ABSTRACT Earthquake catalogs are built from continuous seismic recordings through a series of steps, including signal detection, phase arrival picking, arrival association, event localization, and amplitude-based magnitude determination. Although each step can be automated, ever-present noise in the waveforms often limits reliable processing to larger events or quieter stations or necessitates manual data review to ensure high-quality results. In this study, we introduce Earthquake Seismogram Denoiser (EQS-denoiser), a deep learning–based denoising model trained on earthquake and noise signals recorded in Switzerland and preprocessed to refine label quality. Our objective is to improve automated analysis of continuous seismic waveform streams across the full network, enhancing earthquake monitoring and characterization. We demonstrate its performance in three key tasks essential for catalog curation: (1) signal detection, in comparison with deep learning pickers; (2) signal denoising, against conventional digital filters; and (3) phase arrival picking using denoised data versus raw data. In all cases, the denoiser outperforms baseline methods, particularly under low signal-to-noise ratio conditions. We integrate EQS-denoiser into an end-to-end earthquake monitoring framework to maximize detections and optimally recover signals from continuous data. Furthermore, we assess how existing phase pickers can best leverage denoised data to improve identification of arrival times and estimate associated timing uncertainty using test-time augmentation. We demonstrate through a case study of a small alpine sequence how EQS-denoiser significantly advances the generation of seismicity catalogs compared with both a deep learning picker–based catalog and the manually reviewed Swiss catalog. In automated processing, denoising enables more reliable signal detection, a greater number of phase picks with fewer false picks, more accurate automated peak amplitude estimation for magnitude determination, and enhanced waveforms for other types of data analysis. This results in a deeper seismicity catalog that, after relative relocation, achieves location quality comparable to the manual-reviewed catalog while extending to lower-magnitude earthquakes.
Dahmen et al. (Thu,) studied this question.