Deep learning for denoising High-Rate Global Navigation Satellite System data

Abstract

High-rate global navigation satellite system (HR-GNSS) data records ground displacements and can be used to identify earthquakes and slow slip events. One limitation of such data is the high amplitude, cm-level noise which make it difficult to identify processes that produce surface displacements smaller than these values. Deep learning has proven adept at performing many useful tasks in seismology and geophysics. Here we explore using deep learning to denoise HR-GNSS data. We develop three different convolutional neural networks with similar architectures but different targets. Training data are synthetic HR-GNSS records and actual noise recordings that are superimposed to generate noisy signals. We train each of the three models to output masks that can be used to reconstruct the true signal. We use a set of performance metrics that quantify the models’ ability to denoise the testing data and find that denoising significantly improves the signal-to-noise ratio and the ability to identify first arrivals. Finally, we test the models on HR-GNSS records from the Ridgecrest earthquakes recorded at stations that have nearly colocated strong-motion sites used ground-truth the denoising results. We find that the models greatly improve the signal-to-noise ratios in these records and make the P-wave onset clearly identifiable.