Application of Convolutional Neural Network for Seismic Event Classification: Impact of Dataset Quality, Distribution, and Human-in-the-Loop Feedback

ABSTRACT Convolutional neural networks (CNNs) have gained popularity in geophysical research due to their exceptional performance in various areas. However, achieving reliable results with CNNs typically requires a significant amount of high-quality data for training. In this study, we develop a CNN to classify natural earthquakes, mine collapses, and explosions using seismic waveforms from 287 stations in Shandong Province, China. The dataset comprises 1035 earthquakes, 159 mine collapses, and 586 explosions. To address the impact of unreliable measurements, we employ cross validation to screen, manually correct, or discard measurements with inconsistent labels assigned by human experts and CNN. By refining the dataset through these methods, classification accuracies for the three event types improved substantially, reaching over 95%. Notably, CNN outperforms human classification in this task, with its performance heavily influenced by the quality and distribution of the training dataset. Our research demonstrates the potential of CNNs for classifying seismic events while emphasizing the crucial role of human-in-the-loop feedback and the significance of cross-validation techniques in ensuring the accuracy and reliability of the CNN model.