HESSHydrology and Earth System SciencesHESSHydrol. Earth Syst. Sci.1607-7938Copernicus PublicationsGöttingen, Germany10.5194/hess-21-2595-2017Examining the impacts of precipitation isotope input (δ18Oppt) on distributed, tracer-aided hydrological modellingDelavauCarly J.carly.delavau@gov.mb.caStadnykTriciatricia.stadnyk@umanitoba.cahttps://orcid.org/0000-0002-2145-4963HolmesTeganDepartment of Civil Engineering, University of Manitoba, Winnipeg, R3T 5V6, CanadaTricia Stadnyk (tricia.stadnyk@umanitoba.ca) and Carly J. Delavau (carly.delavau@gov.mb.ca)30May2017215259526148October201625October20161April201711April2017This work is licensed under a Creative Commons Attribution 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by/3.0/This article is available from https://hess.copernicus.org/articles/21/2595/2017/hess-21-2595-2017.htmlThe full text article is available as a PDF file from https://hess.copernicus.org/articles/21/2595/2017/hess-21-2595-2017.pdf
Tracer-aided hydrological models are becoming increasingly popular tools as
they assist with process understanding and source separation, which
facilitates model calibration and diagnosis of model uncertainty (Tetzlaff et
al., 2015; Klaus and McDonnell, 2013). Data availability in high-latitude
regions, however, proves to be a major challenge associated with this type of
application (Tetzlaff et al., 2015). Models require a time series of isotopes
in precipitation (δ18Oppt) to drive simulations, and
throughout much of the world – particularly in sparsely populated
high-latitude regions – these data are not widely available. Here we
investigate the impact that choice of precipitation isotope product
(δ18Oppt) has on simulations of streamflow,
δ18O in streamflow (δ18OSF), resulting
hydrograph separations, and model parameters. In a high-latitude,
data-sparse, seasonal basin (Fort Simpson, NWT, Canada), we assess three
precipitation isotope products of different spatial and temporal resolutions
(i.e. semi-annual static, seasonal KPN43, and daily bias-corrected REMOiso),
and apply them to force the isoWATFLOOD tracer-aided hydrologic model. Total
simulated streamflow is not significantly impacted by choice of
δ18Oppt product; however, simulated isotopes in
streamflow (δ18OSF) and the internal apportionment of
water (driven by model parameterization) are impacted. The highest-resolution
product (REMOiso) was distinct from the two lower-resolution products (KPN43
and static), but could not be verified as correct due to a lack of daily
δ18Oppt observations. The resolution of
δ18Oppt impacts model parameterization and seasonal
hydrograph separations, producing notable differences among simulations
following large snowmelt and rainfall events when event compositions differ
significantly from δ18OSF. Capturing and preserving the
spatial variability in δ18Oppt using distributed
tracer-aided models is important because this variability impacts model
parameterization. We achieve an understanding of tracer-aided modelling and
its application in high-latitude regions with limited
δ18Oppt observations, and the value such models have in
defining modelling uncertainty. In this study, application of a tracer-aided
model is able to identify simulations with improved internal process
representation, reinforcing the fact that tracer-aided modelling approaches
assist with resolving hydrograph component contributions and work towards
diagnosing equifinality.
Introduction
Hydrological models are critical tools for the planning, development, design,
operation, and sustainable
management of water resources (Singh and Frevert, 2006). These models provide
insight into applications such as the prediction of floods, droughts and
water availability, and the effects of climate and land use change on water
resources. Problems arise for calibration and validation of hydrological
models when there is (1) a lack of available data at sufficient resolutions
to force and validate model simulations – especially in remote,
high-latitude locations (in Canada: Coulibaly et al., 2013); (2) issues with
equifinality affecting model parameterization; and (3) uncertainty in model
results (e.g. Beven and Binley, 1992; Kirchner, 2006; Fenicia et al., 2008;
Dunn et al., 2008).
It is now widely accepted that calibration and validation of hydrological
models based solely on streamflow is not a sufficient evaluation measure
(Kuczera, 1983; Beven and Binley, 1992; Kuczera and Mroczkowski, 1998;
Seibert and McDonnell, 2002; Kirchner, 2006; Fenicia et al., 2008; Dunn et
al., 2008). Modellers are focusing on a model's ability to correctly
partition, store, and release water from hydrologic compartments, in addition
to adequately simulating total streamflow response. Conservative tracer data
provide insights into the dominant hydrological processes and integrated
runoff response (in northern catchments: Birks and Gibson, 2009; Tezlaff et
al., 2015), and such data assist with constraining model parameter space
during calibration, reducing model uncertainty, and assisting with selection
of appropriate model structures (e.g. Tetzlaff et al., 2008; Birkel et al.,
2010a; Birkel et al., 2014;
Smith et al., 2016). An increasing number of studies have investigated the
utility of tracer-aided modelling approaches, especially over the past decade
(for a comprehensive overview, see Birkel and Soulsby, 2015).
Although greatly informative, previous tracer-aided modelling studies have
generally been conducted using lumped conceptual rainfall–runoff models in
highly instrumented small-scale experimental catchments
(< 102 km2). This has resulted in distributed studies at the
regional scale (> 103 km2) being left largely unexplored, with
the exception of a few, select applications (Stadnyk et al., 2013). Modelling
at the regional scale typically requires a distributed approach to capture
the heterogeneity in meteorological inputs, basin characteristics, and runoff
response, resulting in more complex, highly parameterized models (e.g.
Michaud and Sorooshian, 1994; Carpenter and Georgakakos, 2006; Her and
Chaubey, 2015). Because it is at these larger scales where models are applied
operationally and management decisions are based, there is a critical need to
understand the abilities, limitations, and uncertainties associated with
distributed tracer-aided modelling at the regional scale.
Although there is an identified need, the issue of data availability,
particularly input data, proves to be a major challenge associated with this
type of application (Birkel and Soulsby, 2015). Tracer-aided hydrological
modelling typically requires a time series of isotopes in precipitation
(δ18Oppt) to drive model simulations. Unfortunately,
throughout much of the world, and particularly in sparsely populated
high-latitude regions (such as the vast majority of Canada), these data are
not widely available. Although automatic samplers are becoming increasingly
common, watersheds in which snow accumulation is substantial will continue to
be fraught with difficulties surrounding the collection and characterization
of precipitation isotopes, particularly during the winter months (Dietermann
and Weiler, 2013; Penna et al., 2014). The lack of spatial and temporal
density of δ18Oppt observations highlights the need for
alternative methods to provide estimates of stable isotopes in precipitation
for tracer-aided model input (termed “δ18Oppt
products”). Options include empirically based models generating gridded time
series of precipitation isotopes (e.g. Lykoudis et al., 2010; Delavau et al.,
2015), and isotope-enabled climate model output (for a comprehensive
overview, see Noone and Sturm, 2010; Xi, 2014).
Small-scale catchment studies rely on continuous records of
δ18Oppt observations at high temporal frequencies
(typically daily, and less commonly, weekly) for model input. At the larger
scale, tracer-aided modelling completed by Stadnyk et al. (2013) in the
remote Fort Simpson region of northern Canada used annual average
compositions of rainfall and snowfall δ18O to drive model
simulations. Their results suggested that utilizing annual, spatially static
oxygen-18 in precipitation forcing has the potential to significantly impact
simulations and, consequently, model parameterization as well. The assumption
that model input is spatially invariant is not preferable, as
δ18Oppt can vary drastically over small space scales
and timescales due to changes in moisture sources and transport processes,
rainout history, and seasonality (e.g. in Canada: Gat et al., 1994; Moran et
al., 2007; Birks and Edwards, 2009).
This study aims to explore how varying spatial and temporal resolutions of
precipitation isotope products, or δ18Oppt input,
impact regional tracer-aided model simulations and parameterization. Forcing
a tracer-aided, distributed hydrological model (isoWATFLOOD) with three
precipitation isotope products, we examine how the different
δ18Oppt products impact the
simulation of total streamflow and its isotopic variability (δ18OSF);
internal apportionment of water, namely the seasonality of hydrograph
separation; and
model parameterization and simulation uncertainty.
We explore the impact that varying the resolution of
δ18Oppt inputs has on the capability of the model to
reproduce observed δ18OSF variability, and the
usefulness of a tracer-aided modelling approach to help inform and quantify
simulation equifinality.
Study area and dataThe Fort Simpson Basin
The Fort Simpson Basin (FSB) is located within the lower Liard River Valley
close to the town of Fort Simpson, Northwest Territories, Canada
(61∘45 N, 121∘14 W; Fig. 1). This region has been the focus
of several tracer-aided hydrological studies (e.g. St Amour et al., 2005;
Stadnyk et al., 2005, 2013; Stadnyk-Falcone, 2008). The FSB is selected for
this study to build upon previous modelling work conducted within the region,
and to follow up on recommendations from Stadnyk et al. (2013) suggesting
further analysis and improvement of isoWATFLOOD δ18Oppt
input. The study period of 1997–1999 is selected based on data availability.
Basin characteristics, including land cover classification, area, and
average basin slope (recreated from data provided in St Amour et al., 2005).
Fort Simpson River Basin (all other tributaries of the Liard and
Mackenzie rivers have been removed for ease of viewing).
This study considers two sub-basins of the greater Fort Simpson Basin: the
Jean-Marie (1310 km2) and Blackstone River (1390 km2) sub-basins
(Fig. 1). The basins vary in relief from 0.3 % in the Jean-Marie sub-basin
to 0.63 % for the Blackstone sub-basin, on average. Differences in wetland
distribution and function, basin physiography, and land cover make-up between
the two watersheds (Table 1) are the primary reasons in selecting these
sub-basins for this study. These marked differences ensure that watersheds of
varying dominant hydrological processes are represented in the modelling, and
therefore the impacts of δ18Oppt input selection on
these processes can be examined.
The land cover classification breakdown (Table 1) shows the primary land
cover type within the sub-basins as transitional, consisting of shrubs,
deciduous varieties, and early generation spruce. The region has a high
proportion of wetlands, with the total wetland percentage in Table 1
representing both bogs (disconnected drainage) and fens (connected drainage),
although the amount of each type within each respective sub-basin varies.
Aylsworth and Kettles (2000) state that Jean-Marie is predominately fen peatlands, while Blackstone is bog-dominated
peatlands, with very few or no fen peatlands present.
The Ecoregions Working Group (1989) classifies the FSB as a sub-humid mid- to
high-boreal ecoclimatic region (Hbs), classified by cool summers
approximately 5 months in length, with moderate (300–500 mm) annual
precipitation. Winters are very cold with persistent snow cover. The
hydrological response is dominated by snowmelt during late April to early
May, while summer and fall runoff events are due to major rainfall, with a
return to baseflow occurring during dry summer periods or towards the
beginning of the ice-on season in October.
Data summary for the study period (SP) and period of record (PoR).
The coefficient of variation (CV) is calculated as the ratio of the standard
deviation to the mean.
* Provided only for the study period, 1997–1999. n/a: not applicable.
Meteorological and hydrometric data
Daily total precipitation, mean daily temperature, and hourly relative
humidity data are obtained from Environment Canada's Fort Simpson Airport
weather station. Observed precipitation is supplemented with ANUSPLIN-derived
daily precipitation extracted at eight locations throughout the Fort Simpson
region (Fig. 1). ANUSPLIN is a multidimensional non-parametric surface
fitting method that has been found to be well suited to the interpolation of
various climate variables, particularity in data-sparse, high-elevation
regions, as the method accounts for spatially varying dependencies on
elevation (McKenney et al., 2011). We have validated ANUSPLIN against
independent station observations (precipitation and temperature) across the
Prairies and Boreal regions of Canada as a precipitation forcing for
hydrologic modelling. It has been found to be adequate (r≥ 0.98) for
the purpose of short-term modelling studies. An inverse-distance weighting
approach is used to spatially distribute the daily ANUSPLIN and observed
precipitation time series across the model domain (Kouwen, 2016). Rainfall
that occurred over the study period, particularly in 1997, was significantly
higher than normal. Additionally, 1998 was above average in temperature,
which is especially prevalent in the first portion of the year. Other
researchers have attributed the increased rainfall and warmer temperatures to
a strong El Niño influence from mid-1997 to mid-1998 (Petrone et al.,
2000; St Amour et al., 2005).
Hydrometric records are obtained from Water Survey of Canada. Jean-Marie was
gauged at Highway No. 1 in 1972 with a period of record of 44 years, whereas
Blackstone was gauged at Highway No. 7 in 1991 having a record length of
25 years. Neither sub-basin is regulated; therefore, all flows are considered
to be natural. During the study period, mean annual discharge was above
normal in both sub-basins in 1997, normal in Jean-Marie, slightly below
normal in Blackstone in 1998, and below normal in both sub-basins in 1999.
Winter (ice-on) flows tend to be very low given highly seasonal,
high-latitude hydrology, underlying discontinuous permafrost, and the absence
of mid-winter melt (St Amour et al., 2005). Averaged winter ice-on flows
from 1997 to 1999 were 0.194 and 0.034 m3 s-1 for the Jean-Marie
and Blackstone sub-basins, respectively. A statistical summary of
observations used in this study is provided in Table 2.
Isotope data
During 1997 to 1999, intensive sampling took place in the Fort Simpson Basin
as part of the Mackenzie Study of the Global Energy and Water Experiment
(GEWEX; Stewart et al., 1998). The campaign sampled δ18O and
δ2H of streamflow, rainfall, snowpack, and surface waters (wetlands
and lakes) during the open water season (May to October). During ice-on
conditions, the isotope stratigraphy of river ice extracted during late March
in 1998 and 1999 was used to reconstruct the isotopic composition of winter
streamflow (Gibson and Prowse, 1999; Prowse et al., 2002; St Amour et al.,
2005). This study uses measured δ18O compositions in streamflow in
the Jean-Marie (n= 71) and Blackstone (n= 69) sub-basins for
model calibration. Although δ18Oppt compositions
(n= 27) were collected as part of the GEWEX sampling campaign, these
data are not preferred for tracer-aided hydrologic model input due to their
spatial uniformity and poor temporal resolution. Observations are
incorporated into this study as the “static” δ18Oppt
input, and as a means to validate the KPN43 and REMOiso products and to
inform the static precipitation product. The number of measurements and their
statistical properties are summarized in Table 2. Isotopic compositions of
δ18O are expressed in delta (δ) notation as a deviation from
VSMOW (Vienna Mean Standard Mean Ocean Water) in units of per
mille (‰), such that
δwater= (Rwater/RVSMOW- 1) × 1000 ‰,
where R is 18O/16O in the sample and standard, respectively.
Isotope samples were analysed at the Environmental Isotope Laboratory at the
University of Waterloo, and St Amour et al. (2005) indicated maximum
analytical uncertainties of ±0.1 ‰ for δ18O.
Precipitation oxygen-18 input
The precipitation isotope products evaluated in this study represent a
variety of spatial and temporal scales, and were selected because they are
commonly available for all tracer-aided hydrologic modelling applications.
The first type of input used in this study is annual average
δ18Oppt compositions of rainfall and snowfall for each
year of simulation (i.e. yearly resolution). Values for the FSB were obtained
by averaging observations of δ18O in rainfall and the snowpack
obtained from the GEWEX study (Tables 2 and 3). δ18Oppt
compositions were assumed constant throughout the study domain (i.e.
spatially uniform). Due to a lack of snowfall data collected during this
study, we assumed the average annual isotopic composition of the snowpack was
representative of the snowfall composition, as has been done in other data
sparse, high-latitude tracer-aided modelling studies (Smith et al., 2015,
2016; Holmes, 2016; Stadnyk et al., 2013). It is well established in the
literature that the isotopic composition of snowfall is not necessarily equal
to the average annual composition of the snowpack (due to sublimation and
snow metamorphism; Zhou et al., 2008; Taylor et al., 2001, 2002). The high
latitude of our study site, however, makes freeze–thaw cycling during the
winter rare, making this assumption more reasonable. Due to the averaged
values and lack of spatial variability, this product is referred to as
“static” throughout the remainder of the paper, and consists of two
constant δ18Oppt values (rain and snow) for each year.
This product is specifically designed and evaluated for remote regions that
lack spatially and temporally varying δ18Oppt
observations.
Static δ18Oppt input compositions of annual rainfall
and snowfall oxygen-18 for isoWATFLOOD.
Parameters included in the Monte Carlo calibration, alongside a
description of what the parameter represents and the algorithm it is used within.
NameDescriptionAlgorithmRouting parameters flzLower zone drainage functionAn exponential groundwater depletion function that graduallypwrLower zone drainage functiondiminishes the base flow. Groundwater is replenished byexponentdrainage of the UZS:QLZ = LZF ⋅ LZS)PWR,where LZS is lower zone storage.QLZ is the baseflow flux.thetaWetland porosityPhysically based wetland routing algorithmkcondConductivity parameter(McKillop et al., 1999)Hydrologic parameters f-ratioInterception capacity multiplierConceptual evaporation algorithm based on Hargreaves andSamani (1982). f-ratio is a multiplier for the interceptioncapacity for each land class.akSurface permeability (bare ground)Conceptual infiltration algorithm (similar to Green and Ampt,akfsSurface permeability1911), but based on Richard's equation, which is physicallybased (Philip, 1954).recInterflow coefficientInterflow is represented by a simple storage–dischargeretnUpper zone retention (mm)relation:DUZ = REC ⋅ (UZS-RETN) ⋅ Si,where UZS = upper zone storageDUZ = depth of upper zone storage released as interflow, andSi = internal land surface slope.ak2Recharge coefficient (bare ground)Upper zone to lower zone drainage is represented by a simplestorage–discharge relation:DRNG = AK2 ⋅ (UZS - RETN),where DRNG is the drainage from UZS to LZS.mfMelt factor (mm ∘C-1 h-1)M= MF(Ta- base)baseBase temperature (∘C)Anderson (1976)subSublimation factorSublimation is modelled by a static sublimation factor.Amount of sublimation is a fraction of the observed snowfall.For new model set-ups, the sublimation factor has beenreplaced by a static sublimation rate.
Times series simulations obtained from the KPN43 model created by Delavau et
al. (2015) are used as the second type of δ18Oppt product
in this study. The KPN43 model uses North American Regional Reanalysis
(NARR; Mesinger et al., 2006) climate variables, teleconnection indices, and
geographic information to produce gridded time series of oxygen-18 in
precipitation at a monthly time step (Delavau, 2017). This product is generated at a 10 km
resolution (to mirror model set-up), and varies spatially throughout the
study domain due to the variation in the climatic predictors and geographic
information required to produce simulations.
The third δ18Oppt product included in this study is
regional climate model output from the isotope-enabled climate model, REMOiso
(Sturm et al., 2005, 2007). Raw REMOiso δ18Oppt output
is available at a 55 km spatial resolution and a 6 h time step. REMOiso
output is averaged in this study, however, to a daily time step, as the range
and variability of sub-daily δ18Oppt are erroneously
large, and the resolution of streamflow oxygen-18 calibration data does not
warrant a temporal frequency of input finer than daily.
MethodsBackground and set-up
The tracer-aided hydrological model used in this study is isoWATFLOOD
(Stadnyk-Falcone, 2008; Stadnyk et al., 2013). isoWATFLOOD is an extension of
the WATFLOOD hydrological model, whereby water and oxygen-18 are
simultaneously budgeted throughout the modelled hydrologic cycle. WATFLOOD is
a distributed model that uses grouped response units (GRUs) to simulate
streamflow in hydrologically distinct land cover units (Kouwen et al., 1993;
Kouwen, 2016). Process representation within WATFLOOD is considered to be a
combination of both conceptually and physically based algorithms, as certain algorithms are conceptually based (e.g. evaporation
and snowmelt), while others are more based in physics (e.g. channel routing).
Due to the coupling of isotopes to each hydrological process simulated in
WATFLOOD, simulation of isotopic composition does not introduce any
additional parameters. A more comprehensive description of isoWATFLOOD's
model structure and governing equations can be found in Stadnyk et
al. (2013), and selected descriptions are provided in Table 4.
isoWATFLOOD requires the δ18O of precipitation (either rain and
snow separately, or total precipitation) and can utilize (though does not
require) distributed relative humidity inputs to force the model.
Additionally, δ18O compositions for hydrologic storages of
river/fen water, soil water, baseflow, and snowpack are needed for model
initialization, which can be obtained from field data or estimated. Here,
regional isotopic storage initializations are derived from measured data
obtained during the GEWEX campaign and reported by St Amour et al. (2005).
These include streamflow (-13.52 ‰), interflow (soil water;
-14.60 ‰), baseflow (-20.00 ‰), and snowpack
(-22.00 ‰) background compositions. Sensitivity analyses have shown
that within 1 month of simulation isoWATFLOOD spin-up is complete and, past
this point, initialization values have no bearing on model output. All other
data required by isoWATFLOOD (e.g. distributed precipitation, temperature,
evaporation, inflows) are passed from WATFLOOD forcings or computations.
The isoWATFLOOD model used in this study is based on a previous version
reported by Stadnyk et al. (2013). The current version used here is an
updated version of isoWATFLOOD code, and the watershed set-up incorporates
various model improvements made since 2013, independent of this study. Based
on findings from Aylsworth and Kettles (2000), we implemented a 90 % bog and 10 % fen split in Blackstone and
a 30 % bog and 70 % fen split in Jean-Marie. The entirety of the FSB is
modelled at a 10 km spatial resolution, and the model is run continuously
from January 1996 to December 1999, whereby 1996 is utilized as a spin-up to
set initial hydrologic and isotopic storage conditions.
Calibration and parameter uncertainty
Being a distributed model, WATFLOOD has a large number of parameters
requiring calibration. For this reason, a sensitivity analysis is first
conducted to identify which parameters have the largest influence on both
streamflow and δ18OSF. A subset of parameters is
identified for inclusion in the calibration based on this sensitivity
analysis, including nine hydrological parameters from each of the five most
prominent land classes (mixed/deciduous, coniferous, transit, bogs, and
fens), and four routing parameters from each of the two modelled sub-basins.
This results in 53 parameters that are incorporated into the parameter
uncertainty assessment (Tables 4 and S1 in the Supplement). Allowable ranges
for each parameter are determined based on published values alongside
personal communications with N. Kouwen (Kouwen, 2016) (Table S1).
This study uses a multi-criteria, multi-objective approach to model
calibration, with the procedure summarized as follows.
A Monte Carlo random sampling approach, assuming uniform parameter
distributions, is used to individually select each parameter from its
allowable range (Table S1). Random parameter sampling is completed 30 000 times,
generating 30 000 unique parameter sets for isoWATFLOOD model evaluation.
For each of the three δ18Oppt inputs (KPN43, REMOiso, and
static), streamflow and δ18OSF are simulated from 1996
to 1999 for all 30 000 parameter sets – as defined in (i).
Simulated streamflow and δ18OSF are assessed statistically
over the period of study (1997–1999, excluding the 1996 spin-up year), and
regionally across the Jean-Marie and Blackstone sub-basins. Simulations are
classified as behavioural (or non-behavioural) (Beven and Binley, 1992) based
on meeting (or not) the following set of efficiency criteria thresholds,
defined in detail below, for simulated streamflow and
δ18OSF.
Streamflow:NSE≥0.5;|%Dv|≤20;and|log(%Dv)|≤20%.
δ18OSF:RMSE≤2.5‰andKGE>=0.3.
Behavioural thresholds used in this study are subjectively defined, but are
arrived at through a review of methods employed in similar studies (e.g.
Moriasi et al., 2007; Birkel et al., 2010a, b, 2011; Smith et al., 2016),
measurement error, and an iterative process exploring the sensitivity between
the set thresholds and resulting behavioural simulations for each input type.
Based on this analysis, the Nash–Sutcliffe efficiency (NSE; Nash and
Sutcliffe, 1970), volumetric error criteria (%Dv), root mean square
error (RMSE), and the Kling–Gupta efficiency criterion (KGE; Gupta et al.,
2009; Kling et al., 2012) are selected. A multi-criteria model evaluation
approach places emphasis on different statistical properties of a simulation.
For example, NSE has a documented bias towards peak flow, and conversely,
log(% Dv) is a more appropriate evaluation measure for periods of low
flow. The NSE, % Dv, and log(% Dv) efficiency are not considered
suitable metrics for δ18OSF assessment due to the
temporal discontinuity of the isotope observations; therefore, RMSE and KGE
are used as isotopic simulation statistics. The KGE statistic puts less
emphasis on peak flow differences by providing a more balanced approach where
error is first summed and then squared at the end, preserving the sign of the
error and enabling a trade-off of error throughout the simulation period
(Gupta et al., 2009). It should also be noted that
δ18OSF observations are not equally distributed through
time, whereby the highest concentration of observations occurs during
snowmelt in the month of May (∼ 25 %), and the fewest observations
during the 6-month ice-on period from November to April (∼ 23 %),
with the remaining 52 % of observations sampled during summer. The sporadic
distribution of observations may result in the calibrations more highly
weighted to certain periods of the year and the
dominant processes occurring at that time,
thereby having the potential to impact model parameterization.
Comparison of raw and corrected REMOiso δ18Oppt
output with CNIP monthly compositions at Snare Rapids, NWT.
REMOiso bias correction
Due to a lack of published studies evaluating REMOiso performance within
Canada, a comparison between REMOiso output and Canadian Network for Isotopes
in Precipitation observations (CNIP; Birks and Gibson, 2009) is completed to
determine whether REMOiso simulations require a regional bias correction.
CNIP data are now part of the Global Network for Isotopes in
Precipitation (GNIP) database and can be accessed at
http://www.iaea.org/water (IAEA/WMO, 2014). This analysis is completed
at Snare Rapids, NWT, the closest CNIP station to the FSB, for the years
of 2000 and 2001. Snare Rapids is located approximately 330 km northeast of
Fort Simpson and has monthly δ18Oppt observations
spanning the years of 1997–2010. A longer time frame of comparison between
CNIP and REMOiso is not possible due to the short overlapping period of
REMOiso simulations and CNIP observations. For bias-correction purposes,
daily REMOiso simulations are averaged to monthly compositions for direct
comparison to CNIP data using the precipitation amount-weighting approach in
Eq. (1):
δ18Opptmonthly=∑Pi⋅δ18Oppti∑Pi,
where Pi is the amount of daily precipitation (mm) obtained from the
Snare Rapids Canadian Air and Precipitation Monitoring Network (CAPMoN)
station operated by Environment Canada, where isotopic compositions are also
sampled under the Canadian Network for Isotopes in Precipitation.
Uncorrected REMOiso simulations exhibit a positive bias in this region
(Fig. 2), which is expected based on the ECHAM4 mean annual
δ18Oppt output (Noone and Sturm, 2010) and personal
communications with Birks and Sturm (2016). Therefore, a seasonal bias
correction is applied to daily REMOiso simulations. The bias correction is
calculated as the average seasonal difference between the monthly
amount-weighted REMOiso output and the CNIP observations. Corrected monthly
and daily REMOiso output at Snare Rapids is displayed in Fig. 2 as the dashed
red and solid orange lines, respectively. For the current study, daily
REMOiso output for the Fort Simpson region is bias-corrected with the
seasonal correction values, ranging from -4.5 ‰ (NDJF) to
-8.9 ‰ (MAM), with an average of -7.0 ‰.
The statistical properties of the corrected daily REMOiso simulations,
alongside the KPN43 monthly simulations and the static seasonal averages, are
summarized in Table 2.
Statistical treatment of data
For discussion and analysis purposes (Sects. 4.2–4.4), results represent
only the behavioural simulations derived from each
δ18Oppt product. Uncertainty bounds are the 5th and
95th percentiles drawn from the ensembles of behavioural simulations, denoted
as the shaded bounds around each model's mean simulation.
Kendall's tau coefficient (τ) is a non-parametric test used to compare
the level of correlation between two variables. We compute Kendall's tau for
the mean daily streamflow and δ18OSF simulations
derived from the three inputs. By computing τ coefficients for pairs of
simulated time series (i.e. REMOiso vs. KPN43, REMOiso vs. static, and KPN43
vs. static), we can statistically evaluate the similarity of model output
derived from different precipitation isotope products.
Parameter probability distributions (Table 4) are arrived at by first
weighting behavioural parameters for each land cover type to their
corresponding percent coverage within the modelled sub-basins. Land cover
weighted parameter values are then ranked and non-exceedance probabilities
determined. Routing parameter distributions for each sub-basin are arrived at
using a similar approach, but are not weighted by coverage. The
non-parametric Kolmogorov–Smirnov (K–S) test is used to assess whether
behavioural parameter distributions are considered to be from the same
distribution.
Spatially distributed precipitation isotope product maps (Fig. S1 in the
Supplement) represent daily precipitation isotopes averaged across seasons
(DJF, MAM, JJA, SON), and are precipitation amount-weighted using WATFLOOD
precipitation input (interpolated Environment Canada Canadian Daily Climate
Data, housed in WATFLOODs radcl.r2c files; Kouwen, 2016). Maps are generated
overlapping the model grid (10 k) for the entire FSB domain, which includes
the Jean-Marie and Blackstone sub-basins.
Average simulation statistics from n
behavioural simulations for streamflow and δ18OSF for
the three model calibrations (using KPN43, REMOiso, and static inputs).
AverageKPN43REMOisoStaticstatisticsfrom nbehaviouralsimulationsn321/30 000268/30 000216/30 000Streamflow (1095 observations for performance evaluation) NSE0.680.680.69|% Dv|13.913.414.2|log(% Dv)|11.58.911.6δ18OSF (140 observations for performance evaluation) RMSE (‰)1.391.322.09KGE0.360.330.35
Input and behavioural simulations for Jean-Marie, including
(a) KPN43, REMOiso, and static δ18Oppt input
time series and daily precipitation; and simulated (b) mean daily
streamflow and uncertainty bounds and (c) mean daily
δ18OSF and uncertainty bounds, for KPN43, REMOiso, and
static driven model calibrations. δ18Oppt
input-specific uncertainty bounds are represented as the shaded regions, with
shading colour corresponding to δ18Oppt type.
Input and behavioural simulations for Blackstone, including
(a) KPN43, REMOiso, and static δ18Oppt input
time series and daily precipitation, and simulated (b) mean daily
streamflow and uncertainty bounds and (c) mean daily
δ18OSF and uncertainty bounds, for KPN43, REMOiso, and
static driven model calibrations. δ18Oppt
input-specific uncertainty bounds are represented as the shaded regions, with
shading colour corresponding to δ18Oppt type.
Results and discussion
Results of the three calibrations indicate that the choice of
δ18Oppt input influences the number of simulations that
meet behavioural criteria thresholds. The KPN43 product results in more
behavioural simulations (n= 321) relative to the REMOiso
(n= 268) or static (n= 216) products (Table 5). This also
implies that choice of δ18Oppt input influences the
models' internal apportionment of water (i.e. hydrograph separations) via
model parameters. Among input types, potentially significant differences in
several parameters were noted (Table S1), and are discussed in Sect. 4.4. In
almost all instances, the ranges of the parameters were not significantly
constrained from the allowable parameter ranges, yielding confidence in our
simulated parameter uncertainty envelopes.
Precipitation oxygen-18 input
Of the three δ18Oppt products, KPN43 input is on
average the most enriched (-20.48 ‰), followed by REMOiso
(-21.78 ‰), and static is the most depleted (-22.82 ‰)
(Figs. 3a and 4a). The KPN43 and static products show similar variation about
their means, with CVs equal to 0.19 and 0.20, respectively. Conversely,
REMOiso has a higher CV (0.25) and much larger range, which is, in part, due
to the finer daily time step of this input. Spatial variability between Jean-Marie and Blackstone is zero for the static product as annual snow and
rainfall compositions are spatially averaged across the domain. Spatial
variation among sub-basins is noted in the KPN43 and REMOiso products. Both
the KPN43 and REMOiso products show, on average, more depleted
δ18Oppt values within Blackstone (-20.79 and
-22.01 ‰, respectively) in comparison to Jean-Marie (-20.17 and
-21.54 ‰, respectively), likely caused by the higher-elevation
headwaters of Blackstone relative to Jean-Marie (a maximum difference of
∼ 215 m). Figure S1 provides seasonally averaged, spatially
distributed maps for each product. Averaged spatial variability is greatest
for the KPN43 forcing, followed by REMOiso, and is constant for the static
product. REMOiso shows less long-term average variability because its
temporal variability is greater, resulting in more chaotic (randomized)
signals of δ18Oppt that produce weaker long-term
signals when averaged over time. KPN43, on the other hand, exhibits more
consistent spatial patterning of δ18Oppt variability,
resulting in stronger signals of long-term variability on a per-grid basis
(Fig. S1). REMOiso input is derived on a 55 km grid, meaning that
approximately five isoWATFLOOD grids are equivalent to one REMOiso grid,
which also results in a terrain (variability) smoothing effect. The static
input exhibits seasonal variability caused by the different compositions of
rain and snow, and mixed shoulder season compositions (MAM and SON) when both
rain and snow occur.
Although there are only 19 rainfall δ18O observations collected
over the study period in Jean-Marie, and eight within Blackstone (hollow
black diamonds in Figs. 3a and 4a), these limited data provide some
information regarding the accuracy of the products. By visual inspection,
each of the three products generates reasonable estimates of
δ18Oppt. This is expected for the static input, which
is derived directly from these observations; however, this provides
qualitative validation for KPN43 and REMOiso. REMOiso is the only product
that can somewhat replicate event-scale variability in
δ18Oppt due to its daily time step. The KPN43 product
appears to represent the average composition of summer rainfall events, and
displays reasonable seasonal variability. There are insufficient observations
to statistically validate these products within the Fort Simpson Basin. The
semi-annual static input perhaps does a reasonable job of reflecting
δ18Oppt seasonality because of the high-latitude
location of the basin, where much shorter shoulder seasons exist.
Modelling streamflow
All calibrations adequately capture variations in total streamflow in both
sub-basins, as emphasized by the regional calibration statistics (Table 5).
On average, behavioural streamflow simulations have a NSE of 0.68, and % Dv
of 13.8 %. Mean daily streamflow and uncertainty bounds for the KPN43,
REMOiso, and static model calibrations are displayed in Figs. 3b and 4b. It
is worth noting that both basins have similar drainage areas and received
comparable precipitation inputs over the study period, which would naturally
result in similar streamflow responses. Comparing normalized (by drainage
area) observed discharge over the study period for the basins reveals the
Blackstone sub-basin generates nearly twice as much runoff as the Jean-Marie
sub-basin, with normalized discharges of 0.56 and 0.31 mm km-2,
respectively. Therefore, differences in hydrograph characteristics (i.e. peak
flows, attenuation, etc.) between Jean-Marie and Blackstone result from
variations in basin physiography, storage mechanisms, and land cover
composition, specifically large differences in average basin slope and
surface water and wetland dynamics (St Amour et al., 2005). Namely, the
higher energy environment of Blackstone River promotes a quicker runoff
response, and the flatter, more surface water dominated Jean-Marie basin
yields a damped runoff response, on average.
Within Jean-Marie, both the timing and volume of peak flows derived from
snowmelt are well captured in 1998; however, volume is underpredicted in 1997
and 1999 for the average streamflow simulation. The parameter uncertainty
bounds generally enclose the observed spring melt hydrograph, except in 1999,
when the timing of the melt peak is simulated later than was observed.
Snowmelt is controlled by a degree-day snowmelt function in WATFLOOD, using
temporally constant snowmelt parameters. Parameter stationarity likely
results in an inadequate description of the inter-annual variability in
energy balance and snowpack ripening dynamics within the model. All
simulations have difficulty capturing the volume of the snowmelt recession
limb, which may be caused by the parameterization of baseflow and fen
responses in this sub-basin. Based on previous studies (Connon et al., 2015),
it has been suggested that bog and fen complexes are likely one of the
primary drivers of hydrograph timing and shape due to their ability to
dynamically alter drainage pathways, particularly in this region. In 1997,
following a significant melt event, all simulations in Jean-Marie exhibited
higher than observed recession limb flows, indicating runoff was slow to
drain and storages were too high. This could be an indication of WATFLOOD's
inability to capture the dynamic flow paths occurring within Jean-Marie's
extensive fen network. This same recession limb discrepancy does not occur in
Blackstone, where there are much fewer fens, and bogs would remain
hydraulically isolated even during wetter conditions (Connon et al., 2015).
In Blackstone, the recession limb hydrograph is well simulated across all
inputs; however, peak flows (with the exception of the 1997 snowmelt) are
generally underestimated. Post-freshet, simulations adequately capture the
timing of rainfall events, but (with the exception of 1997 in Jean-Marie)
consistently underestimate the magnitude of the peaks. This underestimation
is most evident when all simulations generated a very limited streamflow
response to an early October rainfall event in 1998, underestimating the
observed peak flow by approximately 50 % (Jean-Marie) and 75 %
(Blackstone). These results may point to inadequate precipitation forcing due
to the climate station proximity, high spatial variability of rainfall, and
inadequate soil moisture parameterization, or could be an unintended
side-effect of using NSE in our calibration (Gupta et al., 2009).
Most interesting is the similarity of the streamflow simulations among the
different δ18Oppt products, further assessed by
Kendall's tau coefficient (τ). In Jean-Marie, τ ranges between
0.92 (REMOiso vs. static) and 0.97 (KPN43 vs. static). In Blackstone τ
is more tightly constrained, ranging from 0.96 (REMOiso vs. static) to
0.98 (KPN43 vs. static). All τ values are statistically significant. It
should be noted that small deviations between mean streamflow simulations
occur during spring melt, where REMOiso-derived streamflow consistently
results in higher peaks than KPN43 and static-driven simulations. These
differences in mean streamflow, however, fall within overlapping uncertainty
bounds and are not significant outside of parameter uncertainty. Despite
significant changes to model parameters (Table S1), the resultant efficiency
statistics among the mean streamflow simulations remain nearly identical
(Table 5). Based on this analysis, we find that the three precipitation
isotope products generate statistically similar streamflow simulations. Given
the insignificant differences in streamflow response, it is only through
analysis of δ18OSF that the impact of different model
parameterizations is assessed.
Modelling δ18O in streamflow
Mean daily δ18OSF simulations and uncertainty bounds
for the KPN43, REMOiso, and static product model calibrations are displayed
in Figs. 3c and 4c. Each model calibration produces mean simulations that
capture many of the trends, but not particularly the magnitudes, present in
the observed δ18OSF record. Observed
δ18OSF show a depletion due to large influxes of
snowmelt during the spring freshets, and gradual enrichment over the summer
months due to the evaporation of surface and soil waters, occasionally
punctuated by rainfall events that may enrich or deplete
δ18OSF. During late fall and throughout the winter,
δ18OSF tends toward a more depleted, stable groundwater
composition (St Amour et al., 2005).
Though each of the model calibrations results in similar trends relative to
the observed δ18OSF record, there are notable
departures. As simulated δ18OSF
uncertainty envelopes associated with each
δ18Oppt product are, at times, non-overlapping,
differences in δ18OSF simulations can be attributed to
δ18Oppt product and, therefore, are not just an
artefact of parameter uncertainty (unlike streamflow). The dissimilarities
between δ18OSF simulations are also reflected in the
RMSE statistic (Table 5); the RMSE is larger for static-derived simulations
due to increased emphasis on periods with a higher observation density (i.e.
spring freshet), where larger offsets between simulated and observed
δ18OSF exist. The KPN43 and REMOiso calibrations
produce comparable RMSE. The KGE statistic shows only minor differences
between δ18OSF simulations given the statistic puts
more emphasis on long-term bias (Gupta et al., 2009), therefore reflecting
the fit of the overall simulation throughout the study period for this highly
seasonal basin (Kling and Gupta, 2009). Further research is required to
better understand the impacts of sporadic sampling resolution (for
δ18OSF observations) on efficiency criteria, and
consequently the objective functions. It is noted that sampling during peak
freshet was, at times, limited by accessibility during the high water stage;
therefore, some temporal gaps exist in the observed
δ18OSF record (particularly in 1999) during the period
that streamflow compositions are generally most depleted.
Differences in δ18OSF simulations within each sub-basin
are due to a combination of (1) markedly different
δ18Oppt input compositions during large precipitation
events amongst precipitation isotope products, and (2) how new water transits
through the system via the model's hydrological compartments. For this study
area, large precipitation events can be separated into (1) the accumulation
of winter snowfall and corresponding spring freshet (approximately 35 to
40 % of annual precipitation), and (2) major rainfall events occurring
post-freshet (summer and fall) (with rainfall representing approximately
60 to 65 % of annual precipitation).
No single model calibration produces consistently strong simulations of
δ18OSF during the snowmelt period. The KPN43
calibration best captures the timing and magnitude of spring freshet, but
overestimates δ18OSF (i.e. is too enriched) during the
1997 melt in Blackstone. Conversely, the static and REMOiso calibrations
capture the large depletion during the 1997 melt in Blackstone, but produce
overly depleted simulations during the 1998 and 1999 freshets – most notably
within Jean-Marie. There is a tendency for all models to simulate relatively
depleted spring freshet δ18OSF compositions. This can
be attributed to several factors: (1) overly enriched
δ18Oppt during the winter months, (2) unaccounted for
snow metamorphism processes, (3) an overestimation of direct snowmelt runoff
(i.e. streamflow volume), and (4) inaccurate antecedent composition of
δ18OSF simulated by the models just prior to the spring
melt.
Post-freshet, δ18OSF simulations are impacted by
rainfall amount and composition, and the offset between simulated
δ18OSF and δ18Oppt input at the
time of rainfall. As rainfall amount and/or this offset increases, the
resulting impact on simulated δ18OSF increases. This
highlights the importance of capturing the spatial and temporal variability
in rainfall δ18O, particularly for large and isotopically distinct
(from streamflow) events. The threshold defining a large rainfall event
varies depending on basin physiography, land cover, storage capacity, and
antecedent conditions. St Amour et al. (2005) estimate this threshold to be
≥ 40 mm within the Fort Simpson region. Such a large, isotopically
distinct rainfall event occurred on 11–12 June 1998, when approximately
7 mm fell over 2 days with an observed bulk event
δ18Oppt composition of -22.7 ‰. Both the
REMOiso and static products reasonably capture this event (-20.9 and
-20.1 ‰, respectively, across the study domain); however, the KPN43
product predicted an average δ18Oppt composition of
-17.6 ‰. In Jean-Marie, where large fen networks help to moderate
rainfall–runoff response, the observed δ18OSF did not
deplete in response to this event, but rather maintained a similar pre-event
composition of around -19.17 ‰ (Fig. 3c). KPN43-driven simulations
most closely match observed δ18OSF due to the
antecedent composition of δ18OSF prior to the event,
even though the KPN43 input generated the least accurate estimate of the
depleted rainfall δ18Oppt. Conversely, in Blackstone
the 11–12 June rainfall generated a much different response in observed
δ18OSF: a sharp depletion from -19.11 to
-20.98 ‰ (Fig. 4c). In this instance, the REMOiso and static
calibrations most closely match the observed δ18OSF due
to their closer representations of the rainfall event composition. In
Blackstone, this single event results in a significant offset between
KPN43-driven δ18OSF simulations relative to those
driven by REMOiso and static products, maintained throughout 1998 and up
until the 1999 freshet resets the δ18OSF.
Precipitation-weighted δ18O of snowpack
(δ18OSNW) for KPN43, REMOiso, and static inputs from
January to the end of melt for each year of the study period. Snow water
equivalent (SWE), snowfall (gray line), and rainfall (blue line) are also
shown. δ18Oppt input-specific uncertainty bounds are
represented as the shaded regions. Diamond symbols represent
δ18OSNW observations sampled within each respective
sub-basin during the GEWEX campaign.
Throughout much of Canada and in other high-latitude climates, the spring
freshet generates a substantial portion of annual streamflow (and typically
peak annual flow) when the accumulation of solid precipitation from the
winter season melts in late spring over a period of a few weeks. It is
therefore important to understand how differences among the products impact
snowpack (and subsequently snowmelt) isotopic compositions in isoWATFLOOD.
Figure 5 shows the evolution of precipitation-weighted snowpack oxygen-18
(δ18OSNW) throughout each winter of the study period
relative to the observed snowpack compositions (hollow black diamonds). Not
surprisingly, the static snowpack compositions closely match with observed
δ18OSNW, and we note that KPN43 and REMOiso snowpacks
are more enriched. Caution should be used when comparing modelled
vs. observed data here as there is little inter-annual consistency in the
number of samples and the location where sampling took place, and no
information was provided as to how the δ18OSNW were
collected (i.e. depth-integrated or depth-dependent samples). Comparison of
like-forcing pairs between Jean-Marie and Blackstone reveals subtle spatial
differences in simulated δ18OSNW. Dissimilarities
between the three products within each basin are, however, significant.
Interestingly, REMOiso and KPN43 end of winter precipitation-weighted
δ18OSNW compositions differ by less than
0.5 ‰ in 1997–1998 and 1998–1999. REMOiso and KPN43 inputs
consistently generate significantly more enriched snowpacks relative to the
static δ18OSNW product (and much of the observed data).
On average, KPN43 is 3.3 ‰ more enriched, and REMOiso is
3.1 ‰ more enriched than end of season static
δ18OSNW. Differences in simulated
δ18OSNW among the products are partially attributed to
the poor representation of snowpack physics (i.e. fractionation resulting
from sublimation and snow metamorphism) in the current version of the
isoWATFLOOD model. The static input inadvertently accounts for some of these
processes, as the specified compositions are derived from snowpack
observation near end of winter (in late March). Uncertainty in simulated
δ18OSNW among the products is notable as well, with
static δ18OSNW uncertainty remaining relatively
constant over the winter relative to REMOiso, and particularly KPN43, where
uncertainty decreases as snowpack depth increases (Fig. 5). This is an
artefact of the parameterization of sublimation in the models. As the winter
progresses, the snowpack grows and sublimated volumes become a smaller
fraction of the total snowpack, thus decreasing the effect (and uncertainty)
that sublimation has on the volume-weighted δ18Oppt of
the snowpack. This is observed during periods when the simulated snowpack and
snow water equivalent (SWE) are larger, for example, 1998 relative to 1999
(Fig. 5).
These significant differences in simulated snowpack composition are one of
the primary reasons for offsets between KPN43, REMOiso, and static
δ18OSF simulations (Figs. 3c and 4c). Once a
δ18OSF simulation has been offset, it is not possible
to “reset” the composition in late fall as streamflow decreases to
near-zero and mass retained in the system. This can result in compounding
isotopic error (if the offset deviates from observed data) during continuous
simulation, thus highlighting the sensitivity of the tracer as a calibration
tool. Compounding error is also observed for rainfall events, but generally
to a lesser extent due to the relatively smaller durations and magnitudes
(volume contributions) of most rainfall events (relative to snowmelt) in
high-latitude regions.
Since both δ18OSF and δ18OSNW are
significantly different among δ18Oppt products, internal
water apportionment (determined by model parameterization) is also likely
impacted. Differences in hydrograph separations among the calibrated models
are explored to determine the impact δ18Oppt has on
internal water apportionment and simulation uncertainty.
Percent seasonal volume contributing to total streamflow from
surface runoff, interflow, and baseflow storages for each season. Cross
symbols represent the 5th and 95th percentiles for each forcing method, and
circle symbols signify the mean values. The combined uncertainty bounds
representing the 5th and 95th simulations from all three
δ18Oppt input types are shaded in gray.
Hydrograph component analysis and parameter distributions
Component contributions to total streamflow from surface runoff, interflow,
and baseflow storage in each season (DJF: December–January–February; MAM:
March–April–May; JJA: June–July–August; and SON:
September–October–November) derived from each δ18Oppt
product are shown in Fig. 6. Jean-Marie and Blackstone display similar trends
in internal water apportionment throughout the year, indicating generally
similar model parameterizations and hydrograph separations among the two
basins. Some seasonal differences in component separations exist, however,
which are linked to variations in basin physiography, land cover, and storage
characteristics reflected by differences in the baseflow (lzf and pwr) and
wetland parameters (kcond and theta) among basins (Table S1). Freshet and
post-freshet percent contributions to total streamflow in this study are in
agreement with those reported in previous studies. St Amour et al. (2005)
reported significant post-freshet groundwater contributions
(71 % ± 9 % and 64 % ± 10 % for Jean-Marie and
Blackstone, respectively) compared to the mean post-freshet (JJASON)
contributions we report in Fig. 6 (40–70 and 60–70 % for Jean-Marie and
Blackstone, respectively). In agreement with this, Jasechko et al. (2016)
estimate that annually 80–90 % of the Mackenzie River streamflow is
“old” water (i.e. water that has not entered the stream within the last
2.3 ± 0.8 months). Their findings also suggest that the annual
percentage of old streamflow can be higher in mountainous watersheds with
steeper slopes, such as in the FSB and specifically Blackstone, relative to
lower-gradient watersheds. Groundwater as defined by St Amour et al. (2005)
and Jasechko et al. (2016) denotes “old water”, or water residing in the
system prior to an event. In our study, groundwater is defined as baseflow in
isoWATFLOOD (Stadnyk et al., 2013) and is separate from interflow (soil water
in the unsaturated zone) and wetlands. Baseflow contributions in this study
are therefore slightly lower than those estimated from the two-component
hydrograph separation methods. Snowmelt contributions from St Amour et
al. (2005) were 21 % (±2 %) and 40 % (±4 %) of total
streamflow for Jean-Marie and Blackstone, respectively, which are in
agreement with mean spring (MAM) surface runoff contributions in our study
(20–40 %) for both basins.
Comparison of seasonal volume contributions derived from each
δ18Oppt product reveals that during spring (MAM),
REMOiso-driven simulations show more surface flow contribution to total
streamflow, with the mean volume lying above the 95th percentile volumes for
both the KPN43 and static input simulations (Fig. 6). On average, REMOiso
simulations contribute almost twice as much surface runoff to total
streamflow as KPN43 and static simulations during MAM (39 % vs. 25 and
22 %, respectively, for Jean-Marie; and similar, yet slightly larger,
percent contributions for Blackstone).
From the seasonal analysis, no other significant differences in component
contributions outside of parameter uncertainty can be attributed to the
δ18Oppt product. It is important to note, however, that
each δ18Oppt product results in different amounts of
parameter uncertainty, both seasonally and overall, as represented by the
width of the uncertainty bounds (cross symbols in Fig. 6). The variation in
uncertainty bounds between δ18Oppt products is also
visible in Fig. 3 through Fig. 5. The REMOiso input yields the largest amount
of uncertainty in total streamflow, also reflected in the relatively larger
amounts of uncertainty in surface water and baseflow component contributions
(Fig. 6). Conversely, KPN43 and static inputs generate similar or slightly
larger uncertainty in interflow (soil water) contributions relative to
REMOiso and lower uncertainty surrounding surface and baseflow contributions,
and overall total streamflow. These differences in uncertainty are attributed
to the number and characteristics of behavioural parameters retained for each
δ18Oppt input, which originate due to distinctions in
magnitude and variability (both spatial and temporal) among
δ18Oppt products.
Probability distributions for selected parameters (Table 5), as
indicated in the bottom right corner of each panel. Parameters are from
behavioural simulations, and (a), (b), (e), and
(f) have been weighted to the land cover distribution within Jean-Marie and Blackstone, as outlined in Table 1. (c) and
(d) are river class
parameters within Jean-Marie, and (g) and (h) contain river
class parameters for Blackstone.
Further demonstrated by parameter probability distributions (Fig. 7), the
three calibrations result in noteworthy differences in behavioural
parameters. We do not display these distributions to comment definitively on
parameter identifiability because, as previously noted, the number of
evaluations was insufficient for that purpose. Rather, we introduce this
analysis to further explore how model parameterization is impacted by
δ18Oppt input. The selected parameters (Table 4)
influence evaporation (f-ratio), surface runoff during snowmelt (akfs,
base), upper and lower zone storage (retn), interflow (retn), and baseflow
(lzf, pwr). REMOiso parameter distributions more often than not differ from
KPN43 and static parameter distributions. Although dissimilarities between
KPN43 and static parameter distributions exist, these are typically not as
prevalent as differences with REMOiso-derived distributions. This echoes the
findings from Fig. 7 that KPN43 and static-derived component contributions
are more similar than those derived from REMOiso, which may very well be due
to the increased spatial and temporal variability of the REMOiso
δ18Oppt product. Though we cannot verify the
correctness of the REMOiso δ18Oppt given the absence of
daily precipitation isotope observations, differences among inputs imply that
temporal resolution of δ18Oppt plays a role in the
parameterization of a model and the resultant hydrograph separation.
Differences in surface water contributions during snowmelt between REMOiso,
KPN43, and static inputs are likely derived from differences in the akfs and
base parameters. Parameter distributions from REMOiso are significantly
different (as verified through Kolmogorov–Smirnov testing of distributions)
than the KPN43 and static input distributions for these parameters (Fig. 7b
and f). Lower akfs values represent decreased infiltration and increased
surface runoff during snowmelt, which corresponds to REMOiso's increased
surface water contributions to total streamflow during spring (MAM).
Dissimilarities in baseflow contributions among δ18Oppt
inputs are influenced by differences in the lzf and pwr parameters
(Fig. 7c, d, and g–h), which have a large impact on the quantity of baseflow
and the slope of the recession limb of the hydrograph. Wider uncertainty
bounds for REMOiso relative to KPN43 and static calibrations within
Blackstone (Fig. 6f), and for all models during fall and winter within Jean-Marie (Fig. 6c), are likely due to the wider range of behavioural values for
the pwr parameter, specifically the inclusion of lower values which results
in longer, more drawn-out recession limbs. It appears that choice of
precipitation isotope product influences parameter distributions in
isoWATFLOOD, which in turn alters internal water apportionment. In the
tracer-aided modelling community, this has significant implications for
hydrograph separation and any associated transit time analyses, both of which
will be influenced by choice (resolution) of δ18Oppt
product.
Conclusions
This study used three types of precipitation isotope products as
δ18Oppt input to a tracer-aided hydrological
model (isoWATFLOOD) to investigate the impact differing spatial and temporal
resolutions have on simulation of streamflow, isotopic composition of
streamflow, internal hydrograph separations, and model parameterization and
corresponding parameter uncertainty. Our study found that choice of
precipitation isotope product (δ18Oppt)
did not impact simulation of total streamflow, or the achieved efficiencies
of streamflow simulation;
impacted the internal apportionment of water, thereby impacting hydrograph separations;
impacted model parameterization, and therefore simulation uncertainty; and
impacted the variability of simulated δ18OSF, most
noticeably when event compositions differed significantly from streamflow
composition (e.g. snowmelt and large rainfall events).
Of the 30 000 simulations performed for each precipitation isotope product
forcing, only 10 % or less were behavioural for each input type. Due to
the wide range of behavioural parameter values (Table S1), however, we are
confident that the approach used was sufficient to characterize parameter
uncertainty. Not unexpectedly, this finding also indicates that 30 000 model
evaluations were not sufficient to quantify parameter identifiability in this study.
Although total simulated streamflow was not significantly affected by choice
of δ18Oppt input, δ18OSF
simulations and the internal apportionment of water (surface flow, interflow,
and baseflow) were significantly impacted here. Significant differences in
internal water apportionment among the models were diagnosed via
δ18O uncertainty. Variation between models was greatest during the
freshet period, where significantly different snowpack compositions were
simulated among the different precipitation isotope products. The
highest-resolution (REMOiso, daily) input resulted in significantly different
parameter distributions and seasonal hydrograph separations than the other
two (monthly and semi-annual) inputs. These findings have direct implications
for hydrograph separation, and simulated water transit times. In this study,
we found that choice of δ18Oppt input directly impacted
model parameterization, and for this reason, studies should account for both
input and parameter uncertainty. Also highlighted was the significance of the
snowpack and melt dynamics in tracer-aided models applied in high-latitude
regions, resulting in high seasonal uncertainty and indicating more research
is warranted to improve process representation. Use of a tracer-aided model
afforded an examination of internal model dynamics resulting from specific
parameterizations, allowing us to assess the realism of individual
simulations as opposed to their efficacy alone.
This study demonstrated that direct quantification of model equifinality was
possible using tracer-aided models, and furthermore, we demonstrated that
this equifinality was not diagnosable via simulation of streamflow. We have
achieved an understanding of how tracer-aided models, like isoWATFLOOD, can
be used in data-sparse regions, with limited input data (including
δ18Oppt observations), and that despite these
limitations, these models can still be of value. Regarding the usefulness of
precipitation isotope products in regions with limited observations, both the
static and REMOiso inputs require existing δ18Oppt
observations (i.e. from CNIP) to either define or bias-correct the input,
limiting their use for certain applications. If these data are not available,
the KPN43 input provided reasonable results without the need for additional
observations. The existence of CNIP (and other precipitation isotope
networks) was crucial to the development, validation, and bias correction of
existing δ18Oppt products. Attaining an understanding
of how δ18Oppt input uncertainty impacts simulated
model output is important when calibrated models are used to assess
climate-driven or land-use-driven impacts on streamflow in remote,
data-sparse, high-latitude regions.
For use in tracer-aided modelling, precipitation isotope products should
capture both the event-based variability and seasonality of precipitation
isotopes to reproduce realistic δ18OSF variability.
Higher-resolution δ18Oppt inputs (e.g. REMOiso, daily)
were able to capture event-specific compositions that, when significantly
different from δ18OSF, tended to cause significant
deviations from the δ18OSF derived from monthly and
semi-annual (i.e. static) inputs. Unfortunately, we could not verify the
correctness of the higher-resolution product (i.e. REMOiso) in this study due
to the sporadic sampling of isotopes in precipitation observations. Static
and seasonal precipitation isotope products missed event-specific isotopic
variation occurring as a result of heavy rainfall events, which require
increased temporal resolution (e.g. daily δ18Oppt
inputs from REMOiso; but perhaps weekly input would suffice) to resolve
rainfall event compositions. In seasonal environments, precipitation isotope
products must capture the transition from rainfall to snowfall, and from snow
accumulation to snowmelt to sufficiently model δ18OSF.
In this study, both static and monthly inputs adequately captured
δ18OSF variability at the basin outlet, perhaps the
result of the unique seasonality in high-latitude regions. Spatial
variability was detected across the study region in
δ18Oppt inputs, and can be represented by distributed
tracer-aided models, like isoWATFLOOD. There is reason to suspect that the
variability in (both spatial and temporal) precipitation isotope inputs
influences model parameterization; therefore, spatial variability should be
preserved to derive the most representative model of a given region.
This work highlighted the power of tracer-aided modelling to inform and
quantify equifinality in hydrological simulation, helping modellers to work
towards reducing modelling uncertainty. Although more work is required to
assess and understand parameter identifiability, our analysis showed that
selection of precipitation isotope (δ18Oppt) product
had direct implications for model parameterization, and that input
uncertainty should be considered in future studies.
Future directions
Distributed hydrological models, such as WATFLOOD, are complex with large
numbers of parameters; therefore, it is important as a community to work
toward conducting comprehensive studies that focus on input data uncertainty
and parameter identifiability. In the tracer-aided modelling community, this
includes uncertainty from precipitation isotope products and their varying
spatial and temporal resolutions. Ideally, further studies should be
conducted in well-instrumented basins where δ18Oppt
input can be better characterized using observed data at higher spatial, and
most importantly, temporal resolutions. Several key questions warranting more
detailed investigation include the following: (1) are precipitation isotope
products adequate alternatives in place of δ18Oppt
observations; (2) are there specific subsets of model parameters that are
more sensitive to choice of precipitation isotope product; and (3) how do (if
at all) parameters compensate for compounding model error? Unfortunately, at
least within Canada, a well-instrumented watershed at the regional scale does
not yet exist, pointing to the importance of implementing additional (or
enhancing current) iso-hydro-meteorological monitoring networks.
Not unexpectedly, the RCM-driven precipitation isotope product in this study,
REMOiso, exhibited some bias and needed correction prior to application. More
studies are needed to better understand the nature of this bias, and the most
appropriate bias-correction methods, which can be done using observations from the
CNIP database at a monthly resolution. Due to the lack of high-resolution
δ18Oppt observations in Canada, however, daily or
weekly validation is not yet possible. Additionally, the suitability and
performance of other isotope-enabled RCMs for use in Canada, and elsewhere,
should be explored.
Lastly, as a tracer-aided hydrologic community we need to push for the
sustained monitoring of isotopes in precipitation and streamflow that are
required to inform our models and improve uncertainty assessment. This study
elucidated the impact that discontinuous observations can have on quantifying
model uncertainty, which would only be further exasperated by the absence of
observations altogether. In Canada, a concerted effort is needed by the
government to protect and sustain our observation networks, which are
required for improved prediction in remote regions for climate and hydrologic
change detection.
Ensemble
simulations of the time series streamflow (total and components) and
oxygen-18 in precipitation (KPN43), streamflow, interflow, groundwater and
snowpack used in this study are publicly accessible through
http://www.hydroshare.org/resource/365bb56b64f74412b2525b3cfc8d73bc.
Matlab code to generate KPN43 time series can be accessed through
https://mspace.lib.umanitoba.ca/xmlui/handle/1993/31946 or by
contacting Carly Delavau at carly.delavau@gov.mb.ca. δ18O
observations were collected as part of the Mackenzie Global Energy and Water
Experiment (GEWEX) study (Stewart et al., 1998). Please contact the authors
of this study to request access to this dataset. Raw REMOiso output can be
requested by contacting Christophe Sturm at christophe.sturm@geo.su.se.
The Supplement related to this article is available online at doi:10.5194/hess-21-2595-2017-supplement.
Carly J. Delavau developed model code to generate Kpn43 δ18Oppt
input, perform Monte Carlo simulations, and process the corresponding output.
Tricia Stadnyk and Tegan Holmes developed and enhanced isoWATFLOOD code for
the version of isoWATFLOOD used in this study. Carly J. Delavau performed the
analysis presented in this paper, with assistance from Tricia Stadnyk.
Carly J. Delavau prepared the manuscript with contributions from all
co-authors. Tricia Stadnyk edited the manuscript based on reviewer comments
with help from Carly J. Delavau. Carly J. Delavau and Tegan Holmes completed
amendments to figures.
The authors declare that they have no conflict of interest.
Acknowledgements
The authors would like to acknowledge the contributions of the late
Peter Rasmussen to this work, whose invaluable advice and mentorship is
sorely missed but largely shaped the outcome of this research. The authors
would like to acknowledge K. Sturm for provision of the REMOiso data utilized
in this study. Additional thanks go to N. Kouwen for direction and input on
WATFLOOD modelling. CNIP is made possible through the help of the Canadian
Air and Precipitation Monitoring Network (CAPMoN) for sample collection,
Kaz Higuchi and Dave MacTavish in particular, the Environmental Isotope
Laboratory at the University of Waterloo for sample analysis, and Tom Edwards
for initiating and maintaining the network. We would also like to acknowledge
Water Survey of Canada for the hydrometric data, and Dan McKinney for
provision of the ANUSPLIN data used in this study. Finally, we would like to
acknowledge the contributions of our reviewers whose valuable input has
improved this paper significantly. This research was partially funded by a
Natural Sciences and Engineering Research Council (NSERC) Alexander Graham
Bell Canada Graduate Scholarship (CGS-D).
Edited by: C. Stumpp Reviewed by: C. Birkel and one anonymous
referee
References
Aylsworth, J. M. and Kettles, I. M.: Distribution of peatlands, in: The
Physical Environment of the Mackenzie Valley, Northwest Territories: A Base
Line for the Assessment of Environmental Change, edited by: Dyke, L. D. and
Brooks, G. R., Geological Survey of Canada Bulletin 547, Geological Survey of
Canada, Ottawa, 49–55, 2000.Beven, K. and Binley, A.: The future of distributed models: Model calibration
and uncertainty prediction, Hydrol. Process., 6, 279–298, 10.1002/hyp.3360060305, 1992.Birkel, C. and Soulsby, C.: Advancing tracer-aided rainfall–runoff modelling:
a review of progress, problems and unrealised potential, Hydrol. Process.,
29, 5227–5240, 10.1002/hyp.10594, 2015.Birkel, C., Dunn, S. M., Tetzlaff, D., and Soulsby, C.: Assessing the value of
high-resolution isotope tracer data in the stepwise development of a lumped
conceptual rainfall–runoff model, Hydrol. Process., 24, 2335–2348, 10.1002/hyp.7763, 2010a.Birkel, C., Tetzlaff, D., Dunn, S. M., and Soulsby, C.: Towards a simple dynamic
process conceptualization in rainfall–runoff models using multi-criteria
calibration and tracers in temperate, upland catchments, Hydrol. Process.,
24, 260–275, 10.1002/hyp.7478, 2010b.Birkel, C., Tetzlaff, D., Dunn, S. M., and Soulsby, C.: Using time domain and
geographic source tracers to conceptualize streamflow generation processes in
lumped rainfall-runoff models, Water Resour. Res., 47, W02515, 10.1029/2010WR009547, 2011.Birkel, C., Soulsby, C., and Tetzlaff, D.: Developing a consistent process-based
conceptualization of catchment functioning using measurements of internal state
variables, Water Resour. Res., 50, 3481– 3501, 10.1002/2013WR014925, 2014.Birks, S. J. and Edwards, T. W. D.: Atmospheric circulation controls on
precipitation isotope-climate relations in western Canada, Tellus B, 61,
566–576, 10.1111/j.1600-0889.2009.00423.x, 2009.Birks, S. J. and Gibson, J. J.: Isotope hydrology research in Canada, 2003–2007,
Can. Water Resour. J., 34, 163–176, 10.4296/cwrj3402163, 2009.Carpenter, T. M. and Georgakakos, K. P.: Intercomparison of lumped versus
distributed hydrologic model ensemble simulations on operational forecast
scales, J. Hydrol., 329, 174–185, 10.1016/j.jhydrol.2006.02.013, 2006.
Connon, R. F., Quinton, W. L., Craig, J. R., Hanisch, J., and Sonnentag, O.:
The hydrology of interconnected bog complexes in discontinuous permafrost
terrain, Hydrol. Process., 29, 3831–3847, 2015.Coulibaly, P., Samuel, J., Pietroniro, A., and Harvey, D.: Evaluation of
Canadian National Hydrometric Network density based on WMO 2008 standards, Can.
Water Resour. J., 38, 159–167, 10.1080/07011784.2013.787181, 2013.Delavau, C.: Development of Precipitation δ158O Isoscapes for
Canada and Application within a Tracer-Aided Hydrological Model, availaable
at: https://mspace.lib.umanitoba.ca/xmlui/handle/1993/31946 (last access: 23 May 2017), 2017a.
Delavau, C.: Data corresponding to Delavau et al. (2017) study, available at:
http://www.hydroshare.org/resource/365bb56b64f74412b2525b3cfc8d73bc (last access: 23 May 2017), 2017b.Delavau, C., Chun, K. P., Stadnyk, T., Birks, S. J., and Welker, J. M.: North
American precipitation isotope (δ18O) zones revealed in time series
modelling across Canada and northern United States, Water Resour. Res., 51,
10.1002/2014WR015687, 2015.Dietermann, N. and Weiler, M.: Spatial distribution of stable water isotopes
in alpine snow cover, Hydrol. Earth Syst. Sci., 17, 2657–2668, 10.5194/hess-17-2657-2013, 2013.Dunn, S. M., Freer, J., Weiler, M., Kirkby, M. J., Seibert, J., Quinn, P. F.,
Lischeid, G., Tetzlaff, D., and Soulsby, C.: Conceptualization in catchment
modelling: simply learning?, Hydrol. Process., 22, 2389–2393, 10.1002/hyp.7070, 2008.
Ecoregions Working Group: Ecoclimatic regions of Canada, first approximation,
in: Ecoregions Working Group of the Canada Committee on ecological land
classification, Ecological Land Classification Series 23, Sustainable Development
Branch, Canadian Wildlife Service, Environment Canada, Ottawa, Ontario, 1989.Fenicia, F., McDonnell, J. J., and Savenije, H. H. G.: Learning from model
improvement: On the contribution of complementary data to process understanding,
Water Resour. Res., 44, W06419, 10.1029/2007WR006386, 2008.Gat, J. R., Bowser, C. J., and Kendall, C.: The contribution of evaporation
from the Great Lakes to the continental atmosphere: estimate based on stable
isotope data, Geophys. Res. Lett., 21, 557–560, 10.1029/94GL00069, 1994.Gibson J. J. and Prowse, T. D.: Isotopic characteristics of ice cover in a
large northern river basin, Hydrol. Process., 13, 2537–2548, 10.1002/(SICI)1099-1085(199911)13:16<2537::AID-HYP940>3.0.CO;2-A, 1999.
Green, W. H. and Ampt, G. A.: Studies in soil physics 1: Flow of air and
water through soils, Agr. Res., 4, 1–24, 1911.Gupta, H. V., Kling, H., Yilmaz, K. K., and Martinez, G. F.: Decomposition of
the mean squared error and NSE performance criteria: Implications for
improving hydrological modelling, J. Hydrol., 377,
10.1016/j.jhydrol.2009.08.003, 2009.
Hargreaves, G. H. and Samani, Z. A.: Estimating potential evapotranspiration,
Journal of Irrigation and Drainage Division, 108, 225–230, 1982.Her, Y. and Chaubey, I.: Impact of the numbers of observations and calibration
parameters on equifinality, model performance, and output and parameter uncertainty,
Hydrol. Process., 29, 4220–4237, 10.1002/hyp.10487, 2015.
Holmes, T. L.: Assessing the Value of Stable Water Isotopes in Hydrologic
Modeling: A Dual-Isotope Approach, MSc Thesis, Department of Civil Engineering,
University of Manitoba, Manitoba, 2016.IAEA/WMO: Global Network of Isotopes in Precipitation, The GNIP Database,
available at: http://www.iaea.org/water (last access: 4 May 2017),
2014.Jasechko, S., Kirchner, J. W., Welker, J. M., and McDonnell, J. J.: Substantial
proportion of global streamflow less than three months old, Nat. Geosci., 9,
126–129, 10.1038/ngeo2636, 2016.Kirchner, J.: Getting the right answers for the right reasons: Linking
measurements, analyses, and models to advance the science of hydrology, Water
Resour. Res., 42, W03S04, 10.1029/2005WR004362, 2006.
Klaus, J. and McDonnell, J. J.: Hydrograph separation using stable isotopes:
Review and evaluation, J. Hydrol., 505, 47–64, 2013.Kling, H., Fuchs, M., and Paulin, M.: Runoff conditions in the upper Danube
basin under an ensemble of climate change scenarios, J. Hydrol., 424–425,
264–277, 10.1016/j.jhydrol.2012.01.011, 2012.
Kling, H. V. and Gupta, H.: Decomposition of the mean squared error and NSE
performance criteria: Implications for improving hydrological modelling, J.
Hydrol., 377, 80–91, 2009.Kouwen, N.: WATFLOOD/WATROUTE Hydrological Model Routing and Flood
Forecasting System, User's Manual, University of Waterloo, Waterloo, ON,
http://www.watflood.ca (last access: 4 May 2017), 2016.Kouwen, N., Soulis, E. D., Pietroniro, A., Donald, J., and Harrington, R. A.:
Grouped response units for distributed hydrological modelling, J. Water Resour.
Pl. Manage., 119, 289–305, 10.1061/(ASCE)0733-9496(1993)119:3(289), 1993.Kuczera, G.: Improved parameter inference in catchment models: 2. Combining
different kinds of hydrologic data and testing their compatibility, Water
Resour. Res., 19, 1163–1172, 10.1029/WR019i005p01163, 1983.Kuczera, G. and Mroczkowski, M.: Assessment of hydrologic parameter uncertainty
and the worth of multiresponse data, Water Resour. Res., 34, 1481–1489,
10.1029/98WR00496, 1998.Lykoudis, S. P., Argiriou, A. A., and Dotsika, E.: Spatially interpolated time
series of δ18O in Eastern Mediterranean precipitation, Global Planet.
Change, 71, 150–159, 10.1016/j.gloplacha.2009.09.004, 2010.McKenney, D. W., Hutchinson, M. F., Papadopol, P., Lawrence, K., Pedlar, J.,
Campbell, K., and Owen, T.: Customized Spatial Climate Models for North America,
B. Am. Meteorol. Soc., 92, 1611–1622, 10.1175/2011BAMS3132.1, 2011.
McKillop, R., Kouwen, N., and Soulis, E. D.: Modeling the rainfall-
runoff response of a headwater wetland, Water Resour. Res., 35, 1165–1177, 1999.Mesinger, F., DiMego, G., Kalnay, E., Mitchell, K., Shafran, P. C., Ebisuzaki,
W., Jović, D., Woollen, J., Rogers, E., Berbery, E. H., Ek, M. B., Fan, Y., Grumbine,
R., Higgins, W., Li, H., Lin, Y., Manikin, G., Parrish, D., and Shi, W.: North
American regional reanalysis, B. Am. Meteorol. Soc, 87, 343–360,
10.1175/BAMS-87-3-343, 2006.Michaud, J. and Sorooshian, S.: Comparison of simple versus complex distributed
runoff models on a midsized semiarid watershed, Water Resour. Res., 30, 593–605,
10.1029/93WR03218, 1994.
Moran, T. A., Marshall, S. J., Evans, E. C., and Sinclair, K. E.: Altitudinal
gradients of stable isotopes in lee-slope precipitation in the Canadian Rocky
Mountains, Arct. Antarct. Alp. Res., 39, 455–467, 2007.
Moriasi, D. N., Arnold, J. G., Van Liew, M. W., Bingner, R. L., Harmel, R. D.,
and Veith, T. L.: Model evaluation guidelines for systematic quantification of
accuracy in watershed simulations, T. ASABE, 50, 885–900, 2007.Nash, J. E. and Sutcliffe, J. V.: River flow forecasting through conceptual
models part I – A discussion of principles, J. Hydrol., 10, 282–290,
10.1016/0022-1694(70)90255-6, 1970.
Noone, D. and Sturm, C.: Comprehensive dynamical models of global and regional
water isotope distributions, in: Isoscapes: Understanding movement, pattern,
and process on Earth through isotope mapping, edited by: West, J. B., Bowen,
G. J., Dawson, T. E., and Tu, K. P., Springer, London, 195–219, 2010.Penna, D., Ahmad, M., Birks, S. J., Bouchaou, L., Brenčič, M., Butt, S.,
Holko, L., Jeelani, G., Martínez, D. E., Melikadze, G., Shanley, J. B.,
Sokratov, S. A., Stadnyk, T., Sugimoto, A., and Vreča, P.: A new method of
snowmelt sampling for water stable isotopes, Hydrol. Process., 28, 5367–5644,
10.1002/hyp.10273, 2014.Petrone, R. M., Griffis, T. J., and Rouse, W. R.: Synoptic and surface
climatology interactions in the central Canadian subarctic: normal and
El Niño seasons, Phys. Geogr., 21, 368–383,
10.1080/02723646.2000.10642715,
2000.
Philip, J. R.: An infiltration equation with physical significance, J. Soil
Sci., 77, 153–157, 1954.Prowse, T. D., Conly, F. M., Church, M., and English, M. C.: A review of
hydroecological results of the Northern River Basins study, Canada. Part 1.
Peace and Slave Rivers, River Res. Appl., 18, 429–446, 10.1002/rra.681, 2002.Seibert, J. and McDonnell, J. J.: On the dialog between experimentalist and
modeler in catchment hydrology: Use of soft data for multicriteria model
calibration, Water Resour. Res., 38, 1241, 10.1029/2001WR000978, 2002.
Singh, V. P. and Frevert, D. K.: Watershed models, CRC Taylor & Francis, Boca Raton, 2006.
Smith, A., Delavau, C., and Stadnyk, T.: Identification of geographical
influences and flow regime characteristics using regional water isotope
surveys in the lower Nelson River, Canada, Can. Water Resour. J., 40, 23–35, 2015.Smith, A., Welch, C., and Stadnyk, T.: Assessment of a lumped coupled
flow-isotope model in data scarce Boreal catchments, Hydrol. Process., 30,
3871–3884,
10.1002/hyp.10835, 2016.
Stadnyk, T., St Amour, N. A., Kouwen, N., Edwards, T. W. D., Pietroniro, A.,
and Gibson, J. J.: A groundwater separation study in boreal wetland terrain:
The WATFLOOD hydrological model compared with stable isotope tracers, Isotop.
Environ. Health Stud., 41, 49–60, 2005.Stadnyk, T. A., Delavau, C., Kouwen, N., and Edwards, T. W. D.: Towards
hydrological model calibration and validation: Simulation of stable water
isotopes using the isoWATFLOOD model, Hydrol. Process., 27, 3791–3810,
10.1002/hyp.9695, 2013.Stadnyk-Falcone, T. A.: Mesoscale Hydrological Model Validation and Verification
using Stable Water Isotopes: The isoWATFLOOD Model, PhD Thesis, University of
Waterloo, Waterloo, 386 pp., http://hdl.handle.net/10012/3970 (last access: 4 May 2017), 2008.St Amour, N. A., Gibson, J. J., Edwards, T. W. D., Prowse, T. D., and Pietroniro,
A.: Isotopic time-series partitioning of streamflow components in wetland
dominated catchments, lower Liard river basin, Northwest Territories, Canada,
Hydrol. Process., 19, 3357–3381, 10.1002/hyp.5975, 2005.Stewart, R. E., Leighton, H. G., Marsh, P., Moore, G. W. K., Ritchie, H.,
Rouse, W. R., Soulis, E. D., Strong, G. S., Crawford, R. W., and Kochtubajda,
B.: The Mackenzie GEWEX Study: the water and energy cycles of a major North
American river basin, B. Am. Meteorol. Soc., 79, 2665–2684,
10.1175/1520-0477(1998)079<2665:TMGSTW>2.0.CO;2,
1998.Sturm, K., Hoffmann, G., Langmann, B., and Stichler, W.: Simulation of δ18O
in precipitation by the regional circulation model REMOiso, Hydrol. Process.,
19, 3425–3444, 10.1002/hyp.5979, 2005.Sturm, K., Hoffmann, G., and Langmann, B.: Simulation of the stable water
isotopes in precipitation over South America: Comparing regional to global
circulation models, J. Climate, 20, 3730–3750, 10.1175/JCLI4194.1, 2007.Taylor, S., Feng, X., Kirchner, J. W., Osterhuber, R., Klaue, B., and Renshaw,
C. E.: Isotopic evolution of a seasonal snowpack and its melt, Water Resour. Res.,
37, 759–769, 10.1029/2000WR900341, 2001.Taylor, S., Feng, X., Williams, M., and McNamara, J.: How isotopic fractionation
of snowmelt affects hydrograph separation, Hydrol. Process., 16, 3683–3690,
10.1002/hyp.1232, 2002.Tetzlaff, D., Uhlenbrook, S., Eppert, S., and Soulsby, C.: Does the incorporation
of process conceptualization and tracer data improve the structure and performance
of a simple rainfall-runoff model in a Scottish mesoscale catchment?, Hydrol.
Process., 22, 2461–2474, 10.1002/hyp.6841, 2008.
Tetzlaff, D., Buttle, J., Carey, S. K., McGuire, K., Laudon, H., and Soulsby,
C.: Tracer-based assessment of flow paths, storage and runoff generation in
northern catchments: a review, Hydrol. Process., 29, 3475–3490, 10.1002/hyp.10412, 2015.Xi, X.: A Review of Water Isotopes in Atmospheric General Circulation Models:
Recent Advances and Future Prospects, Int. J. Atmos. Sci., 2014,
250920, 16 pp., 10.1155/2014/250920, 2014.Zhou, S., Nakawo, M., Hashimoto, S., and Sakai, A.: The effect of refreezing
on the isotopic composition of melting snowpack, Hydrol. Process., 22, 873–882,
10.1002/hyp.6662, 2008.