Denoising InSAR time series with a convolutional autoencoder and applications to strain rate estimates

We present preliminary results of efforts to denoise interferometric synthetic aperture radar (InSAR) imagery using a convolutional autoencoder-decoder. The rapid increase in availability and acquisition frequency of SAR imagery has led to an increase in image available to the scientific community. These observations present new opportunities to constrain surface deformation that spans various spatial and temporal scales. Recent studies have shown that machine learning approaches are able to successfully classify and locate medium to large deformation signals within InSAR interferograms, such as large co-seismic or volcanic deformation; however, these approaches struggle to consistently classify small or transient deformation signals, such as fault creep events or small co-seismic deformation. Spatial and temporal variations of pressure, temperature, and water vapor in the atmosphere introduce well known spatially coherent noise in interferograms and InSAR time series. This noise contributes to limitations of machine learning algorithms to consistently classify small or transient deformation signals. Here, we present preliminary results designed to denoise interferograms using a convolutional autoencoder-decoder algorithm.

We created our denoising convolutional network by incorporating the convolutional layers of previous deep residual networks as the encoding layers and an inverse of these layers as the decoding layers. We use noisy InSAR time series images as inputs to our network and generic atmospheric correction online services for InSAR (GACOS) corrected time series images as the input labels, returning a denoised time series image. We initially train our network using synthetic time series with simulated ground deformation and atmospheric noise. We develop a set of weights for the network through these synthetic time series, and then we apply the weights to a second training using InSAR time series derived from Sentinel-1 observations over Iran.

We then developed and present a line-of-sight strain map for the validation time series to test the capabilities of our network on an InSAR data product. We present a comparison of the derived strain rates from the uncorrected time series, the GACO-corrected time series, and the machine learning corrected time series.