Skip to content

Mitigating Atmospheric Noise in InSAR Displacement Time-series using a Convolutional Neural Network Machine Learning Model

Rebecca Bussard , Christelle Wauthier

Abstract

Atmospheric artifacts are a common cause of noise in InSAR data that can mask deformation signal related to volcanic activity. While several techniques have been developed over the past few decades to mitigate atmospheric noise, the success of these techniques is highly variable depending on the regional setting, and in some cases can even introduce more noise. We introduce a Convolutional Neural Network (CNN) trained to predict deformation from input consecutive unwrapped InSAR displacement maps. The CNN framework consists of a U-net structure that runs the input data through a series of convolution and deconvolution layers to handle the dimensionality of such a large image dataset, as well as to improve the model's generalization capacity. For model training, we simulate ground deformation from a magma storage region using a Mogi point source at depth undergoing various volume changes (linear, step-wise, sinusoidal) through time. The resulting input data consists of 5000 unique time series of 20 unwrapped deformation maps each that are read into the CNN; we also test whether adding simulated atmospheric noise or downloaded atmospheric phase maps from GACOS to these deformation time series trains the CNN to produce better deformation predictions from CosmoSkyMed unwrapped interferograms over Masaya volcano in Nicaragua.