Skip to contents

Fit a bias correction for ERA5-Land meteorology from local observations

Source: R/fit_met_bias_correction.R
fit_met_bias_correction.Rd

Joins an ERA5 met frame (from extract_era5_hourly_met() / extract_era5_lake_met()) to a matching frame of locally measured meteorology (from prepare_obs_met()) over their common period and fits, per variable, a transfer function that maps ERA5 onto the observations. The fitted object is applied to the full ERA5 record with apply_met_bias_correction().

Usage

fit_met_bias_correction(
  era5,
  obs,
  vars = NULL,
  method = "scale",
  transform = "auto",
  by = c("doy-loess", "month", "none"),
  cv = c("loyo", "none"),
  min_n = 60,
  cv_min_months = 8,
  loess_span = 0.25,
  robust = FALSE,
  nquantiles = 50,
  swr_daytime_wm2 = 5,
  verbose = TRUE
)

Arguments

era5

data frame with Date plus MET_* columns (extractor output).

obs

data frame with Date plus MET_* columns (prepare_obs_met() output). Must be the same resolution and time zone as era5.

vars

character vector of MET_* columns to correct. NULL (default) uses the columns common to both, excluding derived/direction fields (MET_wnddir, MET_wnduvu, MET_wnduvv, MET_cldcvr, MET_prvapr, MET_prmslp).

method

"scale" (default), "variance", "linear", "eqm" or "qdm"; or a named vector, e.g. c(MET_wndspd = "eqm", MET_tmpair = "scale").

transform

"auto" (ratio for MET_wndspd/MET_radswd/ MET_pprain/MET_ppsnow, additive otherwise), "additive" or "ratio"; or a named vector. Only used by method = "scale".

by

"doy-loess" (default), "month" or "none". Coerced to "month" for methods other than "scale".

cv

"loyo" (leave-one-year-out, default) or "none" - controls whether the skill table includes out-of-sample rows.

min_n

minimum paired records per monthly stratum (default 60).

cv_min_months

a year is usable as a leave-one-out fold only if it has at least this many months of overlap (default 8).

loess_span

span for the day-of-year loess (default 0.25).

robust

use MASS::rlm for method = "linear" (default FALSE).

nquantiles

number of quantiles for "eqm"/"qdm" (default 50).

swr_daytime_wm2

ERA5 shortwave threshold (W/m2) above which a record counts as daytime for the shortwave fit (default 5).

verbose

print progress and the skill table.

Value

an object of class met_biascorr: a list with models (per variable), skill (data frame: variable, stage in raw/corrected/cv_raw/cv_corrected, n, bias, mae, rmse, r, kge), training (window, resolution, tz, years, station) and call.

Details

Methods (set globally via method, or per variable with a named vector):

"scale"

(default) a per-month offset - additive (obs - era5) or multiplicative (sum(obs) / sum(era5)), chosen by transform. With by = "doy-loess" the 12 monthly offsets are loess-smoothed across the day-of-year (as nz_climate_projections/climate_shift.R).

"variance"

per-month mean and standard-deviation matching: mean_o + (x - mean_e) * sd_o / sd_e.

"linear"

per-month least squares obs ~ a + b * era5 (robust = TRUE uses MASS::rlm).

"eqm"

empirical quantile mapping per month - remap the ERA5 CDF onto the observed CDF, linear interpolation, constant-tail extrapolation.

"qdm"

quantile delta mapping - as "eqm" but preserves the ERA5 anomaly relative to its training quantile, so trends/among-year signal survive extrapolation.

Any stratum with fewer than min_n pairs falls back to the pooled (all-month) fit. Shortwave is fitted on daytime records only.

See also

prepare_obs_met(), apply_met_bias_correction()