A Neural Network Approach for Improved Seismic Event Detection in the Groningen Gas Field, The Netherlands

Over the past decades, the Groningen Gas Field (GGF) has been increasingly faced by induced earthquakes resulting from gas production. The seismic monitoring network at Groningen has been recently densified to improve the seismic network performance, resulting in increasing amounts of seismic data. Although traditional automated event detection techniques generally are successful in detecting events from continuous data, its detection success is challenged in cases of lower signal-to-noise ratios. The data stream coming from these networks has initiated specific interest in neural networks for automated classification and interpretation. Here, we explore the feasibility of neural networks in detecting the occurrence of seismic events. For this purpose, a three-layered feedforward neural network was trained using public data of a seismic event in the GGF obtained from the Royal Netherlands Meteorological Institute (KNMI) data portal. The first arrival times and duration of earthquake waveforms determined by KNMI for a subset of the station data, were used to detect the arrival times and event duration for the other uninterpreted station data. Subsequently various attributes were used as input for the neural network, that were based on different short term averaging/long term averaging (STA/LTA) and frequency sub-band settings. Using these input data, the network's parameters were iteratively improved to maximize its capability in successfully discriminating seismic events from noise and determine the event duration. Results show an increase of 65 % in accurately detecting seismic events and determining their duration as compared to the reference method. This clears the way for improved interpretation of signal waveforms and automated seismic event classification in the Groningen area.