A computer vision-based approach to enhance seismic catalogues

In recent years, AI and deep learning earthquake detectors, combined with an increasing number of dense seismic networks deployed worldwide, have further contributed to the creation of massive seismic catalogs, significantly lowering their magnitude of completeness. However, these automated catalogs are typically released without systematic quality control, and may contain spurious detections, mislocations, or inconsistent magnitudes. In challenging scenarios, such as microseismic monitoring applications, where weak and closely spaced events often overlap in time, pick-based detection and location approaches often fail to reliably associate phases. This leads to missed detections or degraded location accuracy producing seismic catalogues polluted with false or mislocated events. To address this limitation, we present a computer vision-based workflow that integrates waveform-based seismic location methods with deep learning image classification to discriminate real seismic events from noise directly from coherence matrices. These matrices, computed via waveform stacking, exhibit distinct patterns for real events (single, focused maxima) versus noise (blurred, incoherent patterns) hence the problem of cleaning seismic catalogues can be solved as a binary image classification problem. In addition, the robustness of waveform-based location methods allows to obtain an increased resolution in the location of seismic events. Another advantage of this approach is that the training of neural networks can be based entirely on synthetic data. This synthetic-based training removes the need for large labeled datasets, enabling rapid deployment in newly instrumented areas. We validate our workflow using the publicly available COSEISMIQ dataset from the Hengill geothermal area, in Iceland.