# NetCDF: Network Common Data Form - Vector¶

New in version 2.1.

Driver short name

netCDF

The netCDF driver supports read and write (creation from scratch and in some cases append operations) to vector datasets (you can find documentation for the raster side)

NetCDF is an interface for array-oriented data access and is used for representing scientific data.

The driver handles the “point” and “profile” feature types of the CF 1.6 convention. For CF-1.7 and below (as well as non-CF files), it also supports a more custom approach for non-point geometries.

The driver also supports writing and reading from CF-1.8 convention compliant files that have simple geometry information encoded within them.

## Driver capabilities¶

Supports Create()

This driver supports the GDALDriver::Create() operation

Supports Georeferencing

This driver supports georeferencing

Supports VirtualIO

This driver supports virtual I/O operations (/vsimem/, etc.)

## Conventions and Data Formats¶

The netCDF vector driver supports reading and writing netCDF files following the Climate and Forecast (CF) Metadata Conventions. Vector datasets can be written using the simple geometry specifcation of the CF-1.8 convention, or by using the CF-1.6 convention and by writing non-point geometry items as WKT.

### Distinguishing the Two Formats¶

Upon reading a netCDF file, the driver will attempt to read the global Conventions attribute. If it’s value is CF-1.8 or higher (in this exact format, as specified in the CF convention) then the driver will treat the netCDF file as one that has CF-1.8 geometries contained within it. If the Conventions attribute has a value of CF-1.6, the the file will be treated as following the CF-1.6 convention.

### CF-1.8 Writing Limitations¶

Writing to a CF-1.8 netCDF dataset poses some limitations. Only writing the feature types specified by the CF-1.8 standard (see section Geometry for more details) are supported, and measured features are only partially supported. Other geometries, such as non-simple curve geometries, are not supported in any way.

CF-1.8 datasets also do not support the append access mode.

There are what are considered reserved variable names for CF-1.8 datasets. These variable names are used by the driver to store its metadata. Refrain from using these names as layer names to avoid naming conflicts when writing datasets with multiple layers.

Suppose a layer in a CF-1.8 dataset has the name LAYER with a field with name FIELD. Then the following names would be considered reserved:

• LAYER_node_coordinates: used to store point information

• LAYER_node_count: used to store per shape point count information (not created if LAYER has a geometry type of Point)

• LAYER_part_node_count: used to store per part point count information (only created if LAYER consists of MultiLineStrings, MultiPolygons, or has at least one Polygon with interior rings)

• LAYER_interior_ring: used to store interior ring information (only created if LAYER consists of at least one Polygon with interior rings)

• LAYER_field_FIELD: used to store field information for FIELD.

These names are the only reserved names applying to CF-1.8 datasets.

CF-1.6/WKT datasets are not limited to the aforementioned restrictions.

## Mapping of concepts¶

### Field types¶

On creation of netCDF files, the mapping between OGR field types and netCDF type is the following :

OGR field type

netCDF type

String(1)

char

String

char (bi-dimensional), or string for NC4

Integer

int

Integer(Boolean)

byte

Integer(Int16)

short

Integer64

int64 for NC4, or double for NC3 as a fallback

Real

double

Real(Float32)

float

Date

int (with units=”days since 1970-1-1”)

DateTime

double (with units=”seconds since 1970-1-1 0:0:0”)

The driver also writes the following attributes for each OGR fields / netCDF variables.

• ogr_field_name: OGR field name (useful if the netCDF variable name is different, due to collision)

• ogr_field_type: OGR field type (such as String,Integer,Date,DateTime,etc…)

• ogr_field_width: OGR field width. Only set if it is non-zero, except for strings

• ogr_field_precision: OGR field precision. Only set if it is non-zero

They are written by default (unless the WRITE_GDAL_TAGS dataset creation option is set to NO). They are not required for reading, but may help to better identify field characteristics

On reading, the mapping is the following :

netCDF type

OGR field type

byte

Integer

ubyte (NC4 only)

Integer

char (mono dimensional)

String(1)

char (bi dimensional)

String

string (NC4 only)

String

short

Integer(Int16)

ushort (NC4 only)

Integer

int

Integer

int or double (with units=”days since 1970-1-1”)

Date

uint (NC4 only)

Integer64

int64 (NC4 only)

Integer64

uint64 (NC4 only)

Real

float

Real(Float32)

double

Real

double (with units=”seconds since 1970-1-1 0:0:0”)

DateTime

### Layers¶

In the CF-1.8 compliant driver, a single layer corresponds to a single geometry container within a CF-1.8 compliant netCDF file. A geometry container, per the CF-1.8 specification, is referred to by another variable (presumably a data variable) through the geometry attribute. When reading a CF-1.8 compliant netCDF file, all geometry containers within the netCDF file will be present in the opened dataset as separate layers. Similarily, when writing to a CF-1.8 dataset, each layer will be written to a geometry container whose variable name is that of the source layer. When writing to a CF-1.8 dataset specifically, multiple layers are always enabled and are always in a single netCDF file, regardless of the MULTIPLE_LAYERS option.

When working with files made with older versions of the driver (pre CF-1.8), a single netCDF file generally corresponds to a single OGR layer, provided that it contains only mono-dimensional variables, indexed by the same dimension (or bi-dimensional variables of type char). For netCDF v4 files with multiple groups, each group may be seen as a separate OGR layer. On writing, the MULTIPLE_LAYERS dataset creation option can be used to control whether multiple layers is disabled, or if multiple layers should go in separate files, or separate groups.

### Strings¶

Variable length strings are not natively supported in netCDF v3 format. To work around that, OGR uses bi-dimensional char variables, whose first dimension is the record dimension, and second dimension the maximum width of the string.

By default, OGR implements a “auto-grow” mode in writing, where the maximum width of the variable used to store a OGR string field is extended when needed.

For WKT datasets, this leads to a full rewrite of already written records; although this process is transparent for the user, it can slow down the creation process in non-linear ways. A similar mechanism is used to handle layers with geometry types other than point to store the ISO WKT representation of the geometries.

For CF-1.8 datasets, growing the string width dimension is a relatively inexpensive process which does not involve recopying of records, but involves only a simple integer reassignment. Because of how inexpensive dimension growth is with CF-1.8 datasets, auto growth of the string width dimension is always on.

When using a netCDF v4 output format (NC4), strings will be by default written as netCDF v4 variable length strings.

### Geometry¶

Supported feature types when reading from a CF-1.8 convention compliant netCDF file include OGRPoint, OGRLineString, OGRPolygon, OGRMultiPoint, OGRMultiLineString, and OGRMultiPolygon. Due to slight ambiguities present in the CF-1.8 convention concerning Polygons versus MultiPolygons, the driver will in most cases default to assuming a MultiPolygon for the geometry of a layer with geometry_type polygon. The one exception where a Polygon type will be used is when the attribute part_node_count is not present within that layer’s geometry container. Per convention requirements, the driver supports reading and writing from geometries with X, Y, and Z axes. Writing from source layers with features containing an M axis is also partially supported. The X, Y, and Z information of a measured feature will be able to be captured in a CF-1.8 netCDF file, but the measure information will be lost completely.

When working with a CF-1.6/WKT dataset, layers with a geometry type of Point or Point25D will cause the implicit creation of x,y(,z) variables for a projected coordinate system, or lon,lat(,z) variables for geographic coordinate systems. For other geometry types, a variable “ogc_wkt” ( bi-dimensional char for NC3 output, or string for NC4 output) is created and used to store the geometry as a ISO WKT string.

### “Profile” feature type¶

The driver can handle “profile” feature type, i.e. phenomenons that happen at a few positions along a vertical line at a fixed horizontal position. In that representation, some variables are indexed by the profile, and others by the observation.

More precisely, the driver supports reading and writing profiles organized accordingly with the “Indexed ragged array representation” of profiles.

On reading, the driver will collect values of variables indexed by the profile dimension and expose them as long as variables indexed by the observation dimension, based on a variable such as “parentIndex” with an attribute “instance_dimension” pointing to the profile dimension.

On writing, the FEATURE_TYPE=PROFILE layer creation option must be set and the driver will need to be instructed which OGR fields are indexed either by the profile or by the observation dimension. The list of fields indexed by the profile can be specified with the PROFILE_VARIABLES layer creation options (other fields are assumed to be indexed by the observation dimension). Fields indexed by the profile are the horizontal geolocation (created implicitly), and other user attributes such as the location name, etc. Care should be taken into selecting which variables are indexed by the profile dimension: given 2 OGR features (taking into account only the variables indexed by the profile dimension), if they have different values for such variables, they will be considered to belong to different profiles.

In the below example, the station_name and time variables may be indexed by the profile dimension (the geometry is assumed to be also indexed by the profile dimension), since all records that have the same value for one of those variables have same values for the other ones, whereas temparature and Z should be indexed by the default dimension.

station_name

time

geometry

temperature

Z

Paris

2016-03-01T00:00:00Z

POINT (2 49)

25

100

Vancouver

2016-04-01T12:00:00Z

POINT (-123 49.25)

5

100

Paris

2016-03-01T00:00:00Z

POINT (2 49)

3

500

Vancouver

2016-04-01T12:00:00Z

POINT (-123 49.25)

-15

500

An integer field, with the name of the profile dimension (whose default name is “profile”, which can be altered with the PROFILE_DIM_NAME layer creation option), will be used to store the automatically computed id of profile sites (unless a integer OGR field with the same name exits).

The size of the profile dimension defaults to 100 for non-NC4 output format, and is extended automatically in case of additional profiles (with similar performance issues as growing strings). For NC4 output format, the profile dimension is of unlimited size by default.

## Dataset creation options¶

• GEOMETRY_ENCODING=CF_1.8/WKT: Chooses which geometry encoding to use when creating new layers within the dataset. Default is CF_1.8.

• FORMAT=NC/NC2/NC4/NC4C: netCDF format. NC is the classic netCDF format (compatible of netCDF v3.X and 4.X libraries). NC2 is the extension of NC for files larger than 4 GB. NC4 is the netCDF v4 format, using a HDF5 container, offering new capabilities (new types, concept of groups, etc…) only available in netCDF v4 library. NC4C is a restriction of the NC4 format to the concepts supported by the classic netCDF format. Default is NC.

• WRITE_GDAL_TAGS=YES/NO: Whether to write GDAL specific information as netCDF attributes. Default is YES.

• CONFIG_FILE=string. Path to a XML configuration file (or its content inlined) for precise control of the output.

The following option will only have effect when simultaneously specifying GEOMETRY_ENCODING=WKT:

• MULTIPLE_LAYERS=NO/SEPARATE_FILES/SEPARATE_GROUPS. Default is NO, i.e a dataset can contain only a single OGR layer. SEPARATE_FILES can be used to put the content of each OGR layer in a single netCDF file, in which case the name passed at dataset creation is used as the directory, and the layer name is used as the basename of the netCDF file. SEPARATE_GROUPS may be used when FORMAT=NC4 to put each OGR layer in a separate netCDF group, inside the same file.

## Layer creation options¶

The following option applies to both dataset types:

• USE_STRING_IN_NC4=YES/NO. Whether to use NetCDF string type for strings in NC4 format. If NO, bidimensional char variable are used. Default to YES when FORMAT=NC4.

The following options require a dataset with GEOMETRY_ENCODING=WKT:

• RECORD_DIM_NAME=string. Name of the unlimited dimension that index features. Defaults to “record”.

• STRING_DEFAULT_WIDTH=int. Default width of strings (when using bi-dimensional char variables). Default is 10 in autogrow mode, 80 otherwise.

• WKT_DEFAULT_WIDTH=int. Default width of WKT strings (when using bi-dimensional char variables). Default is 1000 in autogrow mode, 10000 otherwise.

• AUTOGROW_STRINGS=YES/NO. Whether to auto-grow string fields of non-fixed width, or ogc_wkt special field, when serialized as bidimensional char variables. Default is YES. When set to NO, if the string is larger than its maximum initial width (set by STRING_DEFAULT_WIDTH), it is truncated. For a geometry, it is completely discarded.

• FEATURE_TYPE=AUTO/POINT/PROFILE. Select the CF FeatureType. Defaults to AUTO where FeatureType=Point is selected if the layer geometry type is Point, otherwise the custom approach involving the “ogc_wkt” field is used. Can be set to PROFILE so as to select the creation of an indexed ragged array representation of profiles.

• PROFILE_DIM_NAME=string. Name of the profile dimension and variable. Defaults to “profile”. Only used when FEATURE_TYPE=PROFILE.

• PROFILE_DIM_INIT_SIZE=int or string. Initial size of profile dimension, or UNLIMITED for NC4 files. Defaults to 100 when FORMAT != NC4 and to UNLIMITED when FORMAT = NC4. Only used when FEATURE_TYPE=PROFILE.

• PROFILE_VARIABLES=string. Comma separated list of field names that must be indexed by the profile dimension. Only used when FEATURE_TYPE=PROFILE.

The following option requires a dataset with GEOMETRY_ENCODING=CF_1.8:

• BUFFER_SIZE=int. The soft limit of the write buffer in bytes. Larger values generally imply better performance, but values should be comfortably less than that of available physical memory or else thrashing can occur. By default, this value is set at 20% of usable physical memory (usable meaning total physical RAM considering limitations of virtual address space size). Buffer contents are committed between translating features, but not during translating a feature, so this limit does not apply to a single feature. The minimum acceptable size is 4096. If a value lower than this is specified the default will be used.

• GROUPLESS_WRITE_BACK=YES/NO. In order to reduce time used to write data to the target netCDF file, data is often grouped together in arrays and written all at once. Each of these arrays is associated with a variable in the target dataset. Arrays are destroyed as soon as the associated data is written to the netCDF file which in turn occurs as soon as a complete data array for a variable is assembled in memory. For machines with small memory sizes, this optimization may cause issues when writing large datasets with large layers. Turning this option on by specifying “YES” disables array writing and causes data to be written one datum at a time. It is strongly recommended to keep this option off unless out of memory errors or performance issues occur. In the general case, this technique greatly improves translation efficiency. The default value is NO.

## XML configuration file¶

A XML configuration file conforming to the following schema can be used for very precise control on the output format, in particular to set all needed attributes (such as units) to conform to the NetCDF CF-1.6 convention.

It has been designed in particular, but not exclusively, to be usable in use cases involving the MapServer OGR output.

Such a file can be used to :

• set dataset and layer creation options.

• set global netCDF attributes.

• map OGR field names to netCDF variable names.

• set netCDF attributes attached to netCDF variables.

The scope of effect is either globally, when elements are defined as direct children of the root <Configuration> node, or specifically to a given layer, when defined as children of a <Layer> node.

The filename is specified with the CONFIG_FILE dataset creation option. Alternatively, the content of the file can be specifid inline as the value of the option (it must then begin strictly with the “<Configuration” characters)

The following example shows all possibilities and precedence rules:

<Configuration>
<DatasetCreationOption name="FORMAT" value="NC4"/>
<DatasetCreationOption name="MULTIPLE_LAYERS" value="SEPARATE_GROUPS"/>
<LayerCreationOption name="RECORD_DIM_NAME" value="observation"/>
<!-- applies to all layers -->
<Field name="weight">  <!-- edit user field/variable -->
<Attribute name="units" value="kg"/>
<Attribute name="maximum" value="10" type="double"/>
</Field>
<Field netcdf_name="z"> <!-- edit predefined variable -->
<Attribute name="long_name" value="Elevation"/>
</Field>
<!-- start of layer specific definitions -->
<Layer name="1st_layer" netcdf_name="firstlayer"> <!-- OGR layer "1st_layer" is renamed as "firstlayer" netCDF group -->
<LayerCreationOption name="FEATURE_TYPE" value="POINT"/>
<Attribute name="copyright" value="Public domain"/> <!-- override global one -->
<Attribute name="description" value="This is my first layer"/> <!-- additional attribute -->
<Field name="1st_field" netcdf_name="firstfield"/> <!-- rename OGR field "1st_field" as the "firstfield" netCDF variable -->
<Field name="weight"/> <!-- cancel above global customization -->
<Field netcdf_name="lat"> <!-- edit predefined variable -->
<Attribute name="long_name" value=""/> <!-- remove predefined attribute -->
</Field>
</Layer>
<Layer name="sounding">
<LayerCreationOption name="FEATURE_TYPE" value="PROFILE"/>
<Field name="station_name" main_dim="profile"/> <!-- the corresponding netCDF variable will be indexed against the profile dimension, instead of the observation dimension -->
<Field name="time" main_dim="profile"/> <!-- the corresponding netCDF variable will be indexed against the profile dimension, instead of the observation dimension -->
</Layer>
</Configuration>


The effect on the output can be checked by running the ncdump utility