Convolutional neural networks for enhanced classification of VLF rock rupture precursors

Very Low Frequency (VLF) electromagnetic emissions have been investigated as potential precursors to rock rupture in laboratory experiments and in atmospheric observations associated with earthquakes of at least moderate magnitude. In previous work, a deep Bidirectional Long Short-Term Memory (BI-LSTM) neural network achieved an overall classification accuracy of approximately 88% in distinguishing Orderly Impulsive Sequences (OIS) from background noise. Although BI-LSTM models can capture long-range dependencies, their recurrent state propagation is less directly matched to fixed short windows dominated by localized, repetitive micro-impulsive structures. In this work, we discuss why convolutional feature extraction is particularly effective for OIS recognition, as it captures localized temporal patterns and repetitive micro-impulses that define these signals. The optimized Convolutional Neural Network (CNN) achieves a holdout accuracy of 99.14% and a cross-validated accuracy of 95.83% ± 4.39% (4-fold) under a strictly leakage-free protocol. The BI-LSTM, retrained under the same protocol, achieves a holdout accuracy of 94.46%, compared with the approximately 88% reported in the original study.