[gdal-dev] NetCDF driver improvements (including groups support)

Julien Demaria Julien.Demaria at acri-st.fr
Thu Feb 5 08:10:47 PST 2015

Hi GDAL team,

I've implemented several improvements to the NetCDF driver and I would like to provide them to the community.
Main goal of the changes is to add full support of NetCDF-4 including groups.
NetCDF-4 is the future format of ESA Sentinel-3 products (no groups) and NASA Ocean Color team is switching their L2/L3 products to NetCDF-4 with groups (VIIRS has already switched to the new format in December).
With the changes NASA L2 products geolocation is automatically handled as geolocation arrays and can be reprojected using gdalwarp.

I validated with autotests that nothing is broken in tests netcdf.py (excepting test 13 but see my point 5), netcdf_cf.py and hdf5.py, using NetCDF-3 and 4 libraries.
I've also tested the new functionalities on various NetCDF-4 files.
I think the only possible regression could be for marginal cases where a file was seen directly as a dataset and is now seen as multiple subdatasets (for example if a file has only one var in the top group and has nested groups containing variables), but I think this is not very common.

For the moment I have all these changes in local GIT separated commits on the latest gdal-1.11 branch, let me know what changes you want and how can I provide them.

Changes :

1) Implement full support for NetCDF-4 groups on reading:
    - explore recursively all nested groups to create the subdatasets list
    - subdatasets in nested groups use the /group1/group2/.../groupn/var standard
      NetCDF-4 convention, excepting for variables in the root group which do not
      have a leading slash for backward compatibility
    - when accessing a subdataset using NETCDF:$file:$path, the leading slash is optional
    - global attributes of each nested group are also collected in the GDAL dataset
      metadata, using the same convention /group1/group2/.../groupn/NC_GLOBAL#attr_name,
      excepting for the root group which do not have a leading slash for backward compatibility
    - when searching for a variable containing auxiliary information on the selected subdataset,
      like coordinate variables or grid_mapping, we now also search in parent groups (using NCDFResolveVar).
      I now this is something not specified at this time in the CF convention because CF does not know groups,
      but it seems logical to me to support this: NetCDF-4 specifies that dimensions of a group are
      shared to its nested groups, so associated coordinate variables could be defined as the same level of its
      corresponding dimension.
    - reference to coordinate variables using the "coordinates" attribute support now also absolute paths,
      this allow for example to specify coordinate variables located outside the group of the selected variable
      or its parents. Relative paths could be implemented if needed.
      This feature is used to add support for new NASA Ocean Color L2 products.

2) Implement full read/write support for new NetCDF4 types NC_UBYTE, NC_USHORT, NC_UINT and NC_STRING, only if NETCDF_HAS_NC4 is defined (and only if format=NC4 for writing).
    Support implemented for variables and attributes.
    NC_STRING type is supported for reading (scalar and arrays) attributes and is used for writing only for array attributes (scalar are still written as NC_CHAR).
    If NETCDF_HAS_NC4 is not defined or format!=NC4, NC_STRING array attributes are written as a single NC_CHAR string using the GDAL {v1,v2,...} convention.
    Add missing support for NC_BYTE in CreateBandMetadata() and NC_BYTE/SHORT in NCDFPut1DVar().

3) Add support for new NASA Ocean Color L2 products and ESA Sentinel-3 L1 or
    L2 products which use the NetCDF-4 format (with groups for NASA, see
    - NASA products: simulate a "coordinates" variable attribute to detect CF
      geolocation arrays, and set bBottomUp to FALSE
    - ESA products: set bBottomUp to FALSE and disable warning on missing
      Conventions attribute

4) Fix bug #4554 with a more generic solution by disabling the installation of the HDF5 atexit() cleanup routine using H5dont_atexit().
    Previous fix was to call GDALExit() (for the moment only defined gdalwarp.cpp) at the end of every program, which is more painful.

5) Fix implementation of GetScale/Offset to not always return pbSuccess=TRUE.
    Fix CopyMetadata to handle bands with only scale or offset.
    ==> WARNING this commit breaks the autotest netcdf_13 (check for scale/offset = 1.0/0.0 if no scale or offset is available), but for me it is not logical to return always pbSuccess=TRUE

6) Optimize IReadBlock() and CheckData() handling of partial blocks in the x axis by re-using the GDAL block buffer instead of allocating a new temporary buffer for each block.

7) Force block size to 1 scanline for bottom-up datasets if nBlockYSize != 1 instead of raising a fatal error
    ==> Solve a recent problem raised on the mailing list

8) Implement Get/SetUnitType using the standard "units" NetCDF attribute

9) Change default block size to 256x256 instead of scanline (only affect file without NetCDF chunking)
    ==> because I think this is better for a random access to the data, but I'm not sure if the community want this change which could impact performances

10) I've also implemented for my needs support for 1D variables by simulating 2D datasets with only one row (dimensionless variables are not supported for the moment),
       but this breaks backward compatibility because files containing only one variable and associated 1D coordinate variables are now seen as multiple sub datasets...
       and maybe this is not the goal of GDAL to give access to not-2D-raster variables (but sometimes it's useful ;-) )

Thanks for GDAL!


More information about the gdal-dev mailing list