• deep learning
• ground roll
• machine learning
• surface waves

In recent years, scientific interest in artificial intelligence has been growing exponentially.
This thesis aims at exploring the use of deep learning techniques to attenuate the ground roll
in active seismic data. Many studies have shown the application of various techniques to solve
this problem, using different network architectures and degrees of complexity.
The first part of this work provides an overview on the surface waves and the theoretical
background on their generation and propagation. This knowledge is the starting point for developing an automated processing procedure, aimed at attenuating the surface waves as well
as for developing a generalized workflow that can adapt to different situations.
To solve the problem of the ground roll attenuation, an image segmentation algorithm will
be adapted. This algorithm is widely employed in computer vision and enables machines to
visualize images or videos, recognize patterns, and identify different objects.
The one discussed in this work is a typical supervised learning problem where a set of labeled
data, specified by the user, are used to train the machine learning model. The seismic data
used are first transformed in the frequency-wavenumber (fk) domain. Then a fk-filter that
attenuates the ground roll is defined. This is the so called input-output pair of the training
dataset (input: raw fk-spectrum - output: filtered fk-spectrum). The training dataset is built by
using synthetic seismic data, generated with the ’OASES’ software. However, it is important
to mention that these synthetic shots were not representative of real conditions as they do not
consider dispersion and lateral velocity variations. Once generated the synthetic dataset is
then converted to the frequency-wavenumber domain with the 2D Fourier transform, obtaining a new dataset assigned to the label 0 (raw fk-spectrum). The corresponding filter masks are
generated automatically in the frequency-wavenumber domain, defining the steepness of the
masks starting from the minimum phase velocities extracted from the fundamental mode of
the dispersion curves estimated using the Thomson-Haskell method. The filtered dataset was
than assigned to the label 1. The training process involves a U-Net architecture generating
the filters to be applied to the frequency-wavenumber spectra of seismic shots, stopping the
process before the application.
Despite the assumptions, the U-Net training process produced remarkable results, achieving
an Intersection over Union (IoU) accuracy of 93.21% and a low Mean Squared Error loss (L2
loss) value of 0.019680.
The final section of this thesis focuses on testing the approach with unseen synthetic data,
specifically on testing the network’s ability to process both pure and contaminated data with
signal-to-noise ratios of 1 and 2. The aim of this work was to determine whether the network
was reliable to automatically design fk-filter able to efficiently attenuate the ground roll. Results demonstrate that a machine learning-based approach can effectively and automatically
design a filter able to reduce ground roll. It is important to note that the automated creation of
masks starting from the Thomson-Haskell method cannot be compared to the results obtained
by training the network. Creating masks from dispersion curves is possible only for known
subsurface structures and is significantly more costly as compared to using the pre-trained
U-Net network that, on the contrary, does not require any a-priori informations. The main
challenge is that the network must be trained with a wide range of scenarios to accurately
handle real situations.
This study provides a starting point for inducing greater degrees of complexity, such as changes
in lateral velocity and dispersion.