climpred.classes.HindcastEnsemble.remove_bias

HindcastEnsemble.remove_bias(alignment=None, how='additive_mean', train_test_split='unfair', train_init=None, train_time=None, cv=False, **metric_kwargs)

Calculate and remove bias from HindcastEnsemble. Bias is grouped by seasonality set via set_options. When wrapping xclim.sbda.adjustment use group instead.

Parameters

• alignment (str) –

  which inits or verification times should be aligned?

  • ’maximize’: maximize the degrees of freedom by slicing hind and verif to a common time frame at each lead.

  • ’same_inits’: slice to a common init frame prior to computing metric. This philosophy follows the thought that each lead should be based on the same set of initializations.

  • ’same_verif’: slice to a common/consistent verification time frame prior to computing metric. This philosophy follows the thought that each lead should be based on the same set of verification dates.

• how (str) –

  what kind of bias removal to perform. Defaults to ‘additive_mean’. Select from:

  • ’additive_mean’: correcting the mean forecast additively

  • ’multiplicative_mean’: correcting the mean forecast multiplicatively

  • ’multiplicative_std’: correcting the standard deviation multiplicatively

  • ’modified_quantile’: Reference

  • ’basic_quantile’: Reference

  • ’gamma_mapping’: Reference

  • ’normal_mapping’: Reference

  • ’EmpiricalQuantileMapping’: Reference

  • ’DetrendedQuantileMapping’: Reference

  • ’PrincipalComponents’: Reference

  • ’QuantileDeltaMapping’: Reference

  • ’Scaling’: Reference

  • ’LOCI’: Reference

• train_test_split (str) –

  How to separate train period to calculate the bias and test period to apply bias correction to? For a detailed description, see Risbey et al. 2021:

  • fair: no overlap between train and test (recommended).

    Set either train_init or train_time.

  • unfair: completely overlapping train and test

    (climpred default).

  • unfair-cv: overlapping train and test except for current

    init, which is left out (set cv=’LOO’).

• train_init (xr.DataArray, slice) – Define initializations for training when alignment='same_inits/maximize'.

• train_time (xr.DataArray, slice) – Define time for training when alignment='same_verif'.

• cv (bool or str) –

  Only relevant when train_test_split=’unfair-cv’. Defaults to False.

  • True/’LOO’: Calculate bias by leaving given initialization out

  • False: include all initializations in the calculation of bias, which

    is much faster and but yields similar skill with a large N of initializations.

• **metric_kwargs (dict) – passed to xclim.sdba (including group) or XBias_Correction

Returns

bias removed HindcastEnsemble.

Return type

HindcastEnsemble

Example

Skill from raw model output without bias reduction:

>>> HindcastEnsemble.verify(metric='rmse', comparison='e2o',
...     alignment='maximize', dim='init')
<xarray.Dataset>
Dimensions:  (lead: 10)
Coordinates:
  * lead     (lead) int32 1 2 3 4 5 6 7 8 9 10
    skill    <U11 'initialized'
Data variables:
    SST      (lead) float64 0.08359 0.08141 0.08362 ... 0.1361 0.1552 0.1664

Note that this HindcastEnsemble is already bias reduced, therefore train_test_split='unfair' has hardly any effect. Use all initializations to calculate bias and verify skill:

>>> HindcastEnsemble.remove_bias(alignment='maximize',
...     how='additive_mean', test_train_split='unfair'
... ).verify(metric='rmse', comparison='e2o', alignment='maximize',
... dim='init')
<xarray.Dataset>
Dimensions:  (lead: 10)
Coordinates:
  * lead     (lead) int32 1 2 3 4 5 6 7 8 9 10
    skill    <U11 'initialized'
Data variables:
    SST      (lead) float64 0.08349 0.08039 0.07522 ... 0.07305 0.08107 0.08255

Separate initializations 1954 - 1980 to calculate bias. Note that this HindcastEnsemble is already bias reduced, therefore train_test_split='fair' worsens skill here. Generally, train_test_split='fair' is recommended to use for a fair comparison against real-time forecasts.

>>> HindcastEnsemble.remove_bias(alignment='maximize',
...     how='additive_mean', train_test_split='fair',
...     train_init=slice('1954', '1980')).verify(metric='rmse',
...     comparison='e2o', alignment='maximize', dim='init')
<xarray.Dataset>
Dimensions:  (lead: 10)
Coordinates:
  * lead     (lead) int32 1 2 3 4 5 6 7 8 9 10
    skill    <U11 'initialized'
Data variables:
    SST      (lead) float64 0.132 0.1085 0.08722 ... 0.08209 0.08969 0.08732

Wrapping methods how from xclim and providing group for groupby:

>>> HindcastEnsemble.remove_bias(alignment='same_init', group='init',
...     how='DetrendedQuantileMapping', train_test_split='unfair',
...     ).verify(metric='rmse',
...     comparison='e2o', alignment='maximize', dim='init')
<xarray.Dataset>
Dimensions:  (lead: 10)
Coordinates:
  * lead     (lead) int32 1 2 3 4 5 6 7 8 9 10
    skill    <U11 'initialized'
Data variables:
    SST      (lead) float64 0.09823 0.09747 0.08235 ... 0.07742 0.08115 0.08326

Wrapping methods how from bias_correction:

>>> HindcastEnsemble.remove_bias(alignment='same_init',
...     how='modified_quantile', train_test_split='unfair',
...     ).verify(metric='rmse',
...     comparison='e2o', alignment='maximize', dim='init')
<xarray.Dataset>
Dimensions:  (lead: 10)
Coordinates:
  * lead     (lead) int32 1 2 3 4 5 6 7 8 9 10
    skill    <U11 'initialized'
Data variables:
    SST      (lead) float64 0.07628 0.08293 0.08169 ... 0.1577 0.1821 0.2087