[OpendTect_Users] Neural Network Fault Cube Questions

[Friso Brouwer]

Nicola,

For fault characterization I often prefer to work with well designed
attributes and filters rather then the NNFC. The reason for this is that
some of the key fault attributes lateral displaced to each other:
similarity shows anomalies on the middle of the fault plane, curvature on
the sides of the fault plane. Unless you know how to correct for that
(which is tricky) you are better of creating a separate similarity and
curvature volume. I will high level described the workflows I use below:

*Similarity*: shows faults (and other features such as erosional incisions)
defined by seismic discontinuity. Note that in comparison with the
coherency attribute the similarity attribute is sensitive to lateral change
is reflection shape *and* reflection amplitude, whereas the coherency
attribute is *only* sensitive to lateral change in reflection shape.
Moreover the use of the steering allows you to condition the dip
correction, where the coherency uses local dips only. As a result,
similarity often shows more subtle faulting, while coherency shows the
faults with large dip-slips only in an otherwise cleaner background. Note
that the one is not necessarily better then the other but it is useful to
understand the difference.

To create a similarity attribute I use background steering, that is I
create the steering in two steps:

   - BG steering algorithm with 1x1x1 operator and the median filter with
   0x0x4 (or 5)  option to remove noise bands related to thin bed interference.
   - Then filter the steering with an average operator say 5x5x1 to create
   a volume that represents non-local dips.
   - Note that I use two different operators - the median operator to
   denoise, the average operator to smooth.
   - HowTo:
   http://www.dgbes.com/images/stories/PDF/effectivedipsteeringworkflowusingbgsteering_primerodata.pdf,
mind that in this HowTo the median operator is used to create the
   background steering, I recommend to use the average operator instead.

Then the similarity can be improved by applying a:

   - fault enhancement filter, which filters the seismic data to make the
   faults sharper, best applied when the seismic imaging of the faults is low
   quality, see HowTo
   http://www.dgbes.com/images/stories/PDF/faultenhancementfilter.pdf
   - ridge enhancement filter, is a filter to enhance the similarity
   attribute itself. There is a default attribute set in OpendTect available
   that gives you an easy to access REF, but I do not like to much. You can
   also build this by hand (much better), if you planned to do this I gladly
   provide you with a more detailed workflow.

The *curvature* attribute is a good attribute to highlight faults that have
dip-slips very small compared with seismic resolution, these show as
seismic flexures, and are thus highlighted by the curvature. The curvature
attribute also works good if a lot of fault drag has create structural
deformation on the sides of the fault planes.

My usual workflow: I use for fault detection the *maximum* curvature, the
only input volume is steering and I usually use a slightly modified version
of the detailed steering cube.

   - BG steering algorithm with 1x1x1 operator and the median filter with
   0x0x4 (or 5)  option to remove noise bands related to thin bed interference.
   - Then filter the steering with an average operator say 1x1x1 or 2x2x1
   to create a volume that will give a more robust extraction of the curvature.

If you want to work with NNFC I advise to create one with input attributes
such as similarity with short window, similarity with long window,
frequency, short window rms amplitude only,  as these all produced the
anomalies at the fault plane, and then create a separate curvature volume.
In such a case the NNFC will highlight the stronger fault systems, and the
curvature will include the faults that are most subtle.

For more reading: http://dgbes.com/images/stories/PDF/browhuck.pdf and
http://dgbes.com/images/stories/PDF/browetal.pdf

Success,

Friso
On Mon, Apr 23, 2012 at 3:16 AM, Nicola Scarselli <n.scarselli at es.rhul.ac.uk
> wrote:

>
>  Hi Paul,
>
>
> *1) In order to run NNFC you need to provide a steering cube. Which is
> the one to use between background and detailed? Which are the parameters
> suggested for calculating such steering cube?*
>
> Typically we use background steering cubes as inputs for dip-steered
> filters and detailed steering cubes when we want to preserve details in
> attribute calculations. More information on steering cubes and parameters
> can be found in this webinar on the subject:
> http://www.opendtect.org/tutorials/Webinar_Optimal_Steering .
>
> Which one do you recommend in the end for NNFC?
>
>
>
> *2) I noticed that my NNFC fails in areas of very steep reflectors - even
> if these areas are not faulted at all, my NNFC reports very high fault
> probability there. Is there any attribute I can add/remove/change to the
> process in order to address this problem?*
>
> You could try including picks of "non-faults" in steeply dipping areas and
> see if the NN can learn to distinguish these picks from picks at real
> faults. If "polar dip" is not in your attribute set yet, I recommend
> including that.
>
> This is what I was doing - maybe I need to pick more "non-faults" in
> dipping areas?
>
> Thank you very much for your precious advices.
>
> Nicola
>
>
> _______________________________________________
> Users mailing list
> Users at opendtect.org
> http://www.opendtect.org/mailman/listinfo/users
>
> You receive this mail because you are listed on users at opendtect.org To
> unsubscribe please go to http://lists.opendtect.org/mailman/options/users. If you encounter any problems, please contact
> support at opendtect.org .
>



-- 
Friso Brouwer
COO-Americas
dGB Earth Sciences

+1 281 240 3939 (o)
+1 281 240 3944 (f)
friso.brouwer at dgbes.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.opendtect.org/pipermail/users/attachments/20120423/ede69d3d/attachment.html>