Keywords:Machine learning, InSAR
Spatial and temporal variations of pressure, temperature, and water vapor content in the atmosphere introduce significant errors in interferometric synthetic aperture radar (InSAR) observations of ground deformation. Large amplitude and fast deformations produced during earthquakes are routinely measured using InSAR. However, Interseismic and postseismic deformation -- both smaller and slower -- remains extremely challenging to measure due to atmospheric propagation delays, even after correction for known atmospheric conditions, therefore requiring expert interpretation and a priori knowledge of fault systems. To solve this problem, we introduce a deep autoencoder architecture that is tailored to remove atmospheric delays due to turbulences and the layering of the atmoshpere, as well as to identify and extract ground deformation signals of interest. Taking noisy InSAR time series as input and cumulated ground deformation of interest as output, the deep autoencoder we built is purely convolutional, shrinks the time dimension during the encoding, and uses the ground elevation during the decoding. In order to circumvent the lack of training data and the absence of deformation ground truth, we train our model on simulated data in which small synthetic fault deformation signals are heavily corrupted by statistically realistic atmospheric signals. After training on a few million examples of synthetic InSAR time series, application on real InSAR time series reveals very clean and sharp deformation signals on faults known to be deforming, with no a priori knowledge of fault systems. The proposed deep autoencoder approach enables the automation of denoising and signal extraction from InSAR time series, making possible the study of small deformations at a global scale.