Seismic Data Processing Using Deep Learning¶

Zhu, W., Mousavi, S. M., & Beroza, G. C. (2019). Seismic signal denoising and decomposition using deep neural networks. IEEE Transactions on Geoscience and Remote Sensing, 57(11), 9476-9488.

Paper: https://ieeexplore.ieee.org/abstract/document/8802278

Code: https://github.com/AI4EPS/DeepDenoiser

In [1]:

import os, sys
import numpy as np
import matplotlib.pyplot as plt
import obspy
import requests
sys.path.insert(0, os.path.abspath("../"))

import os, sys import numpy as np import matplotlib.pyplot as plt import obspy import requests sys.path.insert(0, os.path.abspath("../"))

In [2]:

# DEEPDENOISER_API_URL = "http://127.0.0.1:8002"
DEEPDENOISER_API_URL = "https://ai4eps-deepdenoiser.hf.space" # gcp
# DEEPDENOISER_API_URL = "http://test.quakeflow.com:8003" # gcp
# DEEPDENOISER_API_URL = "http://131.215.74.195:8003" # local machine

# DEEPDENOISER_API_URL = "http://127.0.0.1:8002" DEEPDENOISER_API_URL = "https://ai4eps-deepdenoiser.hf.space" # gcp # DEEPDENOISER_API_URL = "http://test.quakeflow.com:8003" # gcp # DEEPDENOISER_API_URL = "http://131.215.74.195:8003" # local machine

Prepare seismic waveforms¶

Find more details in obspy's tutorials:

FDSN web service client for ObsPy

Mass Downloader for FDSN Compliant Web Services

In [3]:

import obspy
stream = obspy.read()
stream.plot();

import obspy stream = obspy.read() stream.plot();

In [4]:

## Extract 3-component data
stream = stream.sort()
assert(len(stream) == 3)
data = []
for trace in stream:
    data.append(trace.data)
data = np.array(data).T
assert(data.shape[-1] == 3)
data_id = stream[0].get_id()[:-1]
timestamp = stream[0].stats.starttime.datetime.strftime("%Y-%m-%dT%H:%M:%S.%f")[:-3]

## Add some noise
noisy_data = data + np.random.randn(*data.shape)*np.max(data)/20

## Extract 3-component data stream = stream.sort() assert(len(stream) == 3) data = [] for trace in stream: data.append(trace.data) data = np.array(data).T assert(data.shape[-1] == 3) data_id = stream[0].get_id()[:-1] timestamp = stream[0].stats.starttime.datetime.strftime("%Y-%m-%dT%H:%M:%S.%f")[:-3] ## Add some noise noisy_data = data + np.random.randn(*data.shape)*np.max(data)/20

Get denoised waveforms using DeepDenoiser¶

In [5]:

req = {"id": [data_id],
       "timestamp": [timestamp],
       "vec": [noisy_data.tolist()]}

resp = requests.post(f'{DEEPDENOISER_API_URL}/predict', json=req)
print(resp)

denoised_data = np.array(resp.json()["vec"])

plt.figure(figsize=(10,4))
plt.subplot(331)
plt.plot(data[:,0], 'k', linewidth=0.5, label="E")
plt.legend()
plt.title("Raw signal")
plt.subplot(332)
plt.plot(noisy_data[:,0], 'k', linewidth=0.5, label="E")
plt.title("Nosiy signal")
plt.subplot(333)
plt.plot(denoised_data[0, :,0], 'k', linewidth=0.5, label="E")
plt.title("Denoised signal")
plt.subplot(334)
plt.plot(data[:,1], 'k', linewidth=0.5, label="N")
plt.legend()
plt.subplot(335)
plt.plot(noisy_data[:,1], 'k', linewidth=0.5, label="N")
plt.subplot(336)
plt.plot(denoised_data[0,:,1], 'k', linewidth=0.5, label="N")
plt.subplot(337)
plt.plot(data[:,2], 'k', linewidth=0.5, label="Z")
plt.legend()
plt.subplot(338)
plt.plot(noisy_data[:,2], 'k', linewidth=0.5, label="Z")
plt.subplot(339)
plt.plot(denoised_data[0,:,2], 'k', linewidth=0.5, label="Z")
plt.tight_layout()
plt.show();

req = {"id": [data_id], "timestamp": [timestamp], "vec": [noisy_data.tolist()]} resp = requests.post(f'{DEEPDENOISER_API_URL}/predict', json=req) print(resp) denoised_data = np.array(resp.json()["vec"]) plt.figure(figsize=(10,4)) plt.subplot(331) plt.plot(data[:,0], 'k', linewidth=0.5, label="E") plt.legend() plt.title("Raw signal") plt.subplot(332) plt.plot(noisy_data[:,0], 'k', linewidth=0.5, label="E") plt.title("Nosiy signal") plt.subplot(333) plt.plot(denoised_data[0, :,0], 'k', linewidth=0.5, label="E") plt.title("Denoised signal") plt.subplot(334) plt.plot(data[:,1], 'k', linewidth=0.5, label="N") plt.legend() plt.subplot(335) plt.plot(noisy_data[:,1], 'k', linewidth=0.5, label="N") plt.subplot(336) plt.plot(denoised_data[0,:,1], 'k', linewidth=0.5, label="N") plt.subplot(337) plt.plot(data[:,2], 'k', linewidth=0.5, label="Z") plt.legend() plt.subplot(338) plt.plot(noisy_data[:,2], 'k', linewidth=0.5, label="Z") plt.subplot(339) plt.plot(denoised_data[0,:,2], 'k', linewidth=0.5, label="Z") plt.tight_layout() plt.show();

<Response [200]>

• TODO: Try to use the pretrained model to denoise the waveforms