Application of machine learning algorithms for microseismic detection close to the Costa Rica Rift, Panama Basin

Abstract

The application of automatic learning algorithms machine learning awoke great interest in the area of Geosciences. Currently, the use of these algorithms is quite common in many investigations, particularly in the branch of seismology. In this work we use an earthquake detection method based on a deep learning approach called SCALODEEP, including two essential parts: the continuous wavelet transform (CWT) and a convolutional neural network (CNN). This method will be used to detect microseismic activity near the Costa Rica Rift (CRR) using Ocean Bottom Seismometer (OBS) signals from the OSCAR program (Oceanographic and Seismic Characterization of heat dissipation and alteration by hydrothermal fluids at an Axial Ridge). Due to the lack of generalization of the SCALODEEP model a new model was built with 3360 microseismic events and 3360 background noise time series. To set the output threshold, it was evaluated according to the behavior of the results, which were chaotic with minuscule changes in the threshold. The accuracy on the training dataset peaks at 86.57%, and on the validation set, it peaks at a maximum of 76.62%. This possibly comes out from an limited training dataset. Hence, to perform general results is required to enlarge the learning dataset, modify the training dataset or apply an alternative machine learning algorithm.