Doing GRIB in GrADS

Introduction
What is GRIB?
The Problem
Up-front Limitations
The Solution - Part 1
The Solution - Part 2
Conclusions

One of the most powerful features of GrADS is its ability to work DIRECTLY with GRIB data. Unfortunately, "doing GRIB in GrADS" is not simple; it requires an understanding of GRIB and how GrADS works this type of data.

This note attempts to provide the required understanding to use GRIB data in GrADS.

WHAT IS GRIB?

GRIB (GRIdded Binary) is an international, public, binary format for the efficient storage of meteorological/oceanographic variables. Typically, GRIB data consists of a sequence of 2-D (lon,lat) chunks of a (in most general sense) 4-D variable (e.g., u comp on the wind = f(lon,lat,level,time)). The sequence is commonly organized in files containing all variables at a particular time (i.e., 3-D (lon,lat,level) volume).

The problem for the user is how to "interface" or "display" GRIB file(s) to GrADS. The solution has two components. The first is to see what is in the GRIB file, i.e., variables, grids, times, etc. and the second is to "map" the 2-D GRIB fields to higher dimensional structures available in GrADS.

There are some limitations on the kinds of GRIB data that can be interfaced/displayed in GrADS:
a) lon,lat grids (NOT lat,lon)
b) simple packing
c) grid point data
d) grids must be contiguous (no blocked or octet grids)

Thus, "thinned" grids (non rectangular) and spectral coefficients are not supported. However, GRIB versions 1 AND 0 are supported (GRIB 0 data must be filtered to GRIB 1 for wgrib and version 1.7 of GrADS will support such non-rectilinear grids. Further, it IS possible to display "preprojected" GRIB data (e.g., polar stereo fields, see eta.ctl).

The first, and relatively straightforward, part of the solution is to find out what is in the GRIB file. Two utilities are included in the GrADS release:
1. gribscan
2. wgrib

The big virtue of wgrib is that the code is written in ANSI C and runs on all the major platforms from PC's to Cray. Although wgrib was designed to support the NCEP reanalysis project, it has been extended to handle other sources of GRIB data (e.g., ECMWF) and is my tool of choice. If I can't read the data in wgrib then I change the code and feed the changes to Wesley Ebisuzaki, NCEP.

Once you have wgrib running,

wgrib ncep.reanl.mo.7901.grb

yields,

1:0:d=79010100:UGRD:kpds5=33:kpds6=100:kpds7=850: TR=113:P1=0:P2=6:TimeU=1:850 mb:anl:ave@6hr:NAve=124
2:15852:d=79010100:UGRD:kpds5=33:kpds6=100:kpds7=500: TR=113:P1=0:P2=6:TimeU=1:500 mb:anl:ave@6hr:NAve=124
3:33018:d=79010100:UGRD:kpds5=33:kpds6=100:kpds7=200: TR=113:P1=0:P2=6:TimeU=1:200 mb:anl:ave@6hr:NAve=124
4:51498:d=79010100:VGRD:kpds5=34:kpds6=100:kpds7=850: TR=113:P1=0:P2=6:TimeU=1:850 mb:anl:ave@6hr:NAve=124
5:66036:d=79010100:VGRD:kpds5=34:kpds6=100:kpds7=500: TR=113:P1=0:P2=6:TimeU=1:500 mb:anl:ave@6hr:NAve=124
6:81888:d=79010100:VGRD:kpds5=34:kpds6=100:kpds7=200: TR=113:P1=0:P2=6:TimeU=1:200 mb:anl:ave@6hr:NAve=124
7:99054:d=79010100:PRES:kpds5=1:kpds6=102:kpds7=0: TR=113:P1=0:P2=6:TimeU=1:MSL:anl:ave@6hr:NAve=124

So we see 7 fields in the file valid at 00Z1jan1979 (d=79010100).

for,

wgrib ncep.reanl.mo.7902.grb

we find,

1:0:d=79020100:UGRD:kpds5=33:kpds6=100:kpds7=850: TR=113:P1=0:P2=6:TimeU=1:850 mb:anl:ave@6hr:NAve=112
2:15852:d=79020100:UGRD:kpds5=33:kpds6=100:kpds7=500: TR=113:P1=0:P2=6:TimeU=1:500 mb:anl:ave@6hr:NAve=112
3:33018:d=79020100:UGRD:kpds5=33:kpds6=100:kpds7=200: TR=113:P1=0:P2=6:TimeU=1:200 mb:anl:ave@6hr:NAve=112
4:50184:d=79020100:VGRD:kpds5=34:kpds6=100:kpds7=850: TR=113:P1=0:P2=6:TimeU=1:850 mb:anl:ave@6hr:NAve=112
5:64722:d=79020100:VGRD:kpds5=34:kpds6=100:kpds7=500: TR=113:P1=0:P2=6:TimeU=1:500 mb:anl:ave@6hr:NAve=112
6:80574:d=79020100:VGRD:kpds5=34:kpds6=100:kpds7=200: TR=113:P1=0:P2=6:TimeU=1:200 mb:anl:ave@6hr:NAve=112
7:96426:d=79020100:PRES:kpds5=1:kpds6=102:kpds7=0: TR=113:P1=0:P2=6:TimeU=1:MSL:anl:ave@6hr:NAve=112

or the same fields as before, except they are valid at 00z1feb1979.

To find out about the data grid use,

wgrib -V ncep.reanl.mo.7901.grb

and for the first record you will find:

rec 1:pos 0:date 79020100 UGRD kpds5=33 kpds6=100 kpds7=850
levels=(3,82) grid=2 850 mb anl:ave@6hr:
  timerange 113 P1 0 P2 6  nx 144 ny 73 GDS grid 0 num_in_ave 112 missing 0
  center 7 subcenter 0 process 80
  latlon: lat  90.000000 to -90.000000 by 2.500000
          long 0.000000 to -2.500000 by 2.500000, (144 x 73) scan 0 bdsgrid 1
  min/max data -12.29 16.86  num bits 12  BDS_Ref -1229  DecScale 2 BinScale 0

The grid is the NCEP #2 grid (see the Gospel According to John) which is a global 2.5 deg grid. The grid has 144 points in x (lon) and 73 points in y (lat). The (1,1) point is located at 90N and 0E.

This is the hard part -- creating a relationship between the sequence of 2-D (lon,lat) GRIB fields in a file(s) and the GrADS, 4-D, external-to-the-data, spatio-temporal data volume (lon,lat,level,time). In GrADS, the GRIB-to-4-D volume relationship is defined by the data descriptor or .ctl file. The actual relationship is created using the GrADS utility gribmap which generates an "index" or "map" between GrADS variables in the .ctl file and the GRIB data.

Before describing the details of gribmap, first consider the unix program, grb2ctl, written by Wesley Ebisuzaki. This program uses wgrib to first create a listing of fields in the GRIB data file, parses the listing for times, variables, and levels and then outputs a suitable descriptor file.

In our example,

grb2ctl.pl ncep.reanl.mo.7901.grb

yields to standard output,

dset ^ncep.reanl.mo.7901.grb
dtype grib
options yrev
index ^ncep.reanl.mo.7901.grb.gmp
undef -9.99E+33
title ncep.reanl.mo.7901.grb
xdef 144 linear 0 2.5
ydef 73 linear -90 2.5
zdef 3 levels
850 500 200
tdef 1 linear 00Z01jan79 1mo
vars 3
PRES 0 1 ,102,0 Pressure [Pa]
UGRD 3 33 ,100 u wind [m/s]
VGRD 3 34 ,100 v wind [m/s]
endvars

How did grb2ctl do this? First, only ONE grid was found in the GRIB data file (defined by dset) and the script had the grid geometry built in. Second, variables with multiply levels (3-D or lon,lat,level) where DEFINED to have a "level indicator" of 100 (more below). Third, there was only ONE time in the file and the script SET the time increment to 1mo.

These conditions will often not be present. Thus, the output from grb2ctl will likely have to be "tweaked" by the good 'ol trial and error method. In some cases the .ctl file may have to built "by hand" using the output from wgrib, so you'll probably have to know a lot about GRIB and how gribmap works in many situations.

The key ingredients in the .ctl file are:

1. grid geometry (xdef,ydef)
2. starting time and time increment (tdef)
3. variables and "units" parameter
4. variable type - "level" or 3-D (zdef) or "surface" (2-D)

The units parameter specifies the GRIB parameters of the variable in the .ctl to be used by gribmap for match GrADS variables to the fields in the GRIB files. This parameter consists of up to four, comma-delimited numbers:

VV,LTYPE,(LEVEL),(TRI)

where,

VV - (Required) The GRIB parameter number (33 = u comp of the wind, table 2 in John:section 1 page 27, i.e., GRIB Edition 1 (FM 92))

LTYPE - (Required) The level type indicator (100 = pressure level, in Table 3 in John:section 1 Page 33)

LEVEL - (optional) The value of the LTYPE (LTYPE 102 is mean sea level so LEVEL is 0 for where level is located (1,102,0, 0 is AT mean sea level)

TRI - (optional) The "time range indicator" for special applications (Table 5 in John:section 1 Page 37).

Coming up with the units parameter, the grid geometry and the times is the trick.

Fortunately, gribmap can tell us how well the .ctl mapped the GRIB data to the higher dimensional, GrADS data view. And, more importantly, the gribmap process does NOT depend on how the data are actually ordered in the GRIB file, in either level or variable.

Let's redirect output from grb2ctl.sh to a .ctl file, e.g.,

grb2ctl.sh on ncep.reanl.mo.7901.grb > ncep.reanl.mo.ctl

and then run gribmap. To reiterate, the gribmap utility compares each field in the GRIB file to each variable, at each level and for all times in the .ctl file and creates an index file telling GrADS WHERE the fields are (or are not) located in the GRIB data.

With the verbose option on,

gribmap -v -i ncep.reanl.mo.ctl

we get,

Scanning binary GRIB file(s):
 ncep.reanl.mo.7901.grb
!!!!!MATCH: 1  15852  2  1  0  33 100 850 btim: 1979010100:00 tau: 0 dtim: 1979010100:00
!!!!!MATCH: 2  33018  2  1  0  33 100 500 btim: 1979010100:00 tau: 0 dtim: 1979010100:00
!!!!!MATCH: 3  51498  2  1  0  33 100 200 btim: 1979010100:00 tau: 0 dtim: 1979010100:00
!!!!!MATCH: 4  66036  2  1  0  34 100 850 btim: 1979010100:00 tau: 0 dtim: 1979010100:00
!!!!!MATCH: 5  81888  2  1  0  34 100 500 btim: 1979010100:00 tau: 0 dtim: 1979010100:00
!!!!!MATCH: 6  99054  2  1  0  34 100 200 btim: 1979010100:00 tau: 0 dtim: 1979010100:00
!!!!!MATCH: 7 116220  2  1  0   1 102 0   btim: 1979010100:00 tau: 0 dtim: 1979010100:00
Reached EOF

We have succeeded!!! Each 2-D in the GRIB file has been mapped to a variable in ncep.reanl.mo.ctl. In the case of UGRD, a 3-D (lon,lat,level) variable which we can be sliced in the vertical with GrADS. However, failure to match will NOT stop GrADS from "working." If the data was NOT there, GrADS will return a grid with "undefined" values on display and this state can actually be tested...

The "tweaking" is done by adjusting the .ctl file until we get a !!!!! MATCH for each GRIB field in the data file(s). I have added a number of options that finely control the mapping process in gribmap for NCEP. See the GrADS document for details (ftp://sprite.llnl.gov/grads/doc/gadoc151.*).

Finally, let's adjust the ncep.reanl.mo.ctl file to take advantage of the file naming convention:

dset ^ncep.reanl.mo.%y2%m2.grb
dtype grib
options yrev template
index ^ncep.reanl.mo.gmp
undef -9.99E+33
title ncep.reanl.mo.7901.grb
xdef 144 linear 0 2.5
ydef 73 linear -90 2.5
zdef 3 levels 850 500 200
tdef 2 linear 00Z01jan79 1mo
vars 3

PRES 0 1 ,102,0 Pressure [Pa]
UGRD 3 33 ,100 u wind [m/s]
VGRD 3 34 ,100 v wind [m/s]

endvars

I have changed the number of times to two and have used the template option. This tells GrADS to locate data in TIME by the file name (dset ^ncep.reanl.mo.%y2%m2.grb). Thus, I have two (or more) data files, but only ONE .ctl file and I can now work with the 2-D GRIB data as if they were 4-D (lon,lat,lev,time) in GrADS. I also changed the name of the index file to be reflective of the now 4-D data structure.

To summarize the process:


1. use wgrib (or gribscan) to see if the data can be worked in GrADS;
2. use grb2ctl.sh or the output from wgrib to construct a .ctl file;
3. run gribmap in verbose mode (-v) to relate the GRIB data to the 4-D structure in the .ctl file, and to see how well the map worked; and
4. repeat steps 2) and 3) until you get !!!! MATCH from gribmap.

There's no doubt about it, "Doing GRIB in GrADS" is not very straightforward, but the benefits are, in my opinion, immense for two reasons. First, we have avoided conversion of the GRIB to a form which supports higher dimensional data (e.g., netCDF). We've saved disk space and have minimized potential technical errors (every time you touch the data you have an opportunity to screw it up). Second, from a GrADS performance standpoint, GRIB is nearly as fast as other binary formats -- the cost in decompression on the fly is compensated by reduced I/O.

In the end, GRIB-to-GrADS interface gives us the advantages of GRIB (efficient storage, self description and an open, international format) while overcoming the disadvantages of GRIB (2-D data and no means to organize to a higher dimension) via the GrADS 4-D data model. We get the best of both worlds, but only if we can make the .ctl file. Hopefully this document will help you do this.