SCEC Project Details
SCEC Award Number | 18165 | View PDF | |||||
Proposal Category | Individual Proposal (Integration and Theory) | ||||||
Proposal Title | Machine Learning Based Convolutional Neural Network in Earthquake Detection and Classification and its Application in Southern California | ||||||
Investigator(s) |
|
||||||
Other Participants |
Lijun Zhu Chenyu Li |
||||||
SCEC Priorities | 3a, 3b, 3c | SCEC Groups | Seismology, CS, MSW | ||||
Report Due Date | 03/15/2019 | Date Report Submitted | 05/06/2019 |
Project Abstract |
We developed a convolutional neural network (CNN) based phase detection and picking approach named CPIC (Zhu et al., 2019). The method was first applied on a one-month aftershock sequence of 2008 Mw7.9 Wenchuan earthquake. 20-sec-long waveform frames are cut from 5 seconds before and 15 seconds after the picked P or S arrival times. Trained only on ~40,000 such frames, our CNN-based phase identification classifier (CPIC) achieve 97.6% classification accuracy on unseen 20,000 picked frames. More importantly, the CPIC approach is generally applicable to many seismic active regions, such as southern California, Oklahoma, and New Zealand. Benchmarked on the SCSN dataset (4.8M picked arrival times) released by the Caltech researchers, the CPIC model achieves similar high accuracy (99.5%) with a significantly simpler model and faster execution time. Finally, when tested on a small dataset from a different region (Oklahoma, US), CPIC achieves 97% accuracy after fine-tuning only the fully connected layer of the model. This result suggests that the CPIC developed in this study can be used to identify and pick P/S arrivals in other regions with no or minimum labeled phases. |
Intellectual Merit | This project develops a machine learning based tool for seismic phase picking and detection that requires a small to moderate amount of training data. It is also applicable to other regions. An improved phase picking and event detection could result in many small-magnitude earthquakes being detected/located, which can help to improve our understanding of subsurface fault structures, large earthquake nucleation and earthquake interaction at nearby and long-range distances. |
Broader Impacts | This project supported collaborations of two GT students. Lijun Zhu from School of Electronic and Computer Engineering (ECE), and he is expected to defend his Ph.D. thesis in July 2019. This work is a major component of his Ph.D. thesis. Chenyu Li is a 5th year graduate student from School of Earth and Atmospheric Sciences (EAS). She is expected to graduate in summer 2020. We are in the process of releasing the related package and test dataset online at https://github.com/lijunzh/yews |
Exemplary Figure | Figure 2. Detection example on 15-minute recording on 14 stations with three catalog events for the Wenchuan dataset. Only vertical components are plotted. Blue and green curves show the probabilities of P and S phases. Red and magenta bars indicate the catalog P and S arrivals. Origin times of three catalog events are marked by the dashed vertical lines along with their magnitudes. After Zhu et al. (2019). |