MCA-Unet: Multi-class Attention-aware U-net for Seismic Phase Picking

This study presents a deep learning based algorithm for seismic event detection and simultaneous phase picking in seismic waveforms. U-net structure-based solutions which consists of a contracting path (encoder) to capture feature information and a symmetric expanding path (decoder) that enables precise localization, have proven to be effective in phase picking. The network architecture of these U-net models mainly comprise of 1D CNN, Bi- & Uni-directional LSTM, transformers and self-attentive layers. Althought, these networks have proven to be a good solution, they may not fully harness the information extracted from multi-scales.

 In this study, we propose a simple yet powerful deep learning architecture by combining multi-class with attention mechanism, named MCA-Unet, for phase picking.  Specially, we treat the phase picking as an image segmentation problem, and incorporate the attention mechanism into the U-net structure to efficiently deal with the features extracted at different levels with the goal to improve the performance on the seismic phase picking. Our neural network is based on an encoder-decoder architecture composed of 1D convolutions, pooling layers, deconvolutions and multi-attention layers. This architecture is applied and tested to a field seismic dataset to check its performance.