skip to main content

Brown Bag Seminar

Wednesday, November 6, 2024
12:00pm to 1:00pm
Add to Cal
South Mudd 256 (Benioff Room)
Enhancement of Phase Picking Models Using Deep Learning by Addressing the Label Imbalance Problem
Shinya Kato, Researcher, Nagao Laboratory, Earthquake Research Institute, The University of Tokyo,

In recent years, various deep learning-based models for seismic phase picking have been developed, with models trained on manually picked P- and S-wave onset times. One widely-used model, PhaseNet (Zhu and Beroza, 2019), treats the phase picking task as a semantic segmentation problem. The input consists of three-component waveform data, while the target labels represent Gaussian distributions with a peak at the P- and S-wave arrival times, set to a height of 1 and a standard deviation of 0.1 s. The remaining regions are labeled as noise (N). PhaseNet estimates arrival times based on the point of maximum output probability.

While deep learning-based phase-picking models are known for their high accuracy, detecting all seismic phases becomes challenging when multiple seismic waves exist within a single trace (Park et al., 2023). This issue is not a major concern in routine seismic monitoring; however, it presents significant limitations in scenarios with increased seismicity, such as aftershock sequences following large earthquakes or induced seismicity triggered by fluid injection experiments. Given the need for automated detection in such high-seismicity settings, models must be able to pick arrivals even when multiple events are present within a trace.

We hypothesize that this limitation may be caused by label imbalance (Saini and Susan, 2023). In PhaseNet, the area ratio of labels during training is 0.8% each for P and S labels, while 98.4% is allocated to the N label. Consequently, most of the loss during training is influenced by N labels, rendering the contribution from P- and S-wave labels minimal. This label imbalance may cause the model to overlook P and S signals, potentially reducing its sensitivity in detecting these arrivals. Thus, addressing this imbalance could improve the model's sensitivity to P- and S-waves.

In this study, we tackle the label imbalance by applying weighting to the cross-entropy loss function, creating a weighted cross-entropy loss. For phase picking, we use SegPhase (Katoh et al., 2024), a model that employs a hierarchical Vision Transformer, similar to PhaseNet, and treats phase picking as a semantic segmentation task with P, S, and N labels.

In this presentation, we introduce this method and present its effectiveness when applied to continuous waveforms following the 2019 Ridgecrest earthquake.