A Multi‐Region Deep Learning Framework for Low‐Frequency Earthquake Detection: Application to the Guerrero Seismic Gap

Abstract

Low‐frequency earthquakes (LFEs) provide key constraints on slow earthquake processes but remain difficult to detect because of their low signal‐to‐noise ratios (SNR) and emergent onsets, particularly in offshore environments where seismic networks are sparse. In this study, we develop a deep learning framework for LFE detection trained on a curated multi‐region data set spanning six tectonic environments. To address uncertainties in arrival timing and improve robustness to noisy waveforms, we employ probabilistic Gaussian labels and incorporate multi‐station waveform inputs within a modified U‐Net architecture. The model demonstrates stable training convergence and consistent detection performance on an independent held out test set containing previously unused events from multiple regions. We apply the trained model to continuous ocean‐bottom seismometer (OBS) data in the Guerrero seismic gap in Mexico. After target‐domain threshold calibration, the workflow recovered 21 events from an existing catalog and identified 25 additional offshore LFE candidates. These detections show temporal clustering and a spatial distribution that differed partly from that of the existing catalog. Our results suggest that deep learning may complement matched‐filter detection in regions where template libraries or offshore observations are limited.