Operational Seismic Denoising Workflow to Enhance Seismic Catalogues

Recent studies have demonstrated the potential of deep learning (DL) techniques for denoising seismic signals and improving signal analysis, but they are not yet widely adopted in seismic monitoring. Denoising models are typically applied to short segments of triggered data. To use the full potential of state-of-the-art denoising models for seismicity catalogue generation, the methods need to be applicable to continuous data. Several challenges arise in this case, in particular during dense (aftershock) sequences, as models may fail to consistently detect signals near or across the window edges, or require overlapping windows that lead to several parallel denoised waveform solutions. 

We investigate the optimal integration of denoising approaches into operational network settings to enhance seismic catalogues, focusing on improving detections, phase picks, and peak amplitude measurements. As we are most interested in characterising weak events that are commonly missing or poorly described in existing catalogues,  special attention is given to them. We train and compare a range of promising algorithms, including a method that operates on time-frequency representation of the data and outputs segmentation masks to separate event and noise signals. 

We evaluate the approach on seismic data recorded by the Swiss network, and train a model on recorded noise and about 25k earthquake signals, corresponding to most of the available high-quality, local recordings. 

To assess the benefit of the denoiser, we test it on a dense seismic sequence recorded by different types of seismic sensors under diverse site and noise conditions. We employ the denoiser to detect event signals, and produce continuous denoised data, which then serve as the input for standard phase pickers and event associators. We compare the derived catalogue to those obtained with i) standard and ii) DL tools, both applied on raw data. We demonstrate i) significantly deeper catalogues in the first case, and ii) catalogues comparable to those obtained with DL pickers, but with enhanced characterisation, including event location and magnitudes.