Combining Seismic Scalograms and Phase Detection Probabilities to Improve Generalization of Deep Learning Classification of Local Explosions and Earthquakes

Deep learning models can help mitigate the need for analyst intervention in seismic classification of local earthquake and explosion signals, but optimizing model performance across multiple localities is challenging. Here, we seek to advance multilocality classification through hyperparameter optimization, evaluation of training data selection, and incorporation of phase pick probabilities from a pre-existing deep learning model into the new source type classification models. One-branch and two-branch convolutional neural network model designs are tested with 16 local data sets, each containing earthquakes and shallow borehole explosions observed by the same receivers. The one-branch model uses a conventional input of time–frequency scalograms and the two-branch model adds vector inputs of P-wave and S-wave probabilities output from a pre-existing phase picking model, which are increasingly available in modern event detection workflows. Hyperparameters for the one- and two-branch models were optimized by a guided search algorithm. For in-distribution data tests spanning all 16 data sets, the two-branch model’s single-station balanced accuracy (95.7%) slightly outperforms the one-branch model (94.7%). In addition, the optimized two-branch model is simpler, with about half as many model parameters and faster convergence in training. Relative performance of the two-branch model increases for out-of-distribution tests (using a locality held out from training), with an average increase in single-station balanced accuracy from 84.4% to 87.6% across 16 held out localities. The two-branch model also substantially outperforms P/S amplitude ratio discrimination accuracy by 14% on average for out-of-distribution tests. The results suggest that the added context from phase detection probabilities, which are a common by-product of event catalog building, can provide an efficient means to improve generalization of event-type classification models when applied to new regions.