Detection and Delimitation of Natural Gas in Seismic Images using

MLP-Mixer and U-Net

Carolina L. S. Cipriano

, Domingos A. D. Junior

, Petterson S. Diniz

, Luiz F. Marin

Anselmo C. Paiva

, Jo

ao O. B. Diniz

1,3

and Arist

ofanes C. Silva

Applied Computer Group NCA-UFMA, Federal University of Maranhao (UFMA), Sao Lu

ıs, Brazil

Tecgraf Institute, Pontiﬁcal Catholic University of Rio de Janeiro (PUC-Rio), Rio de Janeiro, Brazil

abrica de Inovac¸

ao, Instituto Federal do Maranh

ao, Graja

u, Brazil

Keywords:

Hydrocarbons, Seismic Images, MLP-Mixer, U-Net, DenseNet, ResNet, Machine Learning.

Abstract:

The seismic data acquired through the seismic reﬂection method is important for hydrocarbon prospecting.

As an example of hydrocarbon, we have natural gas, one of the leading and most used energy sources in the

current scenario. The techniques for analyzing these data are challenging for specialists. Due to the noisy

nature of data acquisition, it is subject to errors and divergences between the specialists. The growth of deep

learning has brought great highlights to tasks of segmentation, classiﬁcation, and detection of objects in images

from different areas. Consequently, the use of machine learning in seismic data has also grown. Therefore,

this work proposes an automatic detection and delimitation of the natural gas region in seismic images (2D)

using MLP-Mixer and U-Net. The proposed method obtained competitive results with an accuracy of 99.6%

(inline) and 99.55% (crossline); speciﬁcity of 99.79% (inline) and 99.73% (crossline).

1 INTRODUCTION

Hydrocarbons are molecules made up of hydrogen

and carbon. They are present in our energy resources,

such as natural gas. The occurrence of hydrocarbons

varies in space and time, as once important producing

regions have already exhausted their reserves, and

new ones are found in other areas (Teixeira et al.,

2009).

Most analysis and prospecting rely on technology

to detect and determine the extent of these deposits.

Geophysical surveys in the gas industry are mainly

conducted using seismic reﬂection techniques (Cox,

1999). Because they are more economical than good

drilling, it is possible to extract data regarding the

geometry and structure of the layers, rock types,

lithology, and physical properties.

Due to the seismic data’s low resolution and noisy

nature, the data interpretation is challenging. The

expert often creates several alternatives of the same

seismic structure when in doubt. Furthermore, it

is not uncommon for the team to disagree with the

interpretation and consider that parts of the data can

be reinterpreted (Patel et al., 2008). In this scenario,

machine learning has been used for the segmentation,

classiﬁcation, and detection of natural gas in 1D, 2D,

and 3D seismic data.

In (Santos et al., 2019), they proposed a new

approach to detect hydrocarbon indicators in seismic

data using seismic trace and a Long Short-Term

Memory (LSTM) neural network. They used a one-

dimensional way along the seismic trace. In this

process, each seismic trace was extracted using forty

samples of window length of one sample overlapping

each window. The public database used for gases

identiﬁcation was the Netherlands F3-Block. Using

accuracy as the primary metric to automatically

delimit gas pocket locations, the model achieved 97%.

In (El Zini et al., 2019) they proposed a bright

spot detection method. Bright spots are strong

indicators of the presence of natural gas. The model

used SeisNet, a convolutional neural network with a

”butterﬂy” architecture. The model also relied on data

augmentation and transfer learning to overcome the

data cap problem. The data used in SeisNet training

is adopted from (Rizk et al., 2017) and consists of 110

grayscale images. As a result, it reached 95.6% of the

F1 score and accuracy with an average absolute error

that did not exceed 0.04% of the total volume.

578

Cipriano, C., Junior, D., Diniz, P., Marin, L., Paiva, A., Diniz, J. and Silva, A.

Detection and Delimitation of Natural Gas in Seismic Images using MLP-Mixer and U-Net.

DOI: 10.5220/0011075000003179

In Proceedings of the 24th International Conference on Enterprise Information Systems (ICEIS 2022) - Volume 1, pages 578-585

ISBN: 978-989-758-569-2; ISSN: 2184-4992

Therefore, this work proposes an automatic

detection method and delimitation of regions where

natural gas accumulation can occur in 2D seismic

images. It consists of MLP-Mixer for classifying

these regions and U-Net for delimiting the extension

of these accumulations.

As contributions of this research, we highlight

1) The use of the MLP-Mixer to detect the regions

of interest, reducing the search ﬁeld of these

accumulations, and 2) An automatic 2D method of

direct detection of hydrocarbons.

This paper is organized as follows: Section 2

presents the proposed method for automatic detection

and delimitation of natural gas. Section 3 presents

and discusses the results. Finally, Section 4 presents

the ﬁnal considerations of this study.

2 MATERIALS AND METHOD

This section details the steps of the proposed method.

After data acquisition and processing, the proposed

method consists of three main steps: 1) Model entry,

where data preparation takes place; 2) Detection of

the region of interest, where the classiﬁcation of

patches between gas and non-gas occurs; and 3)

Gas delimitation, where the segmentation of the gas

extension in the regions of interest occurs. Figure 1

summarizes these steps. And ﬁnally, the results are

evaluated.

2.1 Data Acquisition

For this work, we used the public database called

Netherlands F3-Block. F3 is a block in the Dutch

sector of the North Sea covered by 3D seismic

that was acquired to explore gas (Nubis, 1987).

The F3 has gas markings for both inlines and

crosslines (Tecgraf Institute (PUC-Rio)). Crosslines

are perpendicular lines (Schlumberger, 2021a), while

inlines are in the same direction of the data acquisition

(Schlumberger, 2021b). The 3D seismic cube was

delimited, as in (Santos et al., 2019) by Tecgraf

Institute (PUC-Rio). And ﬁnally, converted to 2D as

shown in Figure 2.

A seismic line within a 3D survey parallels the

direction in which the data was acquired. In marine

seismic data, the inline is one in the recording vessel

that tows the coils.

2.2 Model Input

After acquiring and processing the 3D data into the

2D data, the slices were loaded and saved in grayscale

by the matplotlib python library, as shown in Figure 3.

Due to the small amount of data (242 crossline slices

and 449 inline slices) and the inline and crossline lines

having different dimensions, we extracted 2D patches

with a size of 128 × 128 to increase the number of

samples and standardize these dimensions.

2.3 Gas Region Detection

After extracting patches at the model input, the

samples were presented to the MLP-Mixer (Tolstikhin

et al., 2021) for classifying gas patches and non-gas

patches.

MLP-Mixer is a classiﬁcation network with

architecture based only on multilayer perceptrons. It’s

a competitive alternative that doesn’t use convolutions

or self-attention. And its architecture is based entirely

on multilayer perceptrons (MLPs) that are repeatedly

applied to spatial locations or resource channels.

(Tolstikhin et al., 2021).

Figure 4 presents the MLP-Mixer architecture,

where: input consists of patches extracted from the

input image; mixer layer blocks and sorting head

(Tolstikhin et al., 2021).

Figure 1: Steps of the proposed method.

Detection and Delimitation of Natural Gas in Seismic Images using MLP-Mixer and U-Net

579

Figure 2: Illustrative picture. Seismic cube and 2D slices.

Figure 3: A) grayscale seismic image and B) corresponding

binary marking (Tecgraf Institute (PUC-Rio)).

Figure 4: Adapted from MLP-Mixer architecture

(Tolstikhin et al., 2021).

The input maintains the dimensionality of patches,

which reduces the computational cost. The

mixer layer block has two MLP layers, each

consisting of two fully connected layers and a

GELU nonlinearity: channel-mixing MLPs and

token-mixing MLPs. Channel-mixing layer allows

communication between different channels. Token-

mixing layer allows communication between different

spatial locations (tokens) (Tolstikhin et al., 2021).

The head classiﬁer contains dropout, Global average

pooling, and softmax for classiﬁcation.

2.4 Gas Extension Delimitation

After detecting gas regions by the classiﬁcation

networks, the samples classiﬁed as gas samples are

presented to the Convolutional Neural Network (U-

Net) for segmentation of the gas in the patches.

The U-Net was specially designed to segment

medical images, but as it shows good results in

other ﬁelds, it was chosen to delimit the gas

extension. It is a fast network, relies heavily on data

augmentation, and can be trained end-to-end from

very few images (Ronneberger et al., 2015). U-

Net simply concatenates the encoder feature maps to

decoder feature maps at every step. This architecture

allows the decoder to learn the relevant features lost

when grouped in the encoder. Its concatenation

connections at each step do this. Figure 5 introduces

the U-Net architecture.

Figure 5: U-Net architecture (Ronneberger et al., 2015).

2.5 Validation

We used seven metrics commonly found in

classiﬁcation and segmentation problems to evaluate

the results: Accuracy (ACC), Precision (PRE),

Sensitivity (SEN), F1 score (F1), Speciﬁcity (SPEC),

AUC, and Dice . These metrics evaluate the

intersection between the real (expert marking) and

what is proposed by the method (method result), as

shown in the following equations:

Accuracy =

T P + T N

T P + FP + FN + T N

(1)

Precision =

T P

T P + FP

(2)

Sensitivity =

T P

T P + FN

(3)

F1 − score = 2 ∗

Precision ∗ Sensitivity

Precision + Sensitivity

(4)

Speci f icity =

T N

T N + FP

(5)

AUC =

2T P

2T P + FP + FN

(6)

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

580

Dice =

2T P

2T P + FP + FN

(7)

Metrics are evaluated against image pixels.

Where:

• The True Positive (TP) is the number of pixels

classiﬁed as gas-containing pixels by the method

and the expert.

• The False Negative (FN) is the number of pixels

classiﬁed as pixels that do not contain gas by the

method but were classiﬁed as pixels that contain

gas by the expert.

• The False Positive (FP) is the number of pixels

classiﬁed as pixels that contain gas by the method

but were classiﬁed as pixels that do not contain

gas by the expert.

• The True Negative (TN) is the number of pixels

classiﬁed as gas-free pixels by the method and

also by the expert.

3 RESULTS

This section presents and discusses the results

achieved by the method developed for gas detection

and delimitation in seismic images. The analysis

sequence consists of 1) Training and testing, 2) Model

Input, 3) Gas region detection and 4) Gas extension

delimitation.

3.1 Training and Testing

The implementation was developed in Python 3.7

programming language, with Keras and TensorFlow

frameworks for U-Net and MLP-Mixer. We used the

Numpy package to manipulate the seismic data, and

the OpenCV and Matplotlib libraries to manipulate

the images. The machine used features hardware

consisting of an Intel(R) Core(TM) i7-6700 CPU @

3.40GHz, 16GB of RAM, and a 6GB GeForce GTX

1060 GPU.

3.2 Model Input

The public database called the F3 dataset (Section

2.1) contains 242 crossline and 449 inline lines.

The crossline lines were randomly divided into 162

for training, 40 for validation, and 40 for testing.

Likewise, the inline lines were divided into 329 for

training, 60 for validation, and 60 for testing. We

maintain the proportions according to (Santos et al.,

2019) for comparative purposes..

To input the gas region detection models, 128x128

patches were extracted with an overlap of 10, as

they showed better results in the MLP-Mixer than

the results without overlap. Table 1 presents the

number of samples generated for training, validation,

and testing in the crossline and inline lines.

Table 1: Number of samples generated for gas region

detection models.

Files Gas No Gas Total

Crossline

Training 162 4133 16279 20412

Validation 40 1057 3983 5040

Test 40 64 336 400

Inline

Training 329 10112 51082 61194

Validation 60 1809 9351 11160

Test 60 164 796 960

For the input of the gas extension delimitation

model, we extract patches with 128x128 pixels

without overlappingg, as they did not show signiﬁcant

improvements for U-Net. The number of samples

generated for training, validation, and testing in the

crossline and inline lines are presented in Table 2.

Table 2: Number of samples generated for the gas extension

delimitation models.

Files Gas No Gas Total

Crossline

Training 162 263 1357 1620

Validation 40 68 332 400

Test 40 64 336 400

Inline

Training 329 876 4388 5264

Validation 60 152 808 960

Test 60 164 796 960

3.3 Gas Region Detection

In the gas region detection step, two models were

generated: one for the inline lines and another for

the crossline lines, according to the division in the

Table 1. MLP-Mixer is a supervised classiﬁcation

architecture. Thus, for training the models (crossline

and inline), the extracted patches were labeled as

patches that contained gas and patches that did not

contain gas.

For this work, we used the MLP-Mixer

implementation of (Benjamin-Etheredge, 2021)

for two classes. It has a patch size of 8 (The patch

size must be divisible by the input data size); 4 mixer

layer blocks; 64 MLP token dimension; MLP channel

Detection and Delimitation of Natural Gas in Seismic Images using MLP-Mixer and U-Net

581

dimension of 128; and, hidden dimension of 128, as

shown in Table 3.

Table 3: Settings used in MLP-Mixer.

Input 128x128x1

Patch size 8

MLP token dimension 64

MLP channel dimension 128

Hidden dimension 32

The training conﬁguration was deﬁned for 25

epochs, using sparse categorical cross-entropy as a

loss function and f1 score as a metric. The Table 4

presents the results for gas region detection in seismic

images.

Table 4: Result of the gas region detection step.

MLP-Mixer

Crossline Inline

ACC (%) 96.15 98.55

SEN(%) 87.61 95.24

SPEC(%) 98.27 99.2

AUC (%) 92.94 97.22

PREC(%) 92.61 95.87

F1 (%) 90.04 95.56

For comparative purposes, other models were also

used for the same delimitation task they were: ResNet

(He et al., 2015) and the DenseNet (Huang et al.,

2017). ResNet and DenseNet were used to extract

features, then a classiﬁcation block was added to both

networks, as shown in Table 5.

Table 5: Classiﬁcation block for the gas region detection

comparison models.

Dense 512 + Relu

Dropout 0.5

Dense 256 + Relu

Dense 2 + Softmax

Table 6 presents results for gas region detection

in seismic images with comparison models. As

we can see, the MLP-Mixer achieved the best

gas detection results compared to the ResNet and

DenseNet models. Except for sensitivity and AUC in

the crossline, DenseNet obtained a higher result.

In addition to the results, the MLP-Mixer

presented a signiﬁcantly superior performance

concerning the training time, as shown in Table 7.

This performance is because the MLP-mixer does

not increase depth (channel) along with the layers, as

with convolutional networks.

The MLP-Mixer proved to be an excellent

classiﬁer for seismic images despite not using

Table 6: Comparison of methods for gas detection.

MLP-Mixer ResNet DenseNet

Crossline

ACC (%) 96.15 90.6 95.54

SEN(%) 87.61 59.94 88.91

SPEC(%) 98.27 98.19 97.18

AUC (%) 92.94 79.07 93.04

PREC(%) 92.61 89.15 88.65

F1 (%) 90.04 71.68 88.78

Inline

ACC (%) 98.55 98 97.03

SEN(%) 95.24 94.15 88.35

SPEC(%) 99.2 98.76 98.74

AUC (%) 97.22 96.45 93.54

PREC(%) 95.87 93.69 93.19

F1 (%) 95.56 93.92 90.71

Table 7: Comparison of models in relation to training time.

Time (min)

Crossline Inline

MLP-Mixer 33 101

ResNet 529 1308

DenseNet 837 2421

convolution and attention mechanisms as the main

models currently. It obtained competitive results in

a much shorter processing time.

3.4 Gas Extension Delimitation

In the delimitation step, we generated two models,

one for the inline lines and another for the crossline

lines. As U-Net is a semantic segmentation network,

we extracted 128x128 patches from the respective gas

and non-gas binary masks for training the models.

After some initial experiments with different

parameters, we set the U-Net training conﬁguration

to 50 epochs with a batch size of 4, the Adam

optimization function with a learning rate of 0.00001,

decay of 0.000001, clipvalue = 0.5, and Dice as a loss

and metric function validation evaluation.

Table 8 presents the ﬁnal result of the proposed

method. This result is the combination of the

detection step and the delimitation step. The inline

line method presents superior results to the crossline

method in the detection and delimitation step. The

reason is the outstanding amount of training images

for the inline.

Table 9 presents a comparison of the results

in relation to the work of (Santos et al., 2019).

Improvements in accuracy were observed with

99.6% (inline) and 99.55% (crossline); speciﬁcity

99.79%(inline), and 99.74%(crossline), showing that

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

582

Table 8: Final result of the proposed method.

Base F1 (%) ACC (%) PREC (%) SEN (%) SPEC (%) AUC (%) Dice (%)

Inline 84.18 99.6 84.0 86.85 99.79 93.32 84.18

Crossline 80.3 99.55 80.49 82.74 99.73 91.23 80.3

the proposed model was able to reduce the number of

false positives. However, he could not overcome the

sensitivity.

Table 9: Comparison of the proposed method with another

work in the literature.

Proposed (Santos et al.,2019)

Crossline

F1 (%) 80.3 -

ACC (%) 99.55 96.83

PREC(%) 80.49 -

SEN(%) 82.74 94.77

SPEC(%) 99.73 96.87

AUC (%) 91.23 98.71

Dice (%) 80.3 -

Inline

F1 (%) 84.18 -

ACC (%) 99.6 97.16

PREC(%) 84.0 -

SEN(%) 86.85 97.83

SPEC(%) 99.79 97.15

AUC (%) 93.32 98.8

Dice (%) 84.18 -

Figure 6 presents a case that shows the

contribution of MLP-Mixer in reducing false

positives. As we can see in the picture 6 - C), U-Net

mistakes the prediction of gas in regions similar to

the region of interest.

Figure 6: Crossline 909: A) Input image, B) Specialist

marking, C) Gas extension delimitation isolated prediction

result, D) Proposed model. The expert mark (red), gas

region detection result (purple) and gas range delimitation

result (green).

The approach of (Santos et al., 2019) of working

with seismic trace (1D) has the advance of generating

more data samples. Due to data imbalance, accuracy

alone cannot reﬂect the method’s performance. The

precision shows that the proposed model could reduce

the number of false positives, and the sensitivity

indicates that the delimitation made by the U-Net

suffered from the number of gas samples.

Next, we will see some examples of the results

of the proposed method. The ﬁgure 7 shows the

result where the MLP-Mixer was able to detect the

gas region. And the U-Net precisely segments the gas.

Figure 7: Inline 86 and 597: A) specialist marking; B)

Prediction result for the proposed method. Result of

gas region detection (purple) and result of gas extension

delimitation (green).

Figure 8 also presents a good result from the

MLP-Mixer, but with the difference that there are

false negatives in the U-Net prediction.

Figure 8: Inline 221: A) specialist marking; B) Prediction

result for the proposed method. Result of gas region

detection (purple) and result of gas extension delimitation

(green).

Figure 9 presents a result where the MLP-Mixer

was able to detect the gas region. And the U-Net

precisely segments the gas.

Figure 8 also presents a good result from the

MLP-Mixer, but with the difference that there are

false negatives in the U-Net prediction.

Detection and Delimitation of Natural Gas in Seismic Images using MLP-Mixer and U-Net

583

Figure 9: Crossline 511 and 100: A) specialist marking;

B) Prediction result for the proposed method. Result of

gas region detection (purple) and result of gas extension

delimitation (green).

Figure 10: Crossline 909: A) specialist marking; B)

Prediction result for the proposed method. Result of

gas region detection (purple) and result of gas extension

delimitation (green).

4 CONCLUSIONS

Due to its characteristics and the limited availability

of public data, the automatic detection and

delimitation of natural gas in seismic images

is a difﬁcult task. Gas occurrences have different

locations and sizes, in addition to having a lot of noise

that is derived from the acquisition process. These

characteristics make the detection and delimitation

process challenging since this location is difﬁcult,

even with the naked eye by the specialist.

The MLP-Mixer was very important to the method

in the gas detection regions. It managed to perform

well in the classiﬁcation of regions and, consequently,

contributed to the delimitation of gas only in the

correct regions. The delimitation is a more difﬁcult

task, but even so, U-Net managed to have a relevant

performance considering the imbalance of the gas and

non-gas areas.

In general, the results for detection and

delimitation of regions and gas were promising,

achieving advances in accuracy with 99.6% (inline)

and 99.55% (crossline); speciﬁcity 99.79% (inline)

and 99.74% (crossline), compared to related work.

As future work, we identiﬁed the need to:

1) Use transfer learning techniques, since seismic

data are scarce, deep learning models are sensitive

to the amount of data and data increase had

little contribution in experimental tests; 2) Use

other semantic segmentation techniques for gas

delimitation available in the literature; 3) And, ﬁnally,

the use of seismic bases from other regions to test the

generalization of the proposed model.

ACKNOWLEDGEMENTS

We thank the Conselho Nacional de Desenvolvimento

Cient

ıﬁco e Tecnol

ogico (CNPq), the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior

(CAPES), the N

ucleo de Computac¸

ao Aplicada da

Universidade Federal do Maranh

ao (NCA), and the

Tecgraf Institute (PUC-Rio) for their support to the

research.

REFERENCES

Benjamin-Etheredge (2021). mlp-mixer-keras. https://

github.com/Benjamin-Etheredge/mlp-mixer-keras0.

Cox, M. (1999). Static corrections for seismic reﬂection

surveys. Society of Exploration Geophysicists.

El Zini, J., Rizk, Y., and Awad, M. (2019). A deep transfer

learning framework for seismic data analysis: A case

study on bright spot detection. IEEE Transactions on

Geoscience and Remote Sensing.

He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep

residual learning for image recognition. CoRR,

abs/1512.03385.

Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,

K. Q. (2017). Densely connected convolutional

networks. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 4700–

4708.

Nubis, T. (1987). Project f3 demo 2020.

Patel, D., Giertsen, C., Thurmond, J., Gjelberg, J., and

Grøller, E. (2008). The seismic analyzer: Interpreting

and illustrating 2d seismic data. IEEE Transactions on

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

584

Visualization and Computer Graphics, 14(6):1571–

1578.

Rizk, Y., Partamian, H., and Awad, M. (2017). Toward real-

time seismic feature analysis for bright spot detection:

a distributed approach. IEEE Journal of Selected

Topics in Applied Earth Observations and Remote

Sensing, 11(1):322–331.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In International Conference on

Medical image computing and computer-assisted

intervention, pages 234–241. Springer.

Santos, L. F., Silva, R. M. G. E., Gattass, M., and Silva,

A. C. (2019). Direct hydrocarbon indicators based

on long short-term memory neural network. In SEG

Technical Program Expanded Abstracts 2019, pages

2373–2377. Society of Exploration Geophysicists.

Schlumberger (2021a). Oilﬁeld glossary.

Schlumberger (2021b). Oilﬁeld glossary.

Teixeira, W., Fairchild, T. R., Toledo, M. C. M. d., and

Taioli, F. (2009). Decifrando a terra.

Tolstikhin, I., Houlsby, N., Kolesnikov, A., Beyer, L.,

Zhai, X., Unterthiner, T., Yung, J., Steiner, A.,

Keysers, D., Uszkoreit, J., et al. (2021). Mlp-mixer:

An all-mlp architecture for vision. arXiv preprint

arXiv:2105.01601.

Detection and Delimitation of Natural Gas in Seismic Images using MLP-Mixer and U-Net

585