Analysis of earthquake detection using deep learning: Evaluating reliability and uncertainty in prediction methods

Abstract

This study evaluates the performance and reliability of earthquake detection using the EQTransformer, a novel deep learning program that is widely used in seismological observatories and research for enhancing earthquake catalogs. We test the EQTransformer capabilities and uncertainties using seismic data from the Volcanological and Seismological Observatory of Costa Rica and compare two detection options: the simplified method (MseedPredictor) and the complex method (Predictor), the latter incorporating Monte Carlo Dropout, to assess their reproducibility and uncertainty in identifying seismic events. Our analysis focuses on 24 h-duration data that began on February 18, 2023, following a magnitude 5.5 mainshock. Notably, we observed that sequential experiments with identical data and parametrization yield different detections and a varying number of events as a function of time. The results demonstrate that the complex method, which leverages iterative dropout, consistently yields more reproducible and reliable detections than the simplified method, which shows greater variability and is more prone to false positives. This study highlights the critical importance of method selection in deep learning models for seismic event detection, emphasizing the need for rigorous evaluation of detection algorithms to ensure accurate and consistent earthquake catalogs and interpretations. Our findings provide valuable insights for the application of AI tools in seismology, particularly in enhancing the precision and reliability of seismic monitoring efforts.