Seismic image regression

Randomly blank tracesworkflow:
To train our 2D Unetregression model we create a data set with 33% randomly blanked traces.
From this cube we extract examples for training in a restricted area. The trained model is
applied to the entire volume, whereby the area from which no examples are extracted acts as
blind test area. The real value is of course when we apply the trained model to an area with
real missing traces (which we don’t have in this case). Random blanking (replacing the values
with hard zeros) is done in OpendTect’sAttribute engine and can be done in different ways. In
this case, we will create an attribute set to perform the following tasks:
1. Math attribute with formula: “randg(1)”. This generates random values with a Gaussian
distribution and 1 standard deviation;
2. Apply this attribute to a horizon and save as horizon data;
3. Horizon attribute that retrieves the random values from the saved horizon data. A Horizon
attribute replaces a value at an inline, crossline position with the value extracted from the
given horizon;
4. Math attribute with formula: “abs(value)> 1 ? 0: seis”. We assign the retrieved horizon data
to the variable “value” and the seismic data to “seis”. This attribute assigns values larger
than the absolute value of 1 standard deviation to zero while all other values are given the
value of the seismic data.
5. Additional attributes in the set are used to compare/QCresults before and after prediction.

1. Selectthe 3D Attributes engine icon.
2. Createa new 3D attribute set
Theseattributes that will be explained in the
next steps.
3. Saveas attribute set with the name
‘ML_Random_holes_interpolation’.
Randomly blank tracesworkflow:
Create a new 3D attribute set to randomly blank traces as explained in the following steps.
ML_Random_holes_interpolation

Randomly blank tracesworkflow:
4. Create1stattribute with name
‘random’ as indicated in the attribute
set windowand Hit‘Add as new’.
5. SetMath attribute with formula:
“randg(1)”.
This generates random values with a
Gaussian distribution and 1 standard
deviation;
Apply this attribute to an horizon and
save as horizon data as indicated in
the next step.

6. Createa constant seismic horizon at Z = 1s.
7. Right mouse click on the 3D Horizon < New < With
constant Z.
8. EnterZ value (ms)= 1000. Type an Output 3D Horizon
name e.g. Hrz_1s. HitCreate.
9. Displaythe horizon –attribute ‘random’. Save as
Horizon data.
Randomly blank tracesworkflow:
Createa seismic horizon at Z = 1 s. Then apply the random attribute to this horizon and save this
as horizon data. This horizon data will beused in the attribute that does the actual blanking.

Randomly blank tracesworkflow:
Create an horizon attribute that retrieves the random values from the saved horizon data. The
horizon attribute replaces a value at an inline, crossline position with the value extracted from the
given horizon.
10. Create 2ndattribute ‘’Retrieve horizon data’’
as indicated in the attribute set windowand Hit
‘Add as new’.
11. Selectthe Input Data that will be blanked ‘4
Dip steered Median filter’.
12. Selectthe constant horizon ‘’Hrz_1s’’ created
in the previous step.
13. Select Output ‘’Horizon Data’’ and Horizon
Data ‘’random’’.

Randomly blank tracesworkflow:
Create an attribute that willrandomly blank traces as zeros in the input seismic.
14. Create3rdattribute ‘Seismic with random missing
traces as zeros’ as indicated in the attribute set
windowand Hit‘Add as new’.
15. Seta Math attribute with formula: “abs(value)> 1
? 0: seis”. Thisassign the retrieved horizon data to
the variable “value” and the seismic data to “seis”.
This attribute assigns values larger than the absolute
value of 1 standard deviation to zero while all other
values are given the value of the seismic data.
16. Selectthe previously created attribute ‘Retrieve
Horizon Data’ in the ‘For Horizon to use’.
17. Selectthe seismic you wish to blank in the ‘seis’
(e.g. 4_Dip steered median filter).

Blank traces workflow cont’d:
18. Display the new seismic attribute with
blanked traces. Right mouse clickon the
In-line. Select‘’Add and Select Data’’
19. Selectthe attribute ‘’Seismic with
random missing traces as zeros’’ and HitOk.
Notice that random traces have been
blanked.

Blank traces workflow cont’d :
20. Select, ‘’Create a Seismic Output’’ from the
attribute –Seismic with random missing traces as
zeros.
21. In the ‘’Create Single Attribute Volume’’ window,
keep the default parameters. Typean Output name
(e.g. Seismic with random missing traces as zeros)
and Run.
22. Closethe progress window when the processing
finish
23. Display/QC the created seismic
This seismic will be used as input for the next step, ML
Seismic Image Regression prediction.

Exercise objective:
To fill blank seismic traces using the ‘Seismic Image Regression’’ tool which is part of the machine
learning plugin. The model will have to learn how to recreate an image from example images
containing blank traces.
Workflow:
1. Openthe Machine Learning Control Center with the
icon.
2. Clickon Seismic.
3. Selectthe ‘Seismic Image Regression’and HitGo.

Workflow cont’d:
4. In the ‘Extract Data’ tab, Pressthe Select
button. The “Deep Learning Target Seismic
Definition” window pops up.
5. Pressthe+ iconand Select the target
seismic volume (e.g.4 Dip steered median filter).
And OK.
*Note:it is possible to create a Training Set from
examples extracted from multiple surveys. To do this,
press the + icon again and select the target volume to
add to the table below.
6. PressProceed[Input Data Selection]. The
“Input seismic for prediction” window pops up
* The option to select data from other surveys is available only in commercial projects

Workflow cont’d:
7. Selectthe input seismic data (i.e. the seismic
with the missing traces as zeros) and PressOK.
8. In the “Input Data” window set the dimensions
of the input features.To minimize processing
time for this exercise, Setthe Images dimension
to: 0x128x128, overlap: 0x0x0 and Inline range:
(100 –400).
Note: If the current HW has large amount of GPU and
CPU/computing power, the recommended Image
Dimensions are128x128x128.
9. Specifythe name of theOutput Deep
Learning Example Data(e.g.
ML_Ex_FillMissingTraces_0x128x128 ) and
PressProceedto start the extraction process.
10. When this process is finished Press
Proceedin the “Seismic Image Transformation”
window to continue to the Training tab.

Workflow cont’d:
11. After the training data is selected the
available models are shown. For seismic image
workflows we useKeras(TensorFlow).
12. ChecktheParameterstab to see which
models are supported and which parameters can
be changed.
13. Specify a name for the ‘’Output Deep
Learning Model’’ (e.g.
ML_Unet_FillMissingTraces_0x128x128).
14. Hit Run.
15. Openthe processing log file to follow the
progress. When the log file shows “Finished
batch processing”, the Proceed button turns
green. You can press Proceedor Open
theApplytab.

Workflow cont’d:
16. Once the Training is done, the trained model
can be applied. Select the trained model and
PressProceed.
17. TheApplywindow pops up. Here you can
optionally apply to aVolume subselection. Type
an Output name (e.g.
Seismic_ML_Unet_FillMissingTraces_0x128x128)
Note: You can run on GPU or CPU using thePredict using
GPUtoggle. Running the application on a GPU is many
times faster than running it on a CPU.
18.PressRunto create the desired output.
19. Closethe ‘Progress Viewer’ window when the
processing is finished.

Workflow cont’d:
Compare the original seismic data with the Unetpredicted
filled seismic results. The line is extracted from the blind test
area.
20. Right Mouse Click on In-line > Add and select Data
> Store. Selectthe created Filled Seismic
(e.g. ML_Unet_FillMissingTraces_0x128x128), and HitOK.
21. Type inthe Inline field: 425, and HitEnter.
The same way, add to the display, the original seismic and
seismic with missing traces .
22. Right-Clickon Inline 425 > Add > Attribute
>Stored. Selectthe original seismic
(e.g. 4 Dip steered median filter), and HitOK.
23. Right-Clickon Inline 425 > Add > Attribute
>Stored. Selectthe seismic with missing traces
(e.g. Seismic with random missing traces as zeros), and
HitOK.

24.Comparevisually in the blind test
area the:
- Original seismic (4 Dip steered
median filter)
- Randomly blanked traces seismic
(Seismic with random missing
traces as zeros)
- Unetfilled seismic
(ML_Unet_FillMissingTraces_0x1
28x128)
25. For more accurate comparison,
Setsimilar colour range for the 3
seismic cubes. Highlight the seismic
cube, Set the colour bar range to (-
8000, 8000).
Workflow cont’d:
Inline 425 -Original Seismic
Inline 425 -UnetFilled Seismic
Inline 425 -33% randomly blanked seismic

Workflow cont’d:
For a better quantitative comparison, create a new attribute ‘difference’ that computes the
difference between the predicted and the original seismic.
26. Selectthe 3D attribute icon . Openthe
attribute set . Selectthe attribute set
‘’ML_Random_holes_interpolation’’
27. Createa 4thattribute ‘’difference’’ as
indicated in the attribute set window and Hit
‘Add as new’.
28. Selectthe Original seismic (e.g. 4 Dip
steered median filter) for ‘Seis’, and the
predicted seismic (e.g.
ML_Unet_FillMissingTraces_0x128x128) for
‘pred’

29. Right-Clickon Inline 425 > Add >
Attribute. Selectthe attribute ‘difference’’,
and HitOK.
Notice the small values of the difference,
range (-593, 590).
30. For more accurate comparison,
Modifythe color range to similar range as
the original and predicted seismic [-
8000,8000]
Inline 425 -Difference (UnetFilled –Original) seismic
Workflow cont’d:
Display/QC the attribute ‘’difference’’. Difference = Original seismic (4 Dip steered median filter) –
Predicted seismic (ML_Unet_FillMissingTraces_0x128x128)

Workflow cont’d (Optional):
For more accurate comparison, apply an RMS gain scaled correction to the original and predicted
seismic, than compute the difference.
Create a new Gain correction attributes to be applied on the original and predicted seismic.
31. Selectthe 3D attribute icon . Openthe
attribute set . Selectthe attribute set
‘’ML_Random_holes_interpolation’’
32. Createa 5thattribute
‘’Gain_corrected_Filled_Seis’’ as indicated in the
attribute set windowand Hit ‘Add as new’
33. Selectthe Input Data
‘’Seismic_ML_Unet_FillMissingTraces_0x128x128’’

34. Selectthe 3D attribute icon . Open
the attribute set . Selectthe attribute set
‘’ML_Random_holes_interpolation’’
35. Createa 6thattribute ‘’Gain_corrected
seismic’’ as indicated in the attribute set
windowand Hit ‘Add as new’
36. Selectthe Input Data ‘Gain_corrected
seismic’’
Workflow cont’d (Optional):
For more accurate comparison, apply an RMS gain scaled correction to the original and predicted
seismic, than compute the difference.
Create a new Gain correction attributes to be applied on the original and predicted seismic.

Workflow cont’d (Optional):
Create a new attribute that will compute the difference between the RMS gain corrected original
seismic and ML predicted seismic
37. Selectthe 3D attribute icon .
Openthe attribute set . Selectthe
attribute set
‘’ML_Random_holes_interpolation’’
38. Createa 7thattribute ‘’difference
between gain corrected seis and ML
pred’’ as indicated in the attribute set
windowand Hit‘Add as new’.
39. Selectthe ‘’Gain_correctedseismic’’
as input for ‘seis’ and the
‘’Gain_corrected_Filled_Seis’’ as input
for ‘pred’

40. Right-Clickon Inline 425 > Add > Attribute.
Selectthe attribute ‘’Gain_corrected_Filled_Seis’’,
and HitOK.
41. Right-Clickon Inline 425 > Add > Attribute.
Selectthe attribute ‘Gain_correctedseismic’’, and
HitOK.
RMS scaled Seismic_ML_Unet_FillMissingTraces_0x128x128
RMS scaled 4 Dip steered median filter
Workflow cont’d (Optional):
Display the attribute ‘’Gain_corrected_Filled_Seis’’ (RMS scaled
Seismic_ML_Unet_FillMissingTraces_0x128x128) and the ‘’Gain_correctedseismic’’(RMS
scaled 4 Dip steered median filter).