2D Discrete wavelet transform for denoising magnetic data
Felipe F. Melo*, Valéria C. F. Barbosa and Yolanda Jiménez-Teja, Observatório Nacional
Summary
Discrete wavelet transform (DWT) is a valuable tool in
signal and imaging processing, in particular for denoising.
Its performance in denoising potential-field data has been
proven to be superior to that of traditional techniques. We
analyze the most common thresholding techniques: soft and
hard with cycle spinning, for denoising magnetic data. To
certify the efficiency of denoising and improvement of the
filtered data we use qualitative and quantitative analysis.
Tests on noise-corrupted Bishop model prove that the soft
thresholding changes the amplitude of the data while hard
thresholding with cycle spinning generates better results. We
show the quality of hard thresholding with cycle spinning
applying it to real aeromagnetic anomaly over the Goiás
Alkaline Province, Brazil, and quantifying the improvement
of the denoised data.
Introduction
In 2017, the Brazilian Geological Service (CPRM) made
publicly available a set of 95 airborne surveys called series
1000. These surveys were flight between 1953 and 2006 and
consist of magnetic and gamma-ray data. In order to deal
properly with these datasets some reprocessing and/or
preprocessing is often required. Discrete wavelet transform
(DWT) has been used over the past 20 years to denoise
potential-field data successfully (e.g. Ridsdill-Smith and
Dentith, 1999; Fedi et al., 2000; Leblanc and Morris, 2001;
Zhang et al., 2017). This technique has proven to be a
valuable tool in enhancing the quality of the denoised data.
A common property of the random noise is the presence of
significant power in the high-frequency band. On this case,
the Fourier transform is commonly used to denoise potential-
field data (Naidu and Mathew, 1998). However, the Fourier
transform has the cumbersome that the space resolution is
absent for any given frequency (Fedi and Quarta, 1998). This
means that we cannot directly correlate a signal in Fourier
domain to its position in space/time domain. This
characteristic is because all the data is transformed together
to Fourier domain, thus the high-frequency spectral content
produced by noise maybe not properly separated from the
high-frequency spectral content produced by geologic
bodies (e.g., shallow-seated sources). Namely, the noise
cannot be removed by Fourier domain filtering because it
may have the same spectral content as the signal. One way
to around these limitations is the windowed Fourier
transform but it has the drawback that it uses the same
window size for all frequencies. Hence, the frequency
resolution analysis is the same at all locations (Graps, 1995)
and once more maybe is not possible to separate the high-
frequency spectral contents due to noise from those
produced by geologic bodies. The DWT overcomes these
limitations allowing the decomposition of the data into a
linear combination of scaled and translated versions of the
basic wavelet (Mallat, 1989; Chui, 1992).
In the literature, we can find several works where the DWT
is applied to denoise one-dimensional data: for instance,
Ridsdill-Smith and Dentith (1999) applied it to a magnetic
dataset using hard threshold and Lyrio et al. (2004)
employed it on gravity gradiometry data using an adaptive
thresholding. For denoising two-dimensional gridded
magnetic data sets, Leblanc and Morris (2001) applied the
DWT using soft threshold and Fedi and Florio (2003)
applied it to decorrugate and remove directional trends using
the local threshold parameter with soft thresholding. Fedi et
al. (2000) applied the DWT on vertical derivatives of a
gravity dataset using the local threshold parameter with soft
thresholding and Zhang et al. (2017) applied it to gravity
gradiometry data with an adaptive Bayesian threshold and
mixed thresholding. In addition to denoising of potential-
field data, some authors applied the wavelet transform to
filter undesired anomalies (Fedi and Quarta, 1998; Fedi et
al., 2004; Paoletti et al., 2007) and to perform interpretations
(Chapin, 1997; Oruç and Selim, 2011; Oruç, 2014).
Most of the above-mentioned authors compared the results
of DWT with the Fourier filtering, and other techniques for
denoising, proving that the DWT performed better in all
cases. Therefore, in this work, we focus on the use of the
DWT with different thresholding methods: soft and hard.
For the hard thresholding case, we will consider a variant:
the cycle spinning algorithm (Coifman and Donoho, 1995).
In the literature, the consequences of using noisy data or
improperly denoised data to the interpretation are well
known. For example, Florio et al. (2014) shown that
improperly denoised data lead to wrong depth estimates of
the geologic bodies using Euler deconvolution as the
interpretation method. Therefore, as qualitative quality
control for denoising we analyze the input data, the denoised
data and the noise. Moreover, quantitative enhancements are
measured with the signal-to-noise ratio (SNR) and root mean
square (RMS) of the predicted noise.
Here, we compare the use of DWT with soft thresholding
and hard thresholding using the cycle spinning algorithm for
denoising and noise determination. Applications to
theoretical magnetic data of the Bishop model corrupted by
additive pseudorandom Gaussian noise show that the DWT
with hard thresholding using cycle spinning yields better
results, both quantitatively and qualitatively, than the DWT
10.1190/segam2018-2998295.1
Page 1469
© 2018 SEG
SEG International Exposition and 88th annual Meeting
Downloaded 09/01/18 to 189.122.99.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/