SeismoNet: A Deep Learning Approach for Time–Frequency Analysis of Seismic Data

This work introduces SeismoNet, a deep learning (DL) framework designed to isolate the source earthquake signal from noisy components on the time–frequency representation (TFR) plane. Accurately detecting and separating earthquake signals from noise remains a significant challenge in seismology. By incorporating a custom wavelet layer into SeismoNet, the proposed framework effectively denoises and distinguishes source signals. Unlike conventional convolutional neural networks (CNNs), which rely on standard kernel filters, SeismoNet utilizes a nonstandard wavelet function (NSWF) as its filtering kernel, allowing for the identification of intricate seismic patterns. To further enhance the network performance, a novel bidirectional conjugate gradient (BiDCG) optimization method is introduced. This method iteratively refines the network weights through conjugate gradient updates in both forward and backward directions. The experimental results demonstrate a substantial improvement in isolating the source signal from noisy components, enhanced TFR resolution, and superior signal reconstruction compared to traditional approaches.