Comparative Evaluation of Deep Learning Based Earthquake Detection Pipelines
Year: 2026
Type: article
Abstract: : In this study, we present a systematic comparison of earthquake processing pipelines combining two deep learning phase pickers, EQTransformer and PhaseNet, with the built-in EQTransformer associator and PyOcto associator. The analysis focuses on the 2019 Silivri earthquake sequence using three-component waveform data from 29 stations of the Kandilli Observatory network. For each pipeline, associated phase picks were located using HypoSeisan with a region-specific 1D layered velocity model for Marmara. Performance was evaluated across P-wave probability thresholds (0.2–0.8) to assess the influence of pick confidence. Automatically detected events were matched against a merged reference catalog from Kandilli regional and Borehole-based catalogs, using combined temporal (3s) and spatial (18km) criteria. Performance was quantified using precision, recall, F1-score, ROC analysis, and spatial and temporal error metrics.Across all pipelines, results showed the most balanced performance at intermediate thresholds (0.3–0.5), where F1-scores peaked. In general PhaseNet detected more events with higher recall, especially at lowerthresholds, while EQTransformer produced fewer detections but with lower mean absolute timing errors (0.1–0.3 s) and higher precision. Thechoice of associator impacted the performance also, PyOcto improved precision by 10–20 percent compared to the EQTransformer associator and provided more stable F1-scores. Spatial accuracy showed RMS epicentral errors of 3.5–5.5 km relative to the Kandilli catalog. HypoSeisan RMS residuals were generally below 0.3 s, indicating reliable locations.These results demonstrate that optimal automatic earthquake detection is affected through balanced combination of picker sensitivity, association strategy, and probability thresholding, and that advanced associators such as PyOcto can substantially enhance the robustness of deep learning based seismicmonitoring systems. We thank Özgün Konca and Birsen Can for their support and helpful discussions.Keywords : Deep learning , Phase picking , Seismic data processing ,Seismic event detection
Authors: Köser, Betül, Özakın, Arkadas
Institution: Boğaziçi University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: European Commission
OpenAlex ID: W7214737903