NeXus standard

[Mark Koennecke]

 High Tim,

On Mon, 13 Dec 1999, Tim Mooney wrote:

> >   I'am afraid I do not fully understand the problem. To me a 2D scan is a
> >   2D dataset somehow like this:
> > 
> >        pos1 ->
> >   pos2 ******************************************
> >    |   ******************************************
> >    V   ******************************************
> >        ******************************************
> >        ******************************************
> >        ******************************************
> >                 .
> >                 .
> >        ******************************************
> > 
> >    Why is this bad? ...
> > ...
> >    Any further opinions? 
> > 
> >    
> >    Mark Koennecke
> 
> One problem is that multidimensional data do not always (or even
> usually) have a simple structure.  Let's take the case of storing
> fluourescence spectra as a function of some variable (say, time, for
> simplicity).  The obvious way to store these data is in a 2D array,
> where each row is a multichannel analyzer spectrum.  But along with
> each spectrum are a number of scalar data, such as the MCA live time
> and real time, the beam current, the signal from an ionization chamber
> integrated over the MCA acquire time, the signal that indicates beam
> steering, tuning, or top-off injection occurred during acquisition of
> the spectrum.  These scalar data can't be plotted against the X axis of
> the 2D data set, which is the MCA channel number or something derived
> from it.  They have to be plotted against the Y axis of the 2D array,
> but this axis has declared itself to be a Y axis (i.e., axis = 2), so
> we must make another copy of the same data and have this copy call
> itself an X axis.  In this case it would be better if the data to be
> plotted could indicate what axis level should be its X axis, etc.
> 
> The real case I'm drawing from is of fluourescence microtomography,
> which produces a 4D data set (a series of 2D images at different
> angles, each pixel of each image has a fluourescence spectrum
> associated with it--several, actually, because several x-ray detectors
> are required to get this data in a reasonable time).  The axes are the
> following:
> 
> 4: image number (rotation angle)
> 3: Y pixel address (Y "motor")
> 2: X pixel address (X "motor")
> 1: fluourescence energy (MCA channel number)
> 
> With this data set, we're going to have some 1D arrays plotted against
> axis 4 (e.g., integrated beam during acquisition of each image), and
> others plotted against axes 3 and 2.  Each rotation angle also will have
> conventional x-ray image (e.g., transmission) associated with it, and
> these images want axis 3 to be their Y axis and axis 2 to be their X
> axis.  The spectra themselves, of course, want fluourescence energy to
> be their X axis.
> 

  Huh, this is indeed a difficult problem. However, IMHO there is no real
  conflict with the standard. The NXdata vGroup is provided in order to 
  select a subset of data which can be easily automatically plotted. It
  should give a reasonable first view of the data whatever that is for
  microtomography. I see no problem of storing all the SDS's you mentioned
  in the file. And with a specialized tools you can plot anything against
  anything, disregarding the axis attributes. 

  Another option would be to have different vGroups of type NXdata for
  different views of the data. In your case this would probably require
  a little duplication of data.   

> -------------
> 
> The other problem is 2D scans in which the X axis varies systematically
> from row to row.  The usual reason for doing this is to avoid wasting
> beam time acquiring data far from the peak, when you know approximately
> how the peak will shift as the Y motor varies.  The result is an image
> with blank regions where data were not acquired, and you really would
> like not to have to regard these data as a series of separate scans.
> If you do split the image into separate scans, and you have
> mixed-dimensional data, as described above in the spectra vs. time
> example, it's a hard job collecting scalar data to be plotted against
> the Y axis.

  If each scan is different I think there is a point for handling this as
  separate scans. As I already mentioned you can choose to preserve the
  2D character of the thing by introducing virtual dimensions NPX, NPY
  for the number of points measured in each dimension and have separate
  2D arrays with x- and y-values which tell you where your where when
  the point was measured. You'll probably need that info in order to be
  able to rebin the data some time later. Or you rebin the scans on a even 
  grid right at the writing stage. 

  Has anybody else comments on this?

                 Mark