Type of publication
Article in journal
Year of publication
Bulletin of the Seismological Society of America

Ruben Tous, Leonardo Alvarado, Beatriz Otero, Leonel Cruz and Otilio Rojas


Ruben Tous, Leonardo Alvarado, Beatriz Otero, Leonel Cruz, Otilio Rojas; Deep Neural Networks for Earthquake Detection and Source Region Estimation in North‐Central Venezuela. Bulletin of the Seismological Society of America 2020;; 110 (5): 2519–2529. doi: https://doi.org/10.1785/0120190172

Short summary
Reliable earthquake detection algorithms are necessary to properly analyze and catalog the continuously growing seismic records. We report the results of applying a deep convolutional neural network, called UPC‐UCV (Universitat Politecnica de Catalunya ‐ Universidad Central de Venezuela), over single‐station three‐channel signal windows for P‐wave earthquake detection and source region estimation in north‐central Venezuela. The analysis is performed on a new dataset of handpicked arrivals of P waves from local events, named CARABOBO, built and made public for reproducibility and benchmarking purposes. The CARABOBO dataset consists of three‐channel continuous data recorded by the broadband stations of the Venezuelan Foundation for Seismological Research in the region of 9.5°–11.5°N and 67.0°–69.0°W during the time period from April 2018 to April 2019. During this period, 949 earthquakes were recorded in that area, corresponding to earthquakes with magnitudes in the range from Mw 1.1 to 5.2. To estimate the epicentral source region of a detected event, the proposed network employs geographical distribution of the CARABOBO dataset into K clusters as a basis. This geographical partitioning is automatically performed by the k‐means algorithm, and the optimality of the K‐values for our dataset has been assessed using the elbow (⁠K=5⁠) and silhouette (⁠K=3⁠) methods. For target seismicity, the proposed network achieves 95.27% detection accuracy and 93.36% source region estimation accuracy, when using K=5 geographic clusters. The location accuracy slightly increases to 95.68% in the case of K=3 geographic partitions. The detection capability of this network has also been tested on the OKLAHOMA dataset, which compiles more than 2000 local earthquakes that occurred in this U.S. state. Without any modification, the proposed network yields excellent detection results when trained and evaluated on that dataset (98.21% accuracy; ConvNetQuake, fine‐tuned for this dataset, achieves a 97.32% accuracy), corresponding to a totally different geographical region.