Our study site is located in the Coal Creek Watershed (Hydrologic Unit Code (HUC) 140200010204) with an area of 53.2 km2 located within the larger East Taylor Watershed (HUC 14020001) near Crested Butte, in southwestern Colorado (Fig. ). The Coal Creek Watershed is a high alpine, snow-dominated catchment, characterized as warm summer, humid continental climate in the Köppen classification system . It receives ∼850 mm of precipitation annually, with ∼530 mm as snowfall which was estimated from the long-term Daymet forcing dataset . The primary land cover types are evergreen forest (62.6 %) and shrub (20.5 %). This watershed has strong variations in topography and land cover, which is representative of many headwater, mountainous watersheds in the western US.

ATS is an integrated, distributed hydrologic code that solves the diffusion wave approximation of the St-Venant equations for surface flow coupled to the Richards equation for flow in variably saturated porous media in the subsurface . The Richards equation is described as 1∂∂t(ϕs)+▿⋅q=0, with 2q=-1μkrκ(▿p+ρg), where ϕ is the effective porosity (–), s is the saturation (–), q is the Darcy flux (m s-1), μ is the dynamic viscosity (Pa s-1), kr is the relative permeability (–), κ is the saturated hydraulic permeability (m2), p is the water pressure (Pa), and g is the gravitational constant (m s-2).

The diffusive wave approximation to overland flow is described as 3∂h∂t+▿⋅(hv)=Qw+Qss, with 4v=-h2/3n⋅maxϵ,▿z▿(z+h), where h is the depth of ponded water (m), v is the surface flow velocity (m s-1), Qw are all external source/sink terms (m s-1), Qss is the exchange flux between surface and subsurface systems (m s-1), n is Manning's coefficient (s m-1/3), z is surface elevation (m), and ϵ is a small positive regularization to keep the equations non-singular in places with zero bed slope (m).

The ATS meshes including surface land covers and subsurface structures and properties were developed using the Watershed Workflow package , which brings together a variety of data streams, delineates the catchment, and generates a variable-resolution mesh with refined resolution at the stream network. Resolutions ranged from typical triangle areas of 5000 m2 near the stream network to 50 000 m2 away from the stream network. This triangular surficial mesh was then elevated using a digital elevation model (DEM) from the USGS National Elevation Dataset (NED) 30 m resolution dataset.

On the surface, 14 land cover types were delineated from the National Land Cover Database (NLCD 2016) product for the CONUS. The leaf area index (LAI) seasonal variations for each land cover type were retrieved from MODIS (https://modis.gsfc.nasa.gov/data, last access: 31 March 2021). Some of the plant functional types and their properties such as rooting profile and photosynthetic parameters were adopted from parameters used in the CLM 4.5 technical notes .

In the subsurface, the model was discretized into 19 terrain-following layers with a total thickness of ∼28 m. A total of six soil layers encompassed the top 2 m of the domain. The depth to bedrock (DTB) was determined from SoilGrids that varies from 3 m at its shallowest to 26 m at its deepest. The geologic layers were sandwiched between the soil and bedrock layers. The vertical resolution of the mesh gradually increased from 5 cm at the surface to 2 m at the 2 m depth, and it remained constant at 2 m until the bottom of the model domain at a depth of 28 m. The total number of cells is 171 760.

Based on the National Resources Conservation Service (NRCS) Soil Survey Geographic (SSURGO) soils database, 22 soil types were identified and mapped within the soil layer. Due to the edge-matching issues in the SSURGO soil database , the 22 soil types were regrouped into 9 types to remove the discontinuity of a soil type across soil survey area boundaries. Using a global surface geology dataset from GLobal HYdrogeology MaPS (GLHYMPS) 2.0 , 11 geologic material types were identified and mapped within the geologic layer. The spatial distribution of the soil and geological layers was shown in Fig. . The permeability and porosity for each soil type were retrieved from the SSURGO database, and the van Genuchten parameters were determined using Rosseta v3, a pedotransfer function that relates sand, silt, and clay percentage to van Genuchten parameters . The permeability and porosity for each geology type were retrieved from the GLHYMPS database. Bedrock functions as a confining layer and is assumed to have a very small permeability of 1×10-17 m2.

The model was first run for 1000 years with constant precipitation (∼850 mm yr-1) as the cold spinup that resulted in steady-state model outputs at the final timestep, which was then used as the initial condition for a 10-year (1 October 2004–1 October 2014) transient simulation (i.e., warm spinup) driven by the Daymet forcing. Model state at the end of the 10-year run was used as the initial condition for a 4-year transient run (1 October 2015–1 October 2019) driven by various GMFs. The water year 2015 was treated as a second warm spinup and was discarded from the analysis to avoid any influence from previous spinup runs. The study period features a high snow year (∼709 mm in water year 2017) and a low snow year (∼296 mm in water year 2018), allowing us to demonstrate how different meteorological forcings impact watershed responses under various weather conditions. ATS runs were taken at a sub-hourly timestep determined by the model while outputting streamflow and watershed-averaged variables at an hourly timestep. Due to large file size, spatially distributed variables such as SWE and ET were output at daily timesteps. Each run took ∼17 h wall-clock time using 64 processors on the Cori clusters at the National Energy Research Scientific Computing Center (NERSC). The models were not calibrated because the focus of this study was to evaluate the effect of meteorological forcings on model simulation instead of estimating the optimal parameters used in ATS.

For this comparison, three widely used GMFs were considered: PRISM , NLDAS-2, ), and Daymet v4 . NLDAS-2 and Daymet v4 are hereafter referred to as NLDAS and Daymet, respectively. The detailed comparison between each meteorological dataset can be found in Table .

The Daymet climate forcing is a gridded, daily product with a spatial resolution of 1 km, covering continental North America, Puerto Rico, and Hawaii. It assimilates data from weather stations (primarily GHCN-D stations) and accounts for elevation, prevailing winds, storm tracks, and proximity to large water bodies . Here, the latest Daymet version 4 product is used because this product has gone through significant bias corrections in station observations and the gridded product shows a better match with weather stations compared to the earlier versions .

The PRISM forcing is developed by the PRISM climate group at Oregon State University and is recognized as the official climate dataset for the US Department of Agriculture. It utilizes a wide range of monitoring networks including GHCN-D stations and local/state weather stations to generate daily, spatially continuous climate data for the CONUS. PRISM provides a native grid resolution of 30 arcsec (∼800 m) for a fee but also provides a coarsened 4 km resolution free of charge. We used the native 30 arcsec resolution and downscaled (upscaled) the dataset to obtain finer (coarser) spatial resolutions.

The NLDAS dataset is a gridded, hourly product with a spatial resolution of 1/8th∘ (∼12 km at the study site) for the entire North American region. The non-precipitation forcing variables are primarily derived from the North American Regional Reanalysis (NARR) by spatially interpolating data from the 32 km-resolution NARR grid to the 1/8th∘ NLDAS grid while temporally disaggregated from 3-hourly to hourly frequency . The precipitation is a product of a temporal disaggregation of a gage-only Climate Prediction Center (CPC) analysis of daily precipitation into hourly frequency, performed directly on the NLDAS grid and including an orographic adjustment based on the widely applied PRISM climatology.

All three datasets provide temperature and precipitation as the primary forcing with a few secondary forcing variables. In addition to temperature and precipitation, ATS requires solar radiation (both incoming shortwave radiation (Srad) and longwave radiation – Lrad), relative humidity, and wind speed as forcing inputs. Relative humidity can be estimated based on vapor pressure and mean temperature . Lrad can be estimated from Srad and relative humidity. Because PRISM does not provide Srad and Lrad, we used solar radiation from Daymet instead. Wind speed was assumed to be constant (i.e., 4 m s-1) for both Daymet and PRISM. Compared to PRISM and Daymet, NLDAS provides the most complete set of variables to drive ATS simulations.

Different meteorological forcings have different definitions for a calendar day, and they are often different from the local time used in the observation data (see Table  in the Appendix). Time zone adjustment and lag corrections have been applied to account for the time lag difference between meteorological forcing and local gages. For example, PRISM lags Daymet by 1 d, so PRISM has been shifted forward 1 d to be consistent with Daymet. Both model simulation and gage observation have been converted to the Coordinated Universal Time (UTC) time zone for hourly streamflow comparison. For consistency, all simulated streamflows are in hourly resolution and are compared to hourly USGS streamflows in Sect. .

To study the effect of spatial resolution of meteorological forcing, precipitation and temperature from 800 m PRISM and 1 km Daymet have been downscaled (upscaled) into finer (coarser) spatial resolutions. The downscaling of 800 m PRISM or 1 km Daymet into 400 m used a data-driven downscaling approach . Specifically, random forests were used to extract the relationships between precipitation (or average temperature) and topography. These relationships were developed at 800 m (for PRISM) and 1 km (for Daymet) resolutions and were used as-is to generate the 400 m downscaled estimates. The downscaled precipitation grids were additionally filtered to ensure a smooth field in low-gradient areas without affecting high-gradient areas . The topographic variables considered were elevation, slope, aspect, latitude, and longitude. These variables were extracted from the NED 10 m-resolution product and upscaled to 400 and 800 m (for PRISM) via bilinear interpolation. Upscaling of topographic variables was done in maximum increments of 2× (e.g., 10 m → 20 m → 40 m and so on).

For consistency, spatial upscaling of 800 m PRISM into 1600 and 4000 m was performed using a coarsened function from python package xarray (http://xarray.pydata.org, last access: 1 April 2021) by applying a moving average based on a 2×2 window size. The same approach was used for spatial upscaling of 1 km Daymet to 2 and 4 km. To study the effect of temporal resolution of meteorological forcing, the daily PRISM dataset was disaggregated into hourly resolution using the temporal pattern of NLDAS. The hourly PRISM dataset was then aggregated into 12-hourly temporal resolution by taking the mean (for temperature) or sum (for precipitation) for the aggregated period.

In ATS, meteorological forcing is distributed linearly across its temporal resolution, and each model surface cell gets its meteorological forcing through spatially bilinear interpolation. For example, both Daymet and PRISM apply their meteorological forcing at the daily timescale, whereas NLDAS applies its meteorological forcing at an hourly timescale.

Instantaneous streamflow data (every 15 min) have been available from 1 April through 15 November every year since 2014 at a USGS gage (station number 09111250) located at the watershed outlet. The 15 min streamflow was aggregated to hourly streamflow which was used to compare against model simulations in the Results section. Past Airborne Snow Observatory (ASO) survey has four flights covering this watershed in 2018 and 2019 to survey the snow depth and SWE. Remote sensing products such as the Moderate Resolution Imaging Spectroradiometer (MODIS) 8 d composite ET have been available at a 500 m resolution since 2000. Groundwater measurements and field-observed soil moisture data are not available within the study site.

To compare the accuracy of each meteorological forcing against field observations, all three meteorological forcings at their native resolutions were compared against GHCN-D weather stations within the East Taylor Watershed. In total, there were seven stations with long-term precipitation records and four stations with long-term temperature records (see GHCN-D station locations in Fig. ). Both precipitation and temperature time series were extracted at each GHCN-D gage location from the GMF.

Model-simulated outputs were compared against observation data including hourly streamflow from a USGS gage and spatially distributed SWE from the ASO survey. The modified Kling–Gupta efficiency (KGE) and its three components (r, γ, β) were used to evaluate the model performance in addition to the standard Nash–Sutcliffe efficiency (NSE). The theoretical version of the modified KGE metric is 5KGE=1-(r-1)2+(γ-1)2+(β-1)2, with 6r=cov(S,O)σsσo, 7γ=σs/μsσo/μo,8β=μsμo, where S and O represent simulated and observed values, respectively, r is the correlation coefficient, γ is the variability ratio, β is the bias ratio, cov(S,O) is the covariance between simulated and observed values, σ is the standard deviation, and μ is the mean.

Using the modified KGE avoids the effect of input bias on the variability indicator, which has an advantage over the original KGE , and it also allows diagnostic interpretation of the performance score. KGE decomposes model performance into correlation (r), variability (γ), and bias (β) terms. For example, the correlation measures the temporal dynamics of streamflow (i.e., timing), while the variability and bias measure the flow duration curve (i.e., magnitude). The KGE ranges from -∞ (poorest model skill) to 1 (perfect) when all three terms reach unity. Similarly, the NSE ranges from -∞ (poorest model skill) to 1 (perfect).

A Taylor diagram is used to show how closely a set of patterns (e.g., meteorological forcing) matches observations . In each Taylor diagram, performance metrics such as standard deviation and Pearson's correlation coefficient (r) are shown together. The azimuthal angle represents correlation, and the radial distance represents the standard deviation. Also shown is the centered root mean square error (RMSE) between simulation and observation. The relationship between these statistics is shown below: 9E2=σs2+σo2-2σsσor, where E is the centered RMSE, which is also measured by the geometric distance between simulation and observation data points on the Taylor diagram (unit is the same as the standard deviation). In cases where more than one observation point are plotted on the same diagram, the centered RMSE is omitted. Note that the centered RMSE is a mean-removed RMSE, and thus any bias in the data is not shown.

The closer the distance between simulation and observation data point on a Taylor diagram, the smaller the centered RMSE (observation data point has centered RMSE=0), the more similarity they show in terms of standard deviation, and the higher the correlation coefficient (observation data point has r=1).

A Taylor diagram was used to compare the similarity in precipitation and temperature patterns between meteorological forcing and GHCN-D stations (Fig. ). Compared to temperature, precipitation showed stronger spatial heterogeneity among stations indicated by the larger difference in standard deviation and correlation. The close clustering of temperature data points indicated that the difference between different stations in temperature patterns was small. For precipitation, PRISM showed a strong correlation (r>0.9) with GHCN-D at three stations, whereas Daymet only showed a strong correlation at one location and all NLDAS sites showed a relatively weak correlation (0.5<r<0.85). For temperature, all three meteorological forcings showed a very strong correlation (r>0.95) with GHCN-D, though Daymet was slightly better than PRISM and NLDAS. Previous studies also reported similar findings at different watersheds that Daymet and PRISM showed better agreement with ground-based observational data than NLDAS , and the temperature was more accurately represented than precipitation .

Triple collocation analysis (TCA) was performed for precipitation and temperature using three GMFs over the East Taylor Watershed (see details in Appendix ). Both precipitation and temperature showed strong spatial heterogeneity of the noise standard error in all three GMFs. On average, Daymet and PRISM showed less error compared to NLDAS for both precipitation and temperature, though the temperature errors in some locations of Daymet were higher than those of PRISM and NLDAS.

To compare the effects of different GMF products on the ATS simulations, precipitation and temperature from each GMF were used as the primary forcing variables with the same other variables (e.g., solar radiation, humidity) from Daymet. To further isolate the impact of different meteorological forcing variables, precipitation and temperature from each GMF were systematically varied. For example, different precipitation from each GMF were used while keeping temperature the same in each simulation, and vice versa.

Using different precipitation and temperatures, simulated hourly streamflow forced by Daymet (1 km, daily) showed better performance against USGS hourly streamflow, followed by PRISM (800 m, daily) and NLDAS (12 km, hourly) during a 3-year simulation period (Fig. ). At their native resolutions, Daymet outperformed PRISM and NLDAS with the largest KGE (0.62) and NSE (0.54) against observed streamflow (also see the statistical summary in Table ). The high agreement between observed and simulated streamflow forced by Daymet is remarkable given that the ATS model has not been calibrated. In general, all three models underestimated discharge (β<1) while showing slightly more variability (γ>1), especially in 2018 and 2019. As expected, models using hourly NLDAS showed larger variability (γ=1.54) due to the highly dynamic flow variations in simulated streamflow compared to those using daily Daymet or PRISM. Daymet underestimated peak flows, but it showed larger flow variability during low flows and early spring.

Not surprisingly, precipitation played a more important role than temperature in driving the simulated watershed responses. Using different precipitation sources alone had similar performance compared to using both precipitation and temperature from each GMF. In both cases, Daymet outperformed PRISM and NLDAS in simulated streamflow (Fig. ). The KGE and NSE metrics from Daymet and PRISM were much higher than those from NLDAS (Table ). However, using different temperature sources alone had little effect on the simulated streamflow (Fig. ). Both KGE and NSE metrics were very similar among different GMFs (Table ).

Average SWE across the watershed showed large differences between different meteorological forcings (Fig. ). Daymet had the largest simulated SWE on average, while NLDAS had the smallest simulated SWE. PRISM produced a similar SWE pattern with Daymet, and their magnitude was very close except for the year 2017. The accumulation of snow started around the same time for all three meteorological forcings, though snow disappeared early for NLDAS in the last 2 years.

Spatially, all meteorological forcings significantly underestimated SWE (β≈0) when compared to SWE from ASO (Table ), and the difference becomes larger at higher elevation with more accumulated snow (Fig. ). The large difference may be attributed to the higher spatial resolution (50 m) used in the ASO snow survey than the spatial resolution of the GMF. Simulated SWE by the three models showed a larger variability (γ>1) than the observed SWE, except on 7 April 2019, when variability was smaller than the observed SWE. Interestingly, the largest variability occurred on 24 May 2018 when little snow was accumulated on the surface. Most of the time, PRISM showed a higher correlation between simulated and observed SWE than that from Daymet and NLDAS. In contrast, NLDAS showed the poorest correlation (r<0.2) with observation.

All three GMFs showed similar ET dynamics because they used the same solar radiation from Daymet (Figs.  and ). Compared to the remote-sensed 8 d composite ET from MODIS, all three meteorological forcings showed a consistent seasonal trend with MODIS, with underestimated ET in the spring. Additionally, the simulated 8 d composite ET by Daymet and PRISM was higher than that from NLDAS in the peak growing season in 2017 and 2019.

To evaluate the effects of spatial resolution of meteorological forcing, precipitation and temperature from different spatial resolutions of PRISM and Daymet were used to drive the model. Because the findings from both PRISM and Daymet were similar, only results from PRISM were summarized below (see Appendix  for Daymet comparison in Figs.  and ). As the spatial resolution became finer, the spatial pattern of precipitation and temperature became more heterogeneous and was more strongly associated with local topography, land use, and land cover. By contrast, coarser-resolution meteorological forcing produced more homogeneous and smoother spatial patterns with less accuracy (Figs.  and ). In addition to precipitation and temperature, the effect of spatial resolution of solar radiation (i.e., Srad) was tested and was found to have little impact on watershed hydrologic variables (see Appendix ).

The simulated discharge showed similar performance in terms of KGE and NSE compared to the observation when meteorological forcing spatial resolution was ≤1600 m (Table ). In fact, the KGE and NSE of 400, 800, and 1600 m resolution were almost identical. All four resolutions showed higher variability (γ>1) with relatively high correlation (r>0.6) in simulated streamflow than the observation. The variability became larger (γ=1.20) and the correlation became weaker (r=0.64) as meteorological forcing spatial resolution reached 4 km (Fig. ). The simulated SWE and total water storage changes were almost identical for all spatial resolutions except during the snowmelt period, when the 4 km spatial resolution showed faster snowmelt in early summer across all 3 years (Fig. ). The spatial distribution of SWE when compared to ASO SWE showed a significantly large bias (β≈0) and thus negative KGE and NSE for all spatial resolution at all times (Fig.  and Table ). Generally, the 4 km resolution had the worst performance and became most obvious on 24 May 2018. PRISM at 400 and 800 m resolution showed a similar spatial pattern and thus a similar correlation with ASO SWE.

Soil moisture at the top 5 cm layer showed a similar pattern when spatial resolution was ≤1600 m (Fig. ). The differences between 4 km resolution and finer resolution became obvious during the snowmelt period (24 May 2018 and 10 June 2019), when soil becomes saturated. For example, soil in the northwestern region from the 4 km resolution was wetter on 24 May 2018, whereas soil close to the outlet from the 4 km resolution was wetter on 10 June 2019. Similarly, spatially distributed ET did not show a significant difference until meteorological forcing resolution reached 4 km (Fig. ).

Surface ponded depth showed very little difference between different spatial resolutions. Four locations (labeled as 1–4 in Fig. ) were selected across the watershed to show the ponded depth variations (Fig. ). One was located at the upstream branch, and the other three were located along the Coal Creek main stem. Similarly to the watershed discharge, ponded depth only differed when the spatial resolution was coarsened to 4 km. The 4 km resolution had a faster recession during peak flows compared to results from other finer resolutions, and the 4 km resolution was also less responsive to rainfall. On average, surface ponded depth varied less than 0.2 m during peak flows.

A transect was selected running from mountain top to river valley bottom with four selected observation locations (labeled as A–D in Fig. ), and groundwater table time series was plotted in Fig. . In general, groundwater rose during snowmelt (April to June) and rainfall events and fell in the dry period (July to September). Location A was at the mountain top, and the groundwater table was ∼4 m below the land surface except during the snowmelt period, when the groundwater table rose to the surface. Since location A was dominated by snow and receives less rainfall, the rise in the groundwater table was mainly due to snowmelt. Locations B and C showed similar trends in groundwater table fluctuations; however, they showed more peaks since they were influenced by both snowmelt and rainfall. The groundwater table at location D was mostly close to the surface except during the dry season, when groundwater started to decline. The 4 km resolution behaved very differently from the other finer resolutions at location A, where the groundwater table peaked earlier due to earlier infiltration from snow. In a dry year in 2018, the groundwater table did not even rise during the snowmelt period at location A. In general, the coarser the meteorological forcing resolution, the larger the bias in precipitation and the more the groundwater table buffered from snowmelt and rainfall.

The effect of meteorological forcing (i.e., precipitation and temperature) temporal resolution on watershed hydrologic simulations was evaluated by using hourly, 12-hourly, and daily PRISM datasets. All resolutions outputted hydrological variables in an hourly timestep and were compared to hourly observed streamflow. In addition to precipitation and temperature, the effect of temporal resolution of solar radiation (i.e., Srad) was tested. The temporal resolution of solar radiation slightly changed the dynamics of streamflow, but overall the impact on watershed hydrologic variables was negligible (see Appendix ).

The match between simulated and observed hourly discharge in terms of KGE was better with daily resolution than hourly and 12-hourly temporal resolution (see Fig.  and Table ). The low performance (KGE=0.20) under hourly resolution was mainly due to the relatively lower correlation (r=0.43) and higher variability (γ=1.12) between simulated and observed streamflow. This is not surprising since hourly meteorological forcing has a more dynamic forcing pattern including hourly temperature and precipitation, and thus it yielded a more dynamic overland flow pattern that can be quite different from field observations. To investigate the hydrograph in more detail, discharge time series were zoomed into the high-flow season in 2017 (Fig. ). It was clear that models driven by hourly and 12-hourly PRISM contained sub-daily flow fluctuations that would be absent in models driven by the daily PRISM. Models driven by 12-hourly PRISM had a weaker diurnal flow pattern in terms of magnitude. This is because hourly PRISM retains the diurnal signal of air temperature and thus the diurnal snowmelt pattern, whereas 12-hourly PRISM had a weaker diurnal signal due to only changing air temperature every 12 h (Fig. ). On the other hand, daily PRISM assumes uniform air temperature throughout a day and thus produced streamflow that was much smoother without diurnal fluctuations. In general, discharge from the hourly PRISM peaked earlier and had a less steep recession limb compared to that from 12-hourly and daily PRISM.

Hourly PRISM had the largest snowmelt and ET variations, followed by 12-hourly and daily PRISM (Fig. ). Compared to the remote-sensed 8 d composite ET from MODIS, the simulated 8 d composite ET forced by the different temporal resolution of PRISM showed a consistent seasonal trend with MODIS (Fig. ). However, significant model underestimation was observed in the spring. All three PRISM resolutions showed similar trends of seasonal snowpack accumulation. However, there was a large difference in peak SWE. Hourly PRISM reached a smaller SWE peak, and the snow melted earlier compared to 12-hourly and daily PRISM, which might be the combined effects of faster snowmelt and slightly larger ET.

We compared three GMFs at their native resolution, Daymet (daily, 1 km), PRISM (daily, 800 m), and NLDAS (hourly, ∼12 km), in driving watershed responses including watershed outlet discharge and spatially distributed variables such as SWE and ET. Can we choose the “best” meteorological forcing for integrated watershed modeling based on discharge comparison alone? What are the strengths and weaknesses of each GMF?

In surface hydrologic modeling, hydrologists often judge the performance of a numerical model by the ability to match the streamflow at the watershed outlet . However, streamflow alone is not good enough to evaluate the performance in meteorological forcing because discharge at the watershed outlet has limited information on the spatial distribution of model outputs (e.g., SWE). Even though simulated streamflow from Daymet has the best match (i.e., highest KGE) against observation, the spatially distributed SWE from Daymet has a weaker correlation with the observed SWE from the ASO survey than that from PRISM. Both Daymet and PRISM perform better than NLDAS in simulating discharge and spatial SWE due to their relatively fine spatial resolution. As shown in Fig. , NLDAS hardly captures the spatial heterogeneity of SWE when compared to ASO SWE. The entire watershed area (∼53.2 km2) is smaller than the size of 1 pixel of the NLDAS grid (∼12×12 km or ∼144 km2), making the meteorological forcing almost homogeneous at the watershed scale.

For a watershed simulation that usually has a mesh resolution < 1 km, Daymet or PRISM provides the best spatial resolution available across the CONUS. However, they do not have the complete forcing dataset that could be directly applied to watershed models without filling in the missing dataset. Daymet has most of the forcing variables except wind speed and longwave radiation. PRISM is missing both wind speed and solar radiation (both shortwave and longwave radiation), which is important for calculating surface energy balance and estimating ET. A common approach is to fill in the missing variables using a different source. For example, used solar radiation from the National Aeronautics and Space Administration’s POWER (NASA-POWER) database (12 000 km2 resolution) combined with PRISM's temperature and precipitation to simulate a crop model. In our study, we used Daymet as a source of solar radiation for PRISM, and the model was able to simulate streamflow reasonably well (KGE=0.45) (Table ) while capturing the spatial heterogeneity of distributed variables.

On the other hand, NLDAS provides the most complete forcing dataset with hourly resolution but comes with a spatial resolution of ∼12 km, which is too coarse for watershed simulation, especially watersheds with complex terrain. The spatial resolution may become less important as model resolution becomes coarser, and the focus is on the system-scale water budget. For example, NLDAS at its native resolution has been applied at the continental scale to study transpiration partitioning using ParFlow-CLM at 1 km resolution . Past studies have also attempted downscaling of NLDAS to much finer spatial resolution . For example, downscaled the meteorological forcing variables from the 12 km resolution of NLDAS to 1 km resolution using high-resolution terrain information at the Río Sonora basin. There have been other attempts to merge the high spatial resolution of the PRISM dataset with the NLDAS dataset, which produced a complete forcing dataset of daily, 4 km resolution covering the CONUS .

Ideally, the GMF with a finer spatiotemporal resolution while providing the most complete forcing is desirable. However, none of the three GMFs is perfect. An alternative approach is to use meteorological forcing outputted from climate models such as WRF. It has the flexibility of generating much finer spatial and temporal resolution output while providing all available meteorological forcings. used the nested-domain configuration of WRF to dynamically generate meteorological forcing variables at various spatial resolutions (from 0.5 to 13.5 km) for use with ParFlow-CLM. Although forcing generated by WRF provides a viable option for meteorological input in the hydrologic model, it requires additional expertise and effort to set up and run the WRF model, which is more challenging than directly using the publicly available gridded forcing (e.g., Daymet).

Is there an optimal spatiotemporal resolution of meteorological forcing for driving watershed simulation while producing realistic results? Should we choose finer spatial resolution over finer temporal resolution? Depending on the quantity of interest and the spatial and temporal scale of the study, the choice may differ. In this study, watershed outlet discharge is shown to be less sensitive to both the spatial and temporal resolution of meteorological forcing because it is an accumulative quantity. The simulated discharge is almost identical between PRISM 400, 800, and 1600 m resolution. The simulated discharge only becomes noticeably worse when the spatial resolution of meteorological forcing is coarsened to 4000 m (or 30 % of the watershed area). Similarly, the watershed average SWE, ET, and total water storage do not show significant differences between different spatial resolutions of PRISM.

The spatial resolution becomes more important if the quantity of interests is the spatially distributed hydrologic variables. For example, the SWE distribution under 400 and 800 m PRISM resembles more closely the ASO SWE than results under much coarser spatial resolution. also found the SWE distribution to be sensitive to the spatial resolution of meteorological forcing, with the finer resolution being able to accurately reproduce SWE spatial distribution as well as total SWE volume. In addition, a higher spatial resolution of forcing would preserve the spatial heterogeneity of distributed variables better and may provide better estimates of variables at point locations (e.g., SWE at the NRCS's Snowpack Telemetry (SNOTEL) stations and groundwater table at wells). For example, the groundwater table at high elevations can be quite different under different spatial resolutions of meteorological forcing (Fig. ).

Temporal resolution becomes more important than spatial resolution when simulating storm events or flash floods that happen within several hours, resulting in a sharp increase in stream discharge . The regular, periodic streamflow fluctuations induced by sub-daily snowmelt or ET could also impact the hyporheic exchange between surface water and groundwater , which in turn impacts nutrient cycling in the stream and hyporheic zone biogeochemical processes . By contrast, it is challenging to match the simulated variable from high temporal resolution with field observation. For example, the performance in simulated discharge deteriorates when temporal resolution increases from daily to hourly using PRISM (Fig. ). Additionally, hourly meteorological forcing is difficult to obtain and may be subject to large bias and errors. There are also sparse weather stations that collect hourly or higher-frequency data. Thus it is impossible to obtain sub-daily resolution by direct interpolation across weather stations. The current available hourly meteorological forcing is usually disaggregated from coarse temporal resolution. For example, the hourly NLDAS is disaggregated from NARR 3-hourly frequency. Previous studies have shown that NLDAS had large discrepancies towards SWE at higher elevation where lower SWE was simulated . Air temperature has also been shown to be systematically colder in winter and warmer in the spring months compared to the observations . These biases could be attributed to the ∼12 km spatial resolution that greatly smoothed the local topographic variability.

There is a lack of high-resolution observation data to compare to the simulated variables. For example, the snow survey from ASO has only been conducted a total of four times at this watershed and misses the temporal dynamics of snow depth. There is also no single SNOTEL station within the watershed that we can use to compare simulated SWE at a point location with the observed SWE. In addition to snow, we also do not have high-resolution ET data. Although MODIS provides an 8 d composite ET, it is relatively coarse compared to the temporal resolution examined in the study. In the subsurface, there are no observed groundwater table depth or soil moisture data that can be used for the comparison. The remote-sensed soil moisture product (9 to 36 km resolution) from Soil Moisture Active Passive (SMAP) is likely to be too coarse to have any meaningful comparison.

Uncertainty in the meteorological forcing has not been quantified. There is undoubtedly uncertainty in each GMF that may impact the simulated watershed responses; however, this is not the focus of this study. Precipitation collected from ground-based gages often has measurement uncertainty that could lead to large uncertainty in interpolated gridded datasets, especially in the mountainous region where the gage network is sparse . In general, PRISM is assumed to perform better in matching gage observation in the mountainous regions compared to Daymet because of the relatively complex interpolation method . However, PRISM does not outperform Daymet in all gages (see Fig. ). Further, in areas where the terrain is flat and the gage network is dense, Daymet might perform better than, if not equally well as, PRISM. Therefore, it is important to put the results into context when comparing different meteorological forcings in a watershed setting.

Uncertainty in model parameterization has not been investigated; however, it does not change the conclusions of this study, as all simulations use the same set of model parameters except for meteorological forcing. It is well known that model parameters such as subsurface structure and properties impact surface and subsurface flows and consequently ET and water storage. As shown in this study, models using different meteorological forcings may produce dramatically different watershed responses including streamflow. This has important implications for model calibration, where the objective is to minimize the differences between simulated and observed streamflow, which is true for most watershed hydrologic model calibration studies . The choice of GMF affects the simulated streamflow and in turn the optimal parameters that are calibrated using the simulated streamflow. showed that there were substantial differences in calibrated model parameters and simulated water balance using four different meteorological forcings for the same watershed. As a result, the choice of meteorological forcing plays a critical role in model calibration and thus long-term planning and watershed management using such a calibrated model. Studying the impact of different meteorological forcings on model calibration is the focus of our future work.

In this study, we choose Coal Creek as an illustrative example to show the effects of meteorological forcing spatiotemporal resolution on watershed simulations. The study site has strong variations in topography and land cover, which is an ideal site for testing heterogeneous spatial–temporal pattern of meteorological forcing. Our conclusions would hold for other mountainous headwater watersheds that are dominated by snow because we did not make any site-specific assumptions. However, additional studies are needed to evaluate the GMF in other areas that are not dominated by snow.