Remotely sensed snow cover observations provide an opportunity to improve operational snowmelt and streamflow forecasting in remote regions. This is particularly true in Alaska, where remote basins and a spatially and temporally sparse gaging network plague efforts to understand and forecast the hydrology of subarctic boreal basins and where climate change is leading to rapid shifts in basin function. In this study, the operational framework employed by the United States (US) National Weather Service, including the Alaska Pacific River Forecast Center, is adapted to integrate Moderate Resolution Imaging Spectroradiometer (MODIS) remotely sensed observations of fractional snow cover area (fSCA) to determine if these data improve streamflow forecasts in interior Alaska river basins. Two versions of MODIS fSCA are tested against a base case extent of snow cover derived by aerial depletion curves: the MODIS 10A1 (MOD10A1) and the MODIS Snow Cover Area and Grain size (MODSCAG) product over the period 2000–2010. Observed runoff is compared to simulated runoff to calibrate both iterations of the model. MODIS-forced simulations have improved snow depletion timing compared with snow telemetry sites in the basins, with discernable increases in skill for the streamflow simulations. The MODSCAG fSCA version provides moderate increases in skill but is similar to the MOD10A1 results. The basins with the largest improvement in streamflow simulations have the sparsest streamflow observations. Considering the numerous low-quality gages (discontinuous, short, or unreliable) and ungauged systems throughout the high-latitude regions of the globe, this result is valuable and indicates the utility of the MODIS fSCA data in these regions. Additionally, while improvements in predicted discharge values are subtle, the snow model better represents the physical conditions of the snowpack and therefore provides more robust simulations, which are consistent with the US National Weather Service's move toward a physically based National Water Model. Physically based models may also be more capable of adapting to changing climates than statistical models corrected to past regimes. This work provides direction for both the Alaska Pacific River Forecast Center and other forecast centers across the US to implement remote-sensing observations within their operational framework, to refine the representation of snow, and to improve streamflow forecasting skill in basins with few or poor-quality observations.
Arctic climate change is rapidly transforming the north with a myriad of
impacts on the hydrologic realm, which has important implications for the
largest biome on Earth, the boreal forest. For the northernmost United
States (US) state, Alaska, climate change has affected the hydrology,
ecology, and society in significant ways (Euskirchen et al., 2009; Hinzman
et al., 2005, 2013; Wrona et al., 2016). Alaska has warmed
more than 2 times the rate of the rest of the US since the 1950s (Karl et
al., 2009). Interior boreal Alaska has warmed the most of all regions in the
state, increasing by 4
A challenge for scientists attempting to accurately represent the impacts of climate change on the Alaskan hydrosphere is the vast territory, complex landscape, and sparse observational network. Alaskan hydrologic systems suffer from large uncertainties in various data inputs, and thus require care when attempting to simulate hydrologic water balance components with skill. For example, precipitation measurements are of very poor quality in winter (Cherry et al., 2005, 2007; Groisman et al., 2014) and river stage and discharge measurements by automated gages do not read accurately when ice is present in the river. Reducing these uncertainties is important, as they will reduce the value of model output (Magnusson et al., 2015; Slater et al., 2013; Clark et al., 2017) and the results cannot provide actionable guidance on water resource management (Stocker et al., 2013). In addition, the variability in landscape (i.e., forest cover, topography, discontinuous permafrost) and climate across Alaska requires robust modeling techniques to account for potential climate-driven shifts. This adaptable approach is increasingly important as the National Oceanic and Atmospheric Administration (NOAA) National Weather Service (NWS) develops the National Water Model (NWM) framework, a multi-scale water prediction model in operations over the contiguous US (NOAA, 2017). Temperature index models, based on the most reliable climate forcing, are often presumed to perform better than other models for regions with highly variable landscapes and a sparse network (Hock, 2003; Stahl et al., 2006). Alternatively, a skillfully calibrated conceptual model may provide a better representation of hydrologic responses because the underlying model is reliant upon parameterizations rather than observations that lack spatial and temporal consistency (Franz et al., 2008; Reed et al., 2004).
To deal with the inoperability of stream gages during breakup and in situ snow observations, one technique is to use remotely sensed snow cover areal extent (fSCA) to supplement point observations such as temperature, precipitation, and streamflow commonly used both as model inputs and for model calibration and validation (Parajka and Blöschl, 2008). There are two main ways that these data have been used to date: to either directly insert a time series of fSCA data into the model (McGuire et al., 2006; Rodell et al., 2004) or use complex assimilation procedures to filter the snow series and merge it with observational data (Andreadis and Lettenmaier, 2006; Sun et al., 2004; Zaitchik and Rodell, 2009). There is a concern that direct insertion methods are ineffective at improving streamflow models and do not perform better than uninformed models because melt can occur before snow cover drops below 100 % (Clark et al., 2006). In addition, the melt season duration is often short, transitioning rapidly from snow covered to snow free, although this is largely basin dependent (Clark et al., 2006). Assimilation approaches have yet to be integrated into operational models, in part because of the limited research showing the impacts of assimilation on the hydrologic forecast. Other studies have found calibrating models based solely on fSCA values may not improve skill in estimating discharge, and the improvements for in-catchment distributed fSCA estimates do not always result in improved discharge simulation (Franz and Karsten, 2013; Duethmann et al., 2014). However, Liu et al. (2013), Thirel et al. (2013), and Déry et al. (2005) found marked improvements in land surface model output for basins in Alaska when MODIS data were applied.
One approach to improve streamflow forecasts under climate change is to utilize newly developed frameworks to ingest remotely sensed data on snow cover area into streamflow models. These newer tools have been adopted by the NWS's River Forecast Centers (RFCs) and offer an opportunity for more advanced streamflow forecasting techniques, including ensemble prediction using variable input and/or forcing data. The Community Hydrologic Prediction System (CHPS), brought online in 2012 by the Alaska Pacific River Forecast Center (APRFC), is a test case for this approach. The modeling framework, developed on the Delft-FEWS software platform (Werner et al., 2011), can run many different types of models, but in its current state implements the conceptual Sacramento Soil Moisture Accounting System (SAC-SMA) rainfall–runoff model (Burnash et al., 1973), with snowpack input from the SNOW-17 snow model (Anderson, 2006).
The objective of this paper is to adapt the CHPS operational forecasting modeling framework to ingest Moderate Resolution Imaging Spectroradiometer (MODIS) remotely sensed fSCA data for improved streamflow modeling of the interior boreal forest region of Alaska within sparsely and poorly observed river basins that are experiencing shifts associated with a changing climate. We replace the standard areal depletion curve used in SNOW-17 with preprocessed MODIS fSCA grids for snow depletion. Two different versions of MODIS are applied: the MOD10A1 fractional fSCA product, which is the standard MODIS global snow cover product (Hall et al., 2002), and the MOD-Snow Covered Area and Grain size (MODSCAG) fractional fSCA product, which is a regional product (Painter et al., 2009). The SNOW-17 manual calibration using all model parameters is evaluated, including a tolerance parameter controlling snow cover updates (snow cover tolerance, SCTOL), to simulate a mixed method between direct insertion and more complex data assimilation. Preprocessing, model frameworks, and use of existing parameterizations are thus offered as a means of incorporating remotely sensed information into operational models that can be utilized out of the box by the NWS RFCs. The paper also examines issues around the use of MODIS fSCA in high-latitude boreal forest basins, the interpolation of missing data, and the improvement of streamflow estimates by calibrating model parameters used in streamflow forecasting systems across the US (Woo et al., 2008; Prowse et al., 2016).
Sub-basin characteristics, including name, sub-basin ID, sub-basin
unit, area, elevation mean (range), average monthly temperature,
This study was carried out in five adjoining headwater sub-basins of the
Tanana River, which is a sub-basin of the Yukon River basin (Fig. 1). The
sub-basins include the Chatanika, Upper Chena, Little Chena, Salcha, and
Goodpaster basins. The Chatanika River basin (64
Map of the five study basins with upper and lower divisions shown. Alaska SNOTEL sites are shown with numbered black triangles: (1) Fairbanks International Airport; (2) Little Chena Ridge; (3) Munson Ridge; (4) Mt. Ryan; (5) Monument Creek; (6) Teuchet Creek; (7) Upper Chena (Table 2).
The MODIS satellite product (Terra MOD10A1, version 5) provides daily 500 m
resolution fractional snow cover area (fSCA) data. It was downloaded from
the National Snow and Ice Data Center (Hall and Riggs, 2007; Hall et al.,
2006; Riggs et al., 2006) for 2000–2010, and we used the MODIS Re-projection
Tool (MRT, USGS, 2011) to preprocess imagery into an Alaska equal-area
conic projected GeoTIFF of fractional fSCA for each sub-basin, which
assisted us to correct, in part, the viewing geometry and other issues
related to projections of the original MODIS data and the influence these
projections have on the MODIS data for Alaska (Dwyer and Schmidt, 2006; Tan
et al., 2006). MODSCAG data products were obtained from the NASA Jet
Propulsion Laboratory's Snow Data System Portal (
Both MOD10A1 and MODSCAG fractional products require correction to adjust
the values of fSCA estimates (Raleigh et al., 2013; Rittger et al., 2013),
which do not account for the snow that is blocked from the sensor view. For
the MOD10A1 fSCA product, this calculation is based on the viewable gap
fraction, or the amount of snow-covered ground between trees that the sensor
can see (Liu et al., 2004). This technique, while widely applied, assumes
that the viewable gap fraction remains constant through the snowmelt season,
which is incorrect as the viewable gap fraction can vary based on a complex
number of factors, including forest canopy density, age and class, zenith
angle of the sensor, solar zenith angles, topography, and snow loading (Kane
et al., 2008; Liu et al., 2008; Molotch and Margulis, 2008; Raleigh et al.,
2013; Rittger et al., 2013). To account for some of these issues, rather
than applying a forest cover product to correct the product itself, the
MOD10A1 data are used (Durand et al., 2008). All 2000–2013 1 to 15 March MOD10A1 pixels across interior Alaska are differenced from 100, and
then a composite average of all days (
SNOTEL stations, map identification, length of record, and observed average snow water equivalent (SWE) used for validation of modeled SWE results. Average April SWE is calculated for the entire period of record.
Mean areal values of temperature and precipitation at 6 h increments are
obtained for each sub-basin from the APRFC for the time period 1969 to 2012;
only the 1999–2010 data are utilized in this study. River discharge at each
gage is based on the US Geological Survey (USGS) gaging record database. The
exception to this is the Chatanika River basin, where observed discharge is
generated based on once-a-day stage readings from a Cooperative Network
observer. These daily stage readings are converted to mean daily discharge
using the APRFC's rating curve for the river. Aspect and elevation were
calculated using the 30 m US Geological Survey's National Elevation Dataset
(NED), updated for the region in 2012 (Gesch et al., 2002). Seven snow
telemetry (SNOTEL) sites are utilized to compare simulated snow water
equivalent (SWE) with observed data (Table 2, NRCS 2013). SNOTEL SWE is
downloaded from the National Resource Conservation Service (NRCS) snow
pillow data repository
(
Potential evapotranspiration (PET) estimates are provided by the APRFC based
on an assessment of historical potential evapotranspiration from pan
evaporation data and Thornthwaite estimates (Anderson, 2006). These data are
used to develop a general linear relationship between PET and elevation to
estimate average monthly PET values for a generic low-elevation site. The
APRFC uses the low-elevation PET values to derive monthly estimates for the
mean elevation of each sub-basin as a coefficient,
The SNOW-17 and the SAC-SMA models are run by the APRFC in an operational framework referred to as CHPS. CHPS is built upon the Delft Flood Early Warning System (FEWS), developed by Deltares (Werner at al., 2013). The CHPS system is briefly described in Sect. S1.2.
SNOW-17 model parameters. Sensitivity indicates whether a parameter has a major or minor influence on model output. Minimum (Min) and maximum (Max) parameter values used in the model simulations. When min and max values are the same the parameter did not vary.
The SNOW-17 snow model is a single-layer snow model that calculates snow accumulation and ablation using empirical formulae to estimate heat and liquid water storage, liquid water throughflow, and snowmelt (Anderson, 1976). The model is designed for river forecasting and has been used operationally by the NWS RFCs since the mid-1970s. The only input requirements for SNOW-17 are temperature and precipitation (winds are accounted for but not input as observations), at the model time step (6 h). There are 12 parameters in the SNOW-17 model, including the areal snow depletion curve; sensitive or “major” parameters control the model outputs while less sensitive or “minor” parameters have little impact on the model output (Table 3; He et al., 2011).
SNOW-17 determines the division between rain and snow using the rain–snow
elevation (RSNWELEV) module. RSNWELEV uses a defined lapse rate
(6
A simplified energy balance method is used to calculate melt from
rain on snow using the following assumptions; the Stefan–Boltzmann constant
is used to estimate incoming longwave radiation, negligible shortwave
radiation, and 90 % relative humidity, and wind speed is accounted for by
adjusting for the average value of the wind during rain-on-snow events using
the parameter UADJ (mm hPa
The snow heat deficit is either negative or positive; the rate of heat loss
or gain is based on the amount of energy exchange that occurs when melt is
not taking place at the snow surface (negative melt factor; NMF; mm
SNOW-17 uses an areal depletion curve (ADC) to represent the snow cover area;
the ADC is used to calculate the area of the basin over which surface melt,
changes in heat storage, ground melt, and rainfall on bare ground occur
(Anderson, 2002; Fig. 7.4.3). The ADC not only represents areal extent of
snow cover, but also accounts for slope, aspect, and differences in
vegetative cover (i.e., open versus closed sites; Anderson, 2002; Fig. 7.4.3). In the baseline model simulation, the areal extent of snow cover was
calculated from a lookup table (Anderson, 2002; Fig. 8) that defines the ADC
and relates it to the ratio of SWE to either (a) the maximum value of SWE
that occurred during snow accumulation or (b) a parameter (SI) that
represents the areal SWE at which 100 % snow cover exists (referred to as
the areal index). The ADC in the baseline model simulation is applied as
follows: when snow accumulates, the snow cover is set to 100 %, and it
stays at this value until it falls below SI or the maximum SWE value,
whichever is smaller. If new snow totaling greater than 0.2 mm h
Snow cover area (SCA, %) for the Upper Chena River basin north slope from SNOW-17 and from MODIS in March to May 2010. Large decreases in the MODIS SCA are observed compared to the SNOW-17 SCE.
The SAC-SMA model is a conceptual rainfall–runoff model that simulates streamflow from observed input precipitation and PET (Burnash et al., 1973). SAC-SMA has been widely applied by the NWS to estimate streamflow runoff in basins across the US. The model moves water into either an upper or lower storage zone that conceptually represent soil interception or deep groundwater storage. Interception water in the upper zone flows to the lower zones via downward percolation, or it can run off directly or via interflow when the upper zone layers become saturated and the precipitation rate exceeds downward percolation. Lower zone water can be held in tension storage and contribute to baseflow runoff slowly over time, or can run off more quickly over shorter durations. Drainage from the upper and lower zones follows gravity drainage and is governed in part by both water delivery from the upper zone and soil moisture in the lower zone. Tension water is driven by potential evapotranspiration and diffusion, with a fraction of the lower zone unavailable for potential evapotranspiration as it is considered below the rooting zone.
A unit hydrograph model is used to adjust runoff timing for each lumped basin in the SAC-SMA model. Each sub-basin has its own unit hydrograph to translate the runoff through the channel system to the gage location. Simple routines sum the unit hydrograph outputs to calculate simulated streamflow at the basin outlet. While downstream basins incorporate routing models to move water from upstream to downstream basins, this study focuses on headwater basins so no routing models are needed.
Several calibration procedures were undertaken for this project: the baseline calibration and the two MODIS data set calibrations. The baseline calibration effort updated the SAC-SMA–SNOW-17 model parameters to the 2000–2010 years used in this study, as they had previously been adjusted by APRFC to 1970–2003 historical data. The two MODIS manual calibrations used the updated baseline to adjust parameters and generate statistics. Calibration entailed using both visualizations of streamflow hydrographs from 2006 to 2010 and monthly statistics from the entire period of record for ultimate parameter selection.
To calibrate the MODIS model output, a simple approach is taken to minimize the terms required for calibration. This ensures that it was (a) easy to replicate the model adjustments to the MODIS fSCA data and (b) solely focused on the snow parameterization, as adjustments to the SAC-SMA parameters resulted in only minor improvements to model calibration statistics during the spring ice breakup period. Also, priority was placed on adjusting the empirical parameters towards a physically based realization using basin and sub-basin unit properties, including the topographic aspects and the observed melt trajectory impacted by the MODIS fSCA data. To complete this simple, physically realistic calibration approach only the parameters MFMAX and TAELEV were adjusted. Further details of the calibration efforts are described in Sect. S1.4.
For validation purposes, statistics from 2000 to 2005 are provided for all
basins except the Chatanika. The Chatanika basin was calibrated from
2000 to 2004 and validated from 2005 to 2010 to make use of the better data
quality and availability during the first 5 years of the study.
Statistics used to evaluate model success are based on five main objective
functions and monthly average daily model output. The first two of these
criteria are standard in NWS RFC calibration approaches and are provided in
the CHPS statistical output. These statistics were used for evaluation
during the calibration: total volume bias as a percent (PBIAS, %) and the
correlation coefficient (
April–May–June monthly calibration (Cal), validation (Val), and the
period of record (Per., 1999–2010) statistics (MAE: mean absolute error
(m
The APRFC SAC-SMA–SNOW-17 baseline model estimates of streamflow in the
Alaskan interior river basins for the 11-year period of record indicate that
these basins are captured with skill (Table 4). The Chatanika basin is
problematic given the limited quality and quantity of the observed
streamflow data, as noted in the statistics below for each objective
function. For all of the five basins analyzed, the daily average bias for
the period of record is
SNOW-17 parameters for the MOD10A1 calibration. North (N), south (S), upper (U), and lower (L) sub-basins are described. For each sub-basin, the first column indicates the parameter value in the APRFC calibration and the second column indicates the parameter value used in the MODIS calibration. Bold values indicate where the MODIS value differs from the APRFC value.
Calibrated SNOW-17 parameters for the APRFC and MOD10A1 simulations resulted in increased MFMAX for the north-facing aspect in two sub-basin units and increased TAELEV for the north slopes (Table 5) compared to the baseline APRFC SAC-SMA–SNOW-17 simulation. In some sub-basin units, TAELEV was set to be equal for the north and south slopes. MFMAX for the Chatanika's lowland sub-basin increased and TAELEV at the north sub-basin was increased, while TAELEV was decreased for the south sub-basin unit. MFMAX in the Upper Chena north was unchanged and TAELEV was equalized for both south and north sub-basin units. The Little Chena sub-basin parameters were altered by setting MFMAX equal to its maximum recommended value for forested regions (1.4; Anderson, 2002; Table 7-4-1) for the upper and lower sub-basins and by increasing TAELEV 100 m greater than the elevation for both sub-basins. TAELEV for Salcha and Goodpaster were differenced by 100 m for the north and south sub-basin units, and the northern sub-basin MFMAX for Goodpaster was increased slightly. Goodpaster's lower basin MFMAX was reduced by a small amount. Although these changes may appear minor, MFMAX is highly sensitive during the melt season and therefore these changes have a substantial effect on the MODIS fSCA forced snowmelt trajectory at these sites (Anderson, 2006).
Snow cover area (%) based on MOD10A1 SCA average across all
watersheds divided into elevation zones. The years 2000 to 2010 are shown,
with the mean of all years in the final panel. Elevation zones are
SNOW-17 parameters for the MODSCAG calibration. North (N), south (S), and lower (L) sub-basins are described. For each sub-basin, the first column indicates the parameter value in the APRFC calibration and the second column indicates the parameter value used in the MODIS calibration. Bolded values indicate where the MODIS value differs from the APRFC value.
In the MODSCAG simulations, values for MFMAX were increased slightly for the
north sub-basin units for all basins. TAELEV values were adjusted slightly
in Upper Chena, Salcha, and Little Chena basins (Table 6), but were not
altered from the baseline run in Chatanika. In the Goodpaster basin, the
TAELEV value for the south sub-basin unit was decreased. NMF was altered
slightly for both MODIS simulations to account for different snow densities
and thermal conductivities of snow on south and lowland sites versus north
aspects. Snow density (gm cm
Simulated SWE (mm) versus SNOTEL SWE (grey line) for APRFC (solid black line), MOD10A1 (blue dashed line), and MODSCAG (orange dotted line) for October 2000 to September 2001. The Upper Chena River basin north slope is shown in the left panels, and the south slope is shown in the right panels.
Compared to the APRFC simulations, the MODIS simulations have less snow cover on the north-facing slopes and more on the south-facing slopes (Fig. 4; the average Upper Chena River basin unit results for 2001 plotted against the SNOTEL stations are shown as an example). Differences between the two simulations become discernable in late January as a result of the different calibrations of the SNOW-17 model in the basins (Fig. 4), with larger differences at the north sub-basin units compared to the south sub-basin unit. As soon as the MOD10A1 fSCA begins to alter the weighting factors for outflow from the snow, differences between the SWE generated by APRFC and MODIS simulations are observed. The greatest differences between the model simulations occur during the melt season. All model simulations peak in early April and start a downward melt trajectory, reflecting melt patterns at the upper-elevation SNOTEL sites: Mt. Ryan, Munson, and Upper Chena. The APRFC and MOD10A1 runs melt out later than the MODSCAG fSCA north unit and the MODSCAG estimates are closer to the APRFC simulations in volume, although all simulations terminate on the same approximate day for the northern sub-basins.
The SNOTEL sites are mostly located at upper elevations (Mt. Ryan, 850 m;
Munson, 940 m) compared to the SNOW-17's
The areal extent of snow cover varies across the basins in both simulations. The preprocessed gridded MOD10A1 fSCA illustrated for 15 May 2001 is shown in Fig. 5a and the MODSCAG fSCA is shown in Fig. 5b for the basins. The high-elevation snowpack (blue) is present within the upper basin regions but the pack is largely gone in the valleys and lower basin reaches. This translates into the lumped average fSCA estimates shown in Fig. 5c and d, which illustrate how CHPS ingests and converts the gridded MODIS fSCA for the sub-basin units. North and south sub-basin units are differentiated in the upper sub-basin units (see Table 1) but not at other locations because both aspects have begun to melt by this date (as opposed to early in the melt period when the south slopes would have comparatively less fSCA than the north slopes). MODSCAG has less cloud cover interaction in this scene (Fig. 5b) and this results in slightly higher values of fSCA (Fig. 5d).
Study area snow cover area in the CHPS model framework for
SWE estimates for MOD10A1 (Fig. 6a), MODSCAG (Fig. 6b), and the difference between the MODIS (both versions) and APRFC runs (Fig. 6c and d) is shown for 15 May 2001. Sub-basin units can be clearly differentiated in these plots, which illustrate the range of SWE values from 0 to 25 mm in the lowland regions to 125 mm in the upper headwaters. The MODSCAG data have an average fSCA value of 0.51 (51 %) and SWE is 45 mm, whereas the MOD10A1 has an average of 0.45 (45 %) fSCA and an average of 54 mm SWE, very small differences overall although sub-basin to sub-basin the variation between the products is notable. The difference plots highlight the fact that MODIS tends to have lower SWE values compared to the APRFC SNOW-17 model simulations on the north-facing slopes and higher values on the south-facing slopes. The APRFC tends to have lower SWE estimates for the lowland regions, although this is more true for MOD10A1 than MODSCAG (Fig. 6c, d).
Calibration and validation results are provided for April–May–June (Table 4) for the MODIS and APRFC simulations. For MODIS data, many statistics are similar or nearly identical to the APRFC run with slight declines in model performance and some gains (Chatanika; Little Chena), particularly for the analysis focused on the whole period of record (Table 4). NSE statistics are particularly poor for all simulations in the Chatanika basin, where the lack of continuous and high-quality observations hampers calibration efforts. The MOD10A1 data improve streamflow simulations in the Chatanika and Goodpaster systems during the calibration period, while they perform similarly or slightly worse during the validation and period of record in most of the basins except the Chatanika. The MODSCAG run exhibits better performance compared to the APRFC run during the calibration periods in the Chatanika, Salcha, and Goodpaster basins, while the validation period statistics showed improvement for the Chatanika, Little Chena, and Upper Chena basins. Overall, improvements in skill are observed for the MODIS simulations in the Chatanika and Goodpaster basins, the validation period for Upper Chena, and the calibration period for Goodpaster (Table 4).
Study area basin SWE (mm) estimates in CHPS model framework for
The calibration, validation, and whole period of record results illustrate
that for the poorly performing basins, MODSCAG (and MODSCAG with SCTOL
Monthly RMSE (m
Average (2000–2011) streamflow for each basin in Fig. 9 illustrates the
variations between simulated specific discharge (m
Percent difference between observed streamflow and that modeled
using APRFC plotted against modeled MOD10A1 (blue) and MODSCAG (orange)
for March–June from 2000 to 2010. The APRFC percent difference (
Upper Chena River basin streamflow: observed (grey line),
simulated with APRFC (black dotted line), simulated with MOD10A1 (blue
dashed line), and simulated with MODSCAG (orange dashed line) for all years
(2000–2010, 15th of the month shown on
Two methods were applied to integrate the MODIS data into CHPS. One method
involved interpolating between missing data values, changing the number of
interpolated days from 1 to 11 to investigate how changing the value
impacted model results. Generally, the number of days of interpolation had
little impact, but the longer interpolation period results produced slightly
higher correlations and improved streamflow estimation. We also investigated
the response to altering model parameter SCTOL, which can be used by
forecasters to combine the strength of the ADC and the MODIS data and is
similar to partial rule-based direct insertion approach; however the
parameter can be altered without any additional changes to the CHPS model
framework. Table 7 illustrates the results of setting the SCTOL parameter to
0.25, 0.50, and 0.75 for the MODSCAG run only, while holding the rest of the
parameters constant. No recalibration is performed. NSE and
Comparison between RMSE (%) and NSE (in brackets) for April–May–June using SCTOL values of 0.25, 0.50, and 0.75. Absolute differences are calculated from the MODSCAG base run. Cal: calibration; Val: validation; Per: period of record.
Results illustrate that streamflow in interior Alaska can be simulated with skill using conceptual, semi-lumped hydrologic models, even without the use of gridded observations of MODIS fSCA. However, if the initial streamflow observations are of poor quality (i.e., Chatanika River basin), applying gridded observations of MODIS fSCA in the models will generate streamflow estimates as good as or better than estimates based on SNOW-17's areal depletion curve. However, as the climate shifts, conceptual, semi-lumped models may not be representative of process changes that will likely occur as the Arctic warms (Clark et al., 2017). As fully process-based models are challenging to run in Arctic environments, where high-quality data are temporally and spatially sparse, using conceptual models parameterized with as many observations as possible represents a bridge between the fully process-based models and conceptual approaches to hydrologic modeling.
However, we found there to be major challenges in obtaining improvements in simulated streamflow discharge values when introducing additional observed data sets and their associated uncertainties into models. This result was also found in work performed in the American River basin where the California Nevada RFC lumped model provided the most accurate representation of snow cover area (Franz and Karsten, 2013). As indicated by Franz and Karsten (2013), although the gridded representation of fSCA is improved in their distributed version of SNOW-17, the streamflow simulations and associated statistics did not reflect this improvement. In addition, they found that discharge values had lower skill when estimates of snow cover are included in the calibration even though it is hypothesized that the process representation is improved, which is a finding of a number of other research studies focusing on this topic (Parajka and Blöschl, 2008; Udnæs et al., 2007). These findings are also true for Alaskan interior boreal basins, highlighting the importance of performing this work in remote areas and under monitored systems that are changing quickly due to climate shifts and increased occurrences of extreme events (Bennett and Walsh, 2015; Bennett et al., 2015).
The goal of this work was, in part, to undertake a simple application of inserting preprocessed MODIS fSCA into the CHPS operational framework to simulate streamflow across basins in interior Alaska. The preprocessing of MODIS data for insertion into the model, which included the MOD10A1 and MODSCAG data products, along with the CHPS areal averaging, eliminated some of the issues related to cloud cover and missing data, as noted in results provided in Liu et al. (2013), who assimilated Air Force Weather Agency–National Aeronautics and Space Administration Snow Algorithm (ANSA) fSCA data for similar stations in the region. For example, the findings in Liu et al. (2013) for the best case indicate NSE improvement for Salcha, Little Chena, and Chena at Fairbanks of 0.30, 0.31, and 0.06. Our study reports comparable NSE improvement values for some stations (Chatanika and Goodpaster) for the months impacted by the adjustments, although the Salcha and Little Chena system differences are closer to those values reported for the raw MODIS data in the study of Liu et al. (2013). The averaging approach and use of newly developed tools (ANSA, MODSCAG) applied in both studies appear to produce results slightly superior to those of MOD10A1. Further analysis is required to determine if cloud correction processes, such as those applied in the ANSA study, would act to reduce the impact of pixel shifting that is likely a major problem in Alaska (Arsenault et al., 2014) and improve streamflow estimates further. Both studies indicate improved representation of internal snowpack and improvements in streamflow estimates for some basins for these new iterations of the MODIS data.
Differences in the streamflow improvements provided by Liu et al. (2013) for the Salcha and Little Chena basin highlight some important variations between the two studies that should be considered. The first is that, as noted by the authors, the model-simulated streamflow estimates are biased and thus the improvements reported in the paper are still poor representations of the streamflow (Liu et al., 2013). The question then remains that if a model result without updated observations is already skillful, how much better can the model be with added information (which carries its own uncertainty with it)? Perhaps the differences between the distributed model in Liu et al. (2013) versus the lumped models used in this study are adding a buffer to the data improvements in the case of this study, and limiting the amount of difference or improvement that MODIS fSCA insertion can provide. Snow cover data appear to be improved at interior Alaska locations within the model when compared to five different SNOTEL stations (Fig. 5), particularly for the melt timing. However, the discharge values improved moderately given MODIS input over the different periods analyzed, and in particular smaller changes are noted over the entire period of record (Table 4, Figs. 8, 9). For the Chatanika basin, with limited observed data and poorer streamflow simulations however, the improvements are similar to those shown in the Liu study (Liu et al., 2013). These results suggest that skill can be added by introducing new observations when the models are performing poorly due to inadequate or low-quality records. Considering that there are numerous incomplete and low-quality gages throughout the high-latitude regions of the globe, this result is valuable and indicates the utility of the MODIS fSCA data in this regard.
Calibrations performed on the SAC-SMA model were limited in nature and targeted specifically at two parameters exhibiting the most influence on improving discharge estimates during the melt season: MFMAX and TAELEV. These parameters control the air temperature and impact snow cover depletion by either increasing or retaining melt. Previously, the APRFC parameters were set to lower MFMAX values. The TAELEV parameter was not equal to the true elevation (ELEV) and set to different values for north and south aspects. For north-facing upper elevations, TAELEV was less than ELEV so temperatures were lapsed upward to simulate the slower melt rates and cooler conditions. For south-facing aspects, TAELEV was set to greater than ELEV, so temperatures were lapsed downward to simulate increased melt from solar influence. Our updated parameterization using the MODIS data required an upward adjustment of these values because the areal depletion curve is no longer controlling the melt rate. Thus, fSCA present on northern, upper-elevation slopes in the late spring must have higher melt rates applied to melt the snow with the correct timing. The primary reason that the areal depletion curves in SNOW-17 differ from one that would be derived from actual measurements of fSCA is that melt rates decline as fSCA declines because the remaining snow is usually found in locations where snow melts at a slower rate, such as under canopies or on north-facing slopes (Anderson, 2006).
Adjustments to MFMAX across the north sub-basin units suggest that the modified areal depletion curves within SNOW-17 underestimate snow-covered area. At many of the sites, particularly when using the MODSCAG product, MFMAX for the northern sites had to be increased. This suggests that the APRFC run uses a lower value that attempts to account for cooler temperatures on the northern slopes by retaining the snow on these slopes for longer, thus slowing runoff (Franz and Karsten, 2013). By more accurately representing conditions in the north sub-basin units, the MODIS simulations required an increase in the snowmelt factor to allow for initiation of the melt on these slopes. MFMAX represents the dependency between the melt factor to account for a constant fSCA curve used in the model and the ability of the “standard” fSCA curves used in the APRFC SNOW-17 to replicate the conditions of the melt properties within the basins (Shamir and Georgakakos, 2007). As noted in Shamir and Georgakakos (2007), there is considerable inter-annual variability in snow cover depletion and this variability is not represented when the standard APRFC model is applied. Therefore, by improving the internal physical processes in the model, the snowmelt timing should improve. However, this might not translate into improved discharge estimates because precipitation and temperature inputs could still be incorrect, and errors in forcing data that generate incorrect water equivalents for snow carry larger uncertainty bounds than that which can be addressed by changing the weighting factors and timing of snowmelt by adjusting fSCA, as undertaken in this study.
For the MOD10A1 calibration, fewer parameters were adjusted compared to the MODSCAG simulations. The end result is that the MODSCAG data have improved streamflow simulations compared to the MOD10A1 result. The model parameters require greater adjustment for MODSCAG simulations as a result of the variability between the two data sets compared to the APRFC baseline simulations. The MODSCAG data have a different melt trajectory for northern slopes and hold snow for longer on the south-facing slopes of the Upper Chena River basin, while the MOD10A1 acts similarly to the APRFC melt trajectory for SWE data (Fig. 4). This region is known to have variable melt timing based on south-facing slopes; therefore the north and south slopes should be differentiated to reflect the physical processes occurring on the warmer south-facing slopes compared to the cold and often permafrost-dominated north-facing slopes (Jones and Rinehart, 2010). Although MODSCAG improvement is noted for the Chatanika and Goodpaster basins in the streamflow statistics, the results for both MODIS versions are overall very similar in this region (Fig. 8). This may be due to the different canopy adjustments applied to the data sets or because of the lack of a spectral end member for the boreal forest in MODSCAG (Painter et al., 2009). Regardless, it is not clear that one of these data sets is markedly improving streamflow estimates and it is possible that both approaches could be considerably useful as additional observations of fSCA estimates for the region.
Two other means by which the CHPS framework can be altered to improve streamflow estimates are explored in this work. The interpolation over MODIS missing days can be altered easily in CHPS; however this had only a small effect on the streamflow results. The SCTOL, which allows for interaction between the model and the observed MODIS fSCA data, had an effect on streamflow and therefore may be a useful technique for the RFCs to apply during recalibration efforts to observed snow cover data. An advantage was noted between the MODSCAG with an SCTOL setting greater than to 0.25. However, the basins with the strongest improvement (Chatanika) over the APRFC simulation did not improve using an SCTOL greater than zero, which was because the baseline model performed so poorly given the weakness of the underlying observed discharge data. Therefore, the RFCs may wish to selectively apply this parameter when basins have reliable observed information and the MODIS data can be utilized partially in conjunction with the model ADC and partially on the MODIS fSCA observations.
Although complex tools and distributed models are available from the research community and in the CHPS to integrate observed snow cover area data, the RFCs across the US are not, as of writing this paper, using these features in their operational river forecasting to estimate floods and droughts. This study focuses on developing tools that can, with a minor amount of testing, be brought into the RFC's CHPS modeling framework and used to improve physical estimates of fSCA across basins of interest. The method integrates information such as MODIS remotely sensed snow cover into the model framework using a simple calibration approach for the SNOW-17 model, and also provides some input regarding expected improvements and other possible parameters that may be introduced to enrich forecasting and simulation of streamflow. Our recommendation is to incorporate MODIS data as an interim step. However, in the long run the RFCs should begin to use more complex models and data assimilation tools as the move towards the National Water Model proceeds.
In this work, we answer several outstanding questions regarding the application of MODIS data in the RFC models. Basins with poor-quality streamflow observations benefited from the use of the MODIS fSCA but improvements are also made to the internal snow timing estimates, observed in both the validation against SNOTEL data and also through the calibration that corrected the model parameters to better reflect the physical differences altering processes occurring on north- and south-facing slopes. Overall, minor differences were observed between MOD10A1 and MODSCAG data; however the MODSCAG data provided improvement over MOD10A1 in average streamflow simulations across all basins. We observed limited impact of changing the interpolation length between missing days, although adjustments based on altering the interaction between the model and the observed MODIS fSCA data did alter streamflow and therefore are useful during recalibration efforts.
The utility of the MODIS data in CHPS goes beyond improvements to the streamflow; these tools can be used for a number of internal checks for SWE and fSCA that are currently under way, such as the ingestion of data for ensemble forecasts (NWS, 2012). This study opens the door for insertion of parameters via assimilation alongside developments such as physically based model usage.
The observations of rapid change in the Arctic highlight important alterations to hydrological regimes in the subarctic interior boreal forest of Alaska. These observed, rapid changes and future anticipated alterations introduce a pressing need in Alaska to further understand the anticipated changes through modeling of major climate drivers of streamflow. The sparse observational network in Alaska, along with the magnitude and rate of change, necessitates the use of robust modeling tools to examine these changes and their impacts on hydrology. However, due to the limited high-quality observations, and our lack of understanding of Arctic hydrologic processes, process-based modeling approaches are limited in this environment. Therefore, we must apply available conceptual models with calibrations informed by observations, including remote sensing tools of SWE and fSCA to examine these effects. In this way, we will be able to define and quantify increasing impacts associated with these changes that lead to multi-scale risk to hydro-ecological systems, not only to the local and state resources, but also regionally and globally.
MODIS MOD10A data set is available for download through
The supplement related to this article is available online at:
All authors were involved in the conception, and development of workflow. KB and BB responsible for development of software manipulation of CHPS modeling framework. KB and JEC responsible for writing. KB, JEC, and SL contributed to editing process through various stages of the manuscript development.
The authors declare that they have no conflict of interest.
Katrina E. Bennett acknowledges support for this work from the Alaska Climate Science Center, the Natural Science and Engineering Research Council of Canada, and the DOE Office of Science, Biological and Environmental Research program, Next Generation Ecosystem Experiment, NGEE-Arctic project. Katrina E. Bennett, Ben Balk, and Jessica E. Cherry acknowledge support from the GOES-R High Latitude Proving Ground award NA08OAR432075. Thank you to Sarah Boon for reviewing an earlier version of this work. The authors thank the Alaska Pacific River Forecast Center staff for their collaborative efforts and the Arctic Region Supercomputer Center for the use of their supercomputer.
This paper was edited by Shraddhanand Shukla and reviewed by Stephen Déry, Edward Bair, and one anonymous referee.