Neural Network for Seismic Waves Phase Picking Based on Wavelet Scattering

Recent advances in usage of convolutional neural networks (CNNs) and attention-based models (Transformer) to solve the problem of seismic wave phase picking have generated significant interest in the topic of creating and developing deep learning models for analyzing seismic data. One of the promising directions of development in this area is usage of new approaches to extracting features from the analyzed signal, allowing to preserve its information content and at the same time simplify the process of training the model. This paper compares three methods for extracting features from a signal based on its time–frequency representation (short-time Fourier transform (STFT) on linear and mel scales, as well as the less commonly used wavelet splitting (Wavelet Scattering)) and evaluates their impact on the quality of classification seismic signal plots using the California seismic data set as an example. To do this, three similar deep learning models were created and trained using the encoder from the Transformer architecture. It is shown that Wavelet Scattering is slightly superior to STFT in the normal and mel scales in terms of the Accuracy metric. This makes this decomposition more promising in terms of the quality of seismic signal feature extraction.