Assessing High-Latitude Winter Precipitation from Global Precipitation Analyses Using GRACE (original) (raw)

1. Introduction

Global precipitation analyses incorporate multiple data sources to obtain adequate spatial and temporal resolution. Each data source has advantages and/or disadvantages relative to the others, and by optimally combining datasets, the bias and random error of the final precipitation estimate can be, in principle, minimized. For example, in situ precipitation gauges, corrected for systematic errors, typically have good accuracy and temporal sampling, but they represent a relatively small area as a result of the large spatial heterogeneity of precipitation (Huffman 1997; Milly and Dunne 2002). To enhance the spatial resolution of gauge networks, which can be sparse in areas of low population (especially over the oceans), analyses add information from satellites. By combining data from sun-synchronous, geostationary, and polar-orbiting satellites, near-global coverage may be obtained.

Global precipitation analyses are typically assessed via comparisons with independent gauge networks. Huffman et al. (1997) compared the version 1 precipitation analysis from the Global Precipitation Climatology Project (GPCP) to gauge-derived estimates from fifteen 2.5° × 2.5° cells in five geographic regions between ±40° latitude. They reported bias and rms errors in the monthly averaged precipitation rates, for the period 1987–91, of about 3.6 and 32 mm month−1, respectively. Xie and Arkin (1997), comparing the same gauge network to monthly precipitation estimates from the Climate Prediction Center (CPC) Merged Analysis of Precipitation (CMAP) global analysis, obtained similar results. Adler et al. (2003) reported that the GPCP version 2 product, when compared to a gauge network located in Oklahoma, exhibited a bias of about 1 mm month−1 and an rms error of 18 mm month−1.

Because of decreasing gauge network density (Rudolf and Mächel 2005) and lower-quality satellite estimates (Sapiano et al. 2008), the errors in global precipitation analyses are typically larger at high latitudes, especially during the cold season. For example, Sapiano et al. (2008) compared a global 2.5° × 2.5° gridded dataset called the Multisource Analysis of Precipitation (MSAP), as well as the GPCP version 2 dataset, to a gauge-based dataset composed of networks from the United States, Canada, and the former Soviet Union (Serreze et al. 2005). They found average biases of −14 for MSAP and −5 mm month−1 for GPCP with respect to gauge-derived values. Sapiano et al. (2008) noted that the larger cold-season MSAP bias could be partly explained by the gauge undercatch correction applied to both the gauge data and the GPCP data. Gauge undercatch is a systematic underestimation due to aerodynamic effects (Goodison et al. 1998). Adler et al. (2003) showed that much of the difference between precipitation time series from the GPCP version 2 and two gauge networks in Sweden and Poland for the period 1996–98 was due to the undercatch correction being incorporated into the GPCP analysis. When the same correction was applied to the two gauge networks, the GPCP bias was substantially reduced. More recently, Bolvin et al. (2009) compared a new, corrected GPCP version 2 product to a high-quality, high-density gauge network of the Finnish Meteorological Institute (FMI) spanning the period 1995–2007. They found a result consistent with Adler et al. (2003), such that when the FMI gauge data were adjusted for undercatch, very good agreement between the two datasets was obtained. It is important to note that during the cold season, the magnitude of the undercatch correction could be as large as the uncorrected signal.

Because of the uncertainties in gauge-based precipitation estimates due to low network density and wind-induced undercatch, we propose an alternative method of assessing high-latitude, cold-season precipitation using satellite gravimetric data from the Gravity Recovery and Climate Experiment (GRACE). During winter at high latitudes, much of the precipitation that occurs is stored as snow, while evapotranspiration and runoff are relatively small (Serreze et al. 2003; Lammers et al. 2001). Thus, GRACE estimates of total water storage (TWS) for many regions are primarily due to the accumulated precipitation signal. The advantages of using the GRACE dataset are 1) because it measures mass directly, no undercatch correction is necessary; 2) GRACE has a coarse, but comparable, spatial resolution (∼400 km) to satellite-based precipitation analyses; and 3) it is based on a completely independent technique (gravimetry versus radiometry) that does not include empirical parameterizations, nor does it require ground-based calibration.

Using a water balance framework, this study compares large-scale, regional anomalies of TWS from GRACE to accumulated precipitation estimates from two well-known datasets—GPCP and CMAP—during the cold season. To close the water balance, estimates of evapotranspiration and runoff are derived from output of the Community Land Model (CLM; Oleson et al. 2008). Focusing on the water balance during winter has two benefits: 1) it provides a tool for assessment when both in situ and satellite precipitation estimates are most uncertain, and 2) it minimizes the contribution of uncertainties in model-based evapotranspiration and runoff estimates.

2. Data

a. Multisatellite (MS) precipitation

1) GPCP

The GPCP produces global precipitation analyses at temporal resolutions ranging from daily to monthly and spatial resolutions of 1° × 1° and 2.5° × 2.5°. Here, we examine the 2.5° × 2.5° monthly-mean satellite–gauge (SG) precipitation version 2 product (Adler et al. 2003).

The gauge data used by GPCP are assembled and analyzed by the Global Precipitation Climatology Centre (GPCC; Schneider et al. 2003). An important feature of these data is the application of a correction that is meant to reduce systematic errors due to wind-induced gauge undercatch, wetting losses, and trace precipitation. The dominant component of these systematic errors is undercatch (Goodison et al. 1998). Correction factors, based mainly on gauge–height wind speed but also taking into account temperature and humidity, can be of the order of 100% during the snow season (Yang et al. 2005; Fuchs et al. 2001).

Between 40°N and 40°S latitude, GPCP MS analysis combines infrared and microwave precipitation estimates. Poleward of ±40°, infrared data are unreliable (Huffman et al. 1997). Frozen surfaces further complicate the retrieval of high-latitude terrestrial precipitation because the microwave scattering technique is not reliable when snow or ice is present on the surface (Xie and Arkin 1997; Adler et al. 2003). To compensate for the lack of data at high latitudes during the cold season, data from Television and Infrared Observation Satellite (TIROS) Operational Vertical Sounder (TOVS) is included in the MS product. The final SG precipitation estimate is produced by combining the gauge analysis with the MS analysis (Adler et al. 2003).

2) CMAP

The National Oceanic and Atmospheric Administration’s CPC produces a global precipitation analysis called CMAP. The CMAP dataset uses the same GPCC gauge analysis and infrared and microwave satellite data used by the GPCP; however, an aerodynamic correction to account for gauge undercatch is not used by CMAP. In addition, two more satellite datasets are used: outgoing longwave radiation (OLR) and Microwave Sounding Unit (MSU). OLR data (Xie and Arkin 1998) are used globally, whereas MSU data (Spencer 1993) are used over the oceans between 60°S and 60°N. CMAP further differs from GPCP by incorporating output from numerical weather prediction (NWP) models into the analysis in regions where the satellite precipitation analysis may be unreliable and where gauge networks are sparse (Xie and Arkin 1997). Thus, CMAP estimates of cold-season, high-latitude terrestrial precipitation result from a combination of gauge and OLR data and NWP model output.

3) GRACE

The GRACE satellite mission produces monthly estimates of the earth’s gravity field (Tapley et al. 2004). Because changes in the gravity field and changes in mass at the earth’s surface are directly related, GRACE data have been used to make global estimates of vertically integrated terrestrial water storage with a spatial resolution of a few hundred kilometers and greater, with higher accuracy at larger spatial scales (Wahr et al. 2004; Swenson et al. 2003). GRACE TWS estimates are sensitive to variations in water storage at and below the land surface. By combining GRACE TWS estimates with independent data, terrestrial water stores such as snow, surface water, soil moisture, and groundwater have been studied (Niu et al. 2007; Swenson and Wahr 2007; Swenson et al. 2006; Swenson et al. 2008).

In this study, we use Release 4 (RL04) data produced by the Center for Space Research (CSR). CSR gravity fields are provided as spherical harmonic coefficient sets complete to degree and order 60. The gravity field coefficients contain both random and systematic errors, so before the data are converted to mass (in units of equivalent water thickness) and gridded, a two-step filtering process is applied. Systematic errors, identified by correlations between coefficients, are removed using the filter described by Swenson and Wahr (2006), whereas the random error component, which increases as a function of decreasing wavelength, is reduced by smoothing the data with a Gaussian filter with a half-width corresponding to 250 km (Wahr et al. 1998).

4) CLM

The CLM is the land component of the Community Climate System Model (CCSM). CLM simulates the partitioning of mass and energy from the atmosphere, the redistribution of mass and energy within the land surface, and the export of freshwater and heat to the oceans. To realistically simulate these interactions, CLM includes terrestrial hydrological processes, such as the interception of precipitation by the vegetation canopy, throughfall, infiltration, surface and subsurface runoff, snow and soil moisture evolution, evaporation from soil and vegetation, and transpiration Oleson et al. (2008).

Simulations of the global land surface state were produced by running the model at 2.5° longitude by 1.9° latitude spatial resolution using observed forcing data generated by the Global Land Data Assimilation System (GLDAS; Rodell et al. 2004). GLDAS provides 1° × 1°, 3-hourly, near-surface meteorological data (precipitation, air temperature and pressure, specific humidity, short- and long-wave radiation, and wind speed) for the period 2002 to the present. Land surface state variables were initialized from the land surface state at the end of a multidecade offline run. This study uses the following modeled quantities: total (surface plus subsurface) runoff R, river water storage _S_river, and evapotranspiration ET, as well as near-surface air temperature _T_air from the forcing dataset.

3. Methods

a. Terrestrial water balance

The water balance of an arbitrary region can be expressed as follows:

i1525-7541-11-2-405-e1

i1525-7541-11-2-405-e1

where dS/dt is the change in total water storage with respect to time, P is the precipitation rate, _Q_in is inflow to the region, and _Q_out is outflow from the region. All terms in Eq. (1) represent spatially averaged quantities and are presented in units of millimeters per month.

Because GRACE estimates the total water storage anomaly S, rather than its derivative dS/dt, we first convert Eq. (1) to represent storage anomalies rather than fluxes by integrating over time

i1525-7541-11-2-405-e2

i1525-7541-11-2-405-e2

Rearranging Eq. (2) and introducing the notation

leads to the following expression for the precipitation storage anomaly, that is, the total accumulated precipitation during the integration period:

i1525-7541-11-2-405-e3

i1525-7541-11-2-405-e3

The net lateral flux of water from the region _Q_net = _Q_out − _Q_in is estimated from CLM output. Here _Q_in and _Q_out are related to the model variables R and _S_river by

i1525-7541-11-2-405-e4

i1525-7541-11-2-405-e4

such that

i1525-7541-11-2-405-e5

i1525-7541-11-2-405-e5

Because we are interested in assessing cold-season precipitation, _t_1 and _t_2 are specified as the times bounding the period when the regional average near-surface air temperature is below a threshold temperature, for example, _T_max = 0°C. In this study, _T_max is typically set to between 0° and −6°C. Storage anomalies are computed separately each year, referenced to a value of zero at _t_1.

As an example, Fig. 1 shows the time series of anomalies related via Eq. (3) for a region circumscribed by a 4° box centered at 56°N latitude and 38°E longitude. The top panel of Fig. 1 shows the GRACE TWS storage anomalies, and the middle panel shows the GPCP (blue line) and CMAP (green line) precipitation rates, as well as the combined rates of ET and _Q_net (yellow line). The vertical black lines in the top two panels delineate the period during the winter of 2006/2007 for which the regional average air temperature was below 0°C. (Integration periods are defined similarly for all other years, but are not indicated.) In this example, the GPCP precipitation rates are generally larger than the CMAP rates during winter, whereas the combined ET and _Q_net rates are relatively low.

Differencing the TWS anomalies (top panel) and integrating the rates (middle panel) for each year results in the time series shown in the bottom panel. The change in water storage alone is represented by the orange line, whereas the net change in storage, evapotranspiration, and runoff is shown as the red line. The larger GPCP rates shown in the middle panel are manifested as greater accumulations in the bottom panel. It is also notable that the GRACE storage changes are typically smaller than both the GPCP and CMAP accumulated precipitation, but they agree quite well with the CMAP values when combined with the accumulated ET and _Q_net.

b. Spatial filtering

As discussed in section 3, the raw GRACE data are filtered to reduce the errors in the short-wavelength gravity field coefficients. To eliminate any discrepancies that might occur as a result of this filtering, the same data processing is applied to each of the other datasets. Specifically, the gridded precipitation and modeled datasets are converted to spherical harmonic coefficients complete to degree and order 60. The GRACE filter is then applied, and the coefficients are regridded to a common latitude–longitude grid.

4. Results

a. Maps

In Figs. 2 –6, we compare maps of GPCP and CMAP accumulated precipitation to the sum of GRACE water storage changes and CLM modeled evapotranspiration and runoff, which we will refer to as “potential accumulation.” Each map is constructed by computing the mean accumulation at each grid point for the period of each year when near-surface air temperatures are less than −3°C. The top-left panel shows the GPCP accumulated precipitation, the top-middle panel shows the GRACE/CLM potential accumulation, and the top-right panel shows the CMAP accumulated precipitation. The spatial patterns of the three maps are very similar; spatial correlations of the GPCP and CMAP maps range from about 0.95 to 0.98, which might be expected because of their datasets in common, whereas the spatial correlations between the GRACE/CLM and either GPCP or CMAP maps range from 0.85 to 0.89.

Features common to all three maps can be seen in high accumulations in western and central Siberia, eastern Canada, and along the Rocky Mountains of the western United States and Canada. Low accumulations can be seen in central Canada and eastern Siberia. One obvious difference occurs over Greenland as a result of the lack of a pronounced signal in the GRACE/CLM map. This is due to the use of a temperature-based integration period; because the near-surface air temperatures are always near or below zero, the integration period over Greenland covers both the accumulation and ablation periods. Thus, the storage change seen by GRACE during this period is small. In addition, the land surface model does not realistically represent the mass balance over glaciated regions. Other glaciated locations are apt to show results similar to that seen over Greenland.

Although the spatial patterns of the accumulation maps are similar, the amplitude of the signals varies. The largest accumulations are found in the GPCP map. The CMAP and GPCP accumulations are similar over North America, but the CMAP values are considerably smaller over much of Eurasia. The GRACE/CLM amplitudes typically are smaller than both the GPCP and CMAP values in North America. The GPCP values are larger than GRACE/CLM over Eurasia, whereas the CMAP values are generally of similar magnitude or smaller.

The bottom panels of Figs. 2 –6 compare the GRACE/CLM potential accumulation maps to the individual GRACE and CLM components. The left panel shows the change in total water storage from GRACE, the middle panel, which is identical to the top-middle panel, shows the potential accumulation, and the right panel shows the combined ET and Q map. The GRACE water storage signal is the dominant component of the potential accumulation over much of the Northern Hemisphere, with the largest values along western Canada and the United States and western and central Siberia. GRACE shows very little accumulation in central Canada and between about 120° and 150°E longitude in Siberia. The largest values from the combined ET and Q maps occur in eastern Canada, with moderate values found over the southern portion of Siberia.

b. Time series

To examine more closely the temporal evolution of the accumulation signals, we present here time series for various regions throughout the Northern Hemisphere. Nine regions were chosen in both Eurasia and North America. The regions were chosen in an attempt to highlight areas of good agreement as well as areas where significant differences were found. Figure 7 shows a map of the nine regions selected for Eurasia.

Region 1 is located at the western edge of the Karakoram mountains, and region 2 is centered around the mountains of eastern Turkey, Armenia, and northwestern Iran. Both regions are located at midlatitudes, rather than at high latitudes, and are shown to examine the behavior of the data within the ±40° latitude band where IR data contribute to the precipitation estimates. As shown in the first two panels of Fig. 8, the accumulation estimates from GPCP, CMAP, and GRACE/CLM agree well for both regions. The integration period of region 1 is longer than that of region 2 as a result of the higher mean elevation, and therefore lower mean temperature, of region 1. The time series of both regions exhibit significant interannual variability, with year-to-year differences as large as 100 and 50 mm for regions 1 and 2, respectively. The mean differences, averaged over the five winters, are quite small, however. The average difference between GPCP and GRACE/CLM accumulations ΔGPCP is only 2 mm (<1% of the GRACE/CLM value), and the average difference between CMAP and GRACE/CLM accumulations ΔCMAP is −23 mm (−9%) for region 1. For region 2, ΔGPCP and ΔCMAP are 5 (4%) and −21 mm (−17%), respectively.

The close agreement between the GPCP and CMAP data over regions 1 and 2 is not representative of other locations in Eurasia (Table 1). At higher latitudes, the tendency for GPCP to estimate higher precipitation values than CMAP shown in Figs. 2 –6 can be seen in the regional time series as well. Average accumulations in regions 3–7 range from about 200 to 350 mm, and all of them show GPCP accumulations of the order 100 mm larger than those from CMAP. Except for region 4, and individual years in regions 8 and 9, CMAP and GRACE/CLM agree well; ΔCMAP values for these five regions are −5 mm (−3%), 83 mm (87%), −20 mm (−9%), −22 mm (−11%), and −52 mm (−21%). ΔGPCP values for these five regions are 140 mm (74%), 190 mm (200%), 107 mm (48%), 138 mm (69%), and 66 mm (26%). Region 8, located in eastern Siberia, represents an area of low accumulation, as can be seen in the GRACE maps of Figs. 2 –6. The time series for region 8, shown in Fig. 8, shows average accumulations of about 63 mm for GRACE/CLM, 74 mm for CMAP, and 101 mm for GPCP. The ΔGPCP and ΔCMAP for this region are 38 mm (62%) and 11 mm (18%), respectively. The final Eurasian region, region 9, samples the Kamchatka Peninsula. Again, the GPCP time series is larger than CMAP, with ΔGPCP = 136 mm (80%) and ΔCMAP = 33 mm (19%).

Figure 9 summarizes the average winter accumulations for each of the nine regions. It also shows how the GRACE/CLM accumulation is partitioned between GRACE, ET, and Q. Except for region 4, GRACE is always the largest component of the potential accumulation signal. Evapotranspiration is generally small and often negative (because of frost), except for region 1, whose mean elevation is more than 4000 m. Runoff is generally smaller (30%–90%) than the GRACE values, except for region 4, for which runoff is the largest contributor to the total signal.

Nine regions were also selected for North America and are shown in Fig. 10. Although the GPCP precipitation estimates tend to be slightly larger than the CMAP estimates, the large bias seen in Eurasia is absent. Instead, as shown in Fig. 11, GPCP and CMAP generally agree well, but large regional differences exist between GRACE/CLM and the precipitation estimates. Along the West Coast of North America, and in eastern Canada, GPCP and CMAP are larger than GRACE/CLM, whereas in the central part of the continent, GRACE/CLM is larger than GPCP and CMAP (Table 2). An area of good agreement is represented by region 1, which extends into the region below 40°N latitude where IR data is available. The ΔGPCP and ΔCMAP values for this region are 21 mm (27%) and −7 mm (−10%), respectively. In regions 2 and 3, which correspond to areas of low precipitation accumulation in the maps shown in Figs. 2 –6, GRACE/CLM potential accumulation is significantly higher than the GPCP and CMAP precipitation accumulation. In region 2, mean winter precipitation accumulations are only 20–34 mm, while potential accumulations average 109 mm. GPCP and CMAP values average about 50 mm in region 3, whereas GRACE/CLM averages 76 mm. The ΔGPCP and ΔCMAP values are −75 mm (−69%) and −89 mm (−82%) for region 2, respectively, and −27 mm (−35%) and −25 mm (−33%) for region 3, respectively. Regions 4–6 sample the mountainous region of the West Coast of the United States and Canada. In these regions, where precipitation accumulations range from 200 to 300 mm, GRACE/CLM accumulations are typically smaller than GPCP and CMAP, with the greatest differences between 50° and 60°N latitude. The ΔGPCP values for regions 4–6 are 58 mm (33%), 137 mm (79%), and 89 mm (38%), whereas ΔCMAP values are 35 mm (20%), 107 mm (62%), and 25 mm (11%). Region 7, near Hudson Bay, shows smaller GRACE/CLM values, with ΔGPCP and ΔCMAP equaling 44 mm (34%) and 32 mm (25%), respectively, whereas regions 8 and 9 in eastern Canada show more pronounced differences. In these regions, ΔGPCP has values of 74 mm (53%) and 156 mm (87%), whereas ΔCMAP has values of 70 mm (50%) and 147 mm (82%).

Average winter accumulation values for North America are summarized in Fig. 12. GPCP and CMAP accumulations agree more closely over the North American regions than over the Eurasian regions, with typical differences of 10–20 mm. In areas of low accumulation above 40°N latitude, GRACE/CLM values are larger than GPCP and CMAP. In areas of high accumulation, the opposite is true. In the western and central regions (2–6), GRACE is the largest component of the potential accumulation, whereas runoff is the largest component in the eastern regions (7–9). Evapotranspiration is generally a small term (<10 mm) for all regions.

c. Error estimates

Errors in GRACE TWS data are estimated using the method described in Wahr et al. (2004). The uncertainty in each spherical harmonic coefficient is derived from the root variance about the best-fitting annual cycle for each coefficient. The process of fitting an annual cycle absorbs a portion of the variance of the random errors in the time series, so a Monte Carlo simulation is performed to determine the mean variance reduction that results from the fitting procedure. The error estimate is then scaled to account for this effect. Because the dataset used here spans a period of 5 yr, and contains significant interannual variability, we modify the Wahr et al. (2004) methodology using the variation about a low-pass-filtered time series, rather than about the mean annual cycle. The resulting error estimate is conservative because all high-frequency variations are assumed to result from measurement errors rather than actual geophysical sources.

Errors in the modeled runoff are assessed using observed river discharge for six large Arctic river basins. Measurements of river discharge for three Eurasian rivers (Ob, Yenisey, and Lena) and three North American rivers (Nelson, Mackenzie, and Yukon) were obtained for the study period. Discharge records for the Ob, Yenisey, and Lena were downloaded from the R-ArcticNet database (available online at <www.r-arcticnet.sr.unh.edu/>). The Nelson and Mackenzie discharge records were provided by the Water Survey of Canada, and the U. S. Geological Survey provided the Yukon discharge record. The panels in the left column of Fig. 13 show the modeled (black line) and observed (gray line) monthly discharge, averaged for the 5 yr of the study period. Although significant biases can be seen during the spring and early summer (i.e., the melt period), the low-flow period during the winter months shows good agreement. The right column of Fig. 13 shows the integrated discharge during the cold season, averaged over the 5-yr data period. The magnitude of the differences, represented as a percentage of the observed signal, range from 10% to 40%. We use the average value (25%) of the six basins as an estimate of the uncertainty in the accumulated model runoff at the end of the cold season.

Model evapotranspiration errors are difficult to assess as a result of the absence of observational networks. A 25% error estimate is also assumed for the modeled ET estimates. Because the land surface model enforces conservation of mass and energy, it is likely that errors in the modeled evapotranspiration and runoff are compensating, and therefore the errors in the sum would be smaller. The extent of the possible compensation is unknown, however, so these errors are assumed to be uncorrelated in this study. The errors in the GRACE data and the model output are also assumed to be uncorrelated, and the total error variance of the potential accumulation time series is calculated by summing the individual error variances.

5. Discussion

Global precipitation analyses, not surprisingly, are typically assessed via comparisons to ground-based gauge networks (Bolvin et al. 2009; Huffman et al. 1997; Xie and Arkin 1997). Some disadvantages of such a validation approach exist. Although the validation gauge data and the gauges used in the analyses are mutually exclusive, they both suffer from the same kinds of systematic biases, such as wind-induced undercatch. For large-scale applications like the 2.5° × 2.5° products examined here, the uncertainty in the regional average estimated from the validation gauge network becomes greater for larger areas. Finally, validation studies are not performed globally but are limited to those locations where dense networks exist.

The comparison performed in this study is largely free of these problems, and therefore it provides a complementary means of assessing global precipitation estimates. GRACE estimates of total water storage, which are based on satellite measurements of the earth’s time-variable gravity field, are a type of observation that is almost entirely independent from the radiometric observations used in the precipitation analyses. Land surface model output, which is necessary to directly connect the TWS and precipitation datasets, can have large biases at certain times of the year. However, by confining the comparison to times when the temperature is below freezing, the contribution of the model output to the potential accumulation estimates is minimized. During the cold season, evapotranspiration fluxes are generally very small, and runoff, although larger in magnitude than ET, is also at its lowest value of the season. Moreover, accurate discharge records, representative of large parts of the northern high latitudes, can be used to constrain the modeled runoff parameterization. Thus, the high level of correlation between the precipitation analyses and the GRACE/CLM data, both spatially and temporally, is very encouraging, and it increases confidence in each dataset.

The results of this comparison also show that significant differences exist between the GRACE/CLM potential accumulation and the two precipitation analyses, as well as between the precipitation analyses themselves. The error estimates are relatively conservative, because the GRACE errors include geophysical signals at the monthly time scale and the model output error estimate ignores possible correlations between the errors in the runoff and evapotranspiration components. Despite this fact, the accumulated cold-season precipitation values typically did not fall within the GRACE/CLM error bounds. For most of the regions chosen in Eurasia, CMAP agreed more closely with GRACE/CLM than did GPCP, which was larger than GRACE/CLM for all regions.

In regions 3 and 5–7, CMAP is smaller than GRACE/CLM and GPCP is considerably larger, which may be due to the wind-induced undercatch correction used by GPCP. Gauge undercatch has been studied extensively (Goodison et al. 1998; Fuchs et al. 2001), but its application depends on meteorological data—for example, temperature and wind speed—that may not be reliable or available (Rudolf and Schneider 2005). Furthermore, the distance between stations in Eurasia can be 100 km or more, and the number of gauges has been declining over time (Rudolf and Mächel 2005). This study indicates that the undercatch correction used by GPCP may be an overcorrection, and that for some regions CMAP might benefit from a small undercatch correction.

Region 4 shows the poorest agreement between GRACE/CLM and the two precipitation products, possibly as a result of a low modeled runoff estimate. This region has a much lower total water storage component, and it may reveal a model bias in simulating runoff in a subarctic, maritime climate.

The comparison for Eurasia highlights the differences between the two precipitation analyses. Over North America, the comparison shows greater differences between GRACE/CLM and either precipitation product than between precipitation products. The general trend of GPCP values greater than CMAP values also exists, but the differences are much smaller in North America than in Eurasia, perhaps as a result of denser gauge networks and more accurate meteorological data.

Of the 18 regions chosen for this study, only in regions 2 and 3 for North America is total water storage alone larger than the two accumulated precipitation estimates. In all cases, the net effect of evapotranspiration and runoff is to increase the potential accumulation, thus errors in the model output are unlikely to explain the discrepancy in these two regions. The reason for the difference between GRACE and the precipitation estimates is currently unknown, but in situ gauge data, if available, may help to determine which dataset is in error.

The other regions in North America, where the GRACE/CLM accumulation is smaller than the GPCP and CMAP accumulated precipitation, can be divided into two regimes: where total water storage is the largest component of potential accumulation and where runoff is largest. Regions 4–6, along the west coast of the continent, sample the region of heavy orographic precipitation as a result of the Rocky Mountains. Here GRACE total water storage is the dominant component. In regions 7–9, which sample eastern Canada, modeled runoff is larger than TWS. Regions 4–9 are all in forested regions, which may influence the atmospheric retrievals used in the precipitation estimates, as well as increase the possibility of undersampling the large-scale precipitation field from in situ gauges. It is also possible, as it is for region 4 in Eurasia, that modeled runoff in near-coastal regions is biased low.

The water balance approach used in this study is a valuable tool for assessing the accuracy of global precipitation analyses. This technique complements gauge data by providing a large-scale, independent type of measurement. This study examines various regions of the northern high latitudes, highlighting areas of both good and poor agreement. The purpose of this work is not to provide definitive explanations for the differences but a few possible causes are mentioned. Exploring the ability of these hypotheses to explain the apparent discrepancies may be possible by obtaining additional ground-based precipitation gauge data and river discharge data in these regions. Another strength of the GRACE/CLM dataset is that it is global; if areas where high-quality in situ data exist do not coincide with the regions chosen here, then GRACE/CLM time series for other regions can be generated.

Acknowledgments

This research was supported by the Advanced Study Program of the National Center for Atmospheric Research. The author would like to thank Bob Adler and Phil Arkin for their helpful comments as well as three anonymous reviewers, whose comments helped improve the manuscript.

REFERENCES

Fig. 1.

Fig. 1.

Fig. 1.

(top) GRACE TWS monthly anomalies (mm) for a 4° lat–lon box centered at 56°N, 38°E. (middle) Monthly average precipitation rates (mm month−1) from GPCP (blue line) and CMAP (green line); monthly averaged ET + Q rates (mm month−1) simulated by CLM (yellow line). (bottom) Winter accumulations (mm) for GRACE only (orange), GRACE/CLM (red), GPCP (blue), and CMAP (green). Vertical lines indicate period of accumulation for the winter of 2006/07.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Fig. 2.

Fig. 2.

Fig. 2.

Cold-season accumulations from 2003/04. (top left) GPCP accumulated precipitation. (top middle) GRACE/CLM potential accumulation. (top right) CMAP accumulated precipitation. (bottom left) GRACE TWS change. (bottom middle) Same as top-middle panel. (bottom right) CLM accumulated ET + Q. Contours are millimeters of water.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Fig. 7.

Fig. 7.

Fig. 7.

Map showing nine regions in Eurasia used to construct time series of spatially averaged accumulation.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Fig. 8.

Fig. 8.

Fig. 8.

Accumulation time series for nine regions in Eurasia showing GRACE/CLM in red, GPCP in blue, and CMAP in green. Region number located at top right of each plot.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Fig. 9.

Fig. 9.

Fig. 9.

Mean cold-season accumulations (mm) for Eurasian regions showing GRACE/CLM in red, GRACE-only in orange, GPCP in blue, CMAP in green, CLM ET in light blue, and CLM Q in yellow.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Fig. 10.

Fig. 10.

Fig. 10.

Map showing nine regions in North America used to construct time series of spatially averaged accumulation.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Fig. 13.

Fig. 13.

Fig. 13.

(left) Five-yr average monthly discharge climatologies for six rivers. Observations are shown in gray; model results are shown in black. (right) Winter discharge accumulations.

Citation: Journal of Hydrometeorology 11, 2; 10.1175/2009JHM1194.1

Table 1.

Average cold-season accumulation for Eurasian regions.

Table 1.

Table 1.

Table 2.

Same as Table 1, but for North American regions.

Table 2.

Table 2.