The idea is that you store the most signifant bits of the original REAL
value in a INTEGER*2. The way you do this is to make those bits the integer
part of a real number, then trucate by storing into an integer. EG:
1. If you want to store the top 4 digits of .12345 :
10000 * .12345 = 1234.5
truncate and store in the netcdf file: 1234
then the reading program has to know to divide by the scale:
1234 / 100000 = .1234
which is a close approximation to the original number .12345
2. What should the scale and offset be? Normally you store a group of
numbers, all with the same scale and offset. So first you calculate the min
and max of that group of numbers. Also let max_integer = maximum integer (eg
for INTEGER*2 this equals 32,167).
then
offset = min
scale = max_integer / (max - min)
store this number into netcdf file:
store_x = (x - offset) * scale = max_integer * (x - min) / (max - min)
note that when x = min, store_x = 0, and when x = max, store_x
max_integer.
the reading program should use the formula
x = store_x/scale + offset.
In your example below, I suspect that the scale may be incorrect, but the
actual formula is ok.
----- Original Message -----
Sent: Monday, May 13, 2002 3:13 AM
> Hi
>
> I'm a relatively new netCDF user. Having read the netCDF guide plus
> some entries in the support and user forum, I have a burning question
> regarding data packing using 'scale_factor' and 'add_offset'. I've seen
> fortran code which implement the packing with code like the following
> sample snippet:
>
> INTEGER*2 X_PACKED
> REAL X, OFFSET, SCALE
>
> X = 10.23456
> OFFSET = 9.0
> SCALE = 0.23123
>
> X_PACKED = (X-OFFSET)*SCALE
> ! Then proceed to store X_PACKED in netCDF form.
> . . .
>
> What I don't understand is how this can be "correct". When you assign a
> REAL to an INTEGER, the value of the REAL gets truncated before being
> stored in the INTEGER. Thus, the original value of X cannot be
> 'reconstructed' from the corresponding INTEGER*2 variable (X_PACKED) in
> the netCDF file.
>
> I understand the intent is to store only the 2 bytes that represents the
> scaled-down REAL value. But as far as I know, this is not possible in
> Fortran -- as far as simply assigning values to variables is concerned.
>
> Or is the above fortran implementation of packing in netCDF completely
> off track? Am I missing anything?
> What is the common practice of netCDF packing in Fortran?
> Any sample Fortran code for netCDF data packing?
>
> Thanks for any explanations or tips.
>
> -Kek
> Tropical Marine Sc Institute
> Singapore
>
