High-accuracy discrimination of blasts and earthquakes using neural networks with multiwindow spectral data

The manual separation of natural earthquakes from mine blasts in data sets recorded by local or regional seismic networks can be a labor-intensive process. An artificial neural network (ANN) applied to automate discriminating earthquakes from quarry and mining blasts in eastern Kentucky suggests that the analyst effort in this task can be significantly reduced. Based on a dataset of 152 local and regional earthquake and 4192 blast recordings over a three-year period in and around eastern Kentucky, ANNs of different configurations were trained and tested on amplitude spectra parameters. The parameters were extracted from different time windows of three-component broadband seismograms to learn the general characteristics of analyst-classified regional earthquake and blast signals. There was little variation in the accuracies and precisions of various models and ANN configurations. The best result used a network with two hidden layers of 256 neurons, trained on an input set of 132 spectral amplitudes and extracted from the P-wave time window and three overlapping time windows from the global maximum amplitude on all three components through the coda. For this configuration and input feature set, 97% of all recordings were accurately classified by our trained model. Furthermore, 96.7% of earthquakes in our data set were correctly classified with mean-event probabilities greater than 0.7. Almost all blasts (98.2%) were correctly classified by mean-event probabilities of at least 0.7. Our technique should greatly reduce the time required for manual inspection of blast recordings. Additionally, our technique circumvents the need for an analyst, or automatic locator, to locate the event ahead of time, a task that is difficult due to the emergent nature of P-wave arrivals induced by delay-fire mine blasts.