On 7/25/2011 10:51 AM, Tom Rink wrote:
On 7/25/11 9:51 AM, John Caron wrote:
On 7/21/2011 1:31 PM, Tom Whittaker wrote:
The more I think about this, the more I tend
toward the idea that we probably shouldn't try to merge the
different kinds
of binning into a single all-encompassing coordinate. There is a
point at
which we have to go ahead and split the data into different variables.
I agree. We have the issue, for example, of bands that have different
units. Since "units" is an attribute of a variable, that would mean
that only bands with the same units could be included in the 'array'.
This also means that a given file might have more than one "band
coordinate variable", since there might be, say, 3 emissive channels
and 2 reflective ones with different units.
There is nothing preventing you from defining dimensions and
coordinate
variables for the different binnings of your image data and storing
everything into a single variable (with dimensions for wavelength,
bandwidth, polarization, applied filter, etc). It is just the
question of
whether or not some portion of it will be defined via a CF
convention, and
what that portion is.
Well stated -- getting at the heart of the matter. I look at this
from an application developer's standpoint: when I look at a "data
variable" and see that it has dimensions of, say, "abc x y" I would
like to know whether "abc" is "time" or "altitude" or...."band"
because I'll treat the data differently. And, I may then seek other
variables with specific "standard_names" to define things like
wavelength, etc.
tom
I agree that if different bands have different units, these must be
in a different variable. That reasoning applies to all of the
variable's attributes.
Another way to say the same thing: A coordinate variable must have
values that are a "sampling" along a continuous domain. One sees this
violated in the poorly designed AWIPS netcdf files, eg where all
"temperature" values are crammed into one variable, and the
coordinate values have both pressure and height and potential
temperature, etc.
There are instances where the spectral domain sampling will not be
monotonic increasing.
For example, AIRS L1B, where 3 broad band regions overlap, but are
written into a single
array, and MHS, ATMS, where two adjacent bands have the same central
frequency, but
differ in their respective frequency sampling range. Different
applications may have
different ideas how they want to interpret this. For example, in the
former case we
sort the individual wavenumbers and keep track of the sort indexes.
in CF, auxiliary coordinates do not have to be monotonic increasing.
in CDM, they do have to be unique, but the uniqueness can be in the
bounds rather than the coordinate values.