1. Introduction

Soil moisture is an important component of the earth’s climate system. It influences the partitioning of available surface energy into latent and sensible heat fluxes, which links the surface energy balance and atmospheric circulation. Despite its evident importance, spatially comprehensive and long-term measurements of soil moisture have not yet been established. For example, over the United States, there exist only a 19-station network in Illinois and two catchments in Iowa that have long-term soil moisture observations. Over China, the former Soviet Union (FSU), and Mongolia, there are a total of about 600 stations. However, the sampling of soil moisture at these stations is neither homogeneous in space nor time. Satellite retrievals of soil moisture have recently become available and offer a promising approach for global-scale soil moisture monitoring. Nevertheless, their accuracy and representativeness, particularly in terms of soil moisture at depth and over heavily vegetated areas, are still problematic (Vinnikov et al. 1999b). Global retrievals are found to be biased in their mean and variability with respect to the model data and the ground observations, posing a severe problem for global soil moisture data assimilation (Reichle et al. 2004).

The lack of soil moisture observations has led researchers to depend upon model-estimated soil moisture for various studies, including climate modeling, water resource management, and seasonal prediction. Observed or computed atmospheric forcing is used to drive land surface models (LSMs) to produce time-varying soil moisture over large continental or global domains (examples cited later). The accuracy of soil moisture thus derived is affected by the accuracy of the atmospheric forcing and the LSM, and thus may vary significantly from time to time and from place to place.

In current coupled models, the precipitation forcing, a key forcing for soil moisture evolution, suffers from uncertainty and biases (e.g., Roads 2003). Even precipitation observations are not perfect due to spatial and temporal representativeness (Kanamitsu and Gruber 2003). As long-term observation-based precipitation analyses have become increasingly available, there are expanding efforts to take advantage of these precipitation products for developing land-state-related datasets. Specifically, a number of LSMs are integrated in uncoupled “offline” mode, driven by observation-based meteorological forcing, to produce soil moisture and other hydrological quantities. Examples include the Global Soil Wetness Project (GSWP; Dirmeyer et al. 1999), the North American Land Data Assimilation System (NLDAS; Mitchell et al. 2004a), and the Global Land Data Assimilation System (GLDAS; Rodell et al. 2004).

A more complex procedure is to produce soil moisture in a coupled mode, where LSMs are coupled to atmospheric models, such as global circulation models (GCMs). In this case, the land surface not only responds to the atmospheric state but also influences the atmosphere through the coupled fluxes. This approach can be carried out either in real-time mode, such as in the National Centers for Environmental Prediction (NCEP) Global Data Assimilation System (GDAS; Derber et al. 1991), or in retrospective mode, such as in the global reanalyses examined later here. A drawback for such coupled approaches is the computational and storage costs, which are orders of magnitude greater than the uncoupled approaches like GLDAS. Furthermore, soil moisture in coupled mode is forced with model-generated forcing that may not be accurately simulated, especially radiation and precipitation. Methods for assimilating precipitation and cloud cover observations in coupled analysis systems to improve their precipitation and solar insolation forcing at the surface are under development and application (Meetschen et al. 2004; Mitchell et al. 2004b).

To guide possible future improvements in the coupled systems, it is important to evaluate how well current coupled analysis systems actually describe the observed soil moisture characteristics. Although in situ observations are very limited and their spatial representativeness is unsatisfactory, it is still valuable to conduct such comparisons to better understand the problem of estimating the global distribution of surface water. This study examines the soil moisture fields in the NCEP–National Center for Atmospheric Research (NCAR) global reanalysis (R-1) (Kalnay et al. 1996; Kistler et al. 2001) and the NCEP–Department of Energy (DOE) Atmospheric Model Intercomparison Project (AMIP) global reanalysis (R-2; Kanamitsu et al. 2002).

Kanamitsu et al. (2002; see Fig. 7 therein) evaluated R-1 and R-2 soil moisture using soil moisture measurements taken from the Illinois network for the period 1979–98. R-2 soil moisture evolution was found to show a more realistic interannual variability than R-1. However, a recent study by Dirmeyer et al. (2004) indicates that R-1 yields a better temporal correlation with monthly mean observations than R-2 over several observation networks from the Global Soil Moisture Data Bank (Robock et al. 2000). This unexpected result motivated the present study to evaluate the quality of R-1 and R-2 soil moisture in various aspects. We first describe R-1 and R-2 soil wetness analysis and the Global Soil Moisture Data Bank in sections 2 and 3, respectively. We then present mean climate and annual variations of R-1 and R-2 soil moisture in section 4. Next, intercomparisons among R-1, R-2, and in situ observations for selected regions are presented in section 5. Sections 6 and 7 present the land surface water budget and lagged correlations, respectively. Section 8 presents conclusions.

2. R-1 and R-2 soil moisture analyses

R-2 is a follow-up to R-1 (Kanamitsu et al. 2002), with corrections of some known errors in the processing methods of R-1, as well as improvements to the model physics and the analysis system. Both R-1 and R-2 used exactly the same LSM, namely the Oregon State University (OSU) LSM based on Pan and Mahrt (1987). The soil column in both R-1 and R-2 consists of two soil layers of constant thickness (0–10 and 10–200 cm), yielding a globally constant soil depth (2 m). Some land surface characteristics are also universally fixed at all grid cells in both R-1 and R-2, including the fraction of vegetation cover (70%) and soil properties (including porosity of 0.47 and wilting point of 0.12 in volumetric units). However, as discussed next, there were significant differences in how the soil moisture evolution was constrained in R-1 and R-2, which made the resulting soil moisture evolutions distinctly different.

NCEP assessment of R-1 pilot tests indicated that soil moisture manifested serious drift in certain regions, such as drastic drying over the Amazon basin. Thus NCEP added an artificial nudging term in subsequent pilot testing (and ultimately in the production system) to adjust the soil moisture. The soil moisture climatology to which the R-1 was nudged (referred to as “the derived Mintz–Serafini climatology”) was derived by NCEP from the global monthly climatology of Mintz and Serafini (1992). While the Mintz–Serafini climatology itself is from a bucket model of 15-cm capacity, which corresponds to the plant available water of a 1-m soil column, the derived Mintz–Serafini climatology for R-1 is represented in terms of volumetric soil wetness (the units used in the OSU LSM) over a 2-m depth. The relaxation of a 1-m column to a 2-m column is likely to result in an amplification of the annual cycle of the water mass in the 2-m column (discussed in section 4). Subsequent independent studies showed that the nudging term resulted in spurious seasonal (too large) and interannual (too small) variations in R-1 soil moisture (Roads et al. 1999; Srinivasan et al. 2000; Maurer et al. 2001).

In R-2, the nudging of soil moisture toward a prescribed climatology was removed, and instead, differences between model-generated and observed precipitation are used to adjust the modeled soil moisture. The observed precipitation analysis is the global “pentad” analysis (5-day accumulation) of the Climate prediction Center (CPC) Merged Analysis of Precipitation (CMAP; Xie and Arkin 1997), which is a blend of surface rain gauge observations and satellite estimates. Differences between the model-generated and observed pentad precipitation are used to compute a soil moisture correction term that is applied evenly throughout the subsequent 5-day period. Such a correction of either sign is generally applied to the topmost soil layer (0–10 cm), except for the case where it is necessary to adjust the second soil layer (10–200 cm) when an adjustment to the top layer would produce negative soil moisture. When the ground is frozen (determined by the surface temperature), this correction is skipped. In the case of no runoff, the excess between observed and modeled pentad precipitation is added to the modeled soil moisture while the deficit between observed and modeled pentad precipitation is removed from the modeled soil moisture. In the case of runoff, a correction is made only if the observed precipitation is less than water infiltrating the soil column. Otherwise, the difference between modeled and observed precipitation is assumed only to affect modeled runoff and no adjustment is made.

The R-2 soil moisture is presumably physically more realistic than that of R-1, by virtue of (a) applying an observation-based precipitation analysis and (b) avoiding the use of an external soil moisture climatology—especially a climatology derived from an LSM that is different from the one used in the reanalysis system itself. It should be noted that whereas the R-1 reanalysis spans the period 1948–present, the R-2 spans 1979–present. In this study, the comparison is made from 1981 onward because R-2 soil moisture required a couple of years of spinup (R-2 was initialized from R-1) and because the CMAP pentad product was problematic in R-2’s earlier period (missing data were mistakenly reported as zero).

3. The Global Soil Moisture Data Bank

The Global Soil Moisture Data Bank contains soil moisture observations from a large variety of climatic regions across the Northern Hemisphere midlatitudes, including China, India, Mongolia, FSU, and the United States (Robock et al. 2000). Soil moisture measurements from over 600 stations are assembled, quality controlled, and posted on a Web site (http://climate.envsci.rutgers.edu/soil_moisture). These data have been widely used to study spatial and temporal variability of soil moisture (Vinnikov et al. 1996; Entin et al. 2000), to evaluate model simulations of soil moisture (Vinnikov and Yeserkepova 1991; Robock et al. 1995; Schlosser et al. 1997, 2000; Yang et al. 1997), to calibrate satellite-retrieved soil moisture (Vinnikov et al. 1999b), and to design new soil moisture observation networks (Vinnikov et al. 1999a). There are ongoing efforts to collect new data (e.g., a 45-yr record for 141 stations from the Ukraine) and to update the Chinese, Illinois, and Mongolia datasets (A. Robock 2004, personal communication).

In this study, we excluded stations where soil moisture was sampled only during the growing season (April to October). The rationale for such a strict selection criterion is because the analysis technique used in this work (in particular, the application of time filters to soil moisture anomalies in section 5) requires a continuous record with minimal missing data. The application of this selection criterion eliminated the Mongolia network, catchments in Iowa, and many stations in China. While the FSU datasets cover extensive spatial area and long time periods (e.g., 1952–85 for the RUSWET-50STA network), most FSU datasets were available only up to 1985, and hence the overlap between the observations and the R-2 was too short to warrant a robust comparison. Therefore, only the FSU observations at the Valdai water balance station are used here. Four regions were eventually chosen, including 19 stations over Illinois (referred to as IL), the Valdai water balance station from FSU (VD), and Chinese stations located in south China (CS) and around central China (CC). Figure 1 shows a map of the station distribution for the Global Soil Moisture Data Bank (a reproduction of Fig. 4 of Robock et al. 2000). The locations of the four regions used for this analysis are presented in Fig. 1, and the information regarding the datasets is summarized in Table 1. Note that many stations shown in Fig. 1 were not included in this study, as a result of the site-selection criteria employed in this analysis (i.e., only stations that sample soil moisture continuously and well beyond 1981).

The Illinois dataset, beginning in 1981 and still continuing, has been widely used and is well documented (Hollinger and Isard 1994). Observations are taken two–three times per month during the growing season and once per month the rest of the year. The Valdai water balance research station continuously collected all components of land surface water balance and some atmospheric forcing for the period 1960–90. The data have temporal resolution of about one to three measurements per month. The Chinese dataset consists of an 11-yr record (1981–91) and the temporal resolution is the same as for the Valdai data. [Note: the Chinese data have been recently updated through 1999, and a study using the updated Chinese data for evaluating R-1 and R-2 is given in Li et al. (2005).] In summary, the period of record for this study was 1981–98, 1981–90, and 1981–91 for the Illinois, Valdai, and Chinese stations, respectively. The one to three samples per month were averaged whenever the data were available to get the monthly mean, and monthly means for these stations were then averaged to get the domain-average monthly mean values. Since the field examined here is deep soil moisture (1- and 2-m columns), which evolves slowly with time, the monthly average derived here from only one to three samples per month is likely to give a fair estimate of the actual monthly mean.

To compare in situ observations with reanalysis soil moisture products, some conversion was needed to make them compatible with each other. In particular, observed and reanalysis soil moisture were converted to column soil moisture (in millimeters) for the top 1 m, except for Illinois, where observations over 2 m were measured and compared over 2-m from reanalysis. Over Valdai, total soil moisture (in millimeters) stored in layers of 0–20, 0–50, and 0–100 cm was reported and could be directly used for the comparison. For Chinese stations, percent wetness by mass at 11 levels down to 1-m depth was sampled. These measurements were converted to water contents at each level and then summed to get total soil water content for the top 1 m (in millimeters). Over Illinois, soil water contents were sampled for 11 levels down to the 2-m level. Total soil moisture for the top 2 m (in millimeters) was computed by adding soil water content from the top layer down to the 11th layer.

For R-1 and R-2, the monthly values of soil wetness (fraction by volume) at two layers (0–10 and 10–200 cm) were converted to soil water content (in millimeters) and then summed to get column soil water content for the entire 2-m layer. Over the China and Valdai sites where observations were not available beyond the 1-m depth, the reanalyses’ 2-m soil moisture was scaled to 1-m soil moisture. Three methods were tested as candidates for scaling the model column soil moisture from 2-m depth to 1-m depth: (a) SW (1 m) = 0.5 × SW (2 m); (b) SW (1 m) = s × SW (2 m), where the parameter s is a function of month; and (c) SW (1 m) = a + b × SW (2 m), where parameters a and b are constants. The empirical parameters s, a, and b used in methods b and c were derived from the Illinois soil moisture dataset. The three methods were then used to scale R-2 soil moisture from 2- to 1-m depth over the IL region. Iterative comparisons of methods b and c with the straightforward depth-based scaling method (method a) yield comparable results, when compared to observed 1-m column soil moisture. Therefore, method a was ultimately chosen to scale soil moisture water content for the remainder of the study. While this approach provides reasonable mean values, it has some unfortunate implications on perceived persistent characteristics, which are discussed in section 7.

Prior to comparing in situ observations, which represent a single point, with the global reanalysis products, which provide a large-scale average, the spatial variations of the observed soil moisture in the four regions were examined. For IL, where observations are taken at 19 stations, we conducted two analyses to assess soil moisture variability among stations. First, we computed the temporal correlation between two stations, and then computed the average for all possible pairs of stations to get a regional average of the correlation. Second, we computed the domain-averaged soil moisture from all stations. Temporal correlations between single station and domain average were computed, and 19 sets of station versus domain-average correlation were then averaged. High temporal correlations were found (0.69 for the first method and 0.81 for the second), suggesting that soil moisture sampled at the Illinois network was well represented by the domain average.

The same analysis was conducted for VD, CC, and CS regions. Correlations were generally high (above 0.5) except for the CS region. Stations in the CS region were poorly correlated with each other, reflecting a lack of coherence among the stations. On the other hand, high correlations obtained at VD (0.76 and 0.92 for first and second method, respectively) occur mainly because the VD composite contained three catchments within a very small area. In this study, the domain-average observations serve as the “truth” for evaluating model-estimated fields. The analysis conducted above indicates that the averages derived for the IL and CC regions may adequately represent the large-scale features, while those derived for VD and CS may not. Especially for the CS observations, the number and choice of stations could have a significant impact on the domain average.

4. Reanalyses soil moisture climatology and seasonal variation

Figure 2 shows the annual mean of the column soil moisture (millimeters) over the 2-m depth for R-2 and R-1 across the global domain, averaged for the period 1981–98, along with the corresponding R-2 minus R-1 difference field. Both reanalysis products show the expected large-scale climatological features such as the wet tropical forests and dry deserts, although there are significant regional differences. For instance, R-2 soil moisture is considerably drier than R-1 over Eurasia in mid- and high latitudes and over the eastern two-thirds of North America, while R-1 shows drier conditions than R-2 over Borealia. Generally, the results shown here are consistent with the mean climate of the soil wetness index presented in Dirmeyer et al. (2004).

Inspection over central Asia of the R-2 annual mean and the R-2/R-1 difference field in Fig. 2 reveals a distinct boundary around 50°N in Asia. This characteristic is also found in the R-2 seasonal mean (shown later in Figs. 3 and 4). This edge-like feature is likely a result of a corresponding discontinuity in the CMAP precipitation analysis, which contains a similar “edge” feature (not shown), particularly during the warm season (May to October). Further studies are needed to determine why the CMAP precipitation analysis contains such a boundary near 50°N.

Figure 3 is as in Fig. 2, but for boreal summer [July and August average (JA)]. Compared to R-2, R-1 is generally drier where soil moisture is less and wetter where soil moisture is greater. Seasonal changes in atmospheric forcing are evident. For instance, the JA mean shows wetter conditions than the annual mean over Southeast Asia and drier conditions over the Amazon. Figure 4 shows the mean soil moisture field for boreal winter [January and February average (JF)]. Again R-1 shows stronger wet or dry conditions than R-2 and the soil moisture variations are consistent with the seasonal changes in large-scale meteorological processes. However, R-1 shows widespread wet conditions over most of the mid- and high latitudes of the Northern Hemisphere. We examined the JF soil moisture range in R-1 and R-2 for the period 1981–98 over a midlatitude region (52° to 65°N, 70° to 130°E) and found that R-1 column soil moisture over the 2-m depth is bounded by 723 and 771 mm, while R-2 soil moisture ranges from 284 to 641 mm. Upon scaling the column soil moisture (in millimeters) to relative wetness (in percent) based on porosity and wilting point, the range is 69%–76% for R-1 and 6%–57% for R-2. The corresponding range for boreal summer is 3%–72% for R-1 and 4%–70% for R-2. Clearly, the R-1 mean soil moisture for boreal winter (shown in Fig. 4) stays persistently wet across a very broad expanse of the mid- and high latitudes.

It seems unreasonable for the winter-season soil moisture to reach such a spatially homogeneous wet state (∼70%–75% saturation) over such an extensive area. As shown in Fig. 5, the derived Mintz–Serafini climatology during boreal winter, converted to column soil moisture (millimeters) for the 2-m depth, indeed reaches fairly wet conditions over most of the mid- and high latitudes of the Northern Hemisphere. For the midlatitude region examined here, the JF mean from the derived Mintz–Serafini climatology is 940 mm (reaching porosity for R-1) for all the grid cells considered. Such an exaggerated winter versus summer wet-to-dry contrast is likely caused by the relaxation of soil wetness from a smaller capacity bucket to the 2-m column. Spurious features in R-1 soil moisture have been attributed to the nudging term in previous studies (Roads et al. 1999; Srinivasan et al. 2000; Maurer et al. 2001). The results shown here further indicate that not only was the soil moisture nudging term used in R-1 too large, but the soil moisture climatology to which the R-1 was nudged was also problematic.

To assess the annual variations in the two global reanalysis products, the differences between the two seasons mentioned above were examined (not shown). Soil moisture from R-1 shows large annual variations over most of the regions except for arid regions where soil moisture variability is low. Similar annual variations are found in R-2, but the signals are systematically weaker. There are significant differences between the two reanalysis products over the mid- and high latitudes of the Northern Hemisphere. Strong drying patterns associated with winter-to-summer change are evident in R-1 but not in R-2, again presumably due to the nudging scheme and the climatology employed in R-1.

5. Comparisons of R-1, R-2, and observations

Figure 6 shows the area-average seasonal cycles, averaged for the entire length of each dataset, from observations, R-1, and R-2 for the four regions described in section 3. The observed mean seasonal cycles at IL and VD reflect a typical midlatitude seasonal cycle with soil moisture peaking during late winter and reaching a minimum during late summer. The seasonal cycles from CS and CC observations show different characteristics reflecting that enhanced summer evaporation is compensated by increased summer monsoon precipitation. The seasonal cycles shown here generally agree with Robock et al. (2000), although the regions selected in their work were not the same as ours.

The mean seasonal cycle from R-1 is quite strong over all regions. The presence of a winter maximum and summer minimum in R-1 agrees with observations over IL and VD, whereas large discrepancies exist for the CC and CS regions. For all regions, the R-1 interannual variations (as indicated by the length of the dark bars for the standard deviations and the outer hatches for the extremes) are deficient as they tend to be too small. By contrast, the R-2 interannual variations are more representative of the observations than R-1. The phasing and amplitude of the seasonal cycle in R-2, although imperfect, appear to agree slightly better with the observations than R-1. At IL, R-2 is relatively wet during the cold season and relatively dry during late summer and fall. Over CC and CS, R-2 has milder wet conditions during summer. Perhaps the major R-2 problem is that it has no seasonal cycle at VD (similar weakness is shown in the CMAP data; see section 6) and too large a seasonal cycle at CS.

Temporal correlations between the soil moisture anomalies of the reanalysis products and observations (2 m at IL and 1 m over VD, CC, and CS) are summarized in Fig. 7. The correlations are computed for the raw time series (time filter = 0 month) as well as the smoothed time series using time filters for 4, 6, 12, and 24 months. The results shown here are fairly robust, as only a small set (5 of the 40 cases) is not statistically significant. These include the R-2 correlations over VD (filter = 4, 6, 12, 24 months) and R-1 correlation over CS (filter = 24 months). Generally, the temporal correlations for the two reanalyses are comparable except for VD where the correlations for R-1, although poor, are better than R-2. The response of temporal correlations to time filters varies among the regions. For instance, the temporal correlations between R-2 and observations over IL drop as high-frequency variations are removed from the time series, whereas the opposite is found over CC and CS.

The low interannual variability and pronounced seasonal cycle in R-1 reflects the heavy-handed influence of the nudging method applied in R-1, wherein the soil moisture is forced to an assumed external climatology, which itself has no interannual variability. Furthermore, the external climatology is derived from a bucket model rather than the OSU LSM applied in R-1. The resulting R-1 soil moisture suffers from the likelihood that the inherent soil moisture climatologies of the two LSMs are substantially different—a likelihood illustrated by the studies of Koster and Milly (1997), Dirmeyer et al. (1999), and Schlosser et al. (2000).

Although R-2 does a better job of simulating interannual variability and mean seasonal cycle than R-1, R-2 is sometimes worse than R-1 in terms of temporal correlations with the monthly mean observations. Since the land surface models are the same for both R-1 and R-2, the differences in soil moisture evolution between the two reanalyses can be attributed to two factors: differences in the atmospheric forcing from the parent GCM and the different methods applied to adjust the soil moisture. These two components and their relative contribution to the overall surface water budget are examined in the following section.

6. Analysis of the surface water budget

In general, the surface water budget can be expressed as

View Expanded

The diagnostic budget includes precipitation (P), evaporation (E), runoff (R), and a “nonclosure” term (N). In reality, W is the total surface water storage, including soil moisture and the water content of the snowpack. In this study, the calculation of the tendency term—for example, the left-hand side of Eq. (1)—considered only soil moisture, and thus the storage change in the snowpack is incorporated into the nonclosure term. For both reanalysis systems, the global snowpack is updated once daily from an external snowpack analysis. (Note: the snowpack analyses are not the same for R-1 and R-2, as there were errors in the snow specification in R-1 for most of the years since 1974.) This daily increment of the snowpack acts as source or sink of surface water and is one contribution to the nonclosure term. Additionally, N includes the soil moisture changes imposed by the respective soil moisture nudging/adjustment method, and model formulation errors, which are associated with time schemes, time filters, and programming inconsistencies like the treatment of latent heat of fusion in evaporation.

Prior to examining the relative contribution of each term in (1), we compared precipitation forcing in R-1 and R-2 with observation-based precipitation datasets. Table 2 summaries the root-mean-square differences and temporal correlations among various precipitation fields for our four assessment regions for the period 1979–98, including the model precipitation of R-1 and R-2 and three observations-based precipitation products [the CMAP pentad precipitation used to adjust R-2, the CMAP monthly precipitation (Xie and Arkin 1997), and the Global Precipitation Climatology Project (GPCP) version 1 combined precipitation dataset (Huffman et al. 1997)]. Additionally, for the IL region alone, the monthly version of the daily CPC U.S. 1/4° “unified” precipitation analysis (Higgins et al. 1996, 2000) provided by the Climate Diagnostic Center is included as well. These observation-based precipitation fields vary in character, spatial and temporal resolution, and data sources. For instance, the CPC U.S. unified precipitation analysis is derived solely from gauge measurements, but the GPCP and CMAP analyses represent blends of gauge measurements and satellite estimates and they are dominated by the satellite estimates over regions of sparse gauge data. In addition, the spatial resolution is 0.25° for the CPC unified analysis and 2.5° for GPCP; the temporal resolution for the monthly version and pentad version of CMAP is 1 month and 5 days, respectively. All precipitation fields were converted to monthly accumulated precipitation (millimeters) and then averaged over the four regions accordingly. Table 2 shows that the precipitation forcing used to adjust R-2 (i.e., the CMAP pentad product) agrees better with the CMAP monthly product (as expected), the GPCP precipitation dataset, and the CPC U.S. unified analysis than does the model precipitation in either R-1 and R-2. However, among the observation-based datasets, the CMAP pentad product has higher root-mean-square errors and lower correlation. Although the evolution of R-2 soil moisture is improved as a result of the soil moisture adjustment using the CMAP pentad analysis, the errors inherent in the latter analysis are likely to reduce the full potential of the adjustment approach of R-2.

Figure 8 shows the average seasonal cycle of precipitation, P; evaporation, E; runoff, R; soil moisture tendency, dW/dt; and the nonclosure term, N (in millimeters per month), averaged for the period 1981–98, from R-1, R-2, and observations (when available) for the four regions. The observed precipitation time series depicted for IL (averaged for the period 1981–98) is taken from the CPC 0.25° unified precipitation, and that depicted for VD (averaged for the period 1981–90) is taken from the archive of the Global Soil Moisture Data Bank. The precipitation forcing in R-2 shows better agreement with the observed seasonal cycle than R-1, as R-1 appears to have excess precipitation during the warm season.

Both precipitation and evaporation peak during the warm season. Mixed results, however, are found for runoff. Over CS, runoff peaks during the summer for both reanalyses, corresponding to the summer monsoon. Over IL, runoff peaks during late winter for R-1 and shows a very weak summertime peak for R-2. Over VD, both R-1 and R-2 fail to capture the observed intensity associated with springtime snowmelt. Additionally over VD, the phasing of the R-1 seasonal cycle for runoff shows poor agreement with the observations (temporal correlation 0.29, computed from the full time series for the entire period), with the counterpart runoff phasing in R-2 being acceptable (temporal correlation 0.44). For R-1, precipitation is larger than evaporation and runoff at all sites, except for VD where evaporation becomes comparable to precipitation. For R-2, the same feature (P > E > R) is noted for CC and CS, but evaporation becomes the dominant warm-season hydrometeorological process at VD and IL. In general, the results shown over IL are consistent with the study by Maurer et al. (2001) in which an offline hydrologic model was used to evaluate the land surface water budgets in R-1 and R-2 over the Mississippi River basin.

As noted earlier, the nonclosure term, N, in both R-1 and R-2 is presumably dominated by the respective nudging/adjustment methods applied to soil moisture in R-1 and R-2 (e.g., Roads et al. 2002). The precipitation correction in R-2 is presumably more physically consistent than that in R-1, as the adjustment is derived merely from the difference between modeled and observed pentad precipitation. However, as may be seen in Fig. 8 by the lack of full agreement between the precipitation forcing in R-2 and the residual forcing, there are clearly other problems associated with the surface water balance. We believe these additional problems include the following: the 5-day lag used when making the soil moisture correction, the lack of precipitation correction when runoff is occurring and when the ground is frozen, a primitive treatment of soil properties, the snowpack analysis increment added when correcting the model snow cover with observed snow cover, and the inability of the OSU LSM to reach a balanced surface water budget [e.g., an error in the gravitational drainage term has been identified in Maurer et al. (2001)]. A better surface water balance has now been implemented in a recent R-2 revision and the execution of the revised R-2 is now underway for the Coordinated Enhanced Observing Period (e.g., Roads et al. 2003). However, whether these corrections will improve the correlation with observations remains to be seen.

7. Lagged correlations

In this section, we examine the time scale of the soil moisture variation. Geographical variability in the persistence of soil moisture anomalies has been discussed in previous modeling studies (Delworth and Manabe 1988; Roads et al. 1999; Koster and Suarez 2001; Oglesby et al. 2002). Soil moisture datasets from the Global Soil Moisture Data Bank have also been used to quantify soil moisture persistence. Vinnikov and Yeserkepova (1991) analyzed Russian soil moisture in the upper 1-m soil layer and derived anomaly decay scales of 2–3 months. Vinnikov et al. (1996) and Entin et al. (2000) found a similar time scale when analyzing data for different regions. A dependence of soil moisture variations upon the depth of the soil layer considered was found in Entin et al. (2000). In the deeper layer (1–2 m), the time scales of persistence become larger (in the range of 5–7 months) and temporal variations become smaller (about 75% less than in the top 1-m layer). Such time scales have profound implications for both seasonal prediction as well as remote sensing of soil moisture. Recent modeling studies have indeed found an impact from soil moisture on potential predictability at seasonal time scales (Dirmeyer 2000; Koster et al. 2000; Kanamitsu et al. 2003).

An analysis of lagged correlation provides an approximate measure of how the initial soil moisture anomaly persists in time and is a good indication of the characteristics of the hydrological processes. Figure 9 shows the global distribution of 1-month-lagged autocorrelation of the soil moisture anomaly for the top 2 m. Note that R-1 manifests notably less memory of soil moisture anomalies after 1 month than R-2, which shows much greater persistence. The short memory in R-1 is presumably caused by the heavy-handed nudging in R-1, which damps the persistence of any anomaly. Geographical variations of soil moisture persistence in R-1 and R-2 are consistent with a previous study (Koster and Suarez 2001), but only over certain areas, such as North Africa. The differences are expected since a different GCM and LSM are used in Koster and Suarez (2001).

Figure 10 shows the anomaly lag autocorrelation over IL for the top layer (0–10 cm) and the entire column (0–2 m) during 1981–98. The e-folding times are determined from the autocorrelation function using the scheme that was developed by Lohmann and Wood (2003) and later applied in Cosgrove et al. (2003). For the 10-cm soil layer, the e-folding times derived from R-1, R-2, and observations are 0.89, 1.02, and 0.84 months, respectively, and the e-folding time becomes larger (1.5, 6.1, and 7.8 months for R-1, R-2, and observations, respectively) for the entire 2-m soil column. Hence, temporal scale for R-2 for the top 2 m shows better agreement with the observations while the R-1 lag correlation has an unrealistically fast decay. Considering the vertical dependence of time scales reported in Entin et al. (2000), we did not apply such a lag correlation analysis at our other assessment sites, which observe only to 1-m depths. As noted in section 3, a depth-based method has been used to scale R-1 and R-2 soil moisture from the 2- to 1-m soil layer at the Valdai and Chinese sites. Although such an approach can adequately scale the magnitude of the soil water content, it does not appropriately adjust the vertical variation of temporal persistence. Unfortunately, the temporal behavior derived from the scaled 1-m soil moisture of either R-1 or R-2 is greatly influenced by the temporal characteristics of the nearly 2-m-deep second soil layer (10–200 cm).

8. Conclusions

This study evaluates soil moisture from two NCEP global reanalysis systems, R-1 and R-2. The differences in R-1 and R-2 soil moisture were mainly due to the different nudging/adjustment approaches applied to the soil moisture state. The nudging term applied in R-1 adjusted the soil moisture to a prescribed monthly climatology, which itself has no interannual variability and is derived from a different LSM. The nudging/adjustment method was completely redesigned for the follow-on R-2, which used differences between modeled and observed precipitation to derive the magnitude and timing of the soil moisture adjustment.

In general, climatological means of soil moisture between R-2 and R-1 showed many geographic similarities. There were, however, some major regional differences: R-2 soil moisture was considerably drier than R-1 over Eurasia and eastern North America in mid- and high latitudes, while R-1 was drier than R-2 over Borealia. The seasonal cycle of soil moisture variation between R-2 and R-1 was very different, particularly over the high latitudes of the Northern Hemisphere, where R-1 exaggerated the amplitude of the seasonal variations. These differences may be attributed to the differences in the forecast models used for R-1 and R-2, namely the albedo, formulation of shortwave radiation and cloudiness, the different precipitation forcing, and the different methods for soil moisture nudging/adjustment.

R-1 and R-2 soil moisture fields were evaluated using in situ observations from the Global Soil Moisture Data Bank. The observed soil moisture was a reference for assessment over four regions, namely Illinois (IL), Central China (CC), South China (CS), and Valdai in Russia (VD). Soil moisture observations over other regions were deemed poorly suited for this comparison due to the duration of the measurements and other criteria. Among the four regions, observations in South China were not very homogeneous and Valdai represented only one station; thus comparison over these two regions should be interpreted with caution.

Comparison with in situ observations showed that the mean seasonal cycle from R-1 was too strong over almost all regions studied here. The phasing and amplitude of the seasonal cycle of R-2 appeared to agree slightly better with observations than R-1. In addition, R-1 and R-2 yield comparable temporal correlation with monthly mean observations, except for VD where the correlations for R-1, although poor, are better than R-2. Last, the lag correlation analysis of the 2-m column soil moisture anomaly showed that R-1 had unrealistically weak persistence, at least in terms of the available observations. The lag correlation of R-2 seemed more reasonable, and thus potentially should be more useful for initializing the soil moisture of coupled global models used for seasonal predictions.

In conclusion, because of the lack of large-scale, long-term, and continuous observations of soil moisture, it remains very challenging to evaluate model-generated soil moisture fields. The limited comparisons performed here showed that R-2 is generally more physically consistent and agrees better with observations in terms of seasonal cycle, interannual variations, and anomaly persistence. However, R-1 sometimes performs better at capturing soil moisture variations on monthly (but not longer) time scales, although it was strongly influenced by its respective method of soil moisture nudging/adjustment. For future reanalysis projects, it is desirable to drop soil moisture nudging/adjustment and instead assimilate precipitation forcing via the external precipitation analysis or even directly assimilate satellite soil moisture.