Abstract

Developing a reliable and robust automatic earthquake detection system is quite a challenging and highly necessary task as two conditions can make this task a difficult one. First, earthquake detection systems may perform more poorly if they are employed in a region that is different from the one where the training database corresponds to. Systems trained with local databases are assumed to perform better. Nevertheless, these data-bases are usually limited. Second, the performance of such systems worsens when the SNR of the seismogram signals decreases. This paper proposes an end-to-end DNN-HMM based scheme to address these limitations, i.e. it does not require previous phase-picking, backed by engineered features and combined with duration modeling of states and seismic events. The proposed engine requires 10-or 15-times fewer parameters than state-of-the-art methods and therefore needs a smaller training database. Modeling duration can improve the noise robustness of the detection system significantly, particularly with limited training data; having a negligible increase in the number of training parameters. The system described here provides a F1-score 101% higher on average than schemes published elsewhere with Iquique and North Chile databases. It provides a reduction in F1-score equal to 10% when the average SNR is reduced by approximately 18 dB. This reduction in F1-score is at least half of the one observed with the state-of-the-art schemes in the same testing conditions. With respect to the detection of small earthquakes at short epicenter-station distances, the averaged precision provided by the DNN-HMM system with duration modeling is at least 5% higher than other systems.