def compute_weighted_crps(da_fc, da_true, mean_dims=xr.ALL_DIMS):
da_true = da_true.sel(time=da_fc.time)
assert (da_true.time == da_fc.time).all #checking size.
weights_lat = np.cos(np.deg2rad(da_fc.lat))
weights_lat /= weights_lat.mean()
crps = xs.crps_ensemble(da_true, da_fc)
crps = (crps * weights_lat).mean(mean_dims)
return crps
def _crps(forecast, reference, dim='svd', comparison=None):
"""
Continuous Ranked Probability Score (CRPS) is the probabilistic MSE.
Range:
* perfect: 0
* max: 0
* else: negative
References:
* Matheson, James E., and Robert L. Winkler. “Scoring Rules for
Continuous Probability Distributions.” Management Science 22, no. 10
(June 1, 1976): 1087–96. https://doi.org/10/cwwt4g.
See also:
* properscoring.crps_ensemble
"""
return crps_ensemble(forecast, reference).mean(dim)
def _crps(forecast, reference, **metric_kwargs):
"""
Continuous Ranked Probability Score (CRPS) is the probabilistic MSE.
Range:
* perfect: 0
* min: 0
* max: ∞
References:
* Matheson, James E., and Robert L. Winkler. “Scoring Rules for
Continuous Probability Distributions.” Management Science 22, no. 10
(June 1, 1976): 1087–96. https://doi.org/10/cwwt4g.
See also:
* properscoring.crps_ensemble
* xskillscore.crps_ensemble
"""
# switch positions because xskillscore.crps_ensemble(obs, forecasts)
return crps_ensemble(reference, forecast)
coords=[np.arange(ensemble_len), lat_target, lon_target],
dims=["member", "lat", "lon"],
name='var')
year_sel = year - 2001
year_sel = int(year_sel)
ensembleindx = ensembleindexjan
ensemblerows = np.delete(ensembleindx, year_sel, 0)
ensemble = DS_target['precipitationCal'][ensemblerows.flatten(), :, :]
fct[:, :, :] = ensemble
DS_obs_sel = DS_target.sel(time=obs_day[0])
obs[:, :] = DS_obs_sel["precipitationCal"][:, :]
crps0 = xs.crps_ensemble(obs, fct, dim=[])
for day in range(1, 365):
if day < 59:
ensembleindx = ensembleindexjan + day
ensemblerows = np.delete(ensembleindx, year_sel, 0)
else:
ensembleindx = ensembleindexrest + day
ensemblerows = np.delete(ensembleindx, year_sel, 0)
ensemble = DS_target['precipitationCal'][ensemblerows.flatten(), :, :]
fct[:, :, :] = ensemble
DS_obs_sel = DS_target.sel(time=obs_day[day])
def crpss(o, f_prob, dim=None):
return 1 - xs.crps_ensemble(o, f_prob, dim=dim) / xs.crps_gaussian(
o, o.mean("time"), o.std("time"), dim=[])
dims=["time","lat", "lon"],
name='var'
)
fct = xr.DataArray(
np.random.rand(365, len(lat_target ), len(lon_target ), ensemble_len),
coords=[DS_target.time.values, lat_target , lon_target, np.arange(ensemble_len)],
dims=["time", "lat", "lon", "member"],
name='var'
)
obs[:,:,:] = DS_target["precipitationCal"].values
ensemble_mean = ensemble_ecmwf.mean(dim = 'number', skipna=True)
indices = np.where(np.isnan(ensemble_ecmwf['tp']))
#print('run imputation')
for i in range(len(indices[0])):
ensemble_ecmwf.tp[indices[0][i], indices[1][i],indices[2][i],indices[3][i]] = ensemble_mean.tp[indices[0][i], indices[1][i],indices[2][i]].values
fct[:,:,:,:] = ensemble_ecmwf['tp'].values
# crps averaged over time
crps = xs.crps_ensemble(obs, fct, dim=['time'])
crps.to_netcdf("ecmwf_crps_"+str(year)+".nc")