[Users] unsupervised NN - why should I only use phase-independent attributes?

[Amanda Pouwel]

Colin Hawke wrote:
> I have a set of 3D horizons picked as part of a sequence stratigraphic
> study and these nicely define stratigraphic "intervals" for volume
> based seismic facies analysis. I notice that it is recommended to use
> phase-independent attributes as input to the unsupervised NN. What is
> the reasoning behind this? It seems intuitive to me that attributes
> that contain polarity information should help with the clustering of
> soft vs hard lithology and avo attributes might help for soft shale vs
> soft sand vs coal for example. Is it simply that we want attributes
> that define "layers" rather than "interfaces"? I know there is a good
> reason but am unsure what it is.
>
> Thanks - Colin.
Hi Colin,

Let me try to explain this by evaluating what happens when we perform a
conventional neural network waveform segmentation. In this horizon-based
method a neural network clusters seismic waveforms along a mapped
horizon (waveforms in this context are seismic trace segments). In
OpendTect we use the UVQ network to do this and we generate two possible
outputs: the *segmentation* grid (which reveals patterns of similar
seismic response along the horizon), and the *match* grid (which shows
the confidence in the clustering). A third possible output we can
generate is a display of the *cluster centers*. (This is a bit of a
hidden option in OpendTect: open the neural network window and press
/Info/ button followed by /Display./ Note, that in some early versions
of v2.4 this option does work. In that case please upgrade to the latest
stable release.)

This horizon-based method works well for parallel geology but heavily
depends on the quality of the mapped horizon.  Under these conditions
subtle variations in the waveform can be recognized and the resulting
patterns might reveal meaningful geologic information. However, the
method is extremely sensitive to the quality of the mapped horizon.
Where the horizon is bad we will cut the wrong part of the seismic trace
and feed this to the neural network. Such a time (=phase) shifted
waveform will appear more different to the neural network than the
subtle lateral variations caused by changes in geology.

Now consider a 3D UVQ clustering in which the input also is the waveform
(trace segment) extracted at each evaluation point. Being 3D this means
that the evaluation point shifts along each seismic trace with the
sampling rate. Each time we shift downwards the extracted waveform looks
different. After approx. half a loop the waveform is more or less
opposite polarity to the initial situation and half a loop deeper yet we
are  back to the original polarity. Clustering such waveforms will
result in patterns that are dictated by the vertical wiggle shapes and
not by the lateral variations, which is what we want to detect. So, with
phase dependent attributes we can expect a clustering that follows the
seismic loops.

When we at dGB perform a *3D UVQ* clustering we use *phase-independent
attributes* such as Spectral decomposition components, Energy and
Frequency derivatives. For smooth results we often calculate these
attributes as the *Median* value extracted in a *dip-steered* disk
(Volume statistics attribute with dip-steering plugin).

I hope this helps. Best regards,

Amanda.

-- 
-- Amanda Pouwel    Geoscientist
-- Nijverheidstraat 11-2, 7511 JM Enschede, The Netherlands
-- Mail to: amanda.pouwel at dgb-group.com, www.dgb-group.com
-- Phone; +31(0)53-4315155, Fax; +31(0)53-4315104