Learning

Recent improvements in frequency, type and availability of satellite images mean it is now feasible to routinely monitor volcanoes in both urban and remote areas. In particular, Interferometric Synthetic Aperture Radar (InSAR) data can detect volcanic surface deformation, which has a strong statistical link to eruption. LïCSAR, our automated processing chain for Sentinel-1 data (http://comet.nerc.ac.uk/COMET-LiCS-portal/), have produced more than 30,000 interferograms on 1300 active land-volcanoes. The amount of data is too large to be manually analysed on a global basis. Therefore, machine learning algorithms become necessary to detect automatically volcanic ground deformation signals. The proposed method works on wrapped interferograms with no atmospheric corrections. The ground deformations displayed as fringes provide strong low-level visual features for image classification. We extract spatial characteristics of the interferograms using deep Convolutional Neural Networks (CNN). The algorithm provides as a result the probability that a given interferogram contains surface deformation. The positive results (probability >0.5) are checked manually by an expert and fed back for model retraining. Following the retraining process with Sentinel-1 data, the classifier reduced the number of interferograms flagged to approximately 100, which required further inspection, and at least 39 are considered 'true positives', e.g. the ground deformation unrest at Cerro Azul (Galapagos) in March 2017 was detected with a probability of 1.0. 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 show multiple fringes in short duration interferograms. For future work, we will resolve a lack of positive samples with synthetic data (atmospheric+deformation) and we will also test our algorithm with different InSAR products (unwrapped or stacked 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 volcanic unrest alert systems based on satellite imagery to support risk assessment on volcano observatories.