Earthquake or blast? Classification of local-distance seismic events in Sweden using fully connected neural networks

Distinguishing between different types of seismic events is a task typically performed manually by expert analysts and can thus be both time and resource expensive. Analysts at the Swedish National Seismic Network (SNSN) use four different event types in the routine analysis: natural (tectonic) earthquakes, blasts (e.g. from mines, quarries and construction) and two different types of mining-induced events associated with large, underground mines. In order to aid manual event classification and to classify automatic event definitions, we have used fully connected neural networks to implement classification models which distinguish between the four event types. For each event, we bandpass filter the waveform data in 20 narrow-frequency bands before dividing each component into four non-overlapping time windows, corresponding to the P phase, P coda, S phase and S coda. In each window, we compute the root-mean-square amplitude and the resulting array of amplitudes is then used as the neural network inputs. We compare results achieved using a station-specific approach, where individual models are trained for each seismic station, to a regional approach where a single model is trained for the whole study area. An extension of the models, which distinguishes spurious phase associations from real seismic events in automatic event definitions, has also been implemented. When applying our models to evaluation data distinguishing between earthquakes and blasts, we achieve an accuracy of about 98 per cent for automatic events and 99 per cent for manually analysed events. In areas located close to large underground mines, where all four event types are observed, the corresponding accuracy is about 90 and 96 per cent, respectively. The accuracy when distinguishing spurious events from real seismic events is about 95 per cent. We find that the majority of erroneous classifications can be traced back to uncertainties in automatic phase picks and location estimates. The models are already in use at the SNSN, both for preliminary type predictions of automatic events and for reviewing manually analysed events.