PhaseNet paper reading summary
PhaseNet: a deep-neural-network-based seismic arrival-time pickingmethod
background
-
Earthquake monitoring and location are the foundation of seismology
- The quality of earthquake catalogs depends primarily on the quantity and precision of time-of-arrival measurements
- Phase picking is typically performed by network analysts
- However, there are more and more seismographs, and the data flow is increasing, making manual picking difficult
-
S wave is the most difficult in phase pickup
- S waves emerge from scattered waves of P waves
- S-waves can reduce depth-origin trade-offs based on P-waves for earthquake locations
- S-wave structure is important for strong ground motion predictions
past research
-
Short-Term Average/Long-Term Average (STA/LTA) Method
- This method records the ratio of the energy in the short-term window to the energy in the long-term window
- The peak value higher than the threshold indicates the arrival of P and S
- This method is easily affected by noise, and the accuracy is relatively low
-
statistical model
- A method based on higher-order statistics (kurtosis and skewness) identifies the transition from Gaussian to non-Gaussian, which coincides with the occurrence of seismic events
-
shallow neural network
- A traditional shallow neural network is tested against four manually defined features
- Variance, absolute value of skewness, demeanor, and combination of skewness and kurtosis based on sliding window forecasts
- Most phase selections are more focused on P waves
-
Despite the above work, the accuracy of automatic picking is still not good
- Because seismic wave travel is highly complex due to multiple effects
- Traditional automatic picking algorithms manually define features and require careful data processing
paper idea
- Instead of using manually defined features, deep neural networks learn features from labeled data
- Input: unfiltered three-component seismic wave travel vertically north-south-east
- Output: Three probability distributions: PS noise
- The peaks of the P-wave and S-wave probability distributions were designed to correspond to the predicted arrival times of PS
- High precision and recall
data set
- Northern California
- 779514 records
- Divided into training set verification machine and test data set 623054 77866 78592
- Training set and validation set are used for training and fine-tuning parameter model selection
- The test set is used to evaluate performance
- The dataset covers a wide range of SNRs for various instruments
data preprocessing
- Randomly select a 30S time window containing PS arrival time as phaseNet input
- 100Hz sampling, then this is the most common sampling rate of the original data set, then each component of the input waveform has 3001 data points
- Normalize the waveform of each component by subtracting the mean and dividing by the standard deviation
Model
-
The architecture of PhaseNet is obtained by modifying the U-Net network
-
The U-Net network is a deep neural network method for biomedical image processing, aiming at locating attributes in images
-
Locate properties of time series into three categories: P-wave S-wave noise
-
The input is an earthquake three-component seismogram
-
The output is the PS noise probability distribution
-
Input and output sequence contains 3001 data points 30S 100HZ sampling
-
The input seismic data goes through four downsampling stages and four upsampling stages
-
Using 1D convolution and RELU in one stage
-
Downsampling extracts useful information from seismic data shrinking it to fewer neurons
-
Upsampling expands Qi to the probability distribution of PS noise at each time point
-
The input dimension is 3 x 3001
-
The output is the probability of 3 x 3001 ps noise at each sampling point
experiment
- Evaluation index: precision rate recall rate F1 score
- Distribution of time residuals for automatically and manually labeled P and s arrival times
- The residual distribution of P picks is easier to pick than the narrow P wave of S picks
- PhaseNet's P and S selections have narrower residual distributions compared to AR selectors
-
Tests on different instruments
-
different SNR
-
The test set is divided into 10 different classes according to the value of log10(SNR). Calculate precision, recall and F1-score for each class.
-
Even for low SNR data, the precision of PhaseNet is high, while the recall becomes relatively small.