MFU-Net: a multi-scale fusion U-Net for seismic phase picking

Area of Science:

• Seismology
• Geophysics
• Machine Learning

Background:

• Seismic phase picking is crucial for seismological research, identifying seismic wave arrival times (P-waves, S-waves).
• Existing deep learning methods for seismic phase picking are often complex in their architectural design.
• There is a need for more efficient and accurate deep learning models in seismology.

Purpose of the Study:

• To propose a Multi-Scale Fusion U-Net (MFU-Net) architecture for seismic phase picking.
• To enhance the accuracy and efficiency of seismic phase picking through architectural improvements.
• To improve the recognition of critical seismic features and minority seismic wave classes.

Main Methods:

• Developed a Multi-Scale Fusion U-Net (MFU-Net) by enhancing the traditional U-Net architecture.
• Integrated a multi-scale feature fusion module within U-Net's skip connections for better information integration.
• Incorporated a multi-head attention mechanism in the bottleneck layer and a weighted class-balanced loss function.

Main Results:

• MFU-Net demonstrated improved P-wave picking accuracy by 1.6% and 1.4% on two benchmark datasets.
• MFU-Net achieved enhanced S-wave picking accuracy by 4.1% and 2.7% compared to existing methods.
• The proposed model showed superior performance against GPD, SegPhase, and SEANet.

Conclusions:

• The MFU-Net offers a more straightforward yet effective deep learning approach for seismic phase picking.
• The architectural enhancements significantly improve the accuracy of identifying both P-wave and S-wave arrivals.
• MFU-Net provides a valuable tool for seismological research, enhancing the analysis of seismic waveform data.