LPPN: a lightweight phase picking network

Introduction

The comparison scripts with other DNN-based pickers are in CSNBENCH. The test are based on Diting data.

Summary

We here present one lightweight phase picking network (LPPN) to pick P/S phases from continuous seismic recordings. It first classifies the phase type for consecutive data points with stride S, and then performs regression for accurate phase arrival time. The classification-regression approach and model optimization using deep separable convolution reduce the number of trainable parameters and ensure its computation efficiency.

Features

LPPN can be configured to have different model size and run on a wide range of devices. Here are:

• S can be configure as 8, 16, 32, 64, 128
• n can be any size.

Requirements

Python>=3.6.0 is required including PyTorch>=1.7

Pretrained models

The pretrained models is kept in "ckpt/stride-feature.wave". Here are 12 models with:

• stride = 8,16,32,64
• frature= 4,8,16

Usage

Data

The STEAD can be obtained from https://github.com/smousavi05/STEAD. After download the data, then copy the .csv and .h5 file to the data folder.

Training

Running lppntrain.py will train the model:

python lppntrain.py -f num_of_feature -s num_of_stride -l learning_rate -i data_folder -d device

where

• num_of_feature is the nubmer of features.
• num_of_stride is the total stride
• learning_rate is the learning rate during training the model
• data_folder is the STEAD data folder
• device is the running device.

Infering

Running lppnvalid.py will valid the model:

python lppnvalid.py -f num_of_feature -s num_of_stride -l learning_rate -o output -i data_folder -d device

the output file is num_of_feature-num_of_stride.stat.txt in output folder.

lppnvalid2.py is used to valid multiple models. Such as:

python lppnvalid2.py -f 4,8,16 -s 8,16,32 -l learning_rate -o output -i data_folder -d device

Format of result, a demo result is kept in outdata/4-8.stat.txt, with the flowing format.

#phase,True P,True S(sampling point),Mean SNR 
Type of phase (1 for P and 2 for S),Predicted time,Probability 
....

Picking continous data.

We designed a pipline for picking continous data. The script is pick.py. You can run:

python pick.py -i data_folder -o result -m model -d device 

data_folder is the base folder of the data. LPPN can automatically search all the .mseed data in the folder and all the sub folders. One can pick a countinious data with typical number of phase within 2 seconds on A100 including the IO time and post process time. The name of the data must be NET.STATION.LOCATION.COMPONENT.OTHER.INFORMATION.mseed and the three component of the data must in the same folder. However the 30s model is not suitalbe for the low SNR datas, you can contact the authors for the optimized pretrained models(Trained with ChinArray data).

Format of result, a demo result shown below:

##Data path of one component
##Type,Relative time,Prob.,Picked time(format:%Y-%m-%d %H:%M:%S.%f),SNR,Amplitude,Station,95% quantile (pre signal),95% quantile (after signal),Maximum (pre and after),Standard diviation,peaks
#path/to/data/other/information/X1.45001.01.BHE.D.2011.261.14.38.50.mseed
Pg,3581.051,0.941,2011-09-18 15:38:31.055625,2.801,12711.759765625,X1.45001,112.34023437499997,419.1605468749998,132.240234375,702.439453125,67.18669891357422,188.2140350341797,499,12290.958984375,502,12178.759765625

Contact

[email protected]

Authors

Ziye Yu, Weitao Wang @ CEA-IGP

License

MIT