Provides an object-oriented python interface to the netCDF version 4 library

numpy, netcdf, data, science, network, oceanography, meteorology, climate, netcdf4, python
pip install netCDF4==0.9.8



Python/numpy interface to the netCDF C library.

Build status PyPI package Anaconda-Server Badge DOI


For details on the latest updates, see the Changelog.

6/4/2023: Version 1.6.4 released. Now requires certifi to locate SSL certificates - this allows OpenDAP https URLs to work with linux wheels (issue #1246).

3/3/2023: Version 1.6.3 released.

11/15/2022: Version 1.6.2 released. Fix for compilation with netcdf-c < 4.9.0 (issue #1209).
Slicing multi-dimensional variables with an all False boolean index array now returns an empty numpy array (instead of raising an exception - issue #1197).

09/18/2022: Version 1.6.1 released. GIL now released for all C lib calls, set_alignment and get_alignment module functions added to modify/retrieve HDF5 data alignment properties. Added Dataset methods to query availability of optional compression filters.

06/24/2022: Version 1.6.0 released. Support for quantization (bit-grooming and bit-rounding) functionality in netcdf-c 4.9.0 which can dramatically improve compression. Dataset.createVariable now accepts dimension instances (instead of just dimension names). 'compression' kwarg added to Dataset.createVariable to support szip as well as new compression algorithms available in netcdf-c 4.9.0 through compression plugins (such as zstd, bzip2 and blosc). Working arm64 wheels for Apple M1 Silicon now available on pypi.

10/31/2021: Version 1.5.8 released. Fix Enum bug, add binary wheels for aarch64 and python 3.10.

6/22/2021: Version 1.5.7 released. Fixed OverflowError on Windows when reading data with dimension sizes greater than 2**32-1. Masked arrays no longer returned for vlens.

2/15/2021: Version 1.5.6 released. Added Dataset.fromcdl and Dataset.tocdl, which require ncdump and ncgen utilities to be in $PATH. Removed python 2.7 support.

12/20/2020: Version released. Updated binary wheels for OSX and linux that link latest netcdf-c and hdf5 libs.

12/01/2020: Version 1.5.5 released. Update license wording to be consistent with MIT license.

07/23/2020: Version 1.5.4 released. Now requires numpy >= 1.9.

10/27/2019: Version 1.5.3 released. Fix for issue #972, plus binary wheels for python 3.8.

09/03/2019: Version 1.5.2 released. Bugfixes, no new features.

05/06/2019: Version released. Fixes another slicing regression (issue #922)) introduced in the 1.5.1 release.

05/02/2019: Version released. Fixes incorrect __version__ module variable in 1.5.1 release, plus a slicing bug (issue #919)).

04/30/2019: Version 1.5.1 released. Bugfixes, no new features.

04/02/2019: Version released. Binary wheels for macos x and linux rebuilt with netcdf-c 4.6.3 (instead of Added read-shared capability for faster reads of NETCDF3 files (mode='rs').

03/24/2019: Version 1.5.0 released. Parallel IO support for classic file formats added using the pnetcdf library (contribution from Lars Pastewka, pull request #897).

03/08/2019: Version released. Include missing membuf.pyx file in source tarball. No need to update if you installed from a binary wheel.

03/07/2019: Version released. Fixes bug in implementation of NETCDF4_CLASSIC parallel IO support in 1.4.3.

03/05/2019: Version 1.4.3 released. Issues with netcdf-c 4.6.2 fixed (including broken parallel IO). set_ncstring_attrs() method added, memoryview buffer now returned when an in-memory Dataset is closed.

10/26/2018: Version 1.4.2 released. Minor bugfixes, added Variable.get_dims() method and master_file kwarg for MFDataset.__init__.

08/10/2018: Version 1.4.1 released. The old slicing behavior (numpy array returned unless missing values are present, otherwise masked array returned) is re-enabled via set_always_mask(False).

05/11/2018: Version 1.4.0 released. The netcdftime package is no longer included, it is now a separate package dependency. In addition to several bug fixes, there are a few important changes to the default behaviour to note:

  • Slicing a netCDF variable will now always return masked array by default, even if there are no masked values. The result depended on the slice before, which was too surprising. If auto-masking is turned off (with set_auto_mask(False)) a numpy array will always be returned.
  • _FillValue is no longer treated as a valid_min/valid_max. This was too surprising, despite the fact the thet netcdf docs attribute best practices suggests that clients should to this if valid_min, valid_max and valid_range are not set.
  • Changed behavior of string attributes so that nc.stringatt = ['foo','bar'] produces an vlen string array attribute in NETCDF4, instead of concatenating into a single string (foobar). In NETCDF3/NETCDF4_CLASSIC, an IOError is now raised, instead of writing foobar.
  • Retrieved compound-type variable data now returned with character array elements converted to numpy strings (issue #773). Works for assignment also. Can be disabled using set_auto_chartostring(False). Numpy structured array dtypes with 'SN' string subtypes can now be used to define netcdf compound types in createCompoundType (they get converted to ('S1',N) character array types automatically).
  • valid_min, valid_max, _FillValue and missing_value are now treated as unsigned integers if _Unsigned variable attribute is set (to mimic behaviour of netcdf-java). Conversion to unsigned type now occurs before masking and scale/offset operation (issue #794)

11/01/2017: Version 1.3.1 released. Parallel IO support with MPI! Requires that netcdf-c and hdf5 be built with MPI support, and mpi4py. To open a file for parallel access in a program running in an MPI environment using mpi4py, just use parallel=True when creating the Dataset instance. See examples/ for a demonstration. For more info, see the tutorial section.

9/25/2017: Version 1.3.0 released. Bug fixes for netcdftime and optimizations for reading strided slices. encoding kwarg added to Dataset.__init__ and Dataset.filepath to deal with oddball encodings in filename paths (sys.getfilesystemencoding() is used by default to determine encoding). Make sure numpy datatypes used to define CompoundTypes have isalignedstruct flag set to avoid segfaults - which required bumping the minimum required numpy from 1.7.0 to 1.9.0. In cases where missing_value/valid_min/valid_max/_FillValue cannot be safely cast to the variable's dtype, they are no longer be used to automatically mask the data and a warning message is issued.

6/10/2017: Version 1.2.9 released. Fixes for auto-scaling and masking when _Unsigned and/or valid_min, valid_max attributes present. updated so that pip install works if cython not installed. Now requires setuptools version 18.0 or greater.

6/1/2017: Version 1.2.8 released. From Changelog:

  • recognize _Unsigned attribute used by netcdf-java to designate unsigned integer data stored with a signed integer type in netcdf-3 issue #656.
  • add Dataset init memory parameter to allow loading a file from memory pull request #652, issue #406 and issue #295.
  • fix for negative times in num2date issue #659.
  • fix for failing tests in numpy 1.13 due to changes in issue #662.
  • Checking for _Encoding attribute for NC_STRING variables, otherwise use 'utf-8'. 'utf-8' is used everywhere else, 'default_encoding' global module variable is no longer used. getncattr method now takes optional kwarg 'encoding' (default 'utf-8') so encoding of attributes can be specified if desired. If _Encoding is specified for an NC_CHAR ('S1') variable, the chartostring utility function is used to convert the array of characters to an array of strings with one less dimension (the last dimension is interpreted as the length of each string) when reading the data. When writing the data, stringtochar is used to convert a numpy array of fixed length strings to an array of characters with one more dimension. chartostring and stringtochar now also have an 'encoding' kwarg. Automatic conversion to/from character to string arrays can be turned off via a new set_auto_chartostring Dataset and Variable method (default is True). Addresses issue #654
  • Cython >= 0.19 now required, _netCDF4.c and _netcdftime.c removed from repository.

1/8/2017: Version 1.2.7 released. Python 3.6 compatibility, and fix for vector missing_values.

12/10/2016: Version 1.2.6 released. Bug fixes for Enum data type, and _FillValue/missing_value usage when data is stored in non-native endian format. Add get_variables_by_attributes to MFDataset. Support for python 2.6 removed.

12/1/2016: Version 1.2.5 released. See the Changelog for changes.

4/15/2016: Version 1.2.4 released. Bugs in handling of variables with specified non-native "endian-ness" (byte-order) fixed ([issue #554] (#554)). Build instructions updated and warning issued to deal with potential backwards incompatibility introduced when using HDF5 1.10.x (see Unidata/netcdf-c/issue#250).

3/10/2016: Version 1.2.3 released. Various bug fixes. All text attributes in NETCDF4 formatted files are now written as type NC_CHAR, unless they contain unicode characters that cannot be encoded in ascii, in which case they are written as NC_STRING. Previously, all unicode strings were written as NC_STRING. This change preserves compatibility with clients, like Matlab, that can't deal with NC_STRING attributes. A setncattr_string method was added to force attributes to be written as NC_STRING.

1/1/2016: Version 1.2.2 released. Mostly bugfixes, but with two new features.

  • support for the new NETCDF3_64BIT_DATA format introduced in netcdf-c 4.4.0. Similar to NETCDF3_64BIT (now NETCDF3_64BIT_OFFSET), but includes 64 bit dimension sizes (> 2 billion), plus unsigned and 64 bit integer data types. Uses the classic (netcdf-3) data model, and does not use HDF5 as the underlying storage format.

  • Dimension objects now have a size attribute, which is the current length of the dimension (same as invoking len on the Dimension instance).

The minimum required python version has now been increased from 2.5 to 2.6.

10/15/2015: Version 1.2.1 released. Adds the ability to slice Variables with unsorted integer sequences, and integer sequences with duplicates.

9/23/2015: Version 1.2.0 released. New features:

7/28/2015: Version 1.1.9 bugfix release.

5/14/2015: Version 1.1.8 released. Unix-like paths can now be used in createVariable and createGroup.

    v = nc.createVariable('/path/to/var1', ('xdim', 'ydim'), float)

will create a variable named 'var1', while also creating the groups 'path' and 'path/to' if they do not already exist.


    g = nc.createGroup('/path/to') 

now acts like mkdir -p in unix, creating groups 'path' and '/path/to', if they don't already exist. Users who relied on nc.createGroup(groupname) failing when the group already exists will have to modify their code, since nc.createGroup will now return the existing group instance. Dataset.__getitem__ was also added. nc['/path/to'] now returns a group instance, and nc['/path/to/var1'] now returns a variable instance.

3/19/2015: Version 1.1.7 released. Global Interpreter Lock (GIL) now released when extension module calls C library for read operations. This speeds up concurrent reads when using threads. Users who wish to use netcdf4-python inside threads should read regarding thread-safety in the HDF5 C library. Fixes to now ensure that pip install netCDF4 with export USE_NCCONFIG=0 will use environment variables to find paths to libraries and include files, instead of relying exclusively on the nc-config utility.


The easiest way to install is through pip:

pip install netCDF4

or, if you are a user of the Conda package manager,

conda install -c conda-forge netCDF4

Development installation

  • Clone GitHub repository (git clone

  • Make sure numpy and Cython are installed and you have Python 3.7 or newer.

  • Make sure HDF5 and netcdf-4 are installed, and the nc-config utility is in your Unix PATH.

  • Run python build, then pip install -e ..

  • To run all the tests, execute cd test && python


See the online docs for more details.


Sample iPython notebooks available in the examples directory on reading and writing netCDF data with Python.