Convolutional Neural Networks Versus P/S Amplitude Ratios in Low-Yield Seismic Event Discrimination: An Evaluation Using Earthquakes, Mine Blasts, and Mining-Related Events from the Kiruna Mining Region, Northern Sweden

We advance the use of convolutional neural networks (CNNs) for discriminating low-yield seismic events recorded at local distances by evaluating a CNN approach based on time–frequency representations (scalograms) of seismic records from earthquakes, mine blasts, and mining-related seismic events in the Kiruna mining region of northern Sweden to (1) determine if the CNN approach can outperform the P/S amplitude ratio method in classifying these source types, and (2) examine the regional transportability of a CNN model trained on data from the United States. An accuracy of 90% or greater was obtained for the CNN approach for binary source classification between the three source types (earthquakes, mine blasts, and mining-related events), an accuracy level not achieved by the P/S amplitude ratio method, illustrating superior performance of the CNN approach over the amplitude ratio approach. The CNN model trained on explosions and earthquakes in United States yields poor binary classification performance (accuracy < 90%) when applied to earthquakes and mine blasts in the Kiruna mining region, suggesting limited transportability of the U.S.-trained model. However, the poor performance may arise from differences in the blasting style between the two data sets (single-fired borehole explosions in the United States versus ripple-fired blasts into a mine shaft at the Kiruna mine) and source depths (near surface in United States vs. 800–900 m depth in the Kiruna mine), leaving open the question of whether transportability is more limited by differences in local geologic structure or in explosion source processes.