PATCHUNET: A fully-unsupervised and highly-generalized

deep learning approach for random noise suppression

Omar M. Saad∗ and Yangkang Chen†

∗ Omar M. Saad is with the Key Laboratory of Geoscience Big Data and Deep Resource of

Zhejiang Province, School of Earth Sciences, Zhejiang University, China, 102249, and

ENSN Lab, Seismology Department, National Research Institute of Astronomy and

Geophysics (NRIAG), Helwan 11731, Egypt. Email: engomar91@[Link];

[Link]@[Link]; (Corresponding author)

† Yangkang Chen is with the Key Laboratory of Geoscience Big Data and Deep Resource of

Zhejiang Province, School of Earth Sciences, Zhejiang University, China, 102249. Email:

[Link]@[Link]

(December 9, 2020)

Running head: Noise Suppression.

ABSTRACT

We develop a deep learning algorithm (PATCHUNET) to suppress random noise and pre-

serve the coherent seismic signal. The input data is divided into several patches, each patch

is encoded to extract the meaningful features. Following this, the extracted features are

decompressed to retrieve the seismic signal. Skip connections are used between the encoder

and decoder parts, allowing the PATCHUNET to extract high-order features without los-

ing important information. Besides, dropout layers are used as regularization layers. The

dropout layers preserve the most meaningful features belonging to the seismic signal and

discard the remaining features. The PATCHUNET is an unsupervised approach that does

1
not require prior information about the clean signal. The input patches are divided into 80%

for training and 20% for testing. However, it is interesting to find that the PATCHUNET

can be trained with only 30% of the input patches with an effective denoising performance.

Four synthetic and four field examples are used to evaluate the PATCHUNET performance,

and compared to the f − x deconvolution (F XDECON ) and the f − x singular spectrum

analysis (f − x SSA). The results indicate the great ability of the PATCHUNET in atten-

uating the random noise and preserving the seismic signal effectively despite the existence

of a large amount of random noise, e.g., when the input signal-to-noise-ratio (S/N) is as

low as -14.2 dB.

Key words: Random noise suppression, deep learning.

2
 INTRODUCTION

Enhancing signal-to-noise ratio (S/N) of the seismic data is a key role in seismic signal

processing, which can be accomplished by random noise suppression (Yilmaz et al., 2001;

Chen and Fomel, 2015; Gan et al., 2016; Dagnino et al., 2017). Over the past few decades,

several methods have been introduced to suppress random noise. Prediction based noise

suppression methods use the predictable property of signals to suppress random noise,

for instance, t − x predictive filtering (Canales, 1984; Abma and Claerbout, 1995), f − x

deconvolution (FXDECON) (Gulunay, 2000; Trickett, 2003), and non-stationary predictive

filtering (Liu et al., 2012; Liu and Chen, 2013; Liu et al., 2018). The sparse transform

domain has been utilized for random noise suppression, for example, wavelet transform

(Mousavi et al., 2016; Anvari et al., 2017; Mousavi and Langston, 2017), curvelet transform

(Neelamani et al., 2008; Liu et al., 2016), seislet transform (Fomel and Liu, 2010; Chen

et al., 2014), dictionary learning-based sparse transform (Yu et al., 2016; Siahsar et al.,

2017; Zu et al., 2019), and singular spectrum analysis (f − x SSA) (Sacchi, 2009; Oropeza

and Sacchi, 2011; Huang et al., 2016).

Recently, deep learning approaches have been widely used in several fields due to its

ability to reach superior performance compared with traditional machine learning techniques

(Cheriyadat, 2013; Cheng et al., 2018; Zhang et al., 2018; Wang and Nealon, 2019; Mousavi

et al., 2019; Zhu et al., 2019; Li et al., 2020; Chen et al., 2020; Liu et al., 2020a; Saad

and Chen, 2020a). Deep learning consists of several processing layers, and learns data

representations with several consecutive layers (LeCun et al., 2015). It extracts non-linear

features based on unsupervised techniques (LeCun et al., 2015), and deals with the raw

data without the need for any pre-processing.

3
 The patching technique is used to enhance the performance of the machine learning

approaches (Wang et al., 2020). The patching technique allows machine learning to learn

the spatial coherency of the input data by dividing the 2D data into several 2D overlapped

windows with certain window size and shift samples (Chen et al., 2019). The extracted 2D

overlapped patches are used as an input for the machine learning approach (Chen et al.,

2019). However, the 2D patches can be converted to a 1D vector in case the machine learning

approach deals with 1D data (Wang et al., 2020). The final output can be reconstructed from

the predicted output patches using the reverse patching operation, while the overlapped

samples between each neighbor windows are averaged (Chen et al., 2019; Wang et al.,

2020). This approach has been used to enhance the S/N of seismological datasets by an

unsupervised machine learning approach (Chen et al., 2019) as well to extract the waveform

features of the microseismic data using an adaptive dictionary (Wang et al., 2020).

A deep image prior network has been proposed to deal with several applications, e.g.,

denoising and generic reconstruction, image inpainting, super-resolution imaging, and Flash-

no flash reconstruction (Ulyanov et al., 2017). Deep image prior network is based on Plug-

and-Play concept (Ulyanov et al., 2017), which does not need any prior information about

the input data. Besides, deep image prior network does not require regularization layers,

e.g., L1 regularization, L2 regularization, and dropout layer (Ulyanov et al., 2017; Shi

et al., 2020). It is an unsupervised technique that mainly based on the network architecture

(Ulyanov et al., 2017). Thus, a deep image prior network directly obtains the optimum

network parameters that minimize the loss function in an iterative way (Ulyanov et al.,

2017). Here, the number of epochs is considered as the number of iterations to reach the

best network parameters (Ulyanov et al., 2017). Besides, the idea of the deep image prior

network has been proposed for seismic data denoising and interpolating (Park et al., 2020;

4
Kong et al., 2020; Liu et al., 2020b; Shi et al., 2020).

We develop a fully unsupervised deep learning approach (PATCHUNET) to suppress

random noise based on the idea of the deep image prior. First, a patching technique (Chen

et al., 2019) is utilized to divide the input noisy data into several patches for training and

testing. Following this, each patch is encoded to extract the meaningful features, allowing

the decoder part to retrieve the clean signal and eliminate the random noise. Next, skip

connections are used, between encoder and decoder parts, to extract high-order features.

Finally, regularization layers are added to avoid over-fitting and keep the useful features.

The PATCHUNET is evaluated using eight synthetic and field examples. The local simi-

larity maps (Fomel, 2007; Chen and Fomel, 2015) indicate the ability of the PATCHUNET

to suppress random noise effectively compared to the benchmark algorithms.

METHOD

The Proposed Deep-learning Framework

Representation and residual learning are the main frameworks for PATCHUNET. The

PATCHUNET encodes the input seismic data, s(t), to new representation, allowing the

decoder to map these extracted features to the clean signal, c(t), without random noise,

n(t). For residual learning, skip connections are added between the hidden layers of the en-

coder and decoder parts. This approach supports the PATCHUNET to extract high-order

features without losing important information. Typically, encoder-decoder technique with

skip connections is namely UNET. In addition, we use several dropout layers (Srivastava

et al., 2014) as the regularization layers.

First, the input 2D seismic data, s(t), is divided into P overlapped patches using a

5
patching technique. In the patching technique, a window of size M × N is slid over the

2D seismic data from the top to the bottom, as well as from the left to the right with k

shift samples. Here k denotes shift size between each neighbor patches. Since the proposed

algorithm consists of several fully connected layers, each extracted patch is converted to a

1D vector, such that the UNET input has a 2D shape with size of P × T , where T = M × N .

In the unpatching process, each seismic sample is retrieved from the corresponding patch.

In the overlapping part, the average is obtained from each neighbor patches. The visual

illustration of the aforementioned patching and unpatching techniques is shown in Figure 1.

The PATCHUNET consists of le hidden layers for encoder and ld hidden layers for

decoder, with le = ld . The encoder extracts the meaningful features, F , to represent the

input seismic data, s(t). Whilst, the decoder reconstructs the clean data, c(t), from the

encoder representation. The matrix representation of each hidden layer in the encoder is

shown as follows:
 
 Fd [1]
 W d [11] W d [21] . . . W d [(q−1)1] W d [q1]  Y [1] 
d−1
 Bd [1]

 Fd.[2]   W d.[12] W d [22]
.
.
.
.
. .
d
. W [(q−1)2]
.
d
W [q2] 
.  Yd−1 [2]
.   Bd.[2] 
 .  = . . . . . . . . +  . ,
  
 .   d . . . . . . .   Y .[q−1]   . 
 
Fd [e−1] d
W [1(e−1)] W [2(e−1)] . . d d
. W [(q−1)(e−1)] W [q(e−1)] Bd [e−1]
d−1
Fd [e] W d [1e] W d [2e] . . . W d [(q−1)e] W d [qe] Yd−1 [q] Bd [e]
(1)

where W d [qe] and Bd [e] represent the network parameters of the encoder, i.e., weights and

biases, respectively. d the index of the hidden layer within the encoder which varies from

1 to le . Whilst, q and e are the neuron numbers in the previous and the current hidden

layers, respectively. qe represents the connection between the neurons in different hidden

layers of the encoder. The input of the decoder is the output of the encoder, where each

hidden layer in the decoder has a matrix representation as follows:

6
 Zr [1] W r [11] W r [21] . . . W r [(g−1)1] W r [g1] Ur−1 [1] Br [1]
      
W r [12] W r [22] . W r [(g−1)2] W r [g2]
 Zr.[2]   . .
.
.
.
. . . .   Ur−1 [2]
.   Br.[2] 
 . 
 .  = 

.
.
.
.
.
.
.
.
.
.
.
.
.
.

 
.
.

 +  . ,
 . 
Zr [v−1] r r
W [1(v−1)] W [2(v−1)] . . r r
. W [(g−1)(v−1)] W [g(v−1)] Ur−1 [g−1] Br [v−1]
Zr [v] r
W [1v] r
W [2v] . . . W r [(g−1)v] W r [gv] Ur−1 [g] Br [v]
(2)

where W r [gv] and Br [v] represent the network parameters of the decoder, i.e., weights and

biases, respectively. r is the number of the hidden layers of the decoder which varies from

1 to ld . gv denotes the connection between the neurons in different hidden layers of the

decoder.

We use the exponential linear unit (ELU) activation function after each hidden layer.

The ELU function can be expressed as follows:




x, x>0


ELU(x) = , (3)


α(expx −1), x <= 0



where, α is the smoothing factor. The output of ELU smoothly varies above -α when the

input is less than or equal zero. The default value of α is 1. Following this, dropout layers

are added after each encoder activation function as a regularization layers.

Finally, skip connections are added between the hidden layers of the encoder and the

hidden layers of the decoder. At the decoder, rather than directly learning Zr , it learns the

concatenated features {Zr , Fd }. The optimization of the residual mapping is easier than

the traditional method, which keeps the training error small. The network learns the local

features from the encoder as well as the global features from the decoder. This technique

enables the network to extract high-order features related to the clean seismic signal and

eliminate the features corresponding to the random noise. After the last hidden layer of

the decoder, an extra fully-connected layer is utilized, which has the same size as the input

layer. The output layer has a linear activation function.

7
 Adam optimizer (Kingma and Ba, 2014) is utilized to optimize the network parameters

by minimizing the mean square error between the input seismic data, s(t), and the recon-

structed signal corresponding to the PATCHUNET. The loss function can be determined

as follows:

min .||F (z; θ) − z||22 , (4)

θ

where F is the PATCHUNET, and z is the input of PATCHUNET (the extracted patches

from the noisy data), θ is the network parameters of PATCHUNET, and F (z; θ) denotes

the output patches corresponding to the PATCHUNET. Afterward, each seismic sample is

retrieved from the corresponding output patch, while in the overlapping part, the average

is determined from each neighbor patches.

PATCHUNET has six hidden layers per each encoder and decoder. For the encoder, the

neuron numbers per each layer is 128, 64, 32, 16, 8 and 4, respectively. Whilst, 4, 8, 16, 32,

64 and 128 neurons are used for each hidden layer of the decoder, respectively. The drop

out rate (Dr) is set to be 0.1, and α is set to be 1. Each seismic example is divided into

P patches. Each patch has a size of 48 × 48 with shift size of 1 sample. These patches are

randomly split to 80% for training, and 20% for testing. 50 epochs are used for the training

process. The selection of network parameters is illustrated in the DISCUSSION section.

The PATCHUNET architecture is shown in Figure 2. The optimal network architecture

gives the best denoising performance, which matches the concept of the deep image prior

(Ulyanov et al., 2017). However, the input of the deep image prior network is random tensor

(Ulyanov et al., 2017), while the noisy patches are the input of the PATCHUNET.

8
Advances Over The Previous Work

The PATCHUNET has several advantages over the recently published DDAE method (Saad

and Chen, 2020b):

A) The architecture of PATCHUNET has skip connections, allowing the network to

extract high-order features while the DDAE does not have these skip connections.

B) Several Dropout layers are used in the PATCHUNET, allowing the network to keep

the meaningful features related to the clean data and eliminate the features corresponding

to the random noise. Also, these Dropout layers are considered as regularization layers

which generalize the network to be fitted with any seismic data. However, DDAE does not

have dropout layers.

C) Patch technique is used in PATCHUNET, avoiding the possibility of artifacts oc-

currence which happened in DDAE (Saad and Chen, 2020b) when inappropriate initial

parameters are set.

D) The DDAE is a semi-unsupervised denoising algorithm. DDAE is trained in a super-

vised way using synthetic datasets. The obtained model was used to initialize the network

of the DDAE to denoise the field dataset in an iterative unsupervised scheme using a

correlation-based loss function. However, PATCHUNET is a fully unsupervised technique

that does not need a pre-trained model. PATCHUNET works in an unsupervised way in

both synthetic and field examples.

9
 EXAMPLES

Synthetic examples

Four synthetic examples are used to evaluate the PATCHUNET. First, four linear crossing

events are created. The clean data, containing 256 traces with a sample rate of 4 ms is

shown in Figure 3a. Following this, the noisy data is created by adding random noise

(Figure 3b) as shown in Figure 3c. In this example, the lateral and vertical smoothing radii

are both 25 samples. We define the S/N as follows (Saad and Chen, 2020b):

ksk2
S/N = 10 log . (5)
ks − S Rk2

The S/N of the input noisy data is −3.74. The f − x SSA denoised section is shown in

Figure 3d, while the f −x SSA removed noise section is shown in Figure 3e; it contains some

horizontal and low-dip-angle signals. However, the PATCHUNET denoised section shown

in Figure 3g has horizontal events with stronger amplitude. Moreover, the PATCHUNET

extracted noise section does not contain any coherent signals in the noise section shown in

Figure 3h. The f − x SSA denoised section has a S/N of 4.80 dB, while the S/N of the

PATCHUNET denoised section is 11.1 dB. Furthermore, the local similarity map is calcu-

lated between the removed noise and the denoised sections as an evaluation measurement

of the denoising performance. Small energy value of the local similarity map indicates a

better denoising performance. The local similarity maps of f − x SSA and PATCHUNET

are shown in Figures 3i and 3f, respectively. The local similarity of PATCHUNET has

a smaller value compared to the f − x SSA local similarity map, indicated by arrows in

Figure 3.

Secondly, a synthetic example similar to Liu et al. (2012) is utilized to assess the

10
PATCHUNET. The shot gather contains four hyperbolic events with 501 traces as shown

in Figure 4a, while the random noise and the noisy data are shown in Figures 4b and 4c,

respectively. The denoised section obtained by F XDECON , Figure 4d, contains random

noise as well as the removed noise section, Figure 4e has a lot of signal leakage which is

also visible in the local similarity map in Figure 4f (indicated by arrows). Contrary to

F XDECON , the PATCHUNET removes random noise (Figure 4g) while reconstructing

the seismic signal effectively as shown in Figure 4h. The S/N of the input noisy data is 1.25

dB, while the F XDECON denoised and PATCHUNET denoised section has S/Ns of 10.65

dB and 14.49 dB, respectively. The local similarity map of PATCHUNET is nearly zero

everywhere as shown in Figure 4i, indicating that the PATCHUNET has a good denoising

performance without losing important information.

Thirdly, a synthetic example similar to Chen and Ma (2014) is used to test the PATCHUNET.

In this example, the Ricker wavelet frequency is 40 Hz with a sampling rate of 2 ms, contain-

ing 76 traces. Figures 5a, 5b, and 5c show the clean signal, noise section and the noisy data,

respectively. The input noisy data has a S/N of -0.30 dB. The S/N of the PATCHUNET

denoised section, Figure 5g, is enhanced to 12.40 dB, while the S/N of the F XDECON de-

noised section, Figure 5d, is 6.56 dB. The PATCHUNET removed noise section, Figure 5h,

has a small amount of signal leakage. The F XDECON removed noise section, Figure 5e

has a large amount of signal leakage which clearly appears in the local similarity map as

shown in Figure 5f (indicated by arrows). The PATCHUNET preserves the useful energy

without losing meaningful information as shown in the corresponding local similarity map

in Figure 5i.

Finally, more realistic example simulated from finite-difference wave modeling is utilized,

which is consisted of 80 traces as shown in Figure 6a. The noise section and the noisy data

11
are shown in Figures 6b and 6c, respectively. The F XDECON and PATCHUNET denoised

sections are shown in Figures 6g and 6d, indicating the abilities of both algorithms in ef-

fectively suppressing random noise. However, the F XDECON -removed noise section, Fig-

ure 6e, and its corresponding local similarity map (indicated by rectangle), Figure 6f, show a

loss of useful signal. The signal loss is barely noticeable in the PATCHUNET removed noise

section, Figure 6h, and its corresponding local similarity map, Figure 6i. The S/N of the

input noisy data is -0.96 dB, while the S/N of the F XDECON and PATCHUNET denoised

sections is 6.20 dB and 13.88 dB, respectively. Therefore, the PATCHUNET outperforms

the F XDECON in preserving the useful signals and suppressing the random noise.

Field examples

Four field examples are used to evaluate the PATCHUNET. The first field example is a

deep-water common receiver section containing events that are nearly flat. The clean data,

noise section and the noisy data are shown in Figures 7a, 7b, and 7c, respectively. According

to the PATCHUNET denoised section, Figure 7g, the S/N is enhanced from -2.81 dB to

9.60 dB. However, the S/N of the denoised section obtained by F XDECON , Figure 7d, is

8.03 dB. The removed noise sections and local similarity maps, as shown in Figures 7h,7e,7i

and 7f, indicate a small amount of useful signal loss.

Secondly, a land seismic section is used to test the performance of PATCHUNET, which

was previously used by Liu et al. (2012) and Chen and Fomel (2015). The field example

is shown in Figure 8a, while the denoised and removed sections obtained by F XDECON

are shown in Figures 8b and 8c, respectively. The denoised and the noise sections extracted

by PATCHUNET are shown in Figures 8e and 8f, respectively. Correspondingly, the local

12
similarity maps of the F XDECON and the PATCHUNET are shown in Figures 8d and 8g,

respectively. In this example, the signal leakage of PATCHUNET is lower than the signal

leakage corresponding to the F XDECON (indicated by rectangles in Figures 8d and 8g),

both algorithms suppress random noise effectively.

Thirdly, a time-migrated dataset (Fomel et al., 2013) is tested using PATCHUNET. The

input field data, Figure 9a, contains 310 traces and 700 samples in the time direction with

a sampling rate of 1 ms. Figures 9b and 9e show the denoised section corresponding to

F XDECON and PATCHUNET, respectively. Checking the F XDECON noise section,

Figure 9c, we can notice many coherent signals, indicating a significant leakage of useful

signal energy which appears clearly in the corresponding local similarity map in Figure 9d

(indicated by rectangle). However, the removed noise section after using the PATCHUNET,

Figure 9f, has a smaller leakage of coherent useful signals that are almost negligible in the

local similarity map shown in Figure 9g.

Finally, we test the PATCHUNET on a field seismic data consisted of 128 traces as

shown in Figure 10, which was previously used by Saad and Chen (2020b). Figures 11a,

11d and 11g show the denoised section obtained by F XDECON , DDAE (Saad and Chen,

2020b), and PATCHUNET, respectively. The removed noise sections are shown in Fig-

ures 11b, 11e and 11h, respectively. The results indicate that the PATCHUNET has the

best denoising performance with the least signal leakage as indicated in the local similarity

maps of Figure 11.

Note that, for all the synthetic and field examples, the pre-whitening (mu) and the filter

order for the F XDECON method are 0.01 and 10, respectively. While, for the f − x SSA

method, the number of singular value to be preserved is 10. For the local similarity map,

13
the smoothing radius are 20, 20, and 1, while the number of iterations is 20.

DISCUSSION

PATCHUNET Parameters Selection

To reach the best performance of the PATCHUNET, the number of the hidden layers, the

neuron numbers per layer, the dropout rate, type of activation function, smoothing factor,

patch size, shift size, and the training/testing splitting rate are tuned using the following

procedures:

A) Varying the patch size (M ×N ) and the number of shift size of the patching technique,

fixing other network parameters.

B) Fixing the network parameters, and varying the hidden layers and the number of

neurons per layer.

C) Using the optimal network parameters obtained by the previous steps, varying the

dropout rate.

D) Varying the type of activation function, using the optimal network parameters ob-

tained by the previous steps.

E) Varying the smoothing factor, α, using the optimal network parameters obtained by

the previous steps.

F) Varying the training/testing splitting rate, using the optimal network parameters

obtained by the previous steps.

The third synthetic example is used as a case study for tuning the PATCHUNET pa-

rameters. First, we fix the number of hidden layers to be four hidden layers per each encoder

14
and decoder with 256, 128, 64, 32 neurons, respectively. Also, the activation function type,

smoothing factor, dropout rate, number of epochs, and training/testing splitting rate are

fixed to be ELU , 1, 0.1, 50, and 80/20 (0.8), respectively. Following this, we vary the patch

size (M × N ) and the shift size (k), obtaining the corresponding S/N for each case as shown

in Figure 12. The best performance can be noticed at patch size of 48 × 48 and 64 × 64 with

one shift sample. We choose the patch size of 48 × 48 because it reduces the total number of

network parameters while having almost the same performance of the batch size of 64 × 64.

Secondly, we fix the number of hidden layers, changing the number of neurons per layer.

The number of neurons is halved each encoder-hidden layer and is doubled each decoder-

hidden layer. For instance, using two hidden layers, the first hidden layer has 256 neurons

and the second hidden layer has 128 neurons, and vice versa in the decoder part. Following

this, we increase the number of hidden layers and try several number of neurons. In this

test, we set the patch size, shift size, activation function type, smoothing factor, dropout

rate, number of epochs, and training/testing splitting rate to be 48 × 48, 1, ELU 1, 0.1, 50,

and 80/20 (0.8), respectively. Figure 13 shows the performance of PATCHUNET using a

different number of hidden layers and neurons. The best performance can be noticed when

using six hidden layers or eight hidden layers with 128 neurons in the first hidden layer.

We choose to use six hidden layers rather than eight hidden layers because it reduces the

number of network parameters and has a similar performance to the eight hidden layers.

Accordingly, 128, 64, 32, 16, 8, and 4 neurons are used for the six hidden layers, respectively.

Thirdly, using the optimal PATCHUNET parameters obtained by the previous steps,

we adapt the dropout rate, Dr, and obtain the S/Ns of each case as shown in Figure 14.

The dropout rate of 0.1 denotes the best PATCHUNET performance.

15
 Fourthly, we use different activation function types; sigmoid, tanh, rectified linear unit

(RELU), and ELU. The S/N is enhanced to 11.06 dB, 12.01 dB, 12.12 dB, and 12.41 dB,

corresponding to sigmoid, tanh, ReLU, and ELU activation function, respectively. Accord-

ingly, we find ELU gives the best denoising performance compared to the other activation

function types. We tune the smoothing factor, α, of the ELU activation function, and we

found the default value, 1, gives the best performance.

Finally, we tune the training/testing splitting rate as shown in Figure 15. The PATCHUNET

has the capability to learn with only 30% of the input patches. At a rate of 0.3, the output

S/N is 11.80 dB, which is considered as a good denoising result. We set the training/testing

splitting rate to be 0.8, which gives the maximum S/N. Figure 16 shows a few examples

of the input and output test patches of PATCHUNET. Accordingly, the PATCHUNET

suppresses the random noise effectively and preserves the seismic signal well.

Stopping criterion

For each seismic example, 50 epochs are used to finish the training process. However, the

training process takes a slightly larger time. Each seismic data is divided into training

and testing patches. The patch size is 48 × 48 with shift size of one sample and the

training/testing splitting ratio is 0.8. For example, 77,748, 164,892, 12,852, and 19,905

patches are used to train the first, the second, the third, and the fourth synthetic examples

with a training execution time of 478s, 1001s, 85s, and 151s, respectively. The training

time for the first, the second, the third, and the fourth field examples are 701s, 990s, 201s,

780s with training patches of 111,652, 157,728, 30,132, and 137,391, respectively. We set a

stopping criterion for the training process to reduce the training process time. If the loss

16
value does not decrease for five consecutive epochs, the training process is stopped and the

last model is considered as the best model. The output S/N of the PATCHUNET does not

change but the training time is decreased, as detailed in Table 1. Accordingly, the training

time significantly decreases after applying the stopping criterion.

S/N test

We create 10 versions of the third synthetic example, each version has a different S/N.

The input S/Ns and the output S/Ns obtained by PATCHUNET are shown in Figure 17.

Accordingly, the PATCHUNET attenuates random noise despite the existence of a large

amount of random noise, e.g., when the input S/N is as low as -14.2 dB.

CONCLUSION

In this study, we propose a deep learning approach (PATCHUNET) to suppress random

noise. The input noisy data is split into several patches using a patching technique.

PATCHUNET compresses the input patches to a new representation. Following this, the

new representation is decoded to reconstruct the seismic signal without random noise. Skip

connections are used as residual learning, allowing the PATCHUNET to combine between

the local features extracted by the encoder and the global features obtained by the decoder.

Dropout layers are utilized to maintain the most meaningful features and eliminate the

others. The output patches corresponding to PATCHUNET are unpatched to retrieve the

seismic samples. The PATCHUNET is evaluated using eight synthetic and field examples

and compared to benchmark algorithms. The local similarity map is calculated to evalu-

ate the PATCHUNET denoising performance. Accordingly, the PATCHUNET suppresses

17
random noise and preserves the coherent seismic signal effectively. The PATCHUNET is

learned using 30% of the input patches with a good denoising performance. In addition,

the PATCHUNET can denoise the extremely noisy data, e.g., when S/N is -14.2 dB.

DATA AVAILABILITY STATEMENT

All datasets and source codes associated with this research will be made available online in

a github repository.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

18
 REFERENCES

Abma, R., and J. Claerbout, 1995, Lateral prediction for noise attenuation by t-x and f-x

techniques: Geophysics, 60, no. 6, 1887–1896.

Anvari, R., M. A. N. Siahsar, S. Gholtashi, A. R. Kahoo, and M. Mohammadi, 2017, Seismic

random noise attenuation using synchrosqueezed wavelet transform and low-rank signal

matrix approximation: IEEE Transactions on Geoscience and Remote Sensing, 55, no.

11, 6574–6581.

Canales, L., 1984, Random noise reduction: 54th Annual International Meeting, SEG,

Expanded Abstracts, 525–527.

Chen, W., L. Yang, B. Zha, M. Zhang, and Y. Chen, 2020, Deep learning reservoir porosity

prediction based on multilayer long short-term memory network: Geophysics, 85, no. 4,

WA213WA225.

Chen, Y., and S. Fomel, 2015, Random noise attenuation using local signal-and-noise or-

thogonalization: Geophysics, 80, no. 6, WD1–WD9.

Chen, Y., S. Fomel, and J. Hu, 2014, Iterative deblending of simultaneous-source seismic

data using seislet-domain shaping regularization: Geophysics, 79, no. 5, V179–V189.

Chen, Y., and J. Ma, 2014, Random noise attenuation by f-x empirical-mode decomposition

predictive filtering: Geophysics, 79, no. 3, V81–V91.

Chen, Y., M. Zhang, M. Bai, and W. Chen, 2019, Improving the signal-to-noise ratio of

seismological datasets by unsupervised machine learning: Seismological Research Letters,

90, 1552–1564.

Cheng, G., C. Yang, X. Yao, L. Guo, and J. Han, 2018, When deep learning meets met-

ric learning: Remote sensing image scene classification via learning discriminative cnns:

IEEE Transactions on Geoscience and Remote Sensing, 56, no. 5, 2811–2821.

19
Cheriyadat, A. M., 2013, Unsupervised feature learning for aerial scene classification: IEEE

Transactions on Geoscience and Remote Sensing, 52, no. 1, 439–451.

Dagnino, D., V. Sallarès, and C. R. Ranero, 2017, Waveform-preserving processing flow

of multichannel seismic reflection data for adjoint-state full-waveform inversion of ocean

thermohaline structure: IEEE Transactions on Geoscience and Remote Sensing, 56, no.

3, 1615–1625.

Fomel, S., 2007, Local seismic attributes: Geophysics, 72, no. 3, A29–A33.

Fomel, S., and Y. Liu, 2010, Seislet transform and seislet frame: Geophysics, 75, no. 3,

V25–V38.

Fomel, S., P. Sava, I. Vlad, Y. Liu, and V. Bashkardin, 2013, Madagascar: Open-source

software project for multidimensional data analysis and reproducible computational ex-

periments: Journal of Open Research Software, 1, no. 1, 8.

Gan, S., S. Wang, Y. Chen, S. Qu, and S. Zu, 2016, Velocity analysis of simultaneous-source

data using high-resolution semblancecoping with the strong noise: Geophysical Journal

International, 204, no. 2, 768–779.

Gulunay, N., 2000, Noncausal spatial prediction filtering for random noise reduction on 3-d

poststack data: Geophysics, 65, no. 5, 1641–1653.

Huang, W., R. Wang, Y. Chen, H. Li, and S. Gan, 2016, Damped multichannel singular

spectrum analysis for 3D random noise attenuation: Geophysics, 81, no. 4, V261–V270.

Kingma, D. P., and J. Ba, 2014, Adam: A method for stochastic optimization: arXiv

preprint arXiv:1412.6980.

Kong, F., F. Picetti, V. Lipari, P. Bestagini, and S. Tubaro, 2020, Deep prior-based seismic

data interpolation via multi-res u-net, in SEG Technical Program Expanded Abstracts

2020: Society of Exploration Geophysicists, 3159–3163.

20
LeCun, Y., Y. Bengio, and G. Hinton, 2015, Deep learning: Nature, 521, no. 7553, 436–444.

Li, S., B. Liu, Y. Ren, Y. Chen, S. Yang, Y. Wang, and P. Jiang, 2020, Deep learning

inversion of seismic data: IEEE Transactions on Geoscience and Remote Sensing, 58, no.

3, 2135–2149.

Liu, B., S. Yang, Y. Ren, X. Xu, P. Jiang, and Y. Chen, 2020a, Deep learning seismic

full waveform inversion for realistic structure models: Geophysics, doi: 10.1190/geo2019–

0435.1.

Liu, D., W. Chen, M. D. Sacchi, and H. Wang, 2020b, Should we have labels for deep

learning ground roll attenuation?, in SEG Technical Program Expanded Abstracts 2020:

Society of Exploration Geophysicists, 3239–3243.

Liu, G., and X. Chen, 2013, Noncausal f-x-y regularized nonstationary prediction filtering

for random noise attenuation on 3D seismic data: Journal of Applied Geophysics, 93,

60–66.

Liu, G., X. Chen, J. Du, and K. Wu, 2012, Random noise attenuation using f-x regularized

nonstationary autoregression: Geophysics, 77, no. 2, V61–V69.

Liu, G., Y. Liu, C. Li, and X. Chen, 2018, Weighted multisteps adaptive autoregression for

seismic image denoising: IEEE Geoscience and Remote Sensing Letters, 15, 1342–1346.

Liu, W., S. Cao, Y. Chen, and S. Zu, 2016, An effective approach to attenuate random

noise based on compressive sensing and curvelet transform: Journal of Geophysics and

Engineering, 13, no. 2, 135.

Mousavi, S. M., and C. A. Langston, 2017, Automatic noise-removal/signal-removal based

on general cross-validation thresholding in synchrosqueezed domain and its application

on earthquake data: Geophysics, 82, no. 4, V211–V227.

Mousavi, S. M., C. A. Langston, and S. P. Horton, 2016, Automatic microseismic denoising

21
 and onset detection using the synchrosqueezed continuous wavelet transform: Geophysics,

81, no. 4, V341–V355.

Mousavi, S. M., W. Zhu, W. Ellsworth, and G. Beroza, 2019, Unsupervised clustering of

seismic signals using deep convolutional autoencoders: IEEE Geoscience and Remote

Sensing Letters, 16, 1693–1697.

Neelamani, R., A. I. Baumstein, D. G. Gillard, M. T. Hadidi, and W. L. Soroka, 2008,

Coherent and random noise attenuation using the curvelet transform: The Leading Edge,

27, no. 2, 240–248.

Oropeza, V., and M. Sacchi, 2011, Simultaneous seismic data denoising and reconstruction

via multichannel singular spectrum analysis: Geophysics, 76, no. 3, V25–V32.

Park, M. J., J. Jennings, B. Clapp, and B. Biondi, 2020, Seismic data interpolation using

a pocs-guided deep image prior, in SEG Technical Program Expanded Abstracts 2020:

Society of Exploration Geophysicists, 3154–3158.

Saad, O., and Y. Chen, 2020a, Automatic waveform-based source-location imaging using

deep learning extracted microseismic signals: Geophysics, 85, no. 6, KS171KS183.

Saad, O. M., and Y. Chen, 2020b, Deep denoising autoencoder for seismic random noise

attenuation: Geophysics, 85, no. 4, V367–V376.

Sacchi, M. D., 2009, FX singular spectrum analysis: CSPG CSEG CWLS Convention,

392–395.

Shi, Y., X. Wu, and S. Fomel, 2020, Deep learning parameterization for geophysical inverse

problems: SEG 2019 Workshop: Mathematical Geophysics: Traditional vs Learning,

Beijing, China, 5-7 November 2019, Society of Exploration Geophysicists, 36–40.

Siahsar, M. A. N., V. Abolghasemi, and Y. Chen, 2017, Simultaneous denoising and inter-

polation of 2D seismic data using data-driven non-negative dictionary learning: Signal

22
 Processing, 141, 309–321.

Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, 2014,

Dropout: a simple way to prevent neural networks from overfitting: The journal of

machine learning research, 15, no. 1, 1929–1958.

Trickett, S. R., 2003, F-xy eigen image noise suppression: Geophysics, 68, 751–759.

Ulyanov, D., V. A., and V. S. Lempitsky, 2017, Deep image prior: arXiv, abs/1711.10925.

Wang, E., and J. Nealon, 2019, Applying machine learning to 3D seismic image denoising

and enhancement: Interpretation, 7, SE131–SE139.

Wang, H., Q. Zhang, G. Zhang, J. Fang, and Y. Chen, 2020, Self-training and learning the

waveform features of microseismic data using an adaptive dictionary: Geophysics, 85,

KS51–KS61.

Yilmaz, Ö., I. Tanir, C. Gregory, and F. Zhou, 2001, Interpretive imaging of seismic data:

The Leading Edge, 20, no. 2, 132–144.

Yu, S., J. Ma, X. Zhang, and S. Osher, 2016, Monte carlo data-driven tight frame for seismic

data recovery: Geophysics, 81, no. 4, V327–V340.

Zhang, G., Z. Wang, and Y. Chen, 2018, Deep learning for seismic lithology prediction:

Geophysical Journal International, 215, 13681387.

Zhu, W., S. M. Mousavi, and G. C. Beroza, 2019, Seismic signal denoising and decomposi-

tion using deep neural networks: IEEE Transactions on Geoscience and Remote Sensing,

57, 9476–9488.

Zu, S., H. Zhou, R. Ru, M. Jiang, and Y. Chen, 2019, Dictionary learning based on dip

patch selection training for random noise attenuation: Geophysics, 84, no. 3, V169–V183.

23
 LIST OF FIGURES

1 (a) Patching technique. (b) Unpatching technique.

2 PATCHUNET architecture.

3 First synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) f − x SSA denoised section. (e) Noise section corresponding to f − x SSA. (f)

Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h) Noise

section corresponding to PATCHUNET. (i) Similarity map between panels (g) and (h).

4 Second synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) F XDECON denoised section. (e) Noise section corresponding to F XDECON .

(f) Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h)

Noise section corresponding to PATCHUNET. (i) Similarity map between panels (g) and

(h).

5 Third synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) F XDECON denoised section. (e) Noise section corresponding to F XDECON .

(f) Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h)

Noise section corresponding to PATCHUNET. (i) Similarity map between panels (g) and

(h).

6 Fourth synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) F XDECON denoised section. (e) Noise section corresponding to F XDECON .

(f) Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h)

Noise section corresponding to PATCHUNET. (i) Similarity map between panels (g) and

(h).

7 First field example. (a) Clean data. (b) Random noise section. (c) Noisy data.

(d) F XDECON denoised section. (e) Noise section corresponding to F XDECON . (f)

24
Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h) Noise

section corresponding to PATCHUNET. (i) Similarity map between panels (g) and (h).

8 Second field example. (a) Input field data. (b) F XDECON -denoised section. (c)

Noise section corresponding to F XDECON . (d) Similarity map between panels (b) and

(c). (e) PATCHUNET denoised section. (f) Noise section corresponding to PATCHUNET.

(g) Similarity map between panels (e) and (f).

9 Third field example. (a) Input field data. (b) F XDECON -denoised section. (c)

Noise section corresponding to F XDECON . (d) Similarity map between panels (b) and

(c). (e) PATCHUNET denoised section. (f) Noise section corresponding to PATCHUNET.

(g) Similarity map between panels (e) and (f).

10 Fourth field example.

11 Fourth field example. (a) F XDECON -denoised section. (b) Noise section corre-

sponding to F XDECON . (c) Similarity map between panels (a) and (b). (d) PATCHUNET

denoised section. (e) Noise section corresponding to PATCHUNET. (f) Similarity map be-

tween panels (d) and (e).(g) DDAE-denoised section. (h) Noise section corresponding to

DDAE. (i) Similarity map between panels (g) and (h).

12 The effect of changing the patch size and shift size.

13 The effect of changing the number of hidden layers and neurons per each layer.

14 The effect of changing the dropout rate.

15 The effect of changing the training/testing splitting rate.

16 (a)-(i) Examples of the input and output test patches of PATCHUNET.

17 S/N test of the PATCHUNET.

25
 Example Number of patches Training time Training time

(P ) before applying after applying

the stopping criterion(s) the stopping criterion(s)

First synthetic 77,748 478 300

Second synthetic 164,892 1001 631

Third synthetic 12,852 85 54

Fourth synthetic 19,905 151 49

First field 111,652 701 505

Second field 157,728 990 665

Third field 30,132 201 139

Fourth field 137,391 780 574

Table 1: Training time comparison before and after using the stopping criterion.

26
 Patch size
Shift size

Patching
……

2D Matrix
2D seismic Image

(a)

Average

…… Unpatching

2D Matrix 2D seismic Image

(b)

Figure 1: (a) Patching technique. (b) Unpatching technique.

27
𝟐𝑫 𝑰𝒎𝒂𝒈𝒆
Each 𝒍𝒆 and 𝒍𝒅 Block
Encoder q NR e NR
𝑷𝒂𝒕𝒄𝒉𝒊𝒏𝒈

𝒊/𝒑 𝒍𝒆𝟏 𝒍𝒆𝟐 𝒍𝒆𝟑 𝒍𝒆𝟒 𝒍𝒆𝟓 𝒍𝒆𝟔

128 NR 64 NR 32 NR 16 NR 8 NR 8 NR
2304 NR
Activation
SC SC SC SC SC . . . Function
. .
2304 NR 128 NR 64 NR 32 NR 16 NR 8 NR 8 NR
. . .
. .
𝑭. 𝑪 𝒍𝒅𝟔 𝒍𝒅𝟓 𝒍𝒅𝟒 𝒍𝒅𝟑 𝒍𝒅𝟐 𝒍𝒅𝟏 . . . (ELU)

𝑼𝒏𝒑𝒂𝒕𝒄𝒉𝒊𝒏𝒈
Decoder
𝟐𝑫 𝑰𝒎𝒂𝒈𝒆

SC ~ Skip Connection F.C ~ Fully-Connected NR ~ Neurons Dropout

Figure 2: PATCHUNET architecture.

28
 0 0 0

0.5 0.5 0.5

Time (s)

Time (s)

Time (s)
1 1 1

1.5 1.5 1.5

2 2 2
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Trace Trace Trace

(a) (b) (c)

0 0 0 1

0.5 0.5 0.5 0.5

Amplitude
Time (s)
Time (s)

Time (s)

1 1 1 0

1.5 1.5 1.5 -0.5

2 2 2 -1
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Trace Trace Trace

(d) (e) (f)

0 0 0 1

0.5 0.5 0.5 0.5

Amplitude
Time (s)
Time (s)

Time (s)

1 1 1 0

1.5 1.5 1.5 -0.5

2 2 2 -1
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Trace Trace Trace

(g) (h) (i)

Figure 3: First synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) f − x SSA denoised section. (e) Noise section corresponding to f − x SSA. (f)

Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h) Noise

section corresponding to PATCHUNET. (i) Similarity map between panels (g) and (h).

29
 0 0 0

0.5 0.5 0.5

Time (s)

Time (s)
Time (s)

1 1 1

1.5 1.5 1.5

2 2 2
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Distance (km) Distance (km) Distance (km)

(a) (b) (c)

0 0 0 1

0.2
0.5 0.5 0.5
0.4

Amplitude
Time (s)

Time (s)
Time (s)

1 0.6 1 0

0.8
1.5 1.5 -0.5
1

2 1.2 2 -1
0 1 2 3 4 5 0 20 40 60 0 1 2 3 4 5
Distance (km) Trace Distance (km)

(d) (e) (f)

0 0 0 1

0.5 0.5 0.5 0.5

Amplitude
Time (s)

Time (s)

Time (s)

1 1 1 0

1.5 1.5 1.5 -0.5

2 2 2 -1
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Distance (km) Distance (km) Distance (km)

(g) (h) (i)

Figure 4: Second synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) F XDECON denoised section. (e) Noise section corresponding to F XDECON .

(f) Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h)

Noise section corresponding to PATCHUNET. (i) Similarity map between panels (g) and

(h).

30
 0 0 0

0.2 0.2 0.2

0.4 0.4 0.4

Time (s)

Time (s)

Time (s)
0.6 0.6 0.6

0.8 0.8 0.8

1 1 1

1.2 1.2 1.2

0 20 40 60 0 20 40 60 0 20 40 60
Trace Trace Trace

(a) (b) (c)

0 0 0 1

0.2 0.2 0.2

0.5
0.4 0.4 0.4

Amplitude
Time (s)

Time (s)

Time (s)
0.6 0.6 0.6 0

0.8 0.8 0.8

-0.5
1 1 1

1.2 1.2 1.2 -1

0 20 40 60 0 20 40 60 0 20 40 60
Trace Trace Trace

(d) (e) (f)

0 0 0 1

0.2 0.2 0.2

0.5
0.4 0.4 0.4

Amplitude
Time (s)

Time (s)

Time (s)

0.6 0.6 0.6 0

0.8 0.8 0.8

-0.5
1 1 1

1.2 1.2 1.2 -1

0 20 40 60 0 20 40 60 0 20 40 60
Trace Trace Trace

(g) (h) (i)

Figure 5: Third synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) F XDECON denoised section. (e) Noise section corresponding to F XDECON .

(f) Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h)

Noise section corresponding to PATCHUNET. (i) Similarity map between panels (g) and

(h).

31
 0 0 0

0.5 0.5 0.5

1 1 1
Time (s)

Time (s)

Time (s)
1.5 1.5 1.5

2 2 2

2.5 2.5 2.5

3 3 3
0 20 40 60 0 20 40 60 0 20 40 60
Trace Trace Trace

(a) (b) (c)

0 0 0 1

0.5 0.5 0.5

0.5
1 1 1

Amplitude
Time (s)

Time (s)

Time (s)
1.5 1.5 1.5 0

2 2 2

2.5 2.5 -0.5

2.5

3 3 3
-1
0 20 40 60 0 20 40 60 0 20 40 60
Trace Trace Trace

(d) (e) (f)

0 0 0 1

0.5 0.5 0.5

0.5
1 1 1

Amplitude
Time (s)

Time (s)

Time (s)

1.5 1.5 1.5 0

2 2 2

2.5 2.5 -0.5

2.5

3 3 3
-1
0 20 40 60 0 20 40 60 0 20 40 60
Trace Trace Trace

(g) (h) (i)

Figure 6: Fourth synthetic example. (a) Clean data. (b) Random noise section. (c) Noisy

data. (d) F XDECON denoised section. (e) Noise section corresponding to F XDECON .

(f) Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h)

Noise section corresponding to PATCHUNET. (i) Similarity map between panels (g) and

(h).

32
 4 4 4

5 5 5

6 6 6
Time (s)

Time (s)

Time (s)
7 7 7

8 8 8

9 9 9

0 50 100 0 50 100 0 50 100

Trace Trace Trace

(a) (b) (c)

4 4 4 1

5 5 5
0.5
6 6 6

Amplitude
Time (s)

Time (s)

Time (s)
7 7 0
7

8 8 8
-0.5

9 9 9
-1
0 50 100 0 50 100 0 50 100
Trace Trace Trace

(d) (e) (f)

4 4 4 1

5 5 5
0.5
6 6 6

Amplitude
Time (s)

Time (s)

Time (s)

7 7 0
7

8 8 8
-0.5

9 9 9
-1
0 50 100 0 50 100 0 50 100
Trace Trace Trace

(g) (h) (i)

Figure 7: First field example. (a) Clean data. (b) Random noise section. (c) Noisy data.

(d) F XDECON denoised section. (e) Noise section corresponding to F XDECON . (f)

Similarity map between panels (d) and (e). (g) PATCHUNET denoised section. (h) Noise

section corresponding to PATCHUNET. (i) Similarity map between panels (g) and (h).

33
 0

0.2

0.4

Time (s)
0.6

0.8

1
0 100 200 300 400
Trace

(a)

0 0 0 1

0.2 0.2 0.2

0.5

Amplitude
0.4 0.4
Time (s)

0.4

Time (s)
Time (s)

0
0.6 0.6 0.6

0.8 0.8 0.8 -0.5

1 1 1
-1
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400
Trace Trace Trace

(b) (c) (d)

0 0 0 1

0.2 0.2 0.2

0.5

Amplitude
0.4 0.4
Time (s)

Time (s)

0.4
Time (s)

0
0.6 0.6 0.6

0.8 0.8 0.8 -0.5

1 1 1
-1
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400
Trace Trace Trace

(e) (f) (g)

Figure 8: Second field example. (a) Input field data. (b) F XDECON -denoised section. (c)

Noise section corresponding to F XDECON . (d) Similarity map between panels (b) and

(c). (e) PATCHUNET denoised section. (f) Noise section corresponding to PATCHUNET.

(g) Similarity map between panels (e) and (f).

34
 0

0.2

Time (s)
0.4

0.6

0 2 4 6 8 10
Distance (km)

(a)

0 0 0 1

0.5
0.2 0.2 0.2

Amplitude
Time (s)

Time (s)

Time (s)
0
0.4 0.4 0.4

-0.5
0.6 0.6 0.6

-1
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Distance (km) Distance (km) Distance (km)

(b) (c) (d)

0 0 0 1

0.5
0.2 0.2 0.2

Amplitude
Time (s)

Time (s)

Time (s)

0
0.4 0.4 0.4

-0.5
0.6 0.6 0.6

-1
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Distance (km) Distance (km) Distance (km)

(e) (f) (g)

Figure 9: Third field example. (a) Input field data. (b) F XDECON -denoised section. (c)

Noise section corresponding to F XDECON . (d) Similarity map between panels (b) and

(c). (e) PATCHUNET denoised section. (f) Noise section corresponding to PATCHUNET.

(g) Similarity map between panels (e) and (f).

35
 0

0.5
Time (s)

1.5

2
0 20 40 60 80 100 120
Trace

Figure 10: Fourth field example.

36
 0 0 0 1

0.5 0.5 0.5 0.5

Amplitude
Time (s)

Time (s)

Time (s)
1 1 1 0

1.5 1.5 1.5 -0.5

2 2 2 -1
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 50 100
Trace Trace Trace

(a) (b) (c)

0 0 0 1

0.5 0.5 0.5 0.5

Amplitude
Time (s)

Time (s)

Time (s)
1 1 1 0

1.5 1.5 1.5 -0.5

2 2 2 -1
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 50 100
Trace Trace Trace

(d) (e) (f)

0 0 0 1

0.5 0.5 0.5 0.5

Amplitude
Time (s)

Time (s)

Time (s)

1 1 1 0

1.5 1.5 1.5 -0.5

2 2 2 -1
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 50 100
Trace Trace Trace

(g) (h) (i)

Figure 11: Fourth field example. (a) F XDECON -denoised section. (b) Noise section

corresponding to F XDECON . (c) Similarity map between panels (a) and (b). (d)

PATCHUNET denoised section. (e) Noise section corresponding to PATCHUNET. (f)

Similarity map between panels (d) and (e).(g) DDAE-denoised section. (h) Noise section

corresponding to DDAE. (i) Similarity map between panels (g) and (h).

37
 Input S/N = -0.30 dB
12
Patch Size 8X8
10 Patch Size 16X16
Patch Size 32X32
8 Patch Size 48X48
S/N (dB)

Patch Size 64X64

6

5 10 15 20 25 30
Shift Samples

Figure 12: The effect of changing the patch size and shift size.

38
 Input S/N = -0.30 dB

12

10
Output S/N (dB)

6
2 Hidden Layers
4 4 Hidden Layers
6 Hidden Layers
8 Hidden Layers
2

0
100 200 300 400 500
Number of Neurons in the First Hidden layer

Figure 13: The effect of changing the number of hidden layers and neurons per each layer.

39
 Input S/N = -0.30 dB
12

10
Output S/N (dB)

0
0.2 0.4 0.6 0.8
Dropout Rate

Figure 14: The effect of changing the dropout rate.

40
 Input S/N = -0.30 dB
12

10
Output S/N (dB)

0
0.2 0.4 0.6 0.8
Training/Testing Ratio

Figure 15: The effect of changing the training/testing splitting rate.

41
 Input test patch Input test patch Input test patch

PATCHUNET output PATCHUNET output PATCHUNET output

(a) (b) (c)

Input test patch Input test patch Input test patch

PATCHUNET output PATCHUNET output PATCHUNET output

(d) (e) (f)

Input test patch Input test patch Input test patch

PATCHUNET output PATCHUNET output PATCHUNET output

(g) (h) (i)

Figure 16: (a)-(i) Examples of the input and output test patches of PATCHUNET.

42
 15
Output S/N (dB)

10

0
-10 -5 0 5
Input S/N (dB)

Figure 17: S/N test of the PATCHUNET.

43