1. Introduction
Precipitation forcing is critical for hydrological modeling and has a strong impact on the accuracy of simulated river flows (Fekete et al. 2004; Lopes 1996). Precipitation measured by traditional weather stations often provides relatively accurate estimates at a few locations of a given region. However, measured precipitation is often biased because of measurement errors due to gauge undercatch. Consequently, measured precipitation is known to underestimate the real precipitation by a factor up to 10% for liquid precipitation (Adam and Lettenmaier 2003) and 100% for solid precipitation (Goodison et al. 1998; Yang et al. 1998, 2000). In many parts of the world, weather station density is low. Recent studies have assessed the impact of precipitation station density on the accuracy of predictions by watershed models (Andréassian et al. 2001; Faurès et al. 1995). Duncan et al. (1993) studied the impact of weather station density on the accuracy of the flow predictions of a rural watershed in southern Quebec and found this influence to be very strong, not only in terms of the accuracy of total runoff estimates, but also on the accuracy of peak flow and simulated peak time. Chaplot et al. (2005) studied the effect of the accuracy of spatial rainfall information on river flow modeling for two small watersheds located in the United States and found that a small station density often led to inaccurate discharge estimates. In the light of these studies, it is clear that the density of weather stations has an impact on hydrological model predictions.
The spatial interpolation of individual weather stations provides a good approximation of precipitation in an area of interest. However, in regions with sparse weather station coverage, the interpolated precipitation fields will provide only a rougher estimate of precipitation levels. In fact, spatial interpolation methods always introduce some artifacts in interpolated datasets, and it is difficult to verify their realism in regions with low weather station densities (Daly 2006; Tozer et al. 2012). Therefore, in such regions, interpolated precipitation and temperature are generally considered unreliable for hydrological modeling (Mizukami and Smith 2012).
Reanalyses may represent a good alternative dataset of precipitation and temperature data for regions where weather stations are sparsely distributed or nonexistent. A reanalyses is a meteorological model experiment in which global observation datasets are assimilated to produce a consistent set of meteorological data, typically at the global scale. Reanalyses use a constant data assimilation scheme and numerical forecasting model, which ingests millions of available observations at a given time step over a given period (Dee et al. 2011; Saha et al. 2010). The observation sources include, but are not limited to, radiosonde, satellite, buoy, aircraft, and ship reports. The reanalysis framework provides a dynamically consistent estimate of the climate state at each time step (Mesinger et al. 2006; Rienecker et al. 2011). In addition, most reanalyses offer a global coverage and commonly span three or more decades. They also provide up to hundreds of climate variables (Kalnay et al. 1996; Rienecker et al. 2008; Uppala et al. 2005; Wang et al. 2011). Observations used in reanalyses vary because the observational network is constantly changing. Changes in the mix of observations over the duration of each reanalysis can produce spurious trends and artificial variability. Fortunately, over the years, with improvements to numerical modeling techniques, data assimilation techniques as well as increases in computing power, reanalysis products with increased spatial resolution, improved accuracy, and smaller biases have become available, and this trend is expected to continue in the future. Commonly, reanalysis products are used to validate climate models in regions where weather stations are not available (Shiu et al. 2012; Sillmann et al. 2013a,b).
Recent studies have also examined the potential of precipitation and temperature data from reanalyses for hydrological studies (Choi et al. 2009; Vu et al. 2012; Woo and Thorne 2006). Fuka et al. (2014) tested the value of Climate Forecast System Reanalysis (CFSR) precipitation and temperature as weather inputs for hydrological modeling of five watersheds representing different hydroclimatic regimes in the United States. They found that using CFSR precipitation and temperature data to force a watershed model provides river discharge simulations that are as good as or better than when the river discharge model is forced with traditional weather gauging stations that are more than 10 km away from the watershed. More recently, Essou et al. (2016) used precipitation and temperature series from the North American Regional Reanalysis (NARR), European Centre for Medium-Range Weather Forecasts (ECMWF) interim reanalysis (ERA-Interim), CFSR, and Modern-Era Retrospective Analysis for Research and Applications (MERRA) and from the gridded Santa Clara observations to calibrate a lumped hydrological model and to simulate river flows over 370 watersheds in the continental United States. They found that the simulated river flows using NARR forcing were as good as when gridded observations were used. Moreover, the Nash–Sutcliffe values of the river flows simulated using the other three reanalyses were equal to those from the gridded observations, with the exception of the humid continental and subtropical regions, where precipitation seasonality is not well reproduced by the three reanalyses.
The objective of this study is to compare the accuracy of river flows over Canada simulated by a watershed model using precipitation and temperature estimates from reanalyses and gridded observations, as a function of the density of surface weather stations. To achieve this goal, 316 Canadian watersheds with different weather station densities and located in three climatic regions are considered. Three state-of-the-art atmospheric reanalyses—ERA-Interim, CFSR, and MERRA—and one gridded observations database over Canada—Natural Resources Canada (NRCan)—are used. For watersheds with low weather station densities, reanalysis-based estimates are expected to be more reliable, and therefore more efficient, for hydrological modeling than the gridded observations.
2. Watersheds of interest and data
a. Watersheds of interest
In this study, we consider 316 watersheds derived from the Canadian Model Parameter Experiment (CANOPEX) database (Arsenault et al. 2016). The size of the watersheds varies between 460 and 127 635 km2. For the analysis, the watersheds are distributed into three climatic regions according to the Köppen–Geiger climate classification (Kottek et al. 2006): mountain (144 watersheds), boreal (149 watersheds), and humid continental or Atlantic Canada (23 watersheds; Fig. 1a). The mean annual precipitation of the watersheds varies between 1 and 6.5 mm day−1 and is more abundant over the coastal regions (Fig. 1b). The mean annual temperature varies between −6° and 7°C and decreases as we go from south to north (Fig. 1c).
(a) The 316 watersheds of interest according to three climatic regions, (b) 1979–2010 mean annual precipitation (mm day−1), and (c) 1979–2010 mean annual temperature (°C) based on estimates from NRCan gridded observations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
(a) The 316 watersheds of interest according to three climatic regions, (b) 1979–2010 mean annual precipitation (mm day−1), and (c) 1979–2010 mean annual temperature (°C) based on estimates from NRCan gridded observations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
(a) The 316 watersheds of interest according to three climatic regions, (b) 1979–2010 mean annual precipitation (mm day−1), and (c) 1979–2010 mean annual temperature (°C) based on estimates from NRCan gridded observations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
b. Data
1) Meteorological data
Daily precipitation and temperature data from NRCan and reanalyses are considered. NRCan is a 10-km gridded database developed by Natural Resources Canada from the interpolation of daily precipitation and maximum and minimum air temperature data over the period of 1950–2010 using thin-plate-smoothing splines [Australian National University spline interpolation (ANUSPLIN); Hopkinson et al. 2011; Hutchinson 1995, 2004; Hutchinson et al. 2009]. The number of active Environment and Climate Change Canada weather stations used varies from 2000 to 3000 for precipitation and from 1500 to 3000 for air temperature.
ERA-Interim is the latest global reanalysis produced by the ECMWF (Dee et al. 2011). It covers the period from 1979 to present and uses a four-dimensional variational data assimilation (4DVAR) approach. The observations assimilated before 2002 are derived mainly from the data used for ERA-40 (Uppala et al. 2005). ERA-Interim is updated in near–real time using data from the operational ECMWF forecast system (Dee et al. 2011). The temperature from ERA-Interim results from the assimilated surface temperature, while precipitation is produced by the weather forecast model. The horizontal resolution of ERA-Interim is 0.75° × 0.75° (about 80 km × 80 km).
The CFSR is produced by the National Centers for Environmental Prediction (NCEP). It is the first reanalysis produced from a coupled climate atmosphere–ocean–land surface system with an interactive sea ice component. It covers the period from 1979 to present and uses a three-dimensional variational data assimilation (3DVAR) approach (Saha et al. 2010). CFSR assimilates satellite radiance data rather than estimated temperature and humidity values (Wang et al. 2011). CFSR uses the Noah land surface model, which is forced with the NOAA pentad Climate Prediction Center (CPC) Merged Analysis of Precipitation (Xie and Arkin 1997) and the CPC unified daily gauge analysis (Wang et al. 2011) instead of using the precipitation generated by the atmospheric model, which is considered too biased (Saha et al. 2010). The horizontal resolution of CFSR is 0.31° (longitude) × 0.31° (latitude) (about 35 km × 35 km).
MERRA is developed by the Global Modeling and Assimilation Office (GMAO) of the National Aeronautics and Space Administration (NASA) in order to maximize the use of GMAO satellite observations in a climate context and to improve the closure of the hydrological cycle (Rienecker et al. 2011). MERRA covers the satellites era (from 1979 to present) and is generated from version 5.2.0 of the Goddard Earth Observing System (GEOS) atmospheric model and a data assimilation system based on a 3DVAR approach. The data assimilation system (DAS), the input data flux, and their sources, observations, and error statistics are well documented in Suarez et al. (2008). The main specificity of MERRA consists in the use of an incremental analysis update (IAU) procedure to improve the closure of the water budget. The horizontal resolution of MERRA is ⅔° (longitude) × ½° (latitude) (about 75 km × 55 km).
2) Hydrometric data
Daily mean river flow data for each of the 316 watersheds from the Hydroclimatological Data Retrieval Program (HYDAT) database (Coulibaly et al. 2013; Winkler 1993) are used. The HYDAT database contains data from 7000 hydrometric stations across Canada. A description of the databases used in this work is presented in Table 1.
Description of the databases used in this study.
3. Methodology
The mean annual precipitation and temperature from reanalyses are compared to those from NRCan to determine the systematic biases present. Each database is then used as input data for the calibration of a lumped hydrological model that simulates river flows. The accuracy of the simulated flows is then assessed with observations over a validation period. Simulated river flows are sorted according to the density of weather stations for each watershed. The period analyzed varies from one watershed to another according to the common period between hydrometric and meteorological data within the period of 1979–2010. The shortest period is 1979–89 and the longest one is 1979–2010.
a. Data comparison
The precipitation and temperature data over each watershed are computed using the Thiessen polygon method (Thiessen 1911). Differences between reanalyses and NRCan mean seasonal temperature are assessed using bias statistics. Relative values—relative bias (RBIAS)—are computed for precipitation. NRCan is considered as the reference data in bias calculations. The bias is the difference between a reanalysis and NRCan for a given period. It indicates how much a reanalysis overestimates or underestimates relative to NRCan.
b. Weather station densities
The weather stations considered in calculating station densities are those used by Environment and Climate Change Canada to develop the NRCan database. The station densities of each watershed are calculated in three steps. First, all the stations within an area covering each watershed area plus an extension of 20 km outside each watershed contour were selected in order to have at least one station per basin. The Thiessen polygons method was then used to keep only the stations that are useful to determine the mean weather conditions over each basin. Second, the number of weather stations that provide weather data for each specific day is summed. Then, the total number of days where the weather stations are operational is divided by the total number of days over the period. Third, the density of stations is computed as the ratio between the mean number of operational weather stations over the period analyzed and the area of the watershed. In other words, the density of stations is computed as the mean number of operational weather stations of each watershed per square kilometer of the watershed.
c. Hydrological model and calibration strategy
The lumped conceptual hydrological model HSAMI (Fortin 2000) is used to simulate river discharges. HSAMI has been used to predict the hourly and daily flows of more than 100 watersheds in Quebec. It has also been used operationally by Hydro-Québec over 100 watersheds for more than 30 years, as well as in climate change impact projects (Chen et al. 2012; Poulin et al. 2011). The HSAMI model has 23 calibration parameters: 2 for evapotranspiration, 6 for snowmelt, 10 for infiltration and percolation, and 5 for the routing of surface runoff and interflow (Fig. 2). In HSAMI, four interconnected reservoirs contribute to the vertical water transfer balance: snow on the ground, surface water, unsaturated zones, and saturated zones. The horizontal water transfer is based on two unit hydrographs (one for surface runoff and one for delayed runoff) and one linear reservoir for groundwater flows.
Flowchart of the HSAMI model. Black boxes represent conceptual reservoirs. Adapted from Minville et al. (2014).
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
Flowchart of the HSAMI model. Black boxes represent conceptual reservoirs. Adapted from Minville et al. (2014).
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
Flowchart of the HSAMI model. Black boxes represent conceptual reservoirs. Adapted from Minville et al. (2014).
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
The calibration was performed on the even years of each watershed study period, while the validation was based solely on the odd years. All the calibrations were performed using the Covariance Matrix Adaptation Evolution Strategy (CMAES) algorithm (Hansen and Ostermeier 1996, 2001). Arsenault et al. (2014) showed that CMAES is able to find optimal parameter sets for the HSAMI model.
The Nash–Sutcliffe efficiency (NSE) metric (Nash and Sutcliffe 1970) was computed as an objective function based on the even years, with cross validation on the odd years to take into account the different climatic trends (e.g., natural, decadal, or multidecadal variability). The NSE is commonly used, and despite drawbacks such as heavily weighting peak flows, it is found to be the best objective function for reflecting the overall fit of a hydrograph (Servat and Dezetter 1991). The HSAMI model was calibrated multiple times for each dataset and watershed, and the best parameter set is selected according to the NSE coefficient. This ensures that the optimization algorithm is not confined to a local minimum, as is always possible when doing a single optimization.
The nonparametric Wilcoxon test (Rakotomalala 2008) was performed to statistically evaluate simulated discharge at the 95% confidence level.
d. Evaluation of simulated flows accuracy
The simulated hydrographs from each dataset were compared to the observed hydrographs over the validation period. The performances of the simulations were computed using the NSE metric values. First, the validation NSE values were compared considering all the 316 watersheds. Second, an evaluation as a function of the density of weather stations was carried out. Third, in each of the three climatic regions, performances obtained using NRCan were compared to those using reanalyses for watersheds where the density of weather stations was considered low. In the mountainous region, the spatial variability of precipitation is known to be great (Bailey et al. 1997). It is expected that in such a region, the considerable spatial variability of precipitation will greatly affect the accuracy of the interpolated datasets when the density of the available weather stations is low.
4. Results
a. Data comparison: Temperature and precipitation
The difference between mean seasonal temperatures from reanalyses and NRCan varies and lies around ±2.5°C (Figs. 3a–f). Generally, reanalyses are warmer than NRCan in winter (DJF). Still in winter, ERA-Interim is mainly colder than NRCan in the mountainous region (biases of around −1.5°C) and warmer in the northern part of the mountainous region and in the boreal and humid continental regions. The warmest biases are obtained in the eastern boreal region, where warm biases reach 2.5°C (Fig. 3a). CFSR is warmer than NRCan in the western boreal, where biases reach +2.5°C. Moreover, CFSR is colder than NRCan in the northern part of the mountainous region, but is generally similar to NRCan in the eastern boreal region (biases within ±0.5°C; Fig. 3b). MERRA is warmer than NRCan throughout Canada, and biases are mainly between +2° and +2.5°C (Fig. 3c).
The 1979–2010 (left) DJF and (right) JJA temperature biases (°C) between reanalyses and observed gridded NRCan.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
The 1979–2010 (left) DJF and (right) JJA temperature biases (°C) between reanalyses and observed gridded NRCan.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
The 1979–2010 (left) DJF and (right) JJA temperature biases (°C) between reanalyses and observed gridded NRCan.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
In summer (JJA), ERA-Interim is colder than NRCan in the northern part of the mountainous region (biases around −1°C) and slightly warmer in the southern part of the mountainous region and in the boreal and humid continental regions (biases between 0° and +1°C; Fig. 3d). CFSR is warmer than NRCan in the western boreal region, where biases are between +0.5° and +1°C. In the western part of the mountainous region, CFSR is generally colder than NRCan. However, in the northern part of the mountainous region and in the eastern boreal, CFSR is similar to NRCan (biases within ±0.5°C; Fig. 3e). MERRA is generally colder than NRCan in the mountainous region. However, in the western and northeastern boreal regions, MERRA is warmer than NRCan (biases between +1.0° and +2.5°C). In the southeastern boreal region, the temperature of MERRA is similar to that of NRCan in summer (biases within ±0.5°C; Fig. 3f).
For precipitation, the mean winter biases between the reanalyses and NRCan vary between −20% and +120% (Figs. 4a–c). ERA-Interim is wetter than NRCan in the mountainous region (biases mainly between +30% and +60%) and drier in the eastern boreal region (biases between −20% and −10%). In the western boreal and humid continental regions, ERA-Interim is similar to NRCan, with biases within ±10% (Fig. 4a). CFSR and MERRA are wetter than NRCan for almost all the watersheds. The wettest CFSR biases (bias greater than +90%) are obtained over 21% of the watersheds located in the mountainous and western boreal regions. For about 32% of the watersheds located in the mountainous and boreal regions, the MERRA biases exceed +100%.
The 1979–2010 (left) DJF and (right) JJA precipitation RBIAS (%) between reanalyses and NRCan gridded observations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
The 1979–2010 (left) DJF and (right) JJA precipitation RBIAS (%) between reanalyses and NRCan gridded observations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
The 1979–2010 (left) DJF and (right) JJA precipitation RBIAS (%) between reanalyses and NRCan gridded observations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
In the summer, all three reanalyses are wetter than NRCan over 91% of the watersheds for ERA-Interim and CFSR and over 62% of the watersheds for MERRA. The wettest reanalysis biases are obtained in the mountainous region.
Overall, reanalyses tend to be considerably wetter than NRCan in winter and summer in western Canada, mainly in the mountainous region. The results are in line with those presented in Lorenz and Kunstmann (2012) for ERA-Interim, CFSR, and MERRA.
b. Weather station density
The spatial distributions of the weather stations and their densities are shown in Figs. 5a and 5b, respectively. Results show that densities are always less than 10 stations per 1000 km2. Watersheds with a weather station density greater than 1 station per 1000 km2 are located mainly in southwestern and southeastern Canada. Results in Fig. 5c show that for 67% of the watersheds, the stations density is less than 1 station per 1000 km2, and only 7% of them have 3 stations or more per 1000 km2. The minimum densities of stations recommended by WMO (2008) vary between 1 and 4 stations per 1000 km2, depending on the physiographic unit (mountain, plain, etc.). Thus, on the basis of these recommendations, most of the 316 watersheds have a low weather station density since they typically have less than 1 station per 1000 km2.
(a) Location of the weather stations, (b) spatial distribution of density of weather station for each watershed, and (c) cumulative percentage of the number of watersheds according to the density of weather stations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
(a) Location of the weather stations, (b) spatial distribution of density of weather station for each watershed, and (c) cumulative percentage of the number of watersheds according to the density of weather stations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
(a) Location of the weather stations, (b) spatial distribution of density of weather station for each watershed, and (c) cumulative percentage of the number of watersheds according to the density of weather stations.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
c. Comparison of the simulated river discharge
The NSE values obtained over the validation period (i.e., on the odd years) using different inputs are shown in Fig. 6. The results from Fig. 6a show that when all the 316 watersheds are considered, the median NSE values are 0.8 for NRCan, 0.81 for ERA-Interim, 0.8 for CFSR, and 0.77 for MERRA. Moreover, the NSE values are greater than 0.6 for 90% of the watersheds for NRCan and ERA-Interim, 85% for CFSR, and 81% for MERRA. Altogether, all the datasets perform satisfactorily. A comparison of each watershed shows that the NSE values for ERA-Interim are greater than those for NRCan for only 52% of the watersheds. The CFSR and MERRA NSE values are greater than those for NRCan for 39% of the watersheds. Results from the Wilcoxon statistical test (Table 2) show that the performance of NRCan is generally equivalent to those of the reanalyses.
NSE values of the simulated river flow with HSAMI over the validation period using NRCan, ERA-Interim, CFSR, and MERRA, and differences between the NSE values of reanalyses and NRCan.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
NSE values of the simulated river flow with HSAMI over the validation period using NRCan, ERA-Interim, CFSR, and MERRA, and differences between the NSE values of reanalyses and NRCan.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
NSE values of the simulated river flow with HSAMI over the validation period using NRCan, ERA-Interim, CFSR, and MERRA, and differences between the NSE values of reanalyses and NRCan.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
The p values of the Wilcoxon test.
However, some significant differences appear when the results are sorted according to climatic regions. In the mountainous region, the median NSE values for the 144 watersheds are 0.83 for NRCan, 0.86 for ERA-Interim, 0.83 for CFSR, and 0.85 for MERRA (Fig. 6b). In addition, the NRCan NSE values are lower than those for ERA-Interim, CFSR, and MERRA for 72%, 47%, and 59% of the watersheds, respectively. Thus, except for CFSR, reanalysis NSE values are greater than those for NRCan on most of the watersheds in the mountainous region. However, the results of the statistical test (Table 2) indicate that only the ERA-Interim NSE values are significantly greater than those for NRCan.
In the boreal region, the median NSE values for the 149 watersheds are 0.76 for NRCan and ERA-Interim, 0.72 for CFSR, and 0.70 for MERRA (Fig. 6c). From the NSE values for each watershed, it can be seen that the values for NRCan are greater than those for ERA-Interim, CFSR, and MERRA for 59%, 64%, and 72% of the watersheds, respectively. Although the NRCan NSE values are greater than those for the reanalyses for most of the watersheds, the statistical test shows that differences between NRCan and reanalyses NSE values are not statistically significant, except for MERRA (see Table 2).
The results in Fig. 6d show that for the 23 watersheds in the humid continental region, the NSE median values for NRCan, ERA-Interim, CFSR, and MERRA are 0.79, 0.76, 0.74, and 0.71, respectively. In addition, the NSE values for NRCan are greater than those for the reanalyses for at least 87% of the watersheds. The results of the statistical test (Table 2) show that the NSE values for NRCan are significantly greater than those for the reanalyses, except for ERA-Interim.
Results in Fig. 7 show the influence of the watershed size and the density of weather stations on the NSE, computed using the three reanalyses and NRCan. Overall, the performance of the reanalyses increases with the watershed size, while the performance of NRCan is not dependent on the watershed size (Fig. 7a). On average, the reanalysis products were superior to the NRCan dataset for predicting flows in watersheds larger than 3000 km2, that is, about the size of (or larger than) model grid cell of reanalyses. Results in Fig. 7b show that the performance of NRCan increases (but weakly) with the density of weather stations, while the performance of reanalyses is not influenced by the density, although a downward trend is seen starting from a density of 1.56 stations per 1000 km2 and above.
Distribution of reanalyses and NRCan NSE values according to (a) the size of the watershed and (b) the density of weather stations. The box plots show the distribution of the NRCan NSE values. The bins were selected such that each box plot would include 50 watersheds, except for the one on the extreme right, which includes only 16 watersheds. The median of the NRCan NSE values are connected by the green line. The other lines connect the median of the NSE values for the reanalyses, but their corresponding box plots are not shown in order to avoid overloading the figure.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
Distribution of reanalyses and NRCan NSE values according to (a) the size of the watershed and (b) the density of weather stations. The box plots show the distribution of the NRCan NSE values. The bins were selected such that each box plot would include 50 watersheds, except for the one on the extreme right, which includes only 16 watersheds. The median of the NRCan NSE values are connected by the green line. The other lines connect the median of the NSE values for the reanalyses, but their corresponding box plots are not shown in order to avoid overloading the figure.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
Distribution of reanalyses and NRCan NSE values according to (a) the size of the watershed and (b) the density of weather stations. The box plots show the distribution of the NRCan NSE values. The bins were selected such that each box plot would include 50 watersheds, except for the one on the extreme right, which includes only 16 watersheds. The median of the NRCan NSE values are connected by the green line. The other lines connect the median of the NSE values for the reanalyses, but their corresponding box plots are not shown in order to avoid overloading the figure.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
These results are not surprising since it is known that the accuracy of precipitation from gridded observations depends on the density of weather stations, which has an impact on the hydrological simulation performance (Chaplot et al. 2005; Tozer et al. 2012; Vischel 2006). However, it is important to note that for a given watershed, there is an optimal threshold of the density of weather stations beyond which the performance of the hydrological simulations will not be improved anymore by the density of weather stations. Indeed, recent studies have shown that beyond an optimal threshold of the density of weather stations, the addition of more stations has no impact on the performance of watershed river flows simulated by a hydrological model (Arsenault and Brissette 2014). On the other hand, precipitation from reanalysis datasets is generated within the respective weather forecast model and is therefore independent from precipitation station measurements. This implies that the weather station density has no impact on the accuracy of precipitation from reanalysis. It is rather the forecast model’s spatial resolution and formulation that conditions the accuracy of precipitation from reanalysis and subsequently its hydrological modeling performance. This explains the influence of the size of the watersheds on the hydrological modeling performance of the three reanalyses.
Consequently, the drop in the performance of the reanalyses observed beyond 1.56 stations per 1000 km2 is not actually related to the density of the weather stations, but rather, on the size of the watersheds. Indeed, all the watersheds with at least 1.56 weather stations per 1000 km2 are small in size (with sizes below 32 000 km2, and more than 64% of them are less than 1000 km2). According to the results of Fig. 7a, the weakest reanalyses performance is generally obtained for watersheds of that size.
The performance of the reanalyses and NRCan are compared according to the density of weather stations in the different climatic regions. The results are shown in Fig. 8.
Distribution of the NSE values for reanalyses and NRCan and the differences between the NSE values of reanalyses and NRCan based on the weather stations density and climatic regions.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
Distribution of the NSE values for reanalyses and NRCan and the differences between the NSE values of reanalyses and NRCan based on the weather stations density and climatic regions.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
Distribution of the NSE values for reanalyses and NRCan and the differences between the NSE values of reanalyses and NRCan based on the weather stations density and climatic regions.
Citation: Journal of Hydrometeorology 18, 2; 10.1175/JHM-D-16-0088.1
1) Mountainous region (144 watersheds)
The weather station density is greater than 3 stations per 1000 km2 for seven watersheds in the mountainous region (Fig. 8a). For these watersheds, the median NSE values are 0.87, 0.83, 0.80, and 0.85 for NRCan, ERA-Interim, CFSR, and MERRA, respectively. The NRCan NSE values are greater than those for the reanalyses for five of the seven watersheds. However, there is statistically no significant difference between the performances of NRCan and the reanalyses.
Similarly, there is no significant difference between the NRCan NSE values and those for the reanalyses for the watersheds where the density of weather stations is between 2 and 3 stations per 1000 km2. The median NSE values are 0.76 for NRCan and CFSR, 0.78 for ERA-Interim, and 0.7 for MERRA (Fig. 8b).
Where the station density is between 1 and 2 stations per 1000 km2, the median NSE values are 0.84 for NRCan and MERRA, 0.85 for ERA-Interim, and 0.83 for CFSR, and there is no significant difference between the performance of the reanalyses and that of NRCan (Fig. 8c).
About 67% of the watersheds in the mountainous region have a density of weather stations lower than 1 station per 1000 km2. For these watersheds, the median NSE values are 0.84 for NRCan, 0.88 for ERA-Interim, 0.85 for CFSR, and 0.86 for MERRA (Fig. 8d). The NSE values for ERA-Interim and MERRA are greater than those for NRCan on most of the watersheds (73% and 60%, respectively). However, the NSE values for CFSR are greater than those for NRCan for only 44% of the watersheds. The statistical test shows that the performance of ERA-Interim and MERRA is significantly greater than that for NRCan, whereas CFSR and NRCan perform equally.
Overall, in the mountainous region, all the reanalyses, except for CFSR, significantly outperform NRCan when the station density is lower than 1 station per 1000 km2. For higher weather station densities, the performances of the reanalyses and NRCan are similar.
2) Boreal region (149 watersheds)
In the boreal region, the weather station density is greater than 3 stations per 1000 km2 for nine watersheds, and the median NSE values are 0.8, 0.7, 0.74, and 0.7 for NRCan, ERA-Interim, CFSR, and MERRA, respectively (Fig. 8e). Moreover, the NSE values for NRCan are greater than those for the reanalyses for at least 67% of the watersheds, and the statistical test reveals that the performance of NRCan is significantly better than that of the reanalyses, except for CFSR.
Where the weather station density is between 2 and 3 stations per 1000 km2 (seven watersheds), the median NSE values are 0.82 for NRCan and 0.80 for reanalyses. Moreover, the NRCan NSE values are greater than those of the reanalyses for at least 71% of the watersheds (Fig. 8f). However, the NRCan NSE values are statistically similar to those for reanalyses.
Considering the watersheds where the weather station density is between 1 and 2 stations per 1000 km2 (21 watersheds), the median NSE values are 0.81 for NRCan, 0.77 for ERA-Interim, 0.8 for CFSR, and 0.74 for MERRA (Fig. 8g). Although the NSE values for NRCan are greater than those for the reanalyses for at least 71% of the watersheds, there is no statistically significant difference between the performance of NRCan and those for reanalyses, except for MERRA.
Similarly, when the weather station density is lower than 1 station per 1000 km2 (112 watersheds) the NSE values for NRCan are significantly greater than those for MERRA, but are statistically equivalent to the NSE values for ERA-Interim and CFSR (Fig. 8h). The median NSE values are 0.74 for NRCan, 0.67 for MERRA, 0.75 for ERA-Interim, and 0.71 for CFSR.
It is clear from this analysis that for watersheds considered in the boreal region, NRCan and CFSR have statistically equivalent performances, regardless of the station density. Moreover, NRCan is significantly better than ERA-Interim when the weather station density is high (more than 3 stations per 1000 km2).
3) Humid continental region (23 watersheds)
In the humid continental region, the weather station density is greater than 3 stations per 1000 km2 for seven watersheds (Fig. 8i). For these watersheds, the median NSE values are 0.78 for NRCan, 0.76 for ERA-Interim, 0.74 for CFSR, and 0.74 for MERRA. The NRCan NSE values are statistically equivalent to those for ERA-Interim and CFSR but are significantly greater than those for MERRA.
The same relation is obtained for the six watersheds where the weather station densities vary between 2 and 3 stations per 1000 km2 (Fig. 8j). For these watersheds, the median NSE values are 0.81 for NRCan, 0.76 for ERA-Interim, 0.77 for CFSR, and 0.73 for MERRA.
When the weather station density is between 1 and 2 stations per 1000 km2, NRCan and the reanalyses perform similarly. The NSE values for NRCan, ERA-Interim, CFSR, and MERRA are 0.7, 0.66, 0.67, and 0.61, respectively (Fig. 8k).
Similarly, for weather station densities lower than 1 station per 1000 km2 (five watersheds), the NSE values for NRCan are statistically equivalent to those for ERA-Interim, CFSR, and MERRA (Fig. 8l). The median NSE values are 0.82 for NRCan and ERA-Interim, 0.76 for CFSR, and 0.77 for MERRA.
Thus, in the humid continental region, the performance of NRCan is statistically equivalent to those for ERA-Interim and CFSR, regardless of the weather station densities. However, NRCan performs significantly better than MERRA when the weather station density is greater than 2 stations per 1000 km2.
Table 3 summarizes the findings per region, per watershed size, and as a function of station density.
Summary of the main findings according to the climatic region, watershed size, and weather station density.
5. Discussion
Gridded observation datasets developed from the spatial interpolation of weather stations are usually useful for hydrological modeling. However, their credibility is questionable in regions where weather stations are sparsely distributed. Meteorological data from global reanalyses can represent a good alternative to gridded observations in forcing hydrological models in regions where weather stations are sparsely distributed. Global reanalyses provide a physically consistent estimate of weather events and rely on global observations from multiple sources that are assimilated in a weather forecast model. However, spatial resolutions of reanalyses are relatively coarse, and their data are generally biased when compared to observations. Nevertheless, biases and model resolutions are steadily improving.
This work compares the use of reanalyses instead of gridded observations to force hydrological models in regions with few conventional weather stations, such as in the subarctic. To investigate the potential of the reanalyses as proxies of temperature and precipitation from weather stations, three atmospheric reanalyses—ERA-Interim, CFSR, and MERRA—were evaluated and compared to gridded observations from NRCan. First, temperature and precipitation from reanalyses were spatially averaged over 316 Canadian watersheds and compared to NRCan. Second, for each of the watersheds, the HSAMI lumped hydrological model was calibrated to each reanalysis and NRCan dataset. The river discharges simulated by the hydrological model were evaluated against observed discharges over a validation period. The performances of the discharges simulated using precipitation and temperature from reanalyses and NRCan were compared according to the density of weather stations. About 67% of the watersheds have less than 1 station per 1000 km2 and therefore provide a good representation of regions with a sparse distribution of weather stations.
Results from the temperature and precipitation comparison showed some difference between the reanalyses and NRCan. Overall, mean seasonal temperature differences between the reanalyses and NRCan are relatively low, especially in the summer. Generally, the differences are lower between NRCan and both CFSR and ERA-Interim. This is possibly linked to satellite radiance assimilated by CFSR (Wang et al. 2011) and the land surface temperature assimilated by ERA-Interim (Dee et al. 2011; Simmons et al. 2010). In general, precipitation from ERA-Interim is closer to that of NRCan as compared to the other reanalyses. However, the three reanalyses tended to be wetter than NRCan. Differences between precipitation from NRCan and that from the reanalyses are particularly great in the mountainous region, where orographic precipitation is predominant (Bailey et al. 1997; Gervais et al. 2014). These differences are likely explained by biases in the reanalyses, or in NRCan, or in both. In fact, the reanalyses are possibly unable to adequately represent orographic precipitation because of their coarse resolutions, which smooth topography. On the other hand, the orographic precipitation is possibly smoothed in NRCan by the spatial interpolation of the few available weather stations. Overall, with such uncertainty in the reanalyses and the gridded observations, it is difficult to determine which one is the most accurate. However, hydrological modeling results indirectly highlight the quality of the reanalyses and NRCan datasets.
The calibration of the hydrological model for each dataset by finding the optimum fit of the 23 parameters of the model filters to some extent the errors of the datasets. Indeed, rainfall–runoff models have an ability to offset errors in the meteorological inputs, but this ability has limits (Andréassian et al. 2001; Oudin et al. 2006). However, this ability has limits, as when the seasonality of precipitation is not adequately represented in a dataset, which lead to low performance of hydrological simulation, despite a specific calibration of the hydrological model (Essou et al. 2016). Thus, the accuracy of a simulated discharge is dependent on the quality of the forcing data.
Globally, for the 316 watersheds, similar performances by HSAMI were obtained when forced by NRCan or by the reanalyses. Results showed that the density of weather stations has an impact on the performance of NRCan, but not on the performance of the reanalyses. This is in line with expectations. However, the performance of NRCan is not directly proportional to the weather station density. This is explained by the fact that for a given watershed, there is a threshold beyond which an increase in the density of weather stations will stop improving the performance of hydrological simulations using a lumped hydrological model (Arsenault and Brissette 2014). Results also showed that when the density of weather stations is greater than 3 stations per 1000 km2, the performance of NRCan tends to be statistically similar to or better than those for reanalyses. Conversely, when the weather station density is low (less than 1 station per 1000 km2), the performances of the reanalyses tend to be statistically equal to or greater than those for NRCan; this is particularly the case in the mountainous region, where the differences in precipitation between the reanalyses and NRCan are the largest, and ERA-Interim and MERRA perform significantly better than NRCan.
These results validate the fact that precipitation and temperature from reanalyses are globally more accurate than those from gridded observations, especially in the Canadian mountainous region, when few surface weather stations are available. This means that in such regions, reanalyses could be used instead of gridded observations to force lumped conceptual hydrological models. This might not necessarily be the case with more physically based hydrological models, which might be more sensitive to bias in forcing data. Results also suggest that reanalyses should be of great interest for hydrological modeling in regions such as northern Canada, which are not well covered with surface weather stations (Lindsay et al. 2014). In Canada, most of the hydroelectricity is produced in remote areas where weather stations are sparsely distributed. In these regions, reanalyses could be used as an alternative or a complement to observational data to fill gaps in historical hydroclimatic databases to support hydrological modeling and water resources management efforts.
Although this study was performed on Canadian watersheds, it can be repeated for other regions of the world where weather stations are even less dense, since the three reanalyses that have been used offer a global coverage. In such regions, similar results as the ones of this study are expected. However, one must be careful because the accuracy of reanalysis data varies spatially depending on the quantity and quality of the assimilated data (Lorenz and Kunstmann 2012).
Model- and observation-based databases such as outputs from operational NWP systems and near-real-time QPE products such as stage IV Multisensor Precipitation Estimates (MPE) in the United States and the Canadian Precipitation Analysis (CaPA) in Canada are other alternatives with some potential for hydrological studies in regions where surface weather stations are sparse.
One of the limitations of this study is the use of a lumped hydrological model for river flow simulations. This limitation is due to the large number of watersheds considered and to the high computational costs that would result from the use of a distributed hydrological model. However, if a distributed hydrological model was used, the individual performance of each database may be different for large watersheds, but the general trend of performances and the main findings of this study may not necessarily change (Lauri et al. 2014; Lobligeois 2014).
6. Conclusions
This study compared precipitation and temperature data from ERA-Interim, CFSR, and MERRA to gridded NRCan observations over 316 watersheds located in three climatic regions in Canada. Moreover, these precipitation and temperature data were used to force a lumped hydrological model, and the Nash–Sutcliffe values of the simulated river flows were compared as a function of the density of surface weather stations.
Results show that temperature data from reanalyses are similar to those of the gridded observations of NRCan in the summer. Nevertheless, significant temperature differences were found between reanalyses and NRCan in the winter. Reanalyses tend to be considerably wetter than NRCan during winter and summer in western Canada, mainly in the mountainous region.
Over the 316 watersheds, the Nash–Sutcliffe values of the reanalyses were statistically equivalent to those for NRCan. However, results from the analysis according to the watershed size showed that the performances of the reanalyses were superior to those of NRCan for watersheds that are about the size of (or larger than) a model grid cell of reanalyses. Moreover, we could confirm our expectation since the analysis according to the weather station density showed that in the mountainous region, the performances of the reanalyses, especially ERA-Interim and MERRA, were significantly better than those for NRCan when the surface weather station density is less than 1 station per 1000 km2.
Overall, this study showed that compared to the gridded observations, reanalyses represent a reliable proxy for weather station data in complex terrain regions and where surface weather stations are sparsely distributed.
Acknowledgments
This work was funded through a Natural Science and Engineering Research Council collaborative research grant (435692-12) with Hydro-Québec, Rio-Tinto-Alcan, Ontario Power Generation, and the Ouranos consortium on regional climatology and adaptation to climate change as industrial partners. These industrial research partners are greatly acknowledged for their direct and indirect contributions to this work. We also sincerely thank all the individuals and institutions that developed the datasets used in this work and made them available to the scientific community.
REFERENCES
Adam, J. C., and D. P. Lettenmaier, 2003: Adjustment of global gridded precipitation for systematic bias. J. Geophys. Res., 108, 4257, doi:10.1029/2002JD002499.
Andréassian, V., C. Perrin, C. Michel, I. Usart-Sanchez, and J. Lavabre, 2001: Impact of imperfect rainfall knowledge on the efficiency and the parameters of watershed models. J. Hydrol., 250, 206–223, doi:10.1016/S0022-1694(01)00437-1.
Arsenault, R., and F. Brissette, 2014: Determining the optimal spatial distribution of weather station networks for hydrological modeling purposes using RCM datasets: An experimental approach. J. Hydrometeor., 15, 517–526, doi:10.1175/JHM-D-13-088.1.
Arsenault, R., A. Poulin, P. Côté, and F. Brissette, 2014: Comparison of stochastic optimization algorithms in hydrological model calibration. J. Hydrol. Eng., 19, 1374–1384, doi:10.1061/(ASCE)HE.1943-5584.0000938.
Arsenault, R., R. Bazile, C. Ouellet-Dallaire, and F. Brissette, 2016: CANOPEX: A Canadian hydrometeorological watershed database. Hydrol. Processes, 30, 2734–2736, doi:10.1002/hyp.10880.
Bailey, W. G., T. R. Oke, and W. Rouse, Eds., 1997: Surface Climates of Canada. McGill–Queen’s University Press, 400 pp.
Chaplot, V., A. Saleh, and D. Jaynes, 2005: Effect of the accuracy of spatial rainfall information on the modeling of water, sediment, and NO3–N loads at the watershed level. J. Hydrol., 312, 223–234, doi:10.1016/j.jhydrol.2005.02.019.
Chen, J., F. P. Brissette, and R. Leconte, 2012: Downscaling of weather generator parameters to quantify hydrological impacts of climate change. Climate Res., 51, 185, doi:10.3354/cr01062.
Choi, W., S. J. Kim, P. F. Rasmussen, and A. R. Moore, 2009: Use of the North American Regional Reanalysis for hydrological modelling in Manitoba. Can. Water Resour. J., 34, 17–36, doi:10.4296/cwrj3401017.
Coulibaly, P., J. Samuel, A. Pietroniro, and D. Harvey, 2013: Evaluation of Canadian National Hydrometric Network density based on WMO 2008 standards. Can. Water Resour. J., 38, 159–167, doi:10.1080/07011784.2013.787181.
Daly, C., 2006: Guidelines for assessing the suitability of spatial climate data sets. Int. J. Climatol., 26, 707–721, doi:10.1002/joc.1322.
Dee, D., and Coauthors, 2011: The ERA‐Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553–597, doi:10.1002/qj.828.
Duncan, M., B. Austin, F. Fabry, and G. Austin, 1993: The effect of gauge sampling density on the accuracy of streamflow prediction for rural catchments. J. Hydrol., 142, 445–476, doi:10.1016/0022-1694(93)90023-3.
Essou, G. R. C., F. Sabarly, P. Lucas-Picher, F. Brissette, and A. Poulin, 2016: Can precipitation and temperature from meteorological reanalyses be used for hydrological modeling? J. Hydrometeor., 17, 1929–1950, doi:10.1175/JHM-D-15-0138.1.
Faurès, J.-M., D. Goodrich, D. A. Woolhiser, and S. Sorooshian, 1995: Impact of small-scale spatial rainfall variability on runoff modeling. J. Hydrol., 173, 309–326, doi:10.1016/0022-1694(95)02704-S.
Fekete, B. M., C. J. Vörösmarty, J. O. Roads, and C. J. Willmott, 2004: Uncertainties in precipitation and their impacts on runoff estimates. J. Climate, 17, 294–304, doi:10.1175/1520-0442(2004)017<0294:UIPATI>2.0.CO;2.
Fortin, V., 2000: Le modèle météo-apport HSAMI: Historique, théorie et application. Institut de recherche d’Hydro-Québec, 68 pp.
Fuka, D. R., M. T. Walter, C. MacAlister, A. T. Degaetano, T. S. Steenhuis, and Z. M. Easton, 2014: Using the Climate Forecast System Reanalysis as weather input data for watershed models. Hydrol. Processes, 28, 5613–5623, doi:10.1002/hyp.10073.
Gervais, M., J. R. Gyakum, E. Atallah, L. B. Tremblay, and R. B. Neale, 2014: How well are the distribution and extreme values of daily precipitation over North America represented in the Community Climate System Model? A comparison to reanalysis, satellite, and gridded station data. J. Climate, 27, 5219–5239, doi:10.1175/JCLI-D-13-00320.1.
Goodison, B. E., P. Y. T. Louie, and D. Yang, 1998: WMO solid precipitation intercomparison. Instruments and Observing Methods Rep. 67, WMO/TD-872, 212 pp. [Available online at https://www.wmo.int/pages/prog/www/IMOP/publications/IOM-67-solid-precip/WMOtd872.pdf.]
Hansen, N., and A. Ostermeier, 1996: Adapting arbitrary normal mutation distributions in evolution strategies: The covariance matrix adaptation. Proceedings of IEEE International Conference on Evolutionary Computation 1996, IEEE, 312–317, doi:10.1109/ICEC.1996.542381.
Hansen, N., and A. Ostermeier, 2001: Completely derandomized self-adaptation in evolution strategies. Evol. Comput., 9, 159–195, doi:10.1162/106365601750190398.
Hopkinson, R. F., D. W. McKenney, E. J. Milewska, M. F. Hutchinson, P. Papadopol, and L. A. Vincent, 2011: Impact of aligning climatological day on gridding daily maximum–minimum temperature and precipitation over Canada. J. Appl. Meteor. Climatol., 50, 1654–1665, doi:10.1175/2011JAMC2684.1.
Hutchinson, M. F., 1995: Interpolating mean rainfall using thin plate smoothing splines. Int. J. Geogr. Inf. Syst., 9, 385–403, doi:10.1080/02693799508902045.
Hutchinson, M. F., 2004: ANUSPLIN version 4.3. User guide, Centre for Resources and Environmental Studies, Australian National University, accessed 15 December 2013. [Available online at http://fennerschool.anu.edu.au/research/products.]
Hutchinson, M. F., D. W. McKenney, K. Lawrence, J. H. Pedlar, R. F. Hopkinson, E. Milewska, and P. Papadopol, 2009: Development and testing of Canada-wide interpolated spatial models of daily minimum–maximum temperature and precipitation for 1961–2003. J. Appl. Meteor. Climatol., 48, 725–741, doi:10.1175/2008JAMC1979.1.
Kalnay, E., and Coauthors, 1996: The NCEP/NCAR Reanalysis 40-Year Project. Bull. Amer. Meteor. Soc., 77, 437–471, doi:10.1175/1520-0477(1996)077<0437:TNYRP>2.0.CO;2.
Kottek, M., J. Grieser, C. Beck, B. Rudolf, and F. Rubel, 2006: World map of the Köppen–Geiger climate classification updated. Meteor. Z., 15, 259–263, doi:10.1127/0941-2948/2006/0130.
Lauri, H., T. Räsänen, and M. Kummu, 2014: Using reanalysis and remotely sensed temperature and precipitation data for hydrological modeling in monsoon climate: Mekong River case study. J. Hydrometeor., 15, 1532–1545, doi:10.1175/JHM-D-13-084.1.
Lindsay, R., M. Wensnahan, A. Schweiger, and J. Zhang, 2014: Evaluation of seven different atmospheric reanalysis products in the Arctic. J. Climate, 27, 2588–2606, doi:10.1175/JCLI-D-13-00014.1.
Lobligeois, F., 2014: Mieux connaitre la distribution spatiale des pluies améliore-t-il la modélisation des crues? Diagnostic sur 181 bassins versants français. Ph.D. thesis, AgroParisTech, 310 pp. [Available online at http://infodoc.agroparistech.fr/index.php?lvl=author_see&id=132536.]
Lopes, V. L., 1996: On the effect of uncertainty in spatial distribution of rainfall on catchment modelling. Catena, 28, 107–119, doi:10.1016/S0341-8162(96)00030-6.
Lorenz, C., and H. Kunstmann, 2012: The hydrological cycle in three state-of-the-art reanalyses: Intercomparison and performance analysis. J. Hydrometeor., 13, 1397–1420, doi:10.1175/JHM-D-11-088.1.
Mesinger, F., and Coauthors, 2006: North American Regional Reanalysis. Bull. Amer. Meteor. Soc., 87, 343–360, doi:10.1175/BAMS-87-3-343.
Minville, M., D. Cartier, C. Guay, L. A. Leclaire, C. Audet, S. Le Digabel, and J. Merleau, 2014: Improving process representation in conceptual hydrological model calibration using climate simulations. Water Resour. Res., 50, 5044–5073, doi:10.1002/2013WR013857.
Mizukami, N., and M. B. Smith, 2012: Analysis of inconsistencies in multi-year gridded quantitative precipitation estimate over complex terrain and its impact on hydrologic modeling. J. Hydrol., 428–429, 129–141, doi:10.1016/j.jhydrol.2012.01.030.
Nash, J., and J. Sutcliffe, 1970: River flow forecasting through conceptual models part I—A discussion of principles. J. Hydrol., 10, 282–290, doi:10.1016/0022-1694(70)90255-6.
Oudin, L., C. Perrin, T. Mathevet, V. Andréassian, and C. Michel, 2006: Impact of biased and randomly corrupted inputs on the efficiency and the parameters of watershed models. J. Hydrol., 320, 62–83, doi:10.1016/j.jhydrol.2005.07.016.
Poulin, A., F. Brissette, R. Leconte, R. Arsenault, and J.-S. Malo, 2011: Uncertainty of hydrological modelling in climate change impact studies in a Canadian, snow-dominated river basin. J. Hydrol., 409, 626–636, doi:10.1016/j.jhydrol.2011.08.057.
Rakotomalala, R., 2008: Comparaison de populations: Tests non paramétriques. Université Lumière Lyon 2, 189 pp. [Available online at http://eric.univ-lyon2.fr/~ricco/cours/cours/Comp_Pop_Tests_Nonparametriques.pdf.]
Rienecker, M. M., and Coauthors, 2008: The GEOS-5 Data Assimilation System—Documentation of versions 5.0.1, 5.1.0, and 5.2.0. Tech. Memo. NASA/TM-2008-104606, Vol. 27, 97 pp. [Available online at http://gmao.gsfc.nasa.gov/pubs/docs/Rienecker369.pdf.]
Rienecker, M. M., and Coauthors, 2011: MERRA: NASA’s Modern-Era Retrospective Analysis for Research and Applications. J. Climate, 24, 3624–3648, doi:10.1175/JCLI-D-11-00015.1.
Saha, S., and Coauthors, 2010: The NCEP Climate Forecast System Reanalysis. Bull. Amer. Meteor. Soc., 91, 1015–1057, doi:10.1175/2010BAMS3001.1.
Servat, E., and A. Dezetter, 1991: Selection of calibration objective functions in the context of rainfall–runoff modelling in a Sudanese savannah area. Hydrol. Sci. J., 36, 307–330, doi:10.1080/02626669109492517.
Shiu, C.-J., S. C. Liu, C. Fu, A. Dai, and Y. Sun, 2012: How much do precipitation extremes change in a warming climate? Geophys. Res. Lett., 39, L17707, doi:10.1029/2012GL052762.
Sillmann, J., V. Kharin, X. Zhang, F. Zwiers, and D. Bronaugh, 2013a: Climate extremes indices in the CMIP5 multimodel ensemble: Part 1. Model evaluation in the present climate. J. Geophys. Res. Atmos., 118, 1716–1733, doi:10.1002/jgrd.50203.
Sillmann, J., V. Kharin, F. Zwiers, X. Zhang, and D. Bronaugh, 2013b: Climate extremes indices in the CMIP5 multimodel ensemble: Part 2. Future climate projections. J. Geophys. Res. Atmos., 118, 2473–2493, doi:10.1002/jgrd.50188.
Simmons, A., K. Willett, P. Jones, P. Thorne, and D. Dee, 2010: Low-frequency variations in surface atmospheric humidity, temperature, and precipitation: Inferences from reanalyses and monthly gridded observational data sets. J. Geophys. Res., 115, D01110, doi:10.1029/2009JD012442.
Thiessen, A. H., 1911: Precipitation averages for large areas. Mon. Wea. Rev., 39, 1082–1089, doi:10.1175/1520-0493(1911)39<1082b:PAFLA>2.0.CO;2.
Tozer, C., A. Kiem, and D. Verdon-Kidd, 2012: On the uncertainties associated with using gridded rainfall data as a proxy for observed. Hydrol. Earth Syst. Sci., 16, 1481–1499, doi:10.5194/hess-16-1481-2012.
Uppala, S. M., and Coauthors, 2005: The ERA‐40 re‐analysis. Quart. J. Roy. Meteor. Soc., 131, 2961–3012, doi:10.1256/qj.04.176.
Vischel, T., 2006: Impact de la variabilité pluviométrique de méso-échelle sur la réponse des systèmes hydrologiques sahéliens: Modélisation, simulation et desegregation. Ph.D. thesis, Grenoble, INPG, 279 pp.
Vu, M., S. Raghavan, and S. Liong, 2012: SWAT use of gridded observations for simulating runoff—A Vietnam river basin study. Hydrol. Earth Syst. Sci., 16, 2801–2811, doi:10.5194/hess-16-2801-2012.
Wang, W., P. Xie, S.-H. Yoo, Y. Xue, A. Kumar, and X. Wu, 2011: An assessment of the surface climate in the NCEP Climate Forecast System Reanalysis. Climate Dyn., 37, 1601–1620, doi:10.1007/s00382-010-0935-7.
Winkler, T., 1993: Standardized hydrometric data collection in Canada: The career development program for technician training. Water Int., 18, 217–224, doi:10.1080/02508069308686182.
WMO, 2008: Density of stations for a network. Hydrology—From Measurement to Hydrological Information, Vol. 1, Guide to Hydrological Practices, 6th ed., WMO-168, World Meteorological Organization, I.2-24–I.2-28
Woo, M. K., and R. Thorne, 2006: Snowmelt contribution to discharge from a large mountainous catchment in subarctic Canada. Hydrol. Processes, 20, 2129–2139, doi:10.1002/hyp.6205.
Xie, P., and P. A. Arkin, 1997: Global precipitation: A 17-year monthly analysis based on gauge observations, satellite estimates, and numerical model outputs. Bull. Amer. Meteor. Soc., 78, 2539–2558, doi:10.1175/1520-0477(1997)078<2539:GPAYMA>2.0.CO;2.
Yang, D., B. Goodison, J. Metcalfe, V. Golubev, R. Bates, T. Pangburn, and C. Hanson, 1998: Accuracy of NWS 8″ standard nonrecording precipitation gauge: Results and application of WMO intercomparison. J. Atmos. Oceanic Technol., 15, 54–68, doi:10.1175/1520-0426(1998)015<0054:AONSNP>2.0.CO;2.
Yang, D., and Coauthors, 2000: An evaluation of the Wyoming gauge system for snowfall measurement. Water Resour. Res., 36, 2665–2677, doi:10.1029/2000WR900158.