Preprocessor¶
In this section, each of the preprocessor modules is described, roughly following the default order in which preprocessor functions are applied:
See Preprocessor functions for implementation details and the exact default order.
Overview¶
The ESMValTool preprocessor can be used to perform a broad range of operations on the input data before diagnostics or metrics are applied. The preprocessor performs these operations in a centralized, documented and efficient way, thus reducing the data processing load on the diagnostics side. For an overview of the preprocessor structure see the Recipe section: preprocessors.
Each of the preprocessor operations is written in a dedicated python module and
all of them receive and return an instance of
iris.cube.Cube
, working
sequentially on the data with no interactions between them. The order in which
the preprocessor operations is applied is set by default to minimize
the loss of information due to, for example, temporal and spatial subsetting or
multimodel averaging. Nevertheless, the user is free to change such order to
address specific scientific requirements, but keeping in mind that some
operations must be necessarily performed in a specific order. This is the case,
for instance, for multimodel statistics, which required the model to be on a
common grid and therefore has to be called after the regridding module.
Variable derivation¶
The variable derivation module allows to derive variables which are not in the CMIP standard data request using standard variables as input. The typical use case of this operation is the evaluation of a variable which is only available in an observational dataset but not in the models. In this case a derivation function is provided by the ESMValTool in order to calculate the variable and perform the comparison. For example, several observational datasets deliver total column ozone as observed variable (toz), but CMIP models only provide the ozone 3D field. In this case, a derivation function is provided to vertically integrate the ozone and obtain total column ozone for direct comparison with the observations.
To contribute a new derived variable, it is also necessary to define a name for it and to provide the corresponding CMOR table. This is to guarantee the proper metadata definition is attached to the derived data. Such custom CMOR tables are collected as part of the ESMValCore package. By default, the variable derivation will be applied only if the variable is not already available in the input data, but the derivation can be forced by setting the appropriate flag.
variables:
toz:
derive: true
force_derivation: false
The required arguments for this module are two boolean switches:
derive
: activate variable derivationforce_derivation
: force variable derivation even if the variable is directly available in the input data.
See also esmvalcore.preprocessor.derive()
. To get an overview on
derivation scripts and how to implement new ones, please go to
Deriving a variable.
CMORization and datasetspecific fixes¶
Data checking¶
Data preprocessed by ESMValTool is automatically checked against its cmor definition. To reduce the impact of this check while maintaining it as reliable as possible, it is split in two parts: one will check the metadata and will be done just after loading and concatenating the data and the other one will check the data itself and will be applied after all extracting operations are applied to reduce the amount of data to process.
Checks include, but are not limited to:
Requested coordinates are present and comply with their definition.
Correctness of variable names, units and other metadata.
Compliance with the valid minimum and maximum values allowed if defined.
The most relevant (i.e. a missing coordinate) will raise an error while others (i.e an incorrect long name) will be reported as a warning.
Some of those issues will be fixed automatically by the tool, including the following:
Incorrect standard or long names.
Incorrect units, if they can be converted to the correct ones.
Direction of coordinates.
Automatic clipping of longitude to 0  360 interval.
Minute differences between the required and actual vertical coordinate values
Dataset specific fixes¶
Sometimes, the checker will detect errors that it can not fix by itself. ESMValTool deals with those issues by applying specific fixes for those datasets that require them. Fixes are applied at three different preprocessor steps:
fix_file: apply fixes directly to a copy of the file. Copying the files is costly, so only errors that prevent Iris to load the file are fixed here. See
esmvalcore.preprocessor.fix_file()
fix_metadata: metadata fixes are done just before concatenating the cubes loaded from different files in the final one. Automatic metadata fixes are also applied at this step. See
esmvalcore.preprocessor.fix_metadata()
fix_data: data fixes are applied before starting any operation that will alter the data itself. Automatic data fixes are also applied at this step. See
esmvalcore.preprocessor.fix_data()
To get an overview on data fixes and how to implement new ones, please go to Fixing data.
Fx variables as cell measures or ancillary variables¶
The following preprocessors may require the use of fx_variables
to be able
to perform the computations:
Preprocessor 
Default fx variables 










If the option fx_variables
is not explicitly specified for these
preprocessors, the default fx variables in the second column are automatically
used. If given, the fx_variables
argument specifies the fx variables that
the user wishes to input to the corresponding preprocessor function. The user
may specify these by simply adding the names of the variables, e.g.,
fx_variables:
areacello:
volcello:
or by additionally specifying further keys that are used to define the fx datasets, e.g.,
fx_variables:
areacello:
mip: Ofx
exp: piControl
volcello:
mip: Omon
This might be useful to select fx files from a specific mip
table or from a
specific exp
in case not all experiments provide the fx variable.
Alternatively, the fx_variables
argument can also be specified as a list:
fx_variables: ['areacello', 'volcello']
or as a list of dictionaries:
fx_variables: [{'short_name': 'areacello', 'mip': 'Ofx', 'exp': 'piControl'}, {'short_name': 'volcello', 'mip': 'Omon'}]
The recipe parser will automatically find the data files that are associated with these variables and pass them to the function for loading and processing.
If mip
is not given, ESMValTool will search for the fx variable in all
available tables of the specified project.
Warning
Some fx variables exist in more than one table (e.g., volcello
exists in
CMIP6’s Odec
, Ofx
, Omon
, and Oyr
tables; sftgif
exists
in CMIP6’s fx
, IyrAnt
and IyrGre
, and LImon
tables). If (for
a given dataset) fx files are found in more than one table, mip
needs to
be specified, otherwise an error is raised.
Note
To explicitly not use any fx variables in a preprocessor, use
fx_variables: null
. While some of the preprocessors mentioned above do
work without fx variables (e.g., area_statistics
or mask_landsea
with datasets that have regular latitude/longitude grids), using this option
is not recommended.
Internally, the required fx_variables
are automatically loaded by the
preprocessor step add_fx_variables
which also checks them against CMOR
standards and adds them either as cell_measure
(see CF conventions on cell
measures
and iris.coords.CellMeasure
) or ancillary_variable
(see CF
conventions on ancillary variables
and iris.coords.AncillaryVariable
) inside the cube data. This ensures
that the defined preprocessor chain is applied to both variables
and
fx_variables
.
Note that when calling steps that require fx_variables
inside diagnostic
scripts, the variables are expected to contain the required cell_measures
or
ancillary_variables
. If missing, they can be added using the following functions:
from esmvalcore.preprocessor import (add_cell_measure, add_ancillary_variable)
cube_with_area_measure = add_cell_measure(cube, area_cube, 'area')
cube_with_volume_measure = add_cell_measure(cube, volume_cube, 'volume)
cube_with_ancillary_sftlf = add_ancillary_variable(cube, sftlf_cube)
cube_with_ancillary_sftgif = add_ancillary_variable(cube, sftgif_cube)
Details on the arguments needed for each step can be found in the following sections.
Vertical interpolation¶
Vertical level selection is an important aspect of data preprocessing since it
allows the scientist to perform a number of metrics specific to certain levels
(whether it be air pressure or depth, e.g. the QuasiBiennialOscillation (QBO)
u30 is computed at 30 hPa). Dataset native vertical grids may not come with the
desired set of levels, so an interpolation operation will be needed to regrid
the data vertically. ESMValTool can perform this vertical interpolation via the
extract_levels
preprocessor. Level extraction may be done in a number of
ways.
Level extraction can be done at specific values passed to extract_levels
as
levels:
with its value a list of levels (note that the units are
CMORstandard, Pascals (Pa)):
preprocessors:
preproc_select_levels_from_list:
extract_levels:
levels: [100000., 50000., 3000., 1000.]
scheme: linear
It is also possible to extract the CMIPspecific, CMOR levels as they appear in
the CMOR table, e.g. plev10
or plev17
or plev19
etc:
preprocessors:
preproc_select_levels_from_cmip_table:
extract_levels:
levels: {cmor_table: CMIP6, coordinate: plev10}
scheme: nearest
Of good use is also the level extraction with values specific to a certain
dataset, without the user actually polling the dataset of interest to find out
the specific levels: e.g. in the example below we offer two alternatives to
extract the levels and vertically regrid onto the vertical levels of
ERAInterim
:
preprocessors:
preproc_select_levels_from_dataset:
extract_levels:
levels: ERAInterim
# This also works, but allows specifying the pressure coordinate name
# levels: {dataset: ERAInterim, coordinate: air_pressure}
scheme: linear_horizontal_extrapolate_vertical
By default, vertical interpolation is performed in the dimension coordinate of
the z axis. If you want to explicitly declare the z axis coordinate to use
(for example, air_pressure
’ in variables that are provided in model levels
and not pressure levels) you can override that automatic choice by providing
the name of the desired coordinate:
preprocessors:
preproc_select_levels_from_dataset:
extract_levels:
levels: ERAInterim
scheme: linear_horizontal_extrapolate_vertical
coordinate: air_pressure
If coordinate
is specified, pressure levels (if present) can be converted
to height levels and vice versa using the US standard atmosphere. E.g.
coordinate = altitude
will convert existing pressure levels
(air_pressure) to height levels (altitude);
coordinate = air_pressure
will convert existing height levels
(altitude) to pressure levels (air_pressure).
If the requested levels are very close to the values in the input data, the function will just select the available levels instead of interpolating. The meaning of ‘very close’ can be changed by providing the parameters:
rtol
Relative tolerance for comparing the levels in the input data to the requested levels. If the levels are sufficiently close, the requested levels will be assigned to the vertical coordinate and no interpolation will take place. The default value is 10^7.
atol
Absolute tolerance for comparing the levels in the input data to the requested levels. If the levels are sufficiently close, the requested levels will be assigned to the vertical coordinate and no interpolation will take place. By default, atol will be set to 10^7 times the mean value of of the available levels.
See also
esmvalcore.preprocessor.extract_levels()
.See also
esmvalcore.preprocessor.get_cmor_levels()
.
Note
For both vertical and horizontal regridding one can control the extrapolation mode when defining the interpolation scheme. Controlling the extrapolation mode allows us to avoid situations where extrapolating values makes little physical sense (e.g. extrapolating beyond the last data point). The extrapolation mode is controlled by the extrapolation_mode keyword. For the available interpolation schemes available in Iris, the extrapolation_mode keyword must be one of:
extrapolate
: the extrapolation points will be calculated by extending the gradient of the closest two points;
error
: aValueError
exception will be raised, notifying an attempt to extrapolate;
nan
: the extrapolation points will be be set to NaN;
mask
: the extrapolation points will always be masked, even if the source data is not aMaskedArray
; or
nanmask
: if the source data is a MaskedArray the extrapolation points will be masked, otherwise they will be set to NaN.
Weighting¶
Land/sea fraction weighting¶
This preprocessor allows weighting of data by land or sea fractions. In other words, this function multiplies the given input field by a fraction in the range 01 to account for the fact that not all grid points are completely land or seacovered.
The application of this preprocessor is very important for most carbon cycle variables (and other land surface outputs), which are e.g. reported in units of \(kgC~m^{2}\). Here, the surface unit actually refers to ‘square meter of land/sea’ and NOT ‘square meter of gridbox’. In order to integrate these globally or regionally one has to weight by both the surface quantity and the land/sea fraction.
For example, to weight an input field with the land fraction, the following preprocessor can be used:
preprocessors:
preproc_weighting:
weighting_landsea_fraction:
area_type: land
exclude: ['CanESM2', 'reference_dataset']
Allowed arguments for the keyword area_type
are land
(fraction is 1
for grid cells with only land surface, 0 for grid cells with only sea surface
and values in between 0 and 1 for coastal regions) and sea
(1 for
sea, 0 for land, in between for coastal regions). The optional argument
exclude
allows to exclude specific datasets from this preprocessor, which
is for example useful for climate models which do not offer land/sea fraction
files. This arguments also accepts the special dataset specifiers
reference_dataset
and alternative_dataset
.
Optionally you can specify your own custom fx variable to be used in cases when e.g. a certain experiment is preferred for fx data retrieval:
preprocessors:
preproc_weighting:
weighting_landsea_fraction:
area_type: land
exclude: ['CanESM2', 'reference_dataset']
fx_variables:
sftlf:
exp: piControl
sftof:
exp: piControl
or alternatively:
preprocessors:
preproc_weighting:
weighting_landsea_fraction:
area_type: land
exclude: ['CanESM2', 'reference_dataset']
fx_variables: [
{'short_name': 'sftlf', 'exp': 'piControl'},
{'short_name': 'sftof', 'exp': 'piControl'}
]
More details on the argument fx_variables
and its default values are given
in Fx variables as cell measures or ancillary variables.
See also esmvalcore.preprocessor.weighting_landsea_fraction()
.
Masking¶
Introduction to masking¶
Certain metrics and diagnostics need to be computed and performed on specific domains on the globe. The ESMValTool preprocessor supports filtering the input data on continents, oceans/seas and ice. This is achieved by masking the model data and keeping only the values associated with grid points that correspond to, e.g., land, ocean or ice surfaces, as specified by the user. Where possible, the masking is realized using the standard mask files provided together with the model data as part of the CMIP data request (the socalled fx variable). In the absence of these files, the Natural Earth masks are used: although these are not modelspecific, they represent a good approximation since they have a much higher resolution than most of the models and they are regularly updated with changing geographical features.
Landsea masking¶
In ESMValTool, landseaice masking can be done in two places: in the preprocessor, to apply a mask on the data before any subsequent preprocessing step and before running the diagnostic, or in the diagnostic scripts themselves. We present both these implementations below.
To mask out a certain domain (e.g., sea) in the preprocessor,
mask_landsea
can be used:
preprocessors:
preproc_mask:
mask_landsea:
mask_out: sea
and requires only one argument: mask_out
: either land
or sea
.
Optionally you can specify your own custom fx variable to be used in cases when e.g. a certain experiment is preferred for fx data retrieval. Note that it is possible to specify as many tags for the fx variable as required:
preprocessors:
landmask:
mask_landsea:
mask_out: sea
fx_variables:
sftlf:
exp: piControl
sftof:
exp: piControl
ensemble: r2i1p1f1
or alternatively:
preprocessors:
landmask:
mask_landsea:
mask_out: sea
fx_variables: [
{'short_name': 'sftlf', 'exp': 'piControl'},
{'short_name': 'sftof', 'exp': 'piControl', 'ensemble': 'r2i1p1f1'}
]
More details on the argument fx_variables
and its default values are given
in Fx variables as cell measures or ancillary variables.
If the corresponding fx file is not found (which is the case for some models and almost all observational datasets), the preprocessor attempts to mask the data using Natural Earth mask files (that are vectorized rasters). As mentioned above, the spatial resolution of the the Natural Earth masks are much higher than any typical global model (10m for land and glaciated areas and 50m for ocean masks).
See also esmvalcore.preprocessor.mask_landsea()
.
Ice masking¶
Note that for masking out ice sheets, the preprocessor uses a different
function, to ensure that both land and sea or ice can be masked out without
losing generality. To mask ice out, mask_landseaice
can be used:
preprocessors:
preproc_mask:
mask_landseaice:
mask_out: ice
and requires only one argument: mask_out
: either landsea
or ice
.
Optionally you can specify your own custom fx variable to be used in cases when e.g. a certain experiment is preferred for fx data retrieval:
preprocessors:
landseaicemask:
mask_landseaice:
mask_out: sea
fx_variables:
sftgif:
exp: piControl
or alternatively:
preprocessors:
landseaicemask:
mask_landseaice:
mask_out: sea
fx_variables: [{'short_name': 'sftgif', 'exp': 'piControl'}]
More details on the argument fx_variables
and its default values are given
in Fx variables as cell measures or ancillary variables.
Glaciated masking¶
For masking out glaciated areas a Natural Earth shapefile is used. To mask
glaciated areas out, mask_glaciated
can be used:
preprocessors:
preproc_mask:
mask_glaciated:
mask_out: glaciated
and it requires only one argument: mask_out
: only glaciated
.
Missing values masks¶
Missing (masked) values can be a nuisance especially when dealing with multimodel ensembles and having to compute multimodel statistics; different numbers of missing data from dataset to dataset may introduce biases and artificially assign more weight to the datasets that have less missing data. This is handled in ESMValTool via the missing values masks: two types of such masks are available, one for the multimodel case and another for the single model case.
The multimodel missing values mask (mask_fillvalues
) is a preprocessor step
that usually comes after all the singlemodel steps (regridding, area selection
etc) have been performed; in a nutshell, it combines missing values masks from
individual models into a multimodel missing values mask; the individual model
masks are built according to common criteria: the user chooses a time window in
which missing data points are counted, and if the number of missing data points
relative to the number of total data points in a window is less than a chosen
fractional threshold, the window is discarded i.e. all the points in the window
are masked (set to missing).
preprocessors:
missing_values_preprocessor:
mask_fillvalues:
threshold_fraction: 0.95
min_value: 19.0
time_window: 10.0
In the example above, the fractional threshold for missing data vs. total data is set to 95% and the time window is set to 10.0 (units of the time coordinate units). Optionally, a minimum value threshold can be applied, in this case it is set to 19.0 (in units of the variable units).
Common mask for multiple models¶
To create a combined multimodel mask (all the masks from all the analyzed
datasets combined into a single mask using a logical OR), the preprocessor
mask_multimodel
can be used. In contrast to mask_fillvalues
,
mask_multimodel
does not expect that the datasets have a time
coordinate, but works on datasets with arbitrary (but identical) coordinates.
After mask_multimodel
, all involved datasets have an identical mask.
Minimum, maximum and interval masking¶
Thresholding on minimum and maximum accepted data values can also be performed:
masks are constructed based on the results of thresholding; inside and outside
interval thresholding and masking can also be performed. These functions are
mask_above_threshold
, mask_below_threshold
, mask_inside_range
, and
mask_outside_range
.
These functions always take a cube as first argument and either threshold
for threshold masking or the pair minimum
, maximum
for interval masking.
See also esmvalcore.preprocessor.mask_above_threshold()
and related
functions.
Horizontal regridding¶
Regridding is necessary when various datasets are available on a variety of latlon grids and they need to be brought together on a common grid (for various statistical operations e.g. multimodel statistics or for e.g. direct intercomparison or comparison with observational datasets). Regridding is conceptually a very similar process to interpolation (in fact, the regridder engine uses interpolation and extrapolation, with various schemes). The primary difference is that interpolation is based on sample data points, while regridding is based on the horizontal grid of another cube (the reference grid). If the horizontal grids of a cube and its reference grid are sufficiently the same, regridding is automatically and silently skipped for performance reasons.
The underlying regridding mechanism in ESMValTool uses
iris.cube.Cube.regrid
from Iris.
The use of the horizontal regridding functionality is flexible depending on what type of reference grid and what interpolation scheme is preferred. Below we show a few examples.
Regridding on a reference dataset grid¶
The example below shows how to regrid on the reference dataset
ERAInterim
(observational data, but just as well CMIP, obs4MIPs,
or ana4mips datasets can be used); in this case the scheme is
linear.
preprocessors:
regrid_preprocessor:
regrid:
target_grid: ERAInterim
scheme: linear
Regridding on an MxN
grid specification¶
The example below shows how to regrid on a reference grid with a cell
specification of 2.5x2.5
degrees. This is similar to regridding on
reference datasets, but in the previous case the reference dataset grid cell
specifications are not necessarily known a priori. Regridding on an MxN
cell specification is oftentimes used when operating on localized data.
preprocessors:
regrid_preprocessor:
regrid:
target_grid: 2.5x2.5
scheme: nearest
In this case the NearestNeighbour
interpolation scheme is used (see below
for scheme definitions).
When using a MxN
type of grid it is possible to offset the grid cell
centrepoints using the lat_offset and lon_offset
arguments:
lat_offset
: offsets the grid centers of the latitude coordinate w.r.t. the pole by half a grid step;lon_offset
: offsets the grid centers of the longitude coordinate w.r.t. Greenwich meridian by half a grid step.
preprocessors:
regrid_preprocessor:
regrid:
target_grid: 2.5x2.5
lon_offset: True
lat_offset: True
scheme: nearest
Regridding to a regional target grid specification¶
This example shows how to regrid to a regional target grid specification.
This is useful if both a regrid
and extract_region
step are necessary.
preprocessors:
regrid_preprocessor:
regrid:
target_grid:
start_longitude: 40
end_longitude: 60
step_longitude: 2
start_latitude: 10
end_latitude: 30
step_latitude: 2
scheme: nearest
This defines a grid ranging from 40° to 60° longitude with 2° steps,
and 10° to 30° latitude with 2° steps. If end_longitude
or end_latitude
do
not fall on the grid (e.g., end_longitude: 61
), it cuts off at the nearest
previous value (e.g. 60
).
The longitude coordinates will wrap around the globe if necessary, i.e.
start_longitude: 350
, end_longitude: 370
is valid input.
The arguments are defined below:
start_latitude
: Latitude value of the first grid cell center (start point). The grid includes this value.end_latitude
: Latitude value of the last grid cell center (end point). The grid includes this value only if it falls on a grid point. Otherwise, it cuts off at the previous value.step_latitude
: Latitude distance between the centers of two neighbouring cells.start_longitude
: Latitude value of the first grid cell center (start point). The grid includes this value.end_longitude
: Longitude value of the last grid cell center (end point). The grid includes this value only if it falls on a grid point. Otherwise, it cuts off at the previous value.step_longitude
: Longitude distance between the centers of two neighbouring cells.
Regridding (interpolation, extrapolation) schemes¶
The schemes used for the interpolation and extrapolation operations needed by the horizontal regridding functionality directly map to their corresponding implementations in Iris:
linear
:Linear(extrapolation_mode='mask')
, seeiris.analysis.Linear
.linear_extrapolate
:Linear(extrapolation_mode='extrapolate')
, seeiris.analysis.Linear
.nearest
:Nearest(extrapolation_mode='mask')
, seeiris.analysis.Nearest
.area_weighted
:AreaWeighted()
, seeiris.analysis.AreaWeighted
.unstructured_nearest
:UnstructuredNearest()
, seeiris.analysis.UnstructuredNearest
.
See also esmvalcore.preprocessor.regrid()
Note
For both vertical and horizontal regridding one can control the extrapolation mode when defining the interpolation scheme. Controlling the extrapolation mode allows us to avoid situations where extrapolating values makes little physical sense (e.g. extrapolating beyond the last data point). The extrapolation mode is controlled by the extrapolation_mode keyword. For the available interpolation schemes available in Iris, the extrapolation_mode keyword must be one of:
extrapolate
– the extrapolation points will be calculated by extending the gradient of the closest two points;
error
– aValueError
exception will be raised, notifying an attempt to extrapolate;
nan
– the extrapolation points will be be set to NaN;
mask
– the extrapolation points will always be masked, even if the source data is not aMaskedArray
; or
nanmask
– if the source data is a MaskedArray the extrapolation points will be masked, otherwise they will be set to NaN.
Note
The regridding mechanism is (at the moment) done with fully realized data in
memory, so depending on how fine the target grid is, it may use a rather
large amount of memory. Empirically target grids of up to 0.5x0.5
degrees should not produce any memoryrelated issues, but be advised that
for resolutions of < 0.5
degrees the regridding becomes very slow and
will use a lot of memory.
Multimodel statistics¶
Computing multimodel statistics is an integral part of model analysis and
evaluation: individual models display a variety of biases depending on model
setup, initial conditions, forcings and implementation; comparing model data to
observational data, these biases have a significantly lower statistical impact
when using a multimodel ensemble. ESMValTool has the capability of computing a
number of multimodel statistical measures: using the preprocessor module
multi_model_statistics
will enable the user to ask for either a multimodel
mean
, median
, max
, min
, std
, and / or pXX.YY
with a set
of argument parameters passed to multi_model_statistics
. Percentiles can be
specified like p1.5
or p95
. The decimal point will be replaced by a dash
in the output file.
Restrictive computation is also available by excluding any set of models that
the user will not want to include in the statistics (by setting exclude:
[excluded models list]
argument). The implementation has a few restrictions
that apply to the input data: model datasets must have consistent shapes, apart
from the time dimension; and cubes with more than four dimensions (time,
vertical axis, two horizontal axes) are not supported.
Input datasets may have different time coordinates. Statistics can be computed
across overlapping times only (span: overlap
) or across the full time span
of the combined models (span: full
). The preprocessor sets a common time
coordinate on all datasets. As the number of days in a year may vary between
calendars, (sub)daily data with different calendars are not supported.
Input datasets may have different time coordinates. The multimodel statistics preprocessor sets a common time coordinate on all datasets. As the number of days in a year may vary between calendars, (sub)daily data are not supported.
preprocessors:
multi_model_preprocessor:
multi_model_statistics:
span: overlap
statistics: [mean, median]
exclude: [NCEP]
see also esmvalcore.preprocessor.multi_model_statistics()
.
When calling the module inside diagnostic scripts, the input must be given as a list of cubes. The output will be saved in a dictionary where each entry contains the resulting cube with the requested statistic operations.
from esmvalcore.preprocessor import multi_model_statistics
statistics = multi_model_statistics([cube1,...,cubeN], 'overlap', ['mean', 'median'])
mean_cube = statistics['mean']
median_cube = statistics['median']
Note
The multimodel array operations can be rather memoryintensive (since they are not performed lazily as yet). The Section on Information on maximum memory required details the memory intake for different run scenarios, but as a thumb rule, for the multimodel preprocessor, the expected maximum memory intake could be approximated as the number of datasets multiplied by the average size in memory for one dataset.
Time manipulation¶
The _time.py
module contains the following preprocessor functions:
extract_time: Extract a time range from a cube.
extract_season: Extract only the times that occur within a specific season.
extract_month: Extract only the times that occur within a specific month.
hourly_statistics: Compute intraday statistics
daily_statistics: Compute statistics for each day
monthly_statistics: Compute statistics for each month
seasonal_statistics: Compute statistics for each season
annual_statistics: Compute statistics for each year
decadal_statistics: Compute statistics for each decade
climate_statistics: Compute statistics for the full period
resample_time: Resample data
resample_hours: Convert between Nhourly frequencies by resampling
anomalies: Compute (standardized) anomalies
regrid_time: Aligns the time axis of each dataset to have common time points and calendars.
timeseries_filter: Allows application of a filter to the timeseries data.
Statistics functions are applied by default in the order they appear in the list. For example, the following example applied to hourly data will retrieve the minimum values for the full period (by season) of the monthly mean of the daily maximum of any given variable.
daily_statistics:
operator: max
monthly_statistics:
operator: mean
climate_statistics:
operator: min
period: season
extract_time
¶
This function subsets a dataset between two points in times. It removes all times in the dataset before the first time and after the last time point. The required arguments are relatively self explanatory:
start_year
start_month
start_day
end_year
end_month
end_day
These start and end points are set using the datasets native calendar. All six arguments should be given as integers  the named month string will not be accepted.
See also esmvalcore.preprocessor.extract_time()
.
extract_season
¶
Extract only the times that occur within a specific season.
This function only has one argument: season
. This is the named season to
extract, i.e. DJF, MAM, JJA, SON, but also all other sequentially correct
combinations, e.g. JJAS.
Note that this function does not change the time resolution. If your original data is in monthly time resolution, then this function will return three monthly datapoints per year.
If you want the seasonal average, then this function needs to be combined with the seasonal_mean function, below.
See also esmvalcore.preprocessor.extract_season()
.
extract_month
¶
The function extracts the times that occur within a specific month.
This function only has one argument: month
. This value should be an integer
between 1 and 12 as the named month string will not be accepted.
See also esmvalcore.preprocessor.extract_month()
.
hourly_statistics
¶
This function produces statistics at a xhourly frequency.
 Parameters:
every_n_hours: frequency to use to compute the statistics. Must be a divisor of 24.
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’ and ‘sum’. Default is ‘mean’
daily_statistics
¶
This function produces statistics for each day in the dataset.
 Parameters:
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’, ‘sum’ and ‘rms’. Default is ‘mean’
monthly_statistics
¶
This function produces statistics for each month in the dataset.
 Parameters:
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’, ‘sum’ and ‘rms’. Default is ‘mean’
seasonal_statistics
¶
This function produces statistics for each season (default: [DJF, MAM, JJA,
SON]
or custom seasons e.g. [JJAS, ONDJFMAM]
) in the dataset. Note that
this function will not check for missing time points. For instance, if you are
looking at the DJF field, but your datasets starts on January 1st, the first
DJF field will only contain data from January and February.
We recommend using the extract_time to start the dataset from the following December and remove such biased initial datapoints.
 Parameters:
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’, ‘sum’ and ‘rms’. Default is ‘mean’
seasons: seasons to build statistics. Default is ‘[DJF, MAM, JJA, SON]’
annual_statistics
¶
This function produces statistics for each year.
 Parameters:
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’, ‘sum’ and ‘rms’. Default is ‘mean’
decadal_statistics
¶
This function produces statistics for each decade.
 Parameters:
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’, ‘sum’ and ‘rms’. Default is ‘mean’
climate_statistics
¶
This function produces statistics for the whole dataset. It can produce scalars (if the full period is chosen) or daily, monthly or seasonal statistics.
 Parameters:
operator: operation to apply. Accepted values are ‘mean’, ‘median’, ‘std_dev’, ‘min’, ‘max’, ‘sum’ and ‘rms’. Default is ‘mean’
period: define the granularity of the statistics: get values for the full period, for each month or day of year. Available periods: ‘full’, ‘season’, ‘seasonal’, ‘monthly’, ‘month’, ‘mon’, ‘daily’, ‘day’. Default is ‘full’
seasons: if period ‘seasonal’ or ‘season’ allows to set custom seasons. Default is ‘[DJF, MAM, JJA, SON]’
 Examples:
Monthly climatology:
climate_statistics: operator: mean period: month
Daily maximum for the full period:
climate_statistics: operator: max period: day
Minimum value in the period:
climate_statistics: operator: min period: full
resample_time
¶
This function changes the frequency of the data in the cube by extracting the timesteps that meet the criteria. It is important to note that it is mainly meant to be used with instantaneous data.
 Parameters:
month: Extract only timesteps from the given month or do nothing if None. Default is None
day: Extract only timesteps from the given day of month or do nothing if None. Default is None
hour: Extract only timesteps from the given hour or do nothing if None. Default is None
 Examples:
Hourly data to daily:
resample_time: hour: 12
Hourly data to monthly:
resample_time: hour: 12 day: 15
Daily data to monthly:
resample_time: day: 15
See also esmvalcore.preprocessor.resample_time()
.
resample_hours:
resample_hours
¶
This function changes the frequency of the data in the cube by extracting the timesteps that belongs to the desired frequency. It is important to note that it is mainly mean to be used with instantaneous data
 Parameters:
interval: New frequency of the data. Must be a divisor of 24
offset: First desired hour. Default 0. Must be lower than the interval
 Examples:
Convert to 12hourly, by getting timesteps at 0:00 and 12:00:
resample_hours: hours: 12
Convert to 12hourly, by getting timesteps at 6:00 and 18:00:
resample_hours: hours: 12 offset: 6
See also esmvalcore.preprocessor.resample_hours()
.
anomalies
¶
This function computes the anomalies for the whole dataset. It can compute anomalies from the full, seasonal, monthly and daily climatologies. Optionally standardized anomalies can be calculated.
 Parameters:
period: define the granularity of the climatology to use: full period, seasonal, monthly or daily. Available periods: ‘full’, ‘season’, ‘seasonal’, ‘monthly’, ‘month’, ‘mon’, ‘daily’, ‘day’. Default is ‘full’
reference: Time slice to use as the reference to compute the climatology on. Can be ‘null’ to use the full cube or a dictionary with the parameters from extract_time. Default is null
standardize: if true calculate standardized anomalies (default: false)
seasons: if period ‘seasonal’ or ‘season’ allows to set custom seasons. Default is ‘[DJF, MAM, JJA, SON]’
 Examples:
Anomalies from the full period climatology:
anomalies:
Anomalies from the full period monthly climatology:
anomalies: period: month
Standardized anomalies from the full period climatology:
anomalies: standardized: true
Standardized Anomalies from the 19792000 monthly climatology:
anomalies: period: month reference: start_year: 1979 start_month: 1 start_day: 1 end_year: 2000 end_month: 12 end_day: 31 standardize: true
See also esmvalcore.preprocessor.anomalies()
.
regrid_time
¶
This function aligns the time points of each component dataset so that the Iris
cubes from different datasets can be subtracted. The operation makes the
datasets time points common; it also resets the time
bounds and auxiliary coordinates to reflect the artificially shifted time
points. Current implementation for monthly and daily data; the frequency
is
set automatically from the variable CMOR table unless a custom frequency
is
set manually by the user in recipe.
See also esmvalcore.preprocessor.regrid_time()
.
timeseries_filter
¶
This function allows the user to apply a filter to the timeseries data. This filter may be
of the user’s choice (currently only the lowpass
Lanczos filter is implemented); the
implementation is inspired by this iris example and uses aggregation via iris.cube.Cube.rolling_window
.
 Parameters:
window: the length of the filter window (in units of cube time coordinate).
span: period (number of months/days, depending on data frequency) on which weights should be computed e.g. for 2yearly: span = 24 (2 x 12 months). Make sure span has the same units as the data cube time coordinate.
filter_type: the type of filter to be applied; default ‘lowpass’. Available types: ‘lowpass’.
filter_stats: the type of statistic to aggregate on the rolling window; default ‘sum’. Available operators: ‘mean’, ‘median’, ‘std_dev’, ‘sum’, ‘min’, ‘max’, ‘rms’.
 Examples:
Lowpass filter with a monthly mean as operator:
timeseries_filter: window: 3 # 3monthly filter window span: 12 # weights computed on the first year filter_type: lowpass # lowpass filter filter_stats: mean # 3monthly mean lowpass filter
Area manipulation¶
The area manipulation module contains the following preprocessor functions:
extract_region: Extract a region from a cube based on
lat/lon
corners.extract_named_regions: Extract a specific region from in the region coordinate.
extract_shape: Extract a region defined by a shapefile.
extract_point: Extract a single point (with interpolation)
zonal_statistics: Compute zonal statistics.
meridional_statistics: Compute meridional statistics.
area_statistics: Compute area statistics.
extract_region
¶
This function returns a subset of the data on the rectangular region requested. The boundaries of the region are provided as latitude and longitude coordinates in the arguments:
start_longitude
end_longitude
start_latitude
end_latitude
Note that this function can only be used to extract a rectangular region. Use
extract_shape
to extract any other shaped region from a shapefile.
If the grid is irregular, the returned region retains the original coordinates, but is cropped to a rectangular bounding box defined by the start/end coordinates. The deselected area inside the region is masked.
See also esmvalcore.preprocessor.extract_region()
.
extract_named_regions
¶
This function extracts a specific named region from the data. This function
takes the following argument: regions
which is either a string or a list
of strings of named regions. Note that the dataset must have a region
coordinate which includes a list of strings as values. This function then
matches the named regions against the requested string.
extract_shape
¶
Extract a shape or a representative point for this shape from the data.
 Parameters:
shapefile
: path to the shapefile containing the geometry of the region to be extracted. If the file contains multiple shapes behaviour depends on the decomposed parameter. This path can be relative toauxiliary_data_dir
defined in the User configuration file.method
: the method to select the region, selecting either all pointscontained by the shape or a single representative point. Choose either ‘contains’ or ‘representative’. If not a single grid point is contained in the shape, a representative point will be selected.
crop
: by default extract_region will be used to crop the data to aminimal rectangular region containing the shape. Set to
false
to only mask data outside the shape. Data on irregular grids will not be cropped.
decomposed
: by defaultfalse
, in this case the union of all the regions in the shape file is masked out. Iftrue
, the regions in the shapefiles are masked out separately, generating an auxiliary dimension for the cube for this.ids
: by default,[]
, in this case all the shapes in the file will be used. If a list of IDs is provided, only the shapes matching them will be used. The IDs are assigned from thename
orid
attributes (in that order of priority) if present in the file or from the reading order if otherwise not present. So, for example, if a file has both`name
andid
attributes, the ids will be assigned fromname
. If the file only has theid
attribute, it will be taken from it and if noname
norid
attributes are present, an integer id starting from 1 will be assigned automatically when reading the shapes. We discourage to rely on this last behaviour as we can not assure that the reading order will be the same in different platforms, so we encourage you to modify the file to add a proper id attribute. If the file has an id attribute with a name that is not supported, please open an issue so we can add support for it.
 Examples:
Extract the shape of the river Elbe from a shapefile:
extract_shape: shapefile: Elbe.shp method: contains
Extract the shape of several countries:
extract_shape: shapefile: NaturalEarth/Countries/ne_110m_admin_0_countries.shp decomposed: True method: contains ids:  Spain  France  Italy  United Kingdom  Taiwan
See also esmvalcore.preprocessor.extract_shape()
.
extract_point
¶
Extract a single point from the data. This is done using either nearest or linear interpolation.
Returns a cube with the extracted point(s), and with adjusted latitude and longitude coordinates (see below).
Multiple points can also be extracted, by supplying an array of latitude and/or longitude coordinates. The resulting point cube will match the respective latitude and longitude coordinate to those of the input coordinates. If the input coordinate is a scalar, the dimension will be missing in the output cube (that is, it will be a scalar).
 Parameters:
cube
: the input dataset cube.latitude
,longitude
: coordinates (as floating point values) of the point to be extracted. Either (or both) can also be an array of floating point values.scheme
: interpolation scheme: either'linear'
or'nearest'
. There is no default.
zonal_statistics
¶
The function calculates the zonal statistics by applying an operator along the longitude coordinate. This function takes one argument:
operator
: Which operation to apply: mean, std_dev, median, min, max, sum or rms.
See also esmvalcore.preprocessor.zonal_means()
.
meridional_statistics
¶
The function calculates the meridional statistics by applying an operator along the latitude coordinate. This function takes one argument:
operator
: Which operation to apply: mean, std_dev, median, min, max, sum or rms.
See also esmvalcore.preprocessor.meridional_means()
.
area_statistics
¶
This function calculates the average value over a region  weighted by the cell
areas of the region. This function takes the argument, operator
: the name
of the operation to apply.
This function can be used to apply several different operations in the horizontal plane: mean, standard deviation, median, variance, minimum, maximum and root mean square.
Note that this function is applied over the entire dataset. If only a specific region, depth layer or time period is required, then those regions need to be removed using other preprocessor operations in advance.
The optional fx_variables
argument specifies the fx variables that the user
wishes to input to the function. More details on this are given in Fx variables as cell measures or ancillary variables.
Volume manipulation¶
The _volume.py
module contains the following preprocessor functions:
extract_volume
: Extract a specific depth range from a cube.volume_statistics
: Calculate the volumeweighted average.depth_integration
: Integrate over the depth dimension.extract_transect
: Extract data along a line of constant latitude or longitude.extract_trajectory
: Extract data along a specified trajectory.
extract_volume
¶
Extract a specific range in the zdirection from a cube. This function
takes two arguments, a minimum and a maximum (z_min
and z_max
,
respectively) in the zdirection.
Note that this requires the requested zcoordinate range to be the same sign
as the Iris cube. That is, if the cube has zcoordinate as negative, then
z_min
and z_max
need to be negative numbers.
See also esmvalcore.preprocessor.extract_volume()
.
volume_statistics
¶
This function calculates the volumeweighted average across three dimensions, but maintains the time dimension.
This function takes the argument: operator
, which defines the operation to
apply over the volume.
No depth coordinate is required as this is determined by Iris. This function
works best when the fx_variables
provide the cell volume. The optional
fx_variables
argument specifies the fx variables that the user wishes to
input to the function. More details on this are given in Fx variables as cell measures or ancillary variables.
depth_integration
¶
This function integrates over the depth dimension. This function does a weighted sum along the zcoordinate, and removes the z direction of the output cube. This preprocessor takes no arguments.
extract_transect
¶
This function extracts data along a line of constant latitude or longitude.
This function takes two arguments, although only one is strictly required.
The two arguments are latitude
and longitude
. One of these arguments
needs to be set to a float, and the other can then be either ignored or set to
a minimum or maximum value.
For example, if we set latitude to 0 N and leave longitude blank, it would
produce a cube along the Equator. On the other hand, if we set latitude to 0
and then set longitude to [40., 100.]
this will produce a transect of the
Equator in the Indian Ocean.
extract_trajectory
¶
This function extract data along a specified trajectory.
The three arguments are: latitudes
, longitudes
and number of point
needed for extrapolation number_points
.
If two points are provided, the number_points
argument is used to set a
the number of places to extract between the two end points.
If more than two points are provided, then extract_trajectory
will produce
a cube which has extrapolated the data of the cube to those points, and
number_points
is not needed.
Note that this function uses the expensive interpolate
method from
Iris.analysis.trajectory
, but it may be necessary for irregular grids.
Cycles¶
The _cycles.py
module contains the following preprocessor functions:
amplitude
: Extract the peaktopeak amplitude of a cycle aggregated over specified coordinates.
amplitude
¶
This function extracts the peaktopeak amplitude (maximum value minus minimum
value) of a field aggregated over specified coordinates. Its only argument is
coords
, which can either be a single coordinate (given as str
) or
multiple coordinates (given as list
of str
). Usually, these
coordinates refer to temporal categorised coordinates
iris.coord_categorisation
like year, month, day of year, etc. For example, to extract the amplitude
of the annual cycle for every single year in the data, use coords: year
; to
extract the amplitude of the diurnal cycle for every single day in the data,
use coords: [year, day_of_year]
.
See also esmvalcore.preprocessor.amplitude()
.
Trend¶
The trend module contains the following preprocessor functions:
linear_trend
: Calculate linear trend along a specified coordinate.linear_trend_stderr
: Calculate standard error of linear trend along a specified coordinate.
linear_trend
¶
This function calculates the linear trend of a dataset (defined as slope of an
ordinary linear regression) along a specified coordinate. The only argument of
this preprocessor is coordinate
(given as str
; default value is
'time'
).
See also esmvalcore.preprocessor.linear_trend()
.
linear_trend_stderr
¶
This function calculates the standard error of the linear trend of a dataset
(defined as the standard error of the slope in an ordinary linear regression)
along a specified coordinate. The only argument of this preprocessor is
coordinate
(given as str
; default value is 'time'
). Note that
the standard error is not identical to a confidence interval.
Detrend¶
ESMValTool also supports detrending along any dimension using the preprocessor function ‘detrend’. This function has two parameters:
dimension
: dimension to apply detrend on. Default: “time”method
: It can belinear
orconstant
. Default:linear
If method is linear
, detrend will calculate the linear trend along the
selected axis and subtract it to the data. For example, this can be used to
remove the linear trend caused by climate change on some variables is selected
dimension is time.
If method is constant
, detrend will compute the mean along that dimension
and subtract it from the data
See also esmvalcore.preprocessor.detrend()
.
Unit conversion¶
Converting units is also supported. This is particularly useful in cases where different datasets might have different units, for example when comparing CMIP5 and CMIP6 variables where the units have changed or in case of observational datasets that are delivered in different units.
In these cases, having a unit conversion at the end of the processing will guarantee homogeneous input for the diagnostics.
Note
Conversion is only supported between compatible units! In other
words, converting temperature units from degC
to Kelvin
works
fine, changing precipitation units from a rate based unit to an
amount based unit is not supported at the moment.
See also esmvalcore.preprocessor.convert_units()
.
Bias¶
The bias module contains the following preprocessor functions:
bias
: Calculate absolute or relative biases with respect to a reference dataset
bias
¶
This function calculates biases with respect to a given reference dataset. For
this, exactly one input dataset needs to be declared as reference_for_bias:
true
in the recipe, e.g.,
datasets:
 {dataset: CanESM5, project: CMIP6, ensemble: r1i1p1f1, grid: gn}
 {dataset: CESM2, project: CMIP6, ensemble: r1i1p1f1, grid: gn}
 {dataset: MIROC6, project: CMIP6, ensemble: r1i1p1f1, grid: gn}
 {dataset: ERAInterim, project: OBS6, tier: 3, type: reanaly, version: 1,
reference_for_bias: true}
In the example above, ERAInterim is used as reference dataset for the bias
calculation. For this preprocessor, all input datasets need to have identical
dimensional coordinates. This can for example be ensured with the preprocessors
esmvalcore.preprocessor.regrid()
and/or
esmvalcore.preprocessor.regrid_time()
.
The bias
preprocessor supports 4 optional arguments:
bias_type
(str
, default:'absolute'
): Bias type that is calculated. Can be'absolute'
(i.e., calculate bias for dataset \(X\) and reference \(R\) as \(X  R\)) orrelative
(i.e, calculate bias as \(\frac{X  R}{R}\)).
denominator_mask_threshold
(float
, default:1e3
): Threshold to mask values close to zero in the denominator (i.e., the reference dataset) during the calculation of relative biases. All values in the reference dataset with absolute value less than the given threshold are masked out. This setting is ignored whenbias_type
is set to'absolute'
. Please note that for some variables with very small absolute values (e.g., carbon cycle fluxes, which are usually \(< 10^{6}\) kg m \(^{2}\) s \(^{1}\)) it is absolutely essential to change the default value in order to get reasonable results.
keep_reference_dataset
(bool
, default:False
): IfTrue
, keep the reference dataset in the output. IfFalse
, drop the reference dataset.
exclude
(list
ofstr
): Exclude specific datasets from this preprocessor. Note that this option is only available in the recipe, not when usingesmvalcore.preprocessor.bias()
directly (e.g., in another python script). If the reference dataset has been excluded, an error is raised.
Example:
preprocessors:
preproc_bias:
bias:
bias_type: relative
denominator_mask_threshold: 1e8
keep_reference_dataset: true
exclude: [CanESM2]
See also esmvalcore.preprocessor.bias()
.
Information on maximum memory required¶
In the most general case, we can set upper limits on the maximum memory the analysis will require:
Ms = (R + N) x F_eff  F_eff
 when no multimodel analysis is performed;
Mm = (2R + N) x F_eff  2F_eff
 when multimodel analysis is performed;
where
Ms
: maximum memory for nonmultimodel moduleMm
: maximum memory for multimodel moduleR
: computational efficiency of module; R is typically 23N
: number of datasetsF_eff
: average size of data per dataset whereF_eff = e x f x F
wheree
is the factor that describes how lazy the data is (e = 1
for fully realized data) andf
describes how much the data was shrunk by the immediately previous module, e.g. time extraction, area selection or level extraction; note that for fix_dataf
relates only to the time extraction, if data is exact in time (no time selection)f = 1
for fix_data so for cases when we deal with a lot of datasetsR + N \approx N
, data is fully realized, assuming an average size of 1.5GB for 10 years of 3D netCDF data,N
datasets will require:
Ms = 1.5 x (N  1)
GB
Mm = 1.5 x (N  2)
GB
As a rule of thumb, the maximum required memory at a certain time for
multimodel analysis could be estimated by multiplying the number of datasets by
the average file size of all the datasets; this memory intake is high but also
assumes that all data is fully realized in memory; this aspect will gradually
change and the amount of realized data will decrease with the increase of
dask
use.
Other¶
Miscellaneous functions that do not belong to any of the other categories.
Clip¶
This function clips data values to a certain minimum, maximum or range. The function takes two arguments:
minimum
: Lower bound of range. Default:None
maximum
: Upper bound of range. Default:None
The example below shows how to set all values below zero to zero.
preprocessors:
clip:
minimum: 0
maximum: null