Convolutional autoencoders for earthquake anomaly detection in InSAR

Interferograms created from InSAR satellite data have become a cornerstone of Earth surface deformation analysis due to their wide spatial coverage. However, our ability to detect spatially focused anomalies from earthquakes is largely impacted by atmospheric signal delays and processing errors. This work explores anomaly detection related to earthquake events using convolutional autoencoders. Our approach shows great potential in distinguishing unwanted atmospheric artifacts and enhancing the feature detection of surface deformations. As a demonstration we explore two case studies of recent earthquakes in Mexico: the 2017 Puebla earthquake and the 2018 Oaxaca earthquake. We generated interferograms using Sentinel-1 data and the JPL/Caltech/Stanford InSAR Scientific Computing Environment (ISCE). In contrast to other machine learning techniques such as support vector machines and isolation forests, our analysis shows that convolutional autoencoders are more effective in leveraging spatial correlation information related to an earthquake event. We also show that in cases where deformations are very close to the noise levels, additional data fusion with other instruments might be required to obtain a reliable classification. We acknowledge support from NASA AIST80NSSC17K0125 (PI Pankratius) and NSF ACI1442997 (PI Pankratius).