1. Introduction
The Global Land Data Assimilation System (GLDAS) is a global offline (uncoupled to the atmosphere) terrestrial modeling system (Rodell et al. 2004), developed jointly by the National Aeronautics and Space Administration (NASA) Goddard Space Flight Center (GSFC) and the National Oceanic and Atmospheric Administration (NOAA)/National Centers for Environmental Prediction (NCEP). Two versions of GLDAS datasets are available online: GLDAS, version 1 (GLDAS-1), which drives four land surface models (LSMs)—Noah, the Community Land Model (CLM), the Variable Infiltration Capacity model (VIC), and Mosaic—and GLDAS, version 2 (GLDAS-2), which drives only the Noah model so far. GLDAS provides a unique opportunity for the geoscience community to assess the global and regional environment change at up to 0.25° spatial resolution and 3-hourly temporal resolution. The fields of land surface water states and fluxes provided by GLDAS include rainfall rate, snowfall rate, evapotranspiration (ET), soil moisture in different layers, surface runoff, and subsurface runoff.
GLDAS data products have been used for assessing changes of terrestrial water storage (TWS; e.g., Yang et al. 2013; Huang et al. 2013; Proulx et al. 2013; Yang and Chen 2015), identifying global dryland areas (Ghazanfari et al. 2013), drought monitoring (Hao et al. 2014), and long-term soil moisture changes (Zawadzki and Kȩdzior 2014). Syed et al. (2008) showed that GRACE-based storage changes are in agreement with those obtained from GLDAS simulations. Rodell et al. (2007) used GLDAS outputs (soil moisture and snow water equivalent) as auxiliary information to isolate groundwater storage anomalies from GRACE TWS. Meanwhile, GLDAS data have been used by some authors (e.g., Gao et al. 2014) to assess long-term land surface change without full awareness of the issue of data quality, which may lead to false detection of changes resulting from data problems.
Many studies have been conducted for validating the GLDAS-1 data products. Zaitchik et al. (2010) found that the four LSMs included in GLDAS-1 yield very different estimates of river discharge and that there are distinct geographic patterns in the accuracy of each model as evaluated against gauged discharge. Wang et al. (2011) showed that the 0.25° × 0.25° GLDAS-1/Noah daily and monthly precipitation data are of high quality for a mesoscale basin in northeastern China. Wang and Zeng (2012) evaluated six reanalysis products (i.e., MERRA, NCEP–NCAR, CFSR, ERA-40, ERA-Interim, and GLDAS-1) using in situ measurements at 63 weather stations over the Tibetan Plateau (TP) from the China Meteorological Administration (CMA) for 1992–2001, concluding that GLDAS-1 has the best overall performance in both daily and monthly precipitation. Zhou et al. (2013) found that there are large errors in GLDAS-1 precipitation data globally in 1996, which lead to large errors in the simulated runoff. Chen et al. (2013) showed that the four GLDAS-1 models tend to systematically underestimate the surface soil moisture (0–5 cm) but simulate the soil moisture for a 20–40-cm layer well for the central Tibetan Plateau. Seyyedi et al. (2015) showed that the GLDAS-1 precipitation product exhibits significantly worse error statistics compared to the TRMM Multisatellite Precipitation Analysis 3B42, version 7, product in the northeastern United States. Qi et al. (2015) showed that GLDAS-1/Noah model air temperatures and humidity have high accuracy, while downward solar radiation and wind speed data are overestimated. Ji et al. (2015) demonstrated that GLDAS-1 air temperature estimates are generally accurate globally, but caution should be taken when the data are used in mountainous areas or places with sparse weather stations.
Most studies focus on one or several inputs or simulated outputs of GLDAS-1 so far and lack the analysis about the link between forcing data and simulated data. The purpose of this study is to conduct a systematic evaluation of the monthly scale precipitation and temperature forcing inputs of both GLDAS-1 and GLDAS-2 over China. In addition, the monthly runoff, evapotranspiration, and soil moisture data from GLDAS-1/Noah and GLDAS-2/Noah simulations are also evaluated by comparing against available reference data.
2. Data used and measures for evaluating GLDAS data
a. GLDAS data
GLDAS products include forcing data (e.g., precipitation, near-surface air temperature, downward shortwave and longwave radiation, specific humidity, wind speed, and surface pressure), land surface states (e.g., soil moisture, surface runoff, and subsurface runoff), and flux data (e.g., evaporation and sensible heat flux). Two versions of GLDAS datasets are available online, that is, GLDAS-1 and GLDAS-2 (http://disc.sci.gsfc.nasa.gov/hydrology/data-holdings), who differ mainly in their forcing data. Forcing sources in GLDAS-1 switched several times throughout the record from 1979 to present (Rui 2016), including 1) 1979–93, bias-corrected ECMWF Re-Analysis data; 2) 1994–99, bias-corrected NCEP–NCAR reanalyses data; 3) 2000, NOAA/GDAS atmospheric analysis fields; and 4) from 2001 to present, a combination of NOAA/GDAS atmospheric analysis fields, spatially and temporally disaggregated NOAA Climate Prediction Center Merged Analysis of Precipitation (CMAP) fields, and observation-based downward shortwave and longwave radiation fields derived using the method of the Air Force Weather Agency’s agricultural meteorology modeling system (AGRMET). GLDAS-2 uses the Princeton meteorological forcing dataset (Sheffield et al. 2006) as the only source of forcing data.
GLDAS-1 datasets include 1.0° data products from four land surface models (Noah, CLM, Mosaic, and VIC), covering the period from 1979 to the present, and 0.25° data products from Noah, version 2.7 (Noah 2.7), covering the period from 2000 to the present. GLDAS-2 datasets consist of 1° × 1° and 0.25° × 0.25° data from Noah, version 3.3 (Noah 3.3), covering the period from 1948 to 2010. The temporal resolution for the GLDAS products is 3-hourly. Monthly products are generated through temporal averaging or summation of the 3-hourly products. The GLDAS-1/Noah and GLDAS-2/Noah data of their overlapping period, that is, 1979–2010, are used in this study.
b. Ground-based hydrometeorological data
Gridded monthly data provided by the CMA are used to validate precipitation and air temperature data of GLDAS. The gridded precipitation data are developed based on over 2400 ground-based precipitation gauging sites (see Fig. 1) taking the effect of elevation into account (National Meteorological Information Center of CMA 2012). As no data are available for Taiwan and islands in the South China Sea, the study focuses on mainland China.
Number of meteorological stations in each 0.5° × 0.5° grid cell over China in 2010 and locations of river basins investigated in the present study.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
The method of partial thin-plate smoothing splines is used for interpolating the irregularly scattered gauging data to gridded data with 0.5° × 0.5° spatial resolution. For comparison with GLDAS data, the gridded precipitation and air temperature data are resampled to 1° × 1° by simply taking the average of the four 0.5° × 0.5° grid cells in each 1° × 1° cell.
China has a typical monsoon climate. Most rain falls in the summer, and annual rainfall amounts decrease from the wet southeast to the dry northwest. As shown in Fig. 1, the gauging stations used for generating the gridded dataset are relatively dense and uniformly distributed in the eastern part of China, whereas far fewer stations are located in the western part of China and Inner Mongolia. There is no station located in the western part of the northern Tibetan Plateau (NTP) and only four stations in central Tibetan Plateau (CTP). The coverage of gauging stations in 0.5° × 0.5° grids is 77% in the area to the south of 40°N and east of 100°E, but only about 1% in the western part of TP (between 30° and 36°N, and west of 90°E). As a result, the gridded precipitation and air temperature data in those areas are not of high quality, which consequently makes the evaluation results of GLDAS for those areas not as reliable as for other areas in China.
Monthly streamflow discharge data of three large exorheic drainage basins, that is, the Yangtze River at Datong (drainage area: 1 705 383 km2), the Yellow River at Huayuankou (730 036 km2), and the Songhua River at Jiamusi (528 277 km2) during the period from 2002 to 2010, are used for evaluating the capability of GLDAS data for describing water flux and water balance features. In addition, three endorheic drainage basins are also investigated, that is, the Tarim basin (1 240 777 km2), the NTP (593 838 km2), and the CTP (388 508 km2). The locations of those basins are illustrated in Fig. 1.
c. GRACE water storage data
For evaluating water storage data simulated by GLDAS, GRACE water storage data are used. GRACE observes temporal variations of Earth’s gravitational potential. GRACE allows for the recovery of the global gravity field down to spatial scales of a few thousand square kilometers, with a temporal resolution of monthly intervals. After atmospheric and oceanic effects are accounted for, the remaining signal of GRACE data on monthly to interannual time scales is mostly related to variations of TWS (Landerer and Swenson 2012). Many studies have clearly demonstrated its capacity to monitor changes in TWS over continental areas (Ramillien et al. 2008). In this study, GRACE TWS data are retrieved from the Jet Propulsion Laboratory’s (JPL) Tellus website (http://grace.jpl.nasa.gov/data/; Landerer and Swenson 2012).
GRACE Tellus TWS data are derived based on the Release 05 (RL05) spherical harmonics produced by JPL, the Center for Space Research at the University of Texas at Austin, and Deutsches GeoForschungsZentrum (GFZ) with additional postprocessing alone or in combination with other data. The spatial sampling of all grids is 1° in both latitude and longitude (~111 km at the equator). Because of the sampling and postprocessing of GRACE observations, surface mass variations at small spatial scales tend to be attenuated. Therefore, the GRACE Tellus land data are multiplied by a scaling grid to restore much of the energy removed in the postprocessing. Each monthly GRACE Tellus grid represents the surface mass deviation for that month relative to the baseline average from January 2004 to December 2009.
d. Data quality measures
MBE computes all deviations from the observed data, reflecting systematic model over- or underestimation, while MAE computes deviations from the original data regardless of sign and gives a balanced perspective of the goodness of fit for values at all levels. The correlation coefficient measures how the variability of observed value is captured by the model. CE is widely used for measuring streamflow modeling. CE is an improvement over r in that it is more sensitive to differences in modeled and observed means and variances, but it is overly sensitive to extreme values. Both r and CE can provide useful comparisons between studies since they are independent of the scale of data used; however, they cannot reflect systematic model over- or underestimation.
In addition to those measures, the temporal change rate and the trend of some variables are evaluated using the method of multiple regression with periodic functions and seasonal Kendall trend test method, respectively (Helsel and Hirsch 2002).
The seasonal Kendall trend test is a nonparametric trend test method, which accounts for seasonality by computing the Mann–Kendall test statistics for each month separately and then combining the results. We can identify the trend based on both the value of Kendall’s τ, which is a measure of correlation between GLDAS-simulated variables and time (month in this study), and the p value that indicates the statistical significance. A p value less than 0.05 indicates a significant increase (if τ > 0) or decrease (if τ < 0), whereas a p value larger than 0.05 but less than 0.1 may be considered as a weak trend. Otherwise, no trend is detected.
e. Method of calculating spatial average
3. Evaluation of GLDAS-1 and GLDAS-2 precipitation data
a. Temporal variation of precipitation data
GLDAS precipitation data are composed of rainfall (kg m−2 s−1) and snowfall (kg m−2 s−1). We convert these units to millimeters per month according to the days of each month and sum up rainfall and snowfall as the total precipitation of the month.
The monthly average precipitation time series over China is calculated and shown in Fig. 2. It shows that both GLDAS-1 and GLDAS-2 precipitation data match the observed precipitation data well by visual inspection. However, GLDAS-1 precipitation is much lower in 1996 and much higher in 2000 than the observation. The error in 1996 has been reported in a previous study (Zhou et al. 2013), and because of this precipitation error, GLDAS-1-simulated land surface variables, including runoff, actual ET, and soil moisture have serious erroneous values too. In addition, GLDAS-1 snowfall data have an obvious step change in 2000. The snowfall approximately doubled after 2000, which is clearly an artifact. In contrast, no obvious erroneous values can be identified by visual inspection in GLDAS-2 precipitation data. Because of the large errors of GLDAS-1 data in 1996, GLDAS-1 precipitation data in 1996 are excluded in the following evaluation.
Monthly average (a) precipitation and (b) snowfall over China during the period from 1979 to 2010.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
The seasonal Kendall trend test method is applied to GLDAS-1, GLDAS-2, and observed precipitation data in the period from 1979 to 2010 on a grid basis. The test results are illustrated in Fig. 3. From Fig. 3, observation data show that only some small areas in southwestern and northeastern China exhibit weak drying trends (at the 0.1 significance level), while vast areas in western China exhibit weak wetting trends (at the 0.1 significance level). The trend test result for GLDAS-2 is quite close to that of the observations, but indicates much smaller areas of drying than the ground-observed data, and the distribution of wetting regions in western China is somewhat different from that indicated by the observed data. At the same time, GLDAS-1 data show a quite different picture of the trend in precipitation, indicating a significant drying trend in many parts of China, especially in northeastern China.
Seasonal Kendall trend test results for the GLDAS precipitation compared against ground-observed precipitation during 1979–2010.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
The precipitation change rate over the period from 1979 to 2010 is calculated on a grid basis using the method of multiple regression with periodic functions. The results are presented in Fig. 4. The spatial patterns and the amplitudes of precipitation change rate by GLDAS-2 and observed precipitation data are also similar to each other, whereas GLDAS-1 data give a quite different picture again. According to GLDAS-1 data, the most significant changes occur in the central part of NTP, where the precipitation increases at a rate up to 3.3 mm yr−1. In contrast, the maximum increase of precipitation estimated based on either GLDAS-2 data or observed data is less than 1.2 mm yr−1 over China.
Annual precipitation change rates for GLDAS-1 and GLDAS-2 compared against the change rates from ground-observed data.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
b. Correlation and error analysis for precipitation data
Correlation and error analysis for precipitation data are conducted from both the temporal and spatial perspectives. The correlation coefficient and error measures (MBE and MAE) between all grid cells over China of GLDAS and those of the observations are calculated on a monthly basis to obtain the monthly r, MBE, and MAE time series, which reflect the temporal characteristics of GLDAS data quality. The r and error measures between GLDAS precipitation series and observed precipitation series are also calculated on a grid basis, which reflects the spatial characteristics of GLDAS data quality.
The time series of the monthly r over China during 1979–2010 are given in Fig. 5a. GLDAS-1 data correlate with observations better before 2001 (average r = 0.855) than since 2001 (average r = 0.761), whereas GLDAS-2 data correlate with observations with a generally stable level of correlation (average r = 0.810 before 2001, and average r = 0.795 since 2001); the correlation of GLDAS-1 data is generally better than that of GLDAS-2 data, which means that GLDAS-2 captured less variability in precipitation than GLDAS-1.
The temporal variation of (a) r, (b) MBE, and (c) MAE of GLDAS precipitation compared against observed precipitation over China.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
The time series of monthly MBE and MAE over China during 1979 to 2010 are shown in Figs. 5b and 5c. Temporal variations of MBEs and MAEs show that overall biases are small, and GLDAS-1 precipitation data are slightly positively biased (average MBE = 0.520 mm month−1) while GLDAS-2 data are slightly negatively biased before 2001 (MBE = −0.352 mm month−1). From 2001 to 2010, the situation reversed, that is, average MBE = −0.948 mm month−1 for GLDAS-1 and −0.053 mm month−1 for GLDAS-2. At the same time, the MAE of GLDAS-1 data increases more than 10% (from 15.987 mm month−1 before 2001 to 17.981 mm month−1 since 2001), whereas the MAE of GLDAS-2 is generally stable (from 18.903 mm month−1 before 2001 to 19.151 mm month−1 since 2001).
The spatial distribution of r between GLDAS data and observed data is displayed in Fig. 6 (top). The maps show a common spatial pattern, that is, the correlation decreases from the southeast (>0.9 in most grid cells) to the northwest (<0.5 in many grid cells). GLDAS precipitation in the western part of NTP has the lowest correlation with observed precipitation all over China, even down to less than 0 in some cells. However, it should be noted that as the precipitation in the western part of NTP is poorly observed on the ground, the lowest r does not necessarily mean that GLDAS data has the lowest quality in that area. GLDAS-1 data generally have higher correlations than GLDAS-2 in most parts of China, especially in the southeast.
Grid-based (top) r, (middle) MBE, and (bottom) MAE for GLDAS-1 and GLDAS-2 precipitation compared against observed precipitation during 1979–2010.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
The grid-based MBE is displayed in Fig. 6 (middle). It shows that MBEs of GLDAS-1 and GLDAS-2 are within the range of ±10 mm month−1 for most parts of China, but the biases of both datasets in southwestern TP are above +30 mm month−1, especially for GLDAS-2. At the same time, GLDAS-2 has negative biases of down to −30 mm month−1 in the eastern TP.
The grid-based MAE is displayed in Fig. 6 (bottom). MAEs of GLDAS-1 and GLDAS-2 basically follow the same spatial pattern as precipitation, that is, declining from the southeast to northwest across China, and the MAE of GLDAS-1 data is generally smaller than that of GLDAS-2 data over China.
The average r, MBE, and MAE in each month of the year are calculated for both GLDAS-1 and GLDAS-2 precipitation data, illustrated in Fig. 7. Correlations of GLDAS-1 and GLDAS-2 have a similar seasonal pattern, that is, higher in spring (March–June) and autumn (September–November) than in winter (December and January) and summer (July and August). GLDAS-1 has higher r than GLDAS-2 in most months (except for January–March). While the monthly average MBE shown in Fig. 6 indicates that GLDAS-1 has smaller absolute values of MBE for many areas across China than GLDAS-2, Fig. 7 shows that GLDAS-1 has large negative biases in summer (from −3 to −4 mm month−1) and large positive biases in winter (1–2 mm month−1), whereas MBE of GLDAS-2 has no clear seasonal pattern and mostly oscillate in the range of ±1 mm month−1. That means that GLDAS-2 is much better than GLDAS-1 from the perspective of MBE. But the MAE of GLDAS-1 is generally smaller than that of GLDAS-2, especially in the summer half year (from May to October).
Average (left) r, (center) MBE, and (right) MAE for GLDAS-1 and GLDAS-2 precipitation over China in each month of the year.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
4. Evaluation of GLDAS-1 and GLDAS-2 air temperature data
a. Temporal change of GLDAS monthly temperature data
The seasonal Kendall trend test is applied on a grid basis to GLDAS-1, GLDAS-2, and observed temperature data in the period from 1979 to 2010. Test results are presented in Fig. 8. The temperature change rate over the same period is also calculated on a grid basis using the method of multiple regression with periodic functions, and the results are presented in Fig. 9. In Figs. 8 and 9, observation data show that almost the whole of China exhibits a significant warming trend (at the 0.05 significance level) with a rate of 0°–0.09°C yr−1. GLDAS-1 data show that only the northern half of China exhibits a significant warming trend (at the 0.05 significance level), while southern China (including southern TP) exhibits a significant cooling trend; the amplitudes of temperature change rate vary greatly over China, mostly ranging from −0.4° to 0.2°C yr−1. The spatial patterns and the amplitudes of temperature change rate by GLDAS-2 are close to those indicated by the observed precipitation data, varying from 0° to 0.1°C yr−1.
Seasonal Kendall trend test results for GLDAS temperature compared against observed temperature during 1979–2010.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
Temperature change rates based on different temperature data during 1979–2010.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
b. Correlation and error analysis for temperature data
Similar to the correlation and error analysis for precipitation data, the analysis for temperature data is conducted from both the temporal perspective and the spatial perspective. The monthly r, MBE, and MAE between all grid cells over China of GLDAS and those of the observation were calculated, reflecting the temporal characteristics of GLDAS temperature data quality. The grid-based r and error measures between GLDAS temperature series and observed series are calculated for each grid cell, which reflect the spatial characteristics of GLDAS data quality.
The temporal variations of the monthly r, MBE, and MAE of both GLDAS-1 and GLDAS-2 during 1979–2010 are shown in Fig. 10. The temporal variations reveal obvious temporal discontinuity in the data, which can be roughly divided into four segments: 1) 1979–90, when r is generally above 0.98, MBE is around 1°C, and MAE is mostly between 1° and 1.5°C; 2) 1991–99, when r drops to around 0.97 and both MBE and MAE increase by about 0.5°C; 3) 2000–05, a period of high variability, when MBE becomes negative, down to ~−2°C in summer in 2000–02, r drops to around 0.92, and MAE reaches ~3°C in 2003–05; and 4) 2006–10, when MAE goes back to the level before 1990 and MBE is close to zero, but r is still lower than the level before 1990. In general, throughout the period from 1979 to 2010, r of GLDAS-1 data has a tendency to go down, MBE shifts from positive to negative, and MAE goes up. In contrast, data quality of GLDAS-2 is generally stable from 1979 to 2010, with an overall r of 0.978, MBE of 0.508°C, and MAE of 1.266°C. However, GLDAS-2 data quality deteriorates seriously in 2008–09, when r drops down to ~0.88, MBE goes down to ~−3°C, and MAE reaches up to ~4°C in summer.
Temporal variation of (a) r, (b) MBE, and (c) MAE of GLDAS temperature compared against ground-observed temperature during 1979–2010.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
By inspecting the spatial distributions of r, MBE, and MAE of GLDAS temperature compared against observed temperature, we find that GLDAS-2 temperature highly correlates with the observed temperature, with r higher than 0.99 over most parts of China and higher than 0.9 over TP, whereas GLDAS-1 temperature also has r higher than 0.99 over most parts of China but less than 0.9 in some areas of central and southern TP. We also find that MBEs of GLDAS-1 and GLDAS-2 temperature are within the range of ±1°C for most parts of the eastern half of China but vary a lot in the western half of China, especially in the TP, ranging from up to ~+7°C down to ~−4°C, and MAEs of GLDAS-1 and GLDAS-2 temperature are less than 1°C in most parts of eastern China, but in western China MAEs may go up to 3°–7°C for some grid cells (mostly in the TP), especially for GLDAS-1.
Seasonal variations of data quality for both GLDAS-1 and GLDAS-2 temperature data are illustrated in Fig. 11, from which it is shown that GLDAS-2 temperature data have an overall higher r and smaller MAE than GLDAS-1 temperature data, while having similar levels of MBE, indicating the superiority of GLDAS-2 over GLDAS-1. Both GLDAS-1 and GLDAS-2 temperature data have a slightly lower r in summer months (June–August) than in other months, indicating a lower capability of describing the spatial variability of temperature in summer. On the other hand, MAEs of both GLDAS-1 and GLDAS-2 in summer months are lower than in other months, indicating better quality in summer.
Average (left) r, (center) MBE, and (right) MAE for GLDAS-1 and GLDAS-2 temperature over China in each month of the year.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
5. Evaluation of GLDAS-1 and GLDAS-2 water storage data
GLDAS-1 data have significant discontinuity over the period from 1979 to the present, GLDAS-2 data are available only before 2010, and GRACE water storage data are available only after 2002: therefore, the following data evaluation will focus on the period 2002–10 for both GLDAS-1 and GLDAS-2.
a. Spatial distribution and temporal variation of TWS
TWS is the total amount of water stored on the surface and subsurface of land. As GLDAS does not simulate deep groundwater, it is common (e.g., Rodell et al. 2007; Syed et al. 2008; Yang and Chen 2015) to derive TWS data from GLDAS-1 and GLDAS-2 data by summing up snow accumulation, total canopy water storage, and the soil water content of four layers (from the land surface down to 2 m in depth).
The GLDAS-1 and GLDAS-2 TWS data over the period 2002–10 are averaged for each grid. By comparing the spatial distribution of GLDAS-1 and GLDAS-2 TWS data, it is found that 1) the spatial patterns of GLDAS-1 TWS and GLDAS-2 TWS resemble each other closely, with TWS decreasing from southeastern to northwestern China; and 2) GLDAS-2 TWS is generally larger than GLDAS-1 TWS, but GLDAS-1 TWS has a much larger variability (ranging from 138 to 3194 mm) than GLDAS-2 TWS (ranging from 185 to 766 mm). The grid cells with GLDAS-1 TWS above 1000 mm (with a maximum over 3000 mm) are located at the border between the Tarim basin and NTP, which may be related to ice and snow in that region.
The trend of long-term TWS change for each grid is analyzed using the multiregression method. The calculated change rates of GLDAS-1, GLDAS-2, and GRACE TWS are illustrated in Fig. 12. The average annual TWS change rate over China is −0.051, −0.091, and −0.095 cm yr−1 for GLDAS-1, GLDAS-2, and GRACE, respectively, which is very similar to each other. However, Fig. 12 shows that GLDAS-1 has much larger spatial variability in the change rate than GLDAS-2, and some grids at the border between Tarim basin and NTP have an increase rate of over 10 cm yr−1, which is quite questionable.
The change rates of TWS by (left) GLDAS-1, (center) GLDAS-2, and (right) GRACE during 2002–10. (Note that the white cells for GLDAS-1 data indicate TWS change rates larger than 4 cm yr−1.)
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
b. Monthly water storage change and its correlation and bias with GRACE observations
Because GRACE-derived TWS data and GLDAS-simulated TWS data are not identical in their constituent components, these TWS data cannot be compared directly. Instead, we compare the volume of total water storage change (TWSC), which is estimated as 
The distribution of correlation coefficients between GLDAS TWSC data and GRACE TWSC data are illustrated in Fig. 13, which demonstrates that GLDAS TWSC and GRACE TWSC are better correlated in wet eastern China (mostly exorheic basins) than in dry western China (mostly endorheic basins), and GLDAS-1 is better correlated with GRACE data than GLDAS-2 in eastern China.
Correlation coefficient between (a) GLDAS-1 TWSC and GRACE TWSC and (b) GLDAS-2 TWSC and GRACE TWSC.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
By calculating r, MBE, and MAE for GLDAS TWSC data and GRACE TWSC data on a grid basis and then taking their spatial averages over China, it is found that 1) the overall correlation between GLDAS data and GRACE data is poor, 0.445 between GLDAS-1 and GRACE and 0.424 between GLDAS-2 and GRACE; 2) MBEs are very small, 0.31 mm month−1 between GLDAS-1 and GRACE and 0.27 mm month−1 between GLDAS-2 and GRACE, indicating that the overall biases between GLDAS TWSC data and GRACE-derived TWSC data are very small; and 3) MAEs are large, 20.88 mm month−1 between GLDAS-1 and GRACE and 21.32 mm month−1 between GLDAS-2 and GRACE, that is, the average MAE is about one-third average monthly precipitation (~55 mm month−1) over China.
The reason for small MBEs is that the average TWSC over China of both GLDAS data and GRACE data approaches zero on average in the long term. Meanwhile, large MAEs and small r values indicate that there are considerable differences in the spatial and temporal variability of TWSC derived by GLDAS data and GRACE data.
The average TWSC in each month derived by GLDAS data and GRACE data over China is shown in Fig. 14. The monthly average precipitation change (
Monthly average TWSC by GLDAS-1, GLDAS-2, and GRACE compared against monthly precipitation change 
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
6. Evaluation of GLDAS-1 and GLDAS-2 runoff data
a. Spatial distribution of runoff
First of all, the monthly averages of three simulated runoff components (i.e., surface runoff Qs, subsurface runoff Qsb, and snowmelt runoff Qsm) over China during 2002–10 are calculated for GLDAS-1 and GLDAS-2, and their statistics are presented in Table 1. In comparison with GLDAS-2, GLDAS-1 has a similar amount of surface runoff, much less subsurface runoff, and much more snowmelt runoff.
Statistics of the monthly averages of precipitation and runoff components (mm month−1) by GLDAS-1 and GLDAS-2 over China (2002–10). Note that Q is the sum of Qs, Qsb, and Qsm.
The spatial distribution of those GLDAS runoff components shows that the area with the most Qs is the CTP by GLDAS-1, whereas it is southeastern China that has the most Qs by GLDAS-2. The simulated Qsb is in general much larger than the simulated Qs in eastern China by both GLDAS-1 and GLDAS-2, but less than Qs in western and northeastern China, and GLDAS-1 has a much larger area with Qsb < Qs than GLDAS-2. In CTP runoff is dominated by Qs according to GLDAS-1, and some cells have up to 700 mm yr−1 differences between Qs and Qsb (i.e., Qs − Qsb). On the contrary, GLDAS-2 shows that Qs is less than Qsb in most parts of TP, and in central and southern TP, Qsb dominates the runoff. By comparing the ratio of Qsb to Qs (i.e., Qsb/Qs) during the period 2002–10, we find that the Qsb/Qs for Tarim, NTP, and CTP is 0.092, 0.005, and 0.012 by GLDAS-1 (Noah 2.7) and 1.324, 0.853, and 3.703 by GLDAS-2 (Noah 3.3), respectively. That means there is a significant change in runoff generation mechanism for permafrost areas in the Noah model, from very impervious in GLDAS-1 (Noah 2.7) to very permeable in GLDAS-2 (Noah 3.3). The simulated Qsm by GLDAS-1 is much larger on average than those by GLDAS-2, because GLDAS-1 has much more snowfall than GLDAS-2 after the year 2000 (Fig. 2) and has a much larger spatial variability.
The spatial characteristics of GLDAS runoff simulations are further examined in terms of the runoff coefficient, which is defined here as the ratio between total runoff (i.e., Q; including Qs, Qsb, and Qsm) and total precipitation. Before being investigated in China, the simulated runoff coefficient by GLDAS is first examined globally. While the runoff coefficient by definition should be less than 1, it is found that there are considerable grid cells with runoff coefficients larger than 1 for both GLDAS-1 and GLDAS-2 (about 7.5% and 3.8% of total cells globally, respectively). A close inspection of the location of those cells shows that almost all the cells are located in arctic regions or high mountains, which seems that the extra runoff probably comes from the depletion of permanent snow and ice due to the effect of global warming. However, the depletion of permanent snow and ice is not included in the Noah model, which means that the runoff coefficient higher than 1 is questionable.
The distributions of runoff coefficients over China calculated based on GLDAS-1 and GLDAS-2 data are shown in Fig. 15. The runoff coefficient given by GLDAS-1 (0.2874) over China is much lower than that by GLDAS-2 (0.341; see Table 1), and from Fig. 15, the difference is especially obvious in southeastern China, the southern TP, and northeastern China. Both GLDAS-1 and GLDAS-2 indicate that southeastern China and the southern TP have much higher runoff coefficients than other regions. For some cells at the southern edge of the TP, runoff coefficients are higher than 0.8.
Runoff coefficients calculated based on (left) GLDAS-1 and (right) GLDAS-2 data.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
b. Comparison with gauge observations
GLDAS runoff simulations are evaluated in six catchments by comparing simulated runoff with observations at streamflow gauging sites. At short time scales (e.g., daily), it is necessary to apply a runoff routing scheme to transport runoff simulated at each grid cell to the drainage basin outlet that corresponds to the in situ gauge, before comparing the simulated runoff and measured discharge. In this study, three of the six basins are endorheic, whose outflows are zero. For the three exorheic drainage basins, as the time of concentration for each of the three river basins is less than 1 month, the monthly runoff at the gauging site is calculated by summing up the monthly runoff (including surface runoff, subsurface runoff, and snowmelt runoff) of all cells in the drainage area.
The observed and GLDAS-simulated precipitation and runoff are summarized in Table 2. For the three exorheic basins, GLDAS precipitation data match the observed precipitation well (only 3.3%, 4.5%, and 6.0% underestimation in Yangtze, Yellow, and Songhua for GLDAS-1 and 4.7%, 7.7%, and 1.2% underestimation in Yangtze, Yellow, and Songhua for GLDAS-2, respectively). But as shown in Fig. 16 (top), both GLDAS-1 and GLDAS-2 remarkably underestimate the monthly runoffs (both high flows and low flows) of the Yangtze River and the Songhua River, and the underestimation of GLDAS-2 (~27% underestimation for the Yangtze and ~70% for the Songhua) is less than GLDAS-1 (~43% underestimation for the Yangtze and ~74% for the Songhua). Consequently, runoff coefficients of both GLDAS-1 and GLDAS-2 (especially GLDAS-1) are much lower than those based on observed data for the Yangtze River and Songhua River. Meanwhile, there is no significant bias in simulated runoff by both GLDAS-1 (0.7% overestimation) and GLDAS-2 (10.7% overestimation) for the Yellow River.
Mean values for GLDAS-1 and GLDAS-2 simulations during 2002–10 (mm month−1).
GLDAS-simulated average runoff in each month of the year compared against observed runoff for three exorheic basins and observed precipitation for three endorheic basins.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
The performance of GLDAS runoff simulations is further assessed with r, MBE, and CE, listed in Table 3. The simulations of both GLDAS-1 and GLDAS-2 are poor in terms of CE. Although GLDAS-1 is better correlated with observations for the Yangtze River, its CE is much lower and the bias (MBE) is much larger than GLDAS-2. For the Songhua River, GLDAS-2 is better than GLDAS-1 in terms of all three measures. But for the Yellow River, GLDAS-1 performs better than GLDAS-2.
Comparison of GLDAS-simulated monthly runoff with observed runoff (2002–10).
For the three endorheic basins there is no outflow. However, there are runoffs flowing into endorheic lakes, so runoff can still be simulated for most grid cells. There is no direct measurement available about the volume of inflows to endorheic lakes to assess the runoff simulations in the three endorheic basins, but we can evaluate runoff simulations by analyzing the reasonability of the runoff coefficients. First, we compared the observed precipitation and GLDAS precipitation. It is found that there are significant biases in GLDAS precipitation data for the three endorheic basins, as shown in Table 2. GLDAS-1 precipitation is overestimated by 20%–40% in the three endorheic basins. GLDAS-2 precipitation is slightly underestimated in the Tarim basin and NTP, but greatly overestimated in CTP (~73%). Such biases in precipitation will surely lead to the biases in runoff simulations. Second, we plot GLDAS-simulated average monthly runoff against observed average monthly precipitation for the endorheic basins in Fig. 16 (bottom) and calculate runoff coefficients based on GLDAS precipitation data and runoff data, also presented in Table 2. As shown in Table 2, runoff coefficients are very high for all the three endorheic basins by GLDAS-1, that is, 0.375 for Tarim, 0.442 for NTP, and 0.412 for CTP. It is hard to judge the soundness of the high runoff coefficient in NTP and CTP because of the existence of permafrost and glaciers. For instance, the observed runoff coefficient for a 50.5 km2 headwater catchment (with one-third of the area covered by glaciers) of the Yangtze River in eastern TP was 0.63 during a 5-month period from May to September (Zhang et al. 1997). However, 0.375 for Tarim is very questionable because Tarim basin is mostly desert. Meanwhile, runoff coefficients by GLDAS-2 are much smaller than by GLDAS-1 for the Tarim basin and NTP (~0.16 and ~0.18, respectively), but high for CTP (0.448).
7. Evaluation of GLDAS-1 and GLDAS-2 evapotranspiration data
a. Comparing simulated ET by GLDAS-1 and GLDAS-2
There is no direct observation of actual ET at the catchment scale for validating the GLDAS-simulated ET data. Here, we compare grid-based average annual GLDAS-1 ET estimates and GLDAS-2 ET estimates first. From Fig. 17, it is shown that GLDAS-2 has a similar amount of ET as GLDAS-1 in the northwestern half of China, but much lower ET estimates in southeastern China, which is consistent with the higher runoff coefficient in that region by GLDAS-2 than by GLDAS-1 (shown in Fig. 15). Statistics show that ET simulations by GLDAS-1 have larger spatial variability (standard deviation/mean = 0.62) than by GLDAS-2 (0.58) over China.
Simulated ET by (left) GLDAS-1 and (right) GLDAS-2.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
b. Comparing GLDAS-simulated regional ET and water storage change with observations
According to the principle of water balance, for a specific time period (e.g., a month), we have 
The goodness of fit of the simulated ET + ΔW with P − Q observation series are evaluated using r, MBE, and CE, listed in Table 4. It is shown that, on average, the simulated ET + ΔW matches P − Q observation very well for the exorheic basins by both GLDAS-1 and GLDAS-2, but less well for the endorheic basins. GLDAS-1 simulates ET + ΔW as well as GLDAS-2 in the exorheic basins, but much worse than GLDAS-2 in the endorheic basins.
Comparison of GLDAS-simulated ET + ΔW and observed P − Q (2002–10).
The seasonal variation of GLDAS-simulated average ET + ΔW is compared against observed monthly average P − Q for the three exorheic basins in Fig. 18 (top) and against observed monthly average P for the three endorheic basins, where the overall Q is zero, in Fig. 18 (bottom). It is shown that the simulated ET + ΔW by both GLDAS-1 and GLDAS-2 matches the seasonal variation of P − Q very well by visual inspection for the exorheic basins, despite of the overall positive bias by GLDAS-1 for the Yangtze and some small biases in some months for others. Meanwhile, ET + ΔW simulated by GLDAS-1 does not match the observed precipitation at all for the endorheic basins, whereas ET + ΔW simulated by GLDAS-2 approximately follows the seasonal pattern of observed precipitation, but with much lower goodness of fit than for exorheic basins.
GLDAS-simulated ET + ΔW in each month compared against observed P − Q for three exorheic basins and observed P for three endorheic basins.
Citation: Journal of Hydrometeorology 17, 11; 10.1175/JHM-D-15-0191.1
8. Discussion
a. Causes of discontinuity for GLDAS-1 forcing data and uncertainty in bias correction for GLDAS-2 forcing data
As mentioned in section 2a, in GLDAS-1, forcing sources switched several times throughout the record from 1979 to 2010 (Rui 2016). Such switches introduce erroneous values and unnatural trends in GLDAS-1 forcing data, including the large increase of snowfall after 2000, large precipitation errors in 1996 and 2000, and temperature errors during 2000–05.
Because of issues of data discontinuity, the use of GLDAS-1 forcing data are problematic for detecting long-term changes. For the TP specifically, when Wang and Zeng (2012) compared six reanalysis products and concluded that GLDAS (GLDAS-1) has the best overall performance in both daily and monthly precipitation, they focused on daily and monthly mean variables without considering the temporal characteristics. Based on their conclusion, Gao et al. (2014) used GLDAS-1 data as the “ground truth” for analyzing P–E change in that region. But comparing the precipitation change rate derived by GLDAS-1 data with that by observed data as shown in Fig. 4, we see that both the wetting in the northern and central part and the drying in the southeastern part of the TP are greatly exaggerated by GLDAS-1 data. Because of such issues, we should be cautious about the intensity and spatial extension of changes when concluding that “the climatologically humid southeastern TP is getting drier while the vast arid and semiarid northwestern TP is getting wetter” (Gao et al. 2014).
Meanwhile, GLDAS-2 uses the Princeton meteorological forcing dataset (Sheffield et al. 2006) as the only source of forcing data, which is a reanalysis product that has been bias corrected using observation-based products for the period 1948–2010. The continuity is much better than GLDAS-1. However, bias correction may introduce some uncertainties. As shown in Figs. 6 and 9, while the absolute value of MBE of GLDAS-2 precipitation is significantly smaller than that of GLDAS-1 precipitation, GLDAS-2 precipitation is less correlated with observations than GLDAS-1 precipitation in some periods (e.g., 1990–95) and in most months (except January–March), and the MAE of GLDAS-2 is generally larger than that of GLDAS-1 (MAE = 16.630 mm month−1 for GLDAS-1 and 18.983 mm month−1 for GLDAS-2 during 1979–2010). The results show that the bias-correction process greatly reduced the biases of the GLDAS-2 precipitation data, but at the same time, decreased the capability of describing the variability of precipitation and increased the MAE on average. Therefore, there is significant room for improving the quality of forcing data for improving the variability of precipitation and simulated outputs from GLDAS.
b. Consistency between GLDAS TWS and GRACE TWS
TWS data from GLDAS are commonly derived by summing up snow accumulation, total canopy water storage, and soil water. Syed et al. (2008) showed that GRACE-based storage changes are in agreement with those obtained from GLDAS simulations. Similarly, Yang and Chen (2015) showed that GRACE TWS and GLDAS TWS have good consistency with significant linear relations in the Tianshan Mountains and its adjacent areas, located along the northern border of Tarim basin.
However, by examining the spatial patterns of the change rates of TWS by GLDAS and GRACE shown in Fig. 12, we see that they do not match well in most areas of China. For instance, GRACE data indicate that TWS in NTP and the middle Yangtze River basin experienced increases of up to 3 cm yr−1, whereas TWS in the middle Yellow River basin experienced decreases of less than 1 cm yr−1. While the TWS increase in NTP and decrease in the middle Yellow River basin indicated by GRACE data are discernible in both GLDAS-1 and GLDAS-2, the large increase in the middle Yangtze River basin is not reflected by either GLDAS-1 or GLDAS-2.
In addition to the differences in spatial patterns, GRACE TWSC has larger seasonal variability than GLDAS TWSC, taking all of China into account, as shown in Fig. 14. This inability to sufficiently capture the seasonality in TWSC for GLDAS can be further illustrated by comparing the standard deviation of GLDAS TWSC data and GRACE TWSC data. The standard deviation of GRACE TWSC is significantly larger than that of GLDAS-1 and GLDAS-2, especially in the southeastern part of China (where the standard deviation is up to ~15 cm month−1 at some grid cells for GRACE vs less than 7 mm month−1 for both GLDAS-1 and GLDAS-2), and GLDAS-2 has the smallest standard deviation over all of China. Niu et al. (2011) stated that the soil thickness of 2 m in the Noah model is not capable of capturing the dynamics of critical zone (down to 5 m). But the comparison between the TWSC of GLDAS-1/Noah model with that of GLDAS-1 CLM and Mosaic model (both of which have a soil layer thickness of ~3.5 m) shows that, while Mosaic indeed has much larger seasonal variability in TWSC (standard deviation is up to ~15 cm month−1) than Noah over southeastern China, CLM has much smaller variability (standard deviation is less than 3 cm month−1) than Noah. That indicates that the increase of soil thickness in the model does not ensure better capability of capturing storage dynamics.
Syed et al. (2008) stated that GLDAS output fields can be used to better understand the processes contributing to TWS variations. While it may be true generally, we should be aware of the spatial and seasonal differences between what GLDAS simulated and what GRACE observed.
c. Causes of errors in simulation outputs
Although the average bias of GLDAS precipitation data in terms of MBE is small over China (−0.948 mm month−1 for GLDAS-1 and −0.053 mm month−1 for GLDAS-2 during 2001–10), the MAEs are 17.98 and 19.15 mm month−1 for GLDAS-1 and GLDAS-2, respectively, more than 30% of the average monthly precipitation (~51 mm month−1) over China. Seyyedi et al. (2015) assessed the error propagation of GLDAS precipitation data, finding that the GLDAS product exhibits significant increase of the mean relative error and random error going from precipitation to runoff. According to the results of the present study, the large MAE of GLDAS precipitation is an important cause of poor simulation of runoff in terms of CE (Table 3).
While the precipitation bias is very small over exorheic basins, the annual runoffs simulated by both GLDAS-1 and GLDAS-2 are much lower than the observed runoffs for the Yangtze River and Songhua River (see Table 2), and consequently, runoff coefficients calculated based on both GLDAS-1 and GLDAS-2 data (especially GLDAS-1) are much lower than those based on observed data for the two rivers. At the same time, the simulated runoff for the Yellow River seems to match the observation well. However, the streamflow of the Yellow River is greatly affected by human interventions (especially reservoir impoundment and water diversion to irrigation areas), and it is found that the observed discharge at Huayuankou (with a drainage area of 730 000 km2) at the lower Yellow River is often less than that at Lanzhou (222 500 km2) at the upper Yellow River. This means that the runoff observed at Huayuankou is in fact much lower than it should be naturally, and we believe that both GLDAS-1 and GLDAS-2 underestimate the real quantity of runoff in the Yellow River. Therefore, assuming that GLDAS is designed to simulate natural processes, both GLDAS-1 and GLDAS-2 significantly underestimate runoffs in all the exorheic basins in China.
Zaitchik et al. (2010) found that Noah 2.7 (i.e., GLDAS-1) with standard GLDAS forcing yields lower estimates of runoff than Noah 2.7 with Sheffield et al. (2006) forcing data (i.e., GLDAS-2 forcing data), and results with Sheffield et al. (2006) forcing data tend to more closely resemble observed discharge. According to the present study, using Sheffield et al. (2006) forcing data in GLDAS-2 (Noah 3.3) indeed increased the runoff for the exorheic basins, but this is not true for two endorheic basins (i.e., Tarim basin and NTP), probably because GLDAS-1 has much more snowfall than GLDAS-2 for those two basins, which contribute greatly to the runoff in Tarim basin and NTP. In fact, we believe that the Noah model has the defect of underestimating runoff even if there is little bias in the forcing data.
As runoff is underestimated in the exorheic basins, according to the principle of water balance, we know that ET is overestimated by both GLDAS-1 and GLDAS-2. Taking the Yangtze River, for example, by comparing the observed data with simulated data in Table 2, it is estimated that the annual runoff is underestimated by about 43% and 22% by GLDAS-1 and GLDAS-2, respectively, and the annual ET is overestimated by about 35% and 11% (by neglecting the annual storage change).
One important limitation regarding the evaluation of GLDAS is that the river flows are regulated in three exorheic basins over which runoff evaluations are conducted here. Therefore, the observed runoff data can only be used as a reference to evaluate GLDAS that does not model anthropogenic impacts on water balance components, such as irrigation and groundwater withdrawals.
9. Conclusions
The Global Land Data Assimilation System (GLDAS) is an important data source for global environment change and water cycle research. Using ground-based in situ measurements over continental China, the monthly scale forcing data (precipitation and air temperature) of GLDAS-1 and GLDAS-2 for the period of 1979–2010, and model simulations (runoff, TWS, and ET) during 2002–10 by GLDAS-1/Noah and GLDAS-2/Noah models, were systematically assessed.
a. Forcing data
Because of several switches of forcing data, GLDAS-1 has serious discontinuity issues in its forcing data. GLDAS-1 precipitation data have larger errors in 1996, and the snowfall amount has approximately doubled after 2000. Both GLDAS-1 precipitation data and GLDAS-2 precipitation data have a tendency toward declining quality during the period from 1979 to 2010. GLDAS-2 precipitation has much better temporal continuity than GLDAS-1 precipitation, as GLDAS-2 uses the Princeton meteorological forcing dataset (Sheffield et al. 2006), which has been bias corrected. While the bias correction greatly reduces the biases of the GLDAS-2 precipitation data, the bias correction makes GLDAS-2 precipitation less correlated with observed precipitation and has a larger mean absolute error (MAE) than GLDAS-1 precipitation for most months over the year.
Both GLDAS-1 and GLDAS-2 temperature data have better quality during 1979–90 than after 1990; GLDAS-1 temperature data have large errors during 2000–05, and GLDAS-2 data quality is poor during 2008–09. GLDAS-2 temperature data are superior to GLDAS-1 temperature data in terms of not only the temporal consistency but also the correlation and MAE with respect to ground observations.
GLDAS precipitation and air temperature data match the ground observations better in wet eastern China than in dry western China, and the match is the poorest in the Tibetan Plateau (TP). However, as the ground observation in TP is poor, especially in the western part of TP; the GLDAS data quality for those areas needs further investigation.
b. Model simulations
The change rates of TWS data by GLDAS and GRACE do not match well in most areas of China, and both GLDAS-1 and GLDAS-2 are not capable enough to capture the seasonal variation in monthly TWS change as observed by GRACE. GLDAS-1 has much larger spatial variability in TWS than GLDAS-2. The temporal change rate of TWS by GLDAS-1 is also more spatially variable than GLDAS-2. Some grid cells at the border between Tarim basin and TP have an increase rate of over 10 cm yr−1 according to GLDAS-1, which is quite questionable.
Runoff is underestimated in the exorheic basins in China by both GLDAS-1 and GLDAS-2, whereas the underestimation by GLDAS-2 is less than by GLDAS-1. GLDAS-2 performs much better than GLDAS-1 in simulating runoffs for the Yangtze River in southern China and the Songhua River in northeastern China. GLDAS-2 gives higher runoff coefficients than GLDAS-1 in exorheic regions across China, especially in southeastern China, the southern TP, and northeastern China. As for the endorheic basins, the estimated runoff coefficients are very high in central TP by both GLDAS-1 and GLDAS-2 and very high in Tarim basin and the northern TP by GLDAS-1, but very low in Tarim and the northern TP by GLDAS-2.
Evapotranspiration is overestimated in the exorheic basins in China by both GLDAS-1 and GLDAS-2, whereas the overestimation of ET by GLDAS-2 is less than that by GLDAS-1. GLDAS-2 has lower spatial variability in ET than GLDAS-1 over all of China and has a similar amount of ET to GLDAS-1 in northwestern half of China, but much lower ET estimates than GLDAS-1 in southeastern China.
In summary, this paper demonstrates the need for improving GLDAS forcing data through the reduction in reducing biases and enhanced spatial temporal variability of precipitation. Further, the results of the article demonstrate that model changes are required to improve the runoff generation processes for producing better data products.
Acknowledgments
The data used in this study were acquired as part of the mission of NASA’s Earth Science Division and archived and distributed by the Goddard Earth Sciences (GES) Data and Information Services Center (DISC). The study is financially supported by the National Science Foundation of China (Projects 41371050 and 41571130071) and the Specialized Research Fund for the Doctoral Program of Higher Education (20130094110007).
REFERENCES
Chen, Y., Yang K. , Qin J. , Zhao L. , Tang W. , and Han M. , 2013: Evaluation of AMSR-E retrievals and GLDAS simulations against observations of a soil moisture network on the central Tibetan Plateau. J. Geophys. Res. Atmos., 118, 4466–4475, doi:10.1002/jgrd.50301.
Gao, Y., Cuo L. , and Zhang Y. , 2014: Changes in moisture flux over the Tibetan Plateau during 1979–2011 and possible mechanisms. J. Climate, 27, 1876–1893, doi:10.1175/JCLI-D-13-00321.1.
Ghazanfari, S., Pande S. , Hashemy M. , and Sonneveld B. , 2013: Diagnosis of GLDAS LSM based aridity index and dryland identification. J. Environ. Manage., 119, 162–172, doi:10.1016/j.jenvman.2013.01.040.
Hao, Z., AghaKouchak A. , Nakhjiri N. , and Farahmand A. , 2014: Global integrated drought monitoring and prediction system. Sci. Data, 1, 140001, doi:10.1038/sdata.2014.1.
Helsel, D. R., and Hirsch R. M. , 2002: Statistical methods in water resources. USGS–TWRI Book 4, Chap. A3, 522 pp. [Available online at https://pubs.usgs.gov/twri/twri4a3/pdf/twri4a3-new.pdf.]
Huang, Y., Salama S. , Krol M. , Velde R. , Hoekstra A. , Zhou Y. , and Su Z. , 2013: Analysis of long-term terrestrial water storage variations in the Yangtze River basin. Hydrol. Earth Syst. Sci., 17, 1985–2000, doi:10.5194/hess-17-1985-2013.
Ji, L., Senay G. B. , and Verdin J. P. , 2015: Evaluation of the Global Land Data Assimilation System (GLDAS) air temperature data products. J. Hydrometeor., 16, 2463–2480, doi:10.1175/JHM-D-14-0230.1.
Landerer, F. W., and Swenson S. C. , 2012: Accuracy of scaled GRACE terrestrial water storage estimates. Water Resour. Res., 48, W04531, doi:10.1029/2011WR011453.
National Meteorological Information Center of CMA, 2012: Evaluation of Chinese ground-based precipitation grid dataset (V 2.0) (in Chinese.) Accessed 1 October 2015. [Available online at http://image.data.cma.cn/static/subject/doc/SURF_CLI_CHN_PRE_MON_GRID_0.5_ASSESSMENT.pdf.]
Niu, G.-Y., Yang Z.-L. , and Mitchell K. E. , 2011: The community Noah land surface model with multiparameterization options (Noah‐MP): 1. Model description and evaluation with local‐scale measurements. J. Geophys. Res., 116, D12109, doi:10.1029/2010JD015139.
Proulx, R. A., Knudson M. D. , Kirilenko A. , VanLooy J. A. , and Zhang X. , 2013: Significance of surface water in the terrestrial water budget: A case study in the Prairie Coteau using GRACE, GLDAS, Landsat, and groundwater well data. Water Resour. Res., 49, 5756–5764, doi:10.1002/wrcr.20455.
Qi, W., Zhang C. , Fu G. , and Zhou H. , 2015: Global Land Data Assimilation System data assessment using a distributed biosphere hydrological model. J. Hydrol., 528, 652–667, doi:10.1016/j.jhydrol.2015.07.011.
Ramillien, R., Famiglietti J. S. , and Wahr J. , 2008: Detection of continental hydrology and glaciology signals from GRACE: A review. Surv. Geophys., 29, 361–374, doi:10.1007/s10712-008-9048-9.
Rodell, M., and Coauthors, 2004: The Global Land Data Assimilation System. Bull. Amer. Meteor. Soc., 85, 381–394, doi:10.1175/BAMS-85-3-381.
Rodell, M., Chen J. , Kato H. , Famiglietti J. S. , Nigro J. , and Wilson C. R. , 2007: Estimating groundwater storage changes in the Mississippi River basin (USA) using GRACE. Hydrogeol. J., 15, 159–166, doi:10.1007/s10040-006-0103-7.
Rui, H., 2016: Readme document for Global Land Data Assimilation System version 1 (GLDAS-1). Accessed 6 October 2016. [Available online at http://disc.sci.gsfc.nasa.gov/services/grads-gds/gldas.]
Seyyedi, H., Anagnostou E. N. , Beighley E. , and McCollum J. , 2015: Hydrologic evaluation of satellite and reanalysis precipitation datasets over a mid-latitude basin. Atmos. Res., 164–165, 37–48, doi:10.1016/j.atmosres.2015.03.019.
Sheffield, J., Goteti G. , and Wood E. F. , 2006: Development of a 50-year high-resolution global dataset of meteorological forcings for land surface modeling. J. Climate, 19, 3088–3111, doi:10.1175/JCLI3790.1.
Syed, T. H., Famiglietti J. S. , Rodell M. , Chen J. , and Wilson C. R. , 2008: Analysis of terrestrial water storage changes from GRACE and GLDAS. Water Resour. Res., 44, W02433, doi:10.1029/2006WR005779.
Wang, A., and Zeng X. , 2012: Evaluation of multireanalysis products with in situ observations over the Tibetan Plateau. J. Geophys. Res., 117, D05102, doi:10.1029/2011JD016553.
Wang, F., Wang L. , Koike T. , Zhou H. , Yang K. , Wang A. , and Li W. , 2011: Evaluation and application of a fine-resolution global data set in a semiarid mesoscale river basin with a distributed biosphere hydrological model. J. Geophys. Res., 116, D21108, doi:10.1029/2011JD015990.
Yang, P., and Chen Y. , 2015: An analysis of terrestrial water storage variations from GRACE and GLDAS: The Tianshan Mountains and its adjacent areas, central Asia. Quat. Int., 358, 106–112, doi:10.1016/j.quaint.2014.09.077.
Yang, T., Wang C. , Yu Z. , and Xu F. , 2013: Characterization of spatio-temporal patterns for various GRACE- and GLDAS-born estimates for changes of global terrestrial water storage. Global Planet. Change, 109, 30–37, doi:10.1016/j.gloplacha.2013.07.005.
Zaitchik, B. F., Rodell M. , and Olivera F. , 2010: Evaluation of the Global Land Data Assimilation System using global river discharge data and a source-to-sink routing scheme. Water Resour. Res., 46, W06507, doi:10.1029/2009WR007811.
Zawadzki, J., and Kȩdzior M. A. , 2014: Statistical analysis of soil moisture content changes in central Europe using GLDAS database over three past decades. Cent. Eur. J. Geosci., 6, 344–353, doi:10.2478/s13533-012-0176-x.
Zhang, Y., Yao T. , Pu J. , Xiao C. , Kang S. , Hou S. , Ohta T. , and Yabuki H. , 1997: The Features of Hydrological Processes in the Dongkemadi River basin, Tanggula Pass, Tibetan Plateau (in Chinese). J. Glaciol. Geocryology, 19 (3), 214–222.
Zhou, X., Zhang Y. , Yang Y. , Yang Y. , and Han S. , 2013: Evaluation of anomalies in GLDAS-1996 dataset. Water Sci. Technol., 67, 1718–1727, doi:10.2166/wst.2013.043.
