20010322: netcdf question: NCEP reanalysis: precision loss

Brent,

>Date: Fri, 23 Mar 2001 13:45:59 -0500 (EST)
>From: Brent A McDaniel <address@hidden>
>Organization: Georgia Institute of Technology
>To: Steve Emmerson <address@hidden>
>Subject: Re: 20010322: netcdf question: NCEP reanalysis: precision loss 
>Keywords: 200103222003.f2MK3hL23783

The above message contained the following:

> Perhaps I'm confused about something.  If I go to the CDC's ftpsite and
> download a file from the daily avg pressure level data, the data is shown
> as being type short, yet it's showing precision to 5 decimal places, not
> integer values.  Here's a snip from ncdump -h for one of these files:
> 
> short omega(time, level, lat, lon) ;
>               omega:long_name = "mean Daily Omega" ;
>               omega:valid_range = -3.f, 7.f ;
>               omega:actual_range = -1.120003f, 0.8909988f ;
>               omega:units = "Pascal/s" ;
>               omega:add_offset = 29.765f ;
>               omega:scale_factor = 0.001f ;

The above "add_offset" and "scale_factor" attributes indicate that the
omega values are "packed" into 16-bit values.  The actual omega values
are given by the expression "scale_factor*x + add_offset", where "x" is
the packed, 16-bit value (in this case).

It is the responsibility of the application to notice the attributes and
perform the unpacking or packing.

>               omega:missing_value = 32766s ;
>               omega:precision = 3s ;
>               omega:least_significant_digit = 3s ;
>               omega:GRIB_id = 39s ;
>               omega:GRIB_name = "VVEL" ;
>               omega:var_desc = "Omega (dp/dt)\n",
>     "Z" ;
>               omega:dataset = "NCEP Reanalysis Daily Averages\n",
>     "AJ" ;
>               omega:level_desc = "Multiple levels\n",
>     "F" ;
>               omega:statistic = "Mean\n",
>     "M" ;
>               omega:parent_stat = "Individual Obs\n",
>     "I" ;
> 
> So, is the difference that the ncep data is "packed" whereas my data is
> not?  If I want to keep the same type accuracy (more than integer say) and
> use type float (to save space) do I need to pack the data with add_offset
> and scale_factor?

If the type in the dataset is to be floating-point, then there's no need
to pack the data and the data should not have either a "scale_factor" or
"add_offset" attribute.

Regards,
Steve Emmerson   <http://www.unidata.ucar.edu>