import bmorph
import pandas as pd
import numpy as np
import xarray as xr
from functools import partial
from tqdm.autonotebook import tqdm
from bmorph.util import mizuroute_utils as mizutil
import os
[docs]def apply_bmorph(raw_ts, train_ts, ref_ts,
apply_window, raw_train_window, ref_train_window,
condition_ts=None,
raw_y=None, train_y=None, ref_y=None,
interval=pd.DateOffset(years=1),
overlap=60, n_smooth_long=None, n_smooth_short=5,
bw=3, xbins=200, ybins=10,
rtol=1e-6, atol=1e-8, method='hist', train_cdf_min=1e-4, **kwargs):
"""Bias correction is performed by bmorph on user-defined intervals.
Parameters
----------
raw_ts : pandas.Series
Raw flow timeseries.
train_ts : pandas.Series
Flow timeseries to train the bias correction model with.
ref_ts : pandas.Series
Observed/reference flow timeseries.
condition_ts: pandas.Series
A timeseries with a variable to condition on. This will be
used in place of `raw_y`, `train_y`, and `ref_y`. It is
mainly added as a convenience over specifying the same
timeseries for each of those variables.
apply_window : pandas.date_range
Date range to apply bmorph onto flow timeseries.
raw_train_window : pandas.date_range
Date range to train the bias correction model.
ref_train_window : pandas.date_range
Date range to smooth elements in 'raw_ts' and 'bmorph_ts'.
raw_y : pandas.Series, optional
Raw time series of the second time series variable for conditioning.
train_y : pandas.Series, optional
Training second time series.
ref_y : pandas.Series, optional
Target second time series.
interval : pandas.DateOffset
Difference between bmorph application intervals.
overlap : int
Total number of days overlap CDF windows have with each other,
distributed evenly before and after the application window.
n_smooth_long : int, optional
Number of elements that will be smoothed in `raw_ts` and `bmorph_ts`.
The nsmooth value in this case is typically much larger than the one
used for the bmorph function itself. For example, 365 days.
n_smooth_short : int, optional
Number of elements that will be smoothed when determining CDFs used
for the bmorph function itself.
bw : int, optional
Bandwidth for KernelDensity. This should only be used if `method='kde'`.
xbins : int, optional
Bins for the first time series. This should only be used if `method='hist'`.
ybins : int, optional
Bins for the second time series. This should only be used if `method='hist'`.
rtol : float, optional
The desired relatie tolerance of the result for KernelDensity.
This should only be used if `method='kde'`.
atol : float, optional
The desired absolute tolerance of the result for KernelDensity.
This should only be used if `method='kde'`.
method : str
Method to use for conditioning. Currently 'hist' using hist2D and 'kde'
using kde2D are the only supported methods.
**kwargs:
Additional keyword arguments. Mainly implemented for cross-compatibility with
other methods so that a unified configuration can be used
Returns
-------
bmorph_corr_ts : pandas.Series
Returns a time series of length of an interval in the bmoprh window
with bmorphed values.
bmorph_multipliers : pandas.Series
Returns a time series of equal length to bc_totals used to scale the
raw flow values into the bmorphed values returned in bc_totals.
"""
assert isinstance(interval, pd.DateOffset)
if condition_ts is not None:
raw_y = condition_ts
ref_y = condition_ts
train_y = condition_ts
if interval == pd.DateOffset(days=1):
raise Exception("Please enter an interval greater than 1 day(s)")
raw_train_window = slice(*raw_train_window)
apply_window = slice(*apply_window)
ref_train_window = slice(*ref_train_window)
raw_ts_window = slice(pd.to_datetime(raw_ts.index.values[0]),
pd.to_datetime(raw_ts.index.values[-1]))
overlap_period = int(overlap / 2)
bmorph_ts = pd.Series([])
bmorph_multipliers = pd.Series([])
bmorph_range = pd.date_range(apply_window.start,
apply_window.stop+interval,
freq=interval)
for i in range(0,len(bmorph_range)-1):
bmorph_start = bmorph_range[i]
bmorph_end = bmorph_range[i+1]
raw_apply_window = slice(pd.to_datetime(str(bmorph_start)),
pd.to_datetime(str(bmorph_end)))
raw_cdf_window = slice(
pd.to_datetime(str(bmorph_start - pd.DateOffset(days=overlap_period))),
pd.to_datetime(str(bmorph_end + pd.DateOffset(days=overlap_period))))
# No overlap for the first/last periods
if ((raw_cdf_window.start < raw_ts_window.start) or (raw_cdf_window.stop > raw_ts_window.stop)):
raw_cdf_window = slice(raw_ts_window.start, raw_ts_window.stop)
# Perform the bias correction
bc_total, bc_mult = bmorph.bmorph(
raw_ts, train_ts, ref_ts,
raw_apply_window, raw_train_window, ref_train_window, raw_cdf_window,
raw_y, ref_y, train_y,
n_smooth_short, bw=bw, xbins=xbins, ybins=ybins, rtol=rtol, atol=atol,
method=method, train_cdf_min=train_cdf_min)
bmorph_ts = bmorph_ts.append(bc_total)
bmorph_multipliers = bmorph_multipliers.append(bc_mult)
# Eliminate duplicate timestamps because of overlapping period
bmorph_ts = bmorph_ts.groupby(bmorph_ts.index).mean()
bmorph_multipliers = bmorph_multipliers.groupby(bmorph_multipliers.index).mean()
# Apply the correction to the change in mean
if n_smooth_long:
nrni_mean = ref_ts[ref_train_window].mean()
train_mean = train_ts[raw_train_window].mean()
bmorph_corr_ts, corr_ts = bmorph.bmorph_correct(raw_ts, bmorph_ts, raw_ts_window,
nrni_mean, train_mean, n_smooth_long)
else:
bmorph_corr_ts = bmorph_ts
return bmorph_corr_ts[apply_window], bmorph_multipliers[apply_window]
[docs]def apply_blendmorph(raw_upstream_ts, raw_downstream_ts,
train_upstream_ts, train_downstream_ts,
ref_upstream_ts, ref_downstream_ts,
apply_window, raw_train_window, ref_train_window,
interval, overlap, blend_factor,
raw_upstream_y=None, raw_downstream_y=None,
train_upstream_y=None, train_downstream_y=None,
ref_upstream_y=None, ref_downstream_y=None,
n_smooth_long=None, n_smooth_short=5,
bw=3, xbins=200, ybins=10, rtol=1e-6, atol=1e-8,
method='hist', train_cdf_min=1e-4, **kwargs):
"""Bias correction performed by blending bmorphed flows on user defined intervals.
Blendmorph is used to perform spatially consistent bias correction, this function
does so on a user-defined interval. This is done by performing bmorph bias correction
for each site's timeseries according to upstream and downstream gauge sites
(or proxies) where true flows are known. The upstream and downstream corrected
timeseries are then multiplied by fractional weights, `blend_factor`, that sum
to 1 between them so the corrected flows can be combined, or "blended," into one,
representative corrected flow series for the site. It is thereby important to specify
upstream and downstream values so bias corrections are performed with values that
most closely represent each site being corrected.
Parameters
---------
raw_upstream_ts : pandas.Series
Raw flow timeseries corresponding to the upstream flows.
raw_downstream_ts : pandas.Series
Raw flow timerseries corresponding to the downstream flows.
train_upstream_ts : pandas.Series
Flow timeseries to train the bias correction model with for the upstream flows.
train_downstream_ts : pandas.Series
Flow timeseries to train the bias correction model with for the downstream flows.
ref_upstream_ts : pandas.Series
Observed/reference flow timeseries corresponding to the upstream flows.
ref_downstream_ts : pandas.Series
Observed/reference flow timeseries corresponding to the downstream flows.
raw_train_window : pandas.date_range
Date range to train the bias correction model.
apply_window : pandas.date_range
Date range to apply bmorph onto flow timeseries.
ref_train_window : pandas.date_range
Date range to smooth elements in 'raw_ts' and 'bmorph_ts'.
interval : pandas.DateOffset
Difference between bmorph application intervals.
overlap: int
Total overlap in number of days the CDF windows have with each other,
distributed evenly before and after the application window.
blend_factor : numpy.array
An array determining how upstream and downstream bmorphing is proportioned.
This is determined by the fill_method used in mizuroute_utils. The blend_factor
entries are the proportion of upstream multiplers and totals added with
1-blend_factor of downstream multipliers and totals.
n_smooth_long : int, optional
Number of elements that will be smoothed in `raw_ts` and `bmorph_ts`.
The nsmooth value in this case is typically much larger than the one
used for the bmorph function itself. For example, 365 days.
n_smooth_short : int
Number of elements that will be smoothed when determining CDFs used
for the bmorph function itself.
raw_upstream_y : pandas.Series, optional
Raw time series of the second time series variable for conditioning corresponding
to upstream flows.
raw_downstream_y : pandas.Series, optional
Raw time series of the second time series variable for conditioning corresponding
to downstream flows.
train_upstream_y : pandas.Series, optional
Training second time series variable for conditioning correpsonding to downstream flows.
train_downstream_y : pandas.Series, optional
Training second time series variable for conditioning correpsonding to upstream flows.
ref_upstream_y : pandas.Series, optional
Target second time series variable for conditioning corresponding to upstream flows.
ref_downstream_y : pandas.Series, optional
Target second time series variable for conditioning corresponding to downtream flows.
bw : int, optional
Bandwidth for KernelDensity. This should only be used if `method='kde'`.
xbins : int, optional
Bins for the first time series. This should only be used if `method='hist'`.
ybins : int, optional
Bins for the second time series. This should only be used if `method='hist'`.
rtol : float, optional
The desired relatie tolerance of the result for KernelDensity.
This should only be used if `method='kde'`.
atol : float, optional
The desired absolute tolerance of the result for KernelDensity.
This should only be used if `method='kde'`.
method : str, optional
Method to use for conditioning. Currently 'hist' using hist2D and 'kde'
using kde2D are the only supported methods.
**kwargs:
Additional keyword arguments. Mainly implemented for cross-compatibility with
other methods so that a unified configuration can be used
Returns
-------
bc_totals : pandas.Series
Returns a time series of length of an interval in the bmoprh window
with bmorphed values.
bc_multipliers : pandas.Series
Returns a time series of equal length to bc_totals used to scale the
raw flow values into the bmorphed values returned in bc_totals.
"""
assert isinstance(interval, pd.DateOffset)
if interval == pd.DateOffset(days=1):
raise Exception("Please enter a interval greater than 1 day(s)")
bc_multipliers = pd.Series([])
bc_totals = pd.Series([])
apply_window = slice(*apply_window)
raw_train_window = slice(*raw_train_window)
ref_train_window = slice(*ref_train_window)
raw_ts_window = slice(pd.to_datetime(raw_upstream_ts.index.values[0]),
pd.to_datetime(raw_upstream_ts.index.values[-1]))
# Check if there is enough data input to run conditioning for both upstream
# and downstream bmorphs. Boolean used here instead of later to make certain
# both upstream and downstream use the same method and to minimze checks within
# the for-loop
run_mdcd = False
y_varlist = [raw_upstream_y, train_upstream_y, ref_upstream_y,
raw_downstream_y, train_downstream_y, ref_downstream_y]
if np.any(list(map(lambda x: x is not None, y_varlist))):
run_mdcd = True
# bmorph the series
overlap_period = int(overlap / 2)
bmorph_ts = pd.Series([])
bmorph_multipliers = pd.Series([])
bmorph_range = pd.date_range(apply_window.start, apply_window.stop+interval,
freq=interval)
for i in range(0,len(bmorph_range)-1):
bmorph_start = bmorph_range[i]
bmorph_end = bmorph_range[i+1]
# we don't need ot adjust for overlap if it is the last entry at i+1
if i < len(bmorph_range)-2:
bmorph_end -= pd.DateOffset(days=1)
raw_apply_window = slice(pd.to_datetime(str(bmorph_start)),
pd.to_datetime(str(bmorph_end)))
raw_cdf_window = slice(pd.to_datetime(str(bmorph_start - pd.DateOffset(days=overlap_period))),
pd.to_datetime(str(bmorph_end + pd.DateOffset(days=overlap_period))))
if (raw_cdf_window.start < raw_ts_window.start):
offset = raw_ts_window.start - raw_cdf_window.start
raw_cdf_window = slice(raw_ts_window.start, raw_cdf_window.stop + offset)
if(raw_cdf_window.stop > raw_ts_window.stop):
offset = raw_ts_window.stop - raw_cdf_window.stop
raw_cdf_window = slice(raw_cdf_window.start + offset, raw_ts_window.stop)
# Upstream & Downstream bias correction
if run_mdcd:
bc_up_total, bc_up_mult = bmorph.bmorph(
raw_upstream_ts, train_upstream_ts, ref_upstream_ts,
raw_apply_window, raw_train_window, ref_train_window, raw_cdf_window,
raw_upstream_y, ref_upstream_y, train_upstream_y,
n_smooth_short, bw=bw, xbins=xbins, ybins=ybins, rtol=rtol, atol=atol,
method=method, train_cdf_min=train_cdf_min)
bc_down_total, bc_down_mult = bmorph.bmorph(
raw_downstream_ts, train_downstream_ts, ref_downstream_ts,
raw_apply_window, raw_train_window, ref_train_window, raw_cdf_window,
raw_downstream_y, ref_downstream_y, train_downstream_y,
n_smooth_short, bw=bw, xbins=xbins, ybins=ybins, rtol=rtol, atol=atol,
method=method, train_cdf_min=train_cdf_min)
else:
bc_up_total, bc_up_mult = bmorph.bmorph(
raw_upstream_ts, train_upstream_ts, ref_upstream_ts,
raw_apply_window, raw_train_window, ref_train_window, raw_cdf_window,
n_smooth_short, train_cdf_min=train_cdf_min)
bc_down_total, bc_down_mult = bmorph.bmorph(
raw_downstream_ts, train_downstream_ts, ref_downstream_ts,
raw_apply_window, raw_train_window, ref_train_window, raw_cdf_window,
n_smooth_short, train_cdf_min=train_cdf_min)
bc_multiplier = (blend_factor * bc_up_mult) + ((1 - blend_factor) * bc_down_mult)
bc_total = (blend_factor * bc_up_total) + ((1 - blend_factor) * bc_down_total)
bc_multipliers = bc_multipliers.append(bc_multiplier)
bc_totals = bc_totals.append(bc_total)
bc_totals = bc_totals.groupby(bc_totals.index).mean()
bc_multipliers = bc_multipliers.groupby(bc_multipliers.index).mean()
# Apply the correction to preserve the mean change
if n_smooth_long:
raw_ts = (blend_factor * raw_upstream_ts) + ((1 - blend_factor) * raw_downstream_ts)
ref_ts = (blend_factor * ref_upstream_ts) + ((1 - blend_factor) * ref_downstream_ts)
train_ts = (blend_factor * train_upstream_ts) + ((1 - blend_factor) * train_downstream_ts)
nrni_mean = ref_ts[ref_train_window].mean()
train_mean = train_ts[raw_train_window].mean()
bc_totals, corr_ts = bmorph.bmorph_correct(raw_ts, bc_totals, raw_ts_window,
nrni_mean, train_mean,
n_smooth_long)
bc_multipliers *= corr_ts
return bc_totals[apply_window], bc_multipliers[apply_window]
def _scbc_c_seg(ds, apply_window, raw_train_window, ref_train_window,
interval, overlap, condition_var, **kwargs):
up_raw_ts = ds['IRFroutedRunoff'].to_series()
up_train_ts = ds['up_raw_flow'].to_series()
up_ref_ts = ds['up_ref_flow'].to_series()
up_seg = int( ds['up_ref_seg'].values[()])
up_cond = ds[f'up_{condition_var}'].to_series()
dn_raw_ts = ds['IRFroutedRunoff'].to_series()
dn_train_ts = ds['down_raw_flow'].to_series()
dn_ref_ts = ds['down_ref_flow'].to_series()
dn_seg = int( ds['down_ref_seg'].values[()])
dn_cond = ds[f'down_{condition_var}'].to_series()
blend_factor = ds['cdf_blend_factor'].values[()]
local_flow = ds['dlayRunoff']
scbc_c_flows, scbc_c_mults = apply_blendmorph(
up_raw_ts, dn_raw_ts,
up_train_ts, dn_train_ts,
up_ref_ts, dn_ref_ts,
apply_window, raw_train_window, ref_train_window,
interval, overlap, blend_factor,
raw_upstream_y=up_cond, raw_downstream_y=dn_cond,
train_upstream_y=up_cond, train_downstream_y=dn_cond,
ref_upstream_y=up_cond, ref_downstream_y=dn_cond, **kwargs)
scbc_c_locals = scbc_c_mults * local_flow.sel(time=scbc_c_mults.index)
return scbc_c_flows, scbc_c_mults, scbc_c_locals
def _scbc_u_seg(ds, apply_window, raw_train_window, ref_train_window,
interval, overlap, condition_var=None, **kwargs):
up_raw_ts = ds['IRFroutedRunoff'].to_series()
up_train_ts = ds['up_raw_flow'].to_series()
up_ref_ts = ds['up_ref_flow'].to_series()
up_seg = int( ds['up_ref_seg'].values[()])
dn_raw_ts = ds['IRFroutedRunoff'].to_series()
dn_train_ts = ds['down_raw_flow'].to_series()
dn_ref_ts = ds['down_ref_flow'].to_series()
dn_seg = int( ds['down_ref_seg'].values[()])
blend_factor = ds['cdf_blend_factor'].values[()]
local_flow = ds['dlayRunoff'].copy(deep=True)
scbc_u_flows, scbc_u_mults = apply_blendmorph(
up_raw_ts, dn_raw_ts,
up_train_ts, dn_train_ts,
up_ref_ts, dn_ref_ts,
apply_window, raw_train_window, ref_train_window,
interval, overlap, blend_factor, **kwargs)
scbc_u_locals = scbc_u_mults * local_flow.sel(time=scbc_u_mults.index)
return scbc_u_flows, scbc_u_mults, scbc_u_locals
[docs]def apply_scbc(ds, mizuroute_exe, bmorph_config, client=None, save_mults=False):
"""
Applies Spatially Consistent Bias Correction (SCBC) by
bias correcting local flows and re-routing them through
mizuroute. This method can be run in parallel by providing
a `dask client`.
Parameters
----------
ds: xr.Dataset
An xarray dataset containing `time` and `seg` dimensions and
variables to be bias corrected. This will mostly likely come from
the provided preprocessing utility, `mizuroute_utils.to_bmorph`
mizuroute_exe: str
The path to the mizuroute executable
bmorph_config: dict
The configuration for the bias correction. See the documentation
on input specifications and selecting bias correction techniques
for descriptions of the options and their choices.
client: dask.Client (optional)
A `client` object to manage parallel computation.
save_mults: boolean (optional)
Whether to save multipliers from bmorph for diagnosis. If True,
multipliers are saved in the same directory as local flows. Defaults
as False to not save multipliers.
Returns
-------
region_totals: xr.Dataset
The rerouted, total, bias corrected flows for the region
"""
def unpack_and_write_netcdf(results, segs, file_path, out_varname='scbc_flow', mult_path=None):
flows = [r[0] for r in results]
mults = [r[1] for r in results]
local = [r[2] for r in results]
local_ds = xr.DataArray(np.vstack(local), dims=('seg','time'))
local_ds['seg'] = segs
local_ds['time'] = flows[0].index
local_ds.name = out_varname
local_ds = local_ds.where(local_ds >= 1e-4, other=1e-4)
try:
os.remove(file_path)
except OSError:
pass
local_ds.transpose().to_netcdf(file_path)
if mult_path:
try:
os.remove(mult_path)
except OSError:
pass
mult_ds = xr.DataArray(np.vstack(mults), dims=('seg','time'))
mult_ds['seg'] = segs
mult_ds['time'] = flows[0].index
mult_ds.name = 'mults'
mult_ds.transpose().to_netcdf(mult_path)
if 'condition_var' in bmorph_config.keys() and bmorph_config['condition_var']:
scbc_type = 'conditional'
scbc_fun = partial(_scbc_c_seg, **bmorph_config)
else:
scbc_type = 'univariate'
scbc_fun = partial(_scbc_u_seg, **bmorph_config)
if client:
futures = [client.submit(scbc_fun, ds.sel(seg=seg)) for seg in ds['seg'].values]
results = client.gather(futures)
else:
results = []
for seg in tqdm(ds['seg'].values):
results.append(scbc_fun(ds.sel(seg=seg)))
out_file = (f'{bmorph_config["data_path"]}/input/'
f'{bmorph_config["output_prefix"]}_local_{scbc_type}_scbc.nc')
if save_mults:
mult_file = (f'{bmorph_config["data_path"]}/input/'
f'{bmorph_config["output_prefix"]}_local_mult_{scbc_type}_scbc.nc')
else:
mult_file = None
unpack_and_write_netcdf(results, ds['seg'], out_file, mult_path=mult_file)
config_path, mizuroute_config = mizutil.write_mizuroute_config(
bmorph_config["output_prefix"],
scbc_type, bmorph_config['apply_window'],
config_dir=bmorph_config['data_path']+'/mizuroute_configs/',
topo_dir=bmorph_config['data_path']+'/topologies/',
input_dir=bmorph_config['data_path']+'/input/',
output_dir=bmorph_config['data_path']+'/output/',
)
mizutil.run_mizuroute(mizuroute_exe, config_path)
region_totals = xr.open_mfdataset(f'{mizuroute_config["output_dir"]}{bmorph_config["output_prefix"]}_{scbc_type}_scbc*')
region_totals = region_totals.sel(time=slice(*bmorph_config['apply_window']))
region_totals['seg'] = region_totals['reachID'].isel(time=0)
return region_totals.load()
def bmorph_to_dataarray(dict_flows, name):
da = xr.DataArray(np.vstack(dict_flows.values()), dims=('site', 'time'))
da['site'] = list(dict_flows.keys())
da['time'] = list(dict_flows.values())[0].index
da.name = name
return da.transpose()