ConvDeNoise: A convolutional denoising autoencoder for seismic monitoring - Viens and Van Houtte (2019, GJI)

Updates:

• 27/09/2019, Update 2: About the hyperparameters of ConvDeNoise (ConvDeNoise_core.py file):

  • We recommend to set the number of filters (F_nb) to 30 or 40 to denoise SC functions.
  • We tried different kernel sizes (K_sz) between 100 and 150 and found that this parameter does not really impact the denoising performance.
  • For more details, see the manuscript and the supplementary material.

• 27/09/2019, Update 1: Small changes of the input data normalization and architecture:

  • We changed the amplitude normalization of the SC functions between -1 and 1 (0 and 1 in the 1st version).
  • The new autoencoder only has 4 hidden layers (6 layers in the 1st version).
  • The last activation function is the hyperbolic tangent activation (tanH) function to output the denoised SC function amplitudes between -1 and 1 (Sigmoid function in the 1st version).
  • These three changes decrease the number of parameters to train for the same level of performance.

Description:

• Python codes to reproduce Figure 7 (Figure 8 can also be plotted by changing 1 line of the code) of Viens L. and Van Houtte C. (2019), Denoising ambient seismic field correlation functions with convolutional autoencoders, Geophys. J. Int., 220, 1521–1535. Note that the published paper incorporates the modifications made on Sept. 27, 2019 (See above).

• The ConvDeNoise algorithm is a convolutional denoising autoencoder which is composed of an encoder and a decoder as:

• The Codes folder contains 5 files:

  • The Reproduce_Fig_7.py file is the python code to reproduce Figures 7 and 8 of the paper (Default: plot Figure 7, simply change "fig_choice = 7" to "fig_choice = 8" (L. 71) to plot Figure 8).
  • The functions_for_autoencoders.py file contains functions to bandpass filter the data with a Butterworth filter, denoise the SC functions with the SVDWF method (Moreau et al., 2017), and compute the stretching to retrieve dv/v measurements
  • The ConvDeNoise_NS7M_station.h5 contains the weights of ConvDeNoise trained for the NS7M station (Requires Keras 2.2.4)
  • The Test_data.mat contains 16 days of raw SC functions at the NS7M station, reference waveforms to compute the dv/v,... (e.g., all the data required to reproduce Figure 7).
  • The ConvDeNoise_core.py file is the convolutional denoising autoencoder main code that was used to compute the ConvDeNoise_NS7M_station.h5 file (requires the raw SC functions, please email me for the training set, file is too big for Github)