4.1. Spatial Patterns Compared With the NLDAS Products
 Soil moisture, although crucial to climate, has even less observed data than soil temperature suitable for model validation. In particular, representative data for geographic distributions over the U.S. are lacking. Here we compare the CMM5 simulation with the NLDAS products. Schaake et al.  demonstrated that total soil water storage from each of the four NLDAS LSMs is highly correlated with observations at a few limited regions with available data, although inter-model differences are substantial.
 Figure 7 presents seasonal mean distributions of soil moisture in the top 1 m simulated by the CMM5 and NLDAS/Mosaic averaged during 1997–2002. Since the Mosaic soil moisture is systematically wetter than the CMM5, the annual mean difference (221 minus 214 mm) between the Mosaic and CMM5 averaged over the entire land domain is removed from the Mosaic patterns in Figure 7, emphasizing the comparison in spatial patterns and seasonal variations. In both models, the general seasonal mean soil moisture patterns are dry over the Rockies/Great Plains and Florida while wet over the rest of the U.S., especially in the Northeast, Northwest, and Midwest. The wettest conditions are simulated in spring, which may be attributed to the integrated excess of precipitation over evapotranspiration during the cold season. The driest conditions are produced in fall, which results primarily from accumulated high evapotranspiration rates during the warm season.
Figure 7. Geographic distributions of soil moisture (mm) for the top 1-m layer simulated by the CMM5 (left) and NLDAS/Mosaic (right) in winter (DJF), spring (MAM), summer (JJA) and fall (SON) averaged during 1997–2002.
Download figure to PowerPoint
 A good spatial agreement between the CMM5 and Mosaic is found over most regions, except where the latter is wetter than the former over the Ohio River basin and lower Mississippi River basin, especially in spring. Note that the Mosaic soil moisture is a product of the NLDAS driven by observed meteorological forcing and precipitation over the entire domain, while the CMM5 values result from the dynamic relaxation of meteorological LBCs only within the narrow domain edges. One may expect that the CMM5 and Mosaic soil moisture differences are caused by CMM5 precipitation biases. Our analyses show that such a correspondence is not robust. For example, the CMM5 simulates larger (smaller) amounts of precipitation than observations over the Midwest (eastern U.S.) during spring (fall) [Liang et al., 2004b]; in contrast, the CMM5 soil moisture is drier (wetter) than the Mosaic over the respective regions. We speculate that deficiencies in LSM representations of both the CMM5 and NLDAS may likely be responsible for the result differences.
4.2. Temporal Variations in Illinois and Iowa
 Table 1 compares the general statistics of the soil moisture annual cycles in Illinois and Iowa between observations and simulations by the CMM5, NLDAS/Mosaic and R-2. The Illinois annual means for the Mosaic, CMM5 and observations are, respectively, 334, 272 and 296 mm in the top 1-m layer and 596, 559 and 686 mm in the top 2-m layer. The amplitude of the annual cycle produced by the Mosaic is greater than the CMM5 and observations. The corresponding standard deviations are 39, 27 and 33 mm in the top 1-m layer and 76, 40 and 38 mm in the top 2-m layer. The R-2 has a much drier annual bias and smaller seasonal variation than others, producing an annual mean of only 501 mm with a standard deviation of 18 mm in the top 2-m layer. The comparisons in Iowa are qualitatively similar.
Table 1. Comparison of Annual Cycle of Soil Moisture (mm) Simulated by the CMM5, Mosaic, and R-2 Outputs With Observationsa
| ||Illinois (Annual)||Iowa (Apr–Nov)|
|R-2|| || ||500.6||18.4|| ||456.7|
 Figure 8 compares the annual cycles of soil moisture in the upper 1-m and 2-m layers and precipitation averaged over Illinois and Iowa between observations, CMM5 simulations, NLDAS/Mosaic outputs, and R-2 products. The annual (April–November) means of soil moisture for both layers listed in Table 1 are removed from the respective cycle in Illinois (Iowa). In Illinois, observations showed that soil moisture in both layers reaches the highest in spring and the lowest in fall, while winter is relatively wet and summer is relatively dry. This seasonality of soil moisture is associated with snow accumulation in winter and subsequent melting in spring, enhanced evapotranspiration by growing vegetation in summer, and decreasing rainfall in fall. The CMM5 realistically simulates this annual cycle except for a systematic dry bias. This dry bias cannot be fully explained by the CMM5 precipitation which is greater than observations on an annual basis. It may partially be attributed to soil moisture overestimation by monitoring sensors that tend to include the contribution from emissions of non-water substances (such as organic material). On the other hand, the annual cycle is relatively flat in the R-2 and much sharper in the Mosaic; both of which are worse than the CMM5 as compared with observations.
Figure 8. Monthly variations of precipitation (mm day−1) (a, d), soil moisture (mm) for top 1-m (b, e) and 2-m (c, f) layer for observations (OBS, thick solid) and simulation of the CMM5 (thick dashed), R-2 (thin dashed) and NLDAS/Mosaic (thin dot-dashed) averaged over Illinois (left) during 1984–2002 and Iowa (right) averaged during 1982–1994 except for the Mosaic during 1997–2002. Annual (April–November) means listed in Table 1 are removed from the respective cycles of soil moisture for both layers in Illinois (Iowa).
Download figure to PowerPoint
 Another possible explanation for the CMM5 soil moisture deficiency may be the model difference in evapotranspiration from observations. Direct measurements of actual evapotranspiration are rare. Such measurements have been continuously conducted since 1997 on a crop field near Champaign, Illinois, with a 2-yr rotation of corn and soybean [Meyers and Hollinger, 2004]. In addition, there are estimates for potential evapotranspiration, following Rosenberg et al. , from surface radiation, temperature, humidity and wind measurements at 19 stations of the Illinois Climate Network (ICN [Hollinger et al., 1994]). Most ICN stations are also soil moisture monitoring sites. A scaling coefficient for each calendar month is first calculated from the actual versus potential evapotranspiration data near Champaign, as averaged during 1997–2002, and then applied identically to all other ICN stations to obtain the approximation for the observed evapotranspiration over Illinois. Figure 9 compares the 1997–2002 annual cycle difference of the CMM5 simulated from the observed evapotranspiration as well as those of concurrent soil moisture in the top 1-m layer, precipitation, and evapotranspiration minus precipitation. Clearly, the soil moisture bias is not well explained by the deficiency in either evapotranspiration or precipitation alone, but more attributable to that in evapotranspiration minus precipitation with a strong positive correlation except for June–August when crops are fully developed. More comprehensive data and analysis are needed to understand this summer breakdown of the relationship.
Figure 9. Monthly variations of the CMM5 differences (mm) from observations in evapotranspiration (ET; thin dashed), accumulated precipitation (PR; thin solid), ET minus PR (thick dashed), and soil moisture in the top 1-m layer (SM; thick solid) as averaged during 1997–2002 over Illinois.
Download figure to PowerPoint
 The CMM5 soil moisture annual cycle in Illinois, after removing the annual mean bias, has an approximate one month delay in relation to observations. This delay may likely be caused by the surface land cover difference where the measurements were taken under short grass while the CMM5 values are represented by the dominant cropland. At the planting and initial growing stages in April–May, croplands have more extensive area of bare ground and smaller leaf coverage, thus less evapotranspiration to keep more moisture in soil than grasslands. On the other hand, developed crops (mainly soybeans and corns) have deeper roots, taller canopies with more leaves, all of which produce greater evapotranspiration and thus more loss of soil moisture than grasses. These two factors collectively lead to the delayed cycle and dry biases under croplands relative to observations under short grass. Figure 10 compares the annual cycle of the CMM5 soil moisture at the top 1-m under cropland and grassland averaged over the two neighboring areas identical to those for soil temperature discussed previously (Figure 4). From March to May, soil moisture in cropland is slightly higher than that in grassland, but much lower during the rest of the year. Thus, the CMM5 result likely represents the actual soil moisture of croplands.
Figure 10. Monthly variations of soil moisture (mm) for the top 1-m layer simulated by the CMM5 between two neighboring areas of cropland (thick dashed) and grassland (thick solid) in the southwest corner of Illinois averaged during 1982–2002.
Download figure to PowerPoint
 Compared to the result of Illinois, the CMM5 performed better in Iowa, where the observed soil moisture variations during April–November are relatively weak in both the top 1-m and 2-m layers. The CMM5 captures this characteristic well but with about one month delay in fall. The Mosaic is too wet and has a larger deviation with a delay of about 1 and 2 months in the top 1-m and 2-m layers, respectively. On the other hand, the R-2 is too dry with a much smaller seasonal variation in the top 2-m layer.
 The differences between the NLDAS/Mosaic and others in the annual mean (Table 1) and the annual cycle (Figure 8) of soil moisture are not likely caused by the different averaging periods. During the common period 1997–2002, the CMM5 simulated soil moisture in Illinois has an annual mean (deviation) of 279 (21) and 561 (33) mm in the top 1-m and 2-m layer, respectively; these differences from the 1984–2002 averages are relatively small as compared with the Mosaic values. A similar conclusion is reached by comparing CMM5 averages between 1997–2002 and 1982–1994 in Iowa. The result implies that the Mosaic product contains soil moisture biases likely inherent from its LSM deficiency.
 Table 2 lists the general statistics and Figure 11 illustrates the time evolution of interannual variations in Illinois and Iowa, where the respective monthly means of the soil moisture annual cycle for both layers (Figure 8) are removed from all quantities. In Illinois, the correlation coefficients between the CMM5 and observations during 1984–2002 are 0.52 and 0.58 in the top 1-m and 2-m layers, respectively. The precipitation is realistically simulated with an even higher correlation of 0.76 for the same period. The observed and CMM5 simulated interannual deviations of soil moisture are, respectively, 14 and 21 mm in the top 1-m layer and 19 and 36 mm in the top 2-m layer, while the R-2 value in the top 2-m layer is 27 mm.
Figure 11. Interannual variations during 1982–2002 of precipitation (mm day−1) (a, d), soil moisture (mm) for top 1-m (b, e) and 2-m (c, f) for observations (OBS, thick solid) and simulations of the CMM5 (thick dashed) and R-2 (thin dashed) over Illinois (left) and Iowa (right). The respective seasonal cycles of soil moisture for both layers in Figure 8 are removed.
Download figure to PowerPoint
Table 2. Comparison of Interannual Variations of Soil Moisture (mm) Simulated by the CMM5 and R-2 as Well as Their Correlations With Observations
| ||Illinois (1984–2002 Annual)||Iowa (1982–1994 Apr–Nov)|
|Deviation, mm||Correlation||Deviation, mm||Correlation||Deviation, mm||Correlation||Deviation, mm||Correlation|
|OBS||13.9|| ||18.9|| ||22.6|| ||36.9|| |
|R-2|| || ||26.9||0.59|| || ||22.9||0.81|
 The CMM5 simulated Iowa soil moisture in the top 1-m and 2-m layers are close to observations during the 1982–1994 April–November. The R-2 soil moisture in the top 2-m layer follows the observed interannual variations well, albeit with a much drier annual bias. The difference between the R-2 and CMM5 is obvious from 1996 onward, during which no observation was available. Although the CMM5 has a high precipitation correlation (0.63) with observations, the simulation of soil moisture in Iowa is not as good as in Illinois.
 Soil moisture in the upper layer seems to be more closely related with precipitation than the lower layer. This is reflected in both the annual cycle and interannual variation. The positive precipitation biases in spring and summer may account for the less dry biases of soil moisture while negative precipitation biases in fall may cause larger dry biases (Figure 8). The interannual correlation coefficients between precipitation and soil moisture for observations and CMM5 are, respectively, 0.54 and 0.80 in the top 1-m layer, which are greater than 0.33 and 0.76 in the top 2-m layer.