a. Overview of global soil moisture by JMA-SiB GSWP

Under the Global Soil Wetness Project (GSWP; Dirmeyer 1999), the deca-day (three times per month) temporal-mean values of 2-yr (1987–88) 1° × 1° global soil moisture were produced by integrating one-way uncoupled LSMs under the observed atmospheric forcing. The atmospheric forcing data and soil and vegetation properties input to the LSMs were the products of the first International Satellite Land Surface Climatology Project Initiative I CD-ROM (ISLSCP; Sellers et al. 1995). Although the majority of LSMs simulated the seasonal cycle of soil moisture (Entin et al. 1999) or seasonal runoff pattern (Oki et al. 1999) well, no model could quantitatively reproduce the observed soil moisture (Entin et al. 1999). Since hydrology treatments such as drainage, runoff, and lateral flow differ markedly from one LSM to the next, the differences in soil moisture produced by different LSMs usually exceeded the range of the seasonal-variation of soil moisture. In order to minimize the effect of specific model parameters on soil moisture, the soil moisture in this paper is expressed in a soil wetness index (SWI) defined as

View Expanded

where W is the volumetric soil within a grid, W_lp is the soil moisture at the wilting point, and W_fc is the field capacity of the soil moisture. SWI ranges from 0 (wilting point) to 1 (field capacity), and larger values indicate a wetter state. SWI can exceed 1, or it can be negative in certain instances. Using SWI makes it possible to apply the relationship for soil moisture obtained from one specific LSM to another LSM, relatively easily, as discussed in section 3.

Of several GSWP soil moisture products, the soil moisture derived with JMA-SiB was selected for the purpose of investigating the relationship between surface and root-zone soil moisture, because JMA-SiB was the only dataset that provided both surface and root-zone soil moisture at the GSWP data center at that time (http://www.tkl.iis.u-tokyo.ac.jp:8080/DV/gswp/index.html). JMA-SiB is an SiB-type (Sellers et al. 1986) LSM that was used in the previous JMA-GCM, with some modification of the snow melting process. The hydrological process in the JMA-SiB scheme takes into account water pumping from the root-zone soil layer by vegetation, evapotranspiration from leaf stomata, and interception loss (Sato et al. 1999, manuscript submitted to J. Meteor. Soc. Japan). This scheme divides the soil into three layers: 1) the upper 5 cm; 2) from 5 cm down to root depth, according to vegetation type; and 3) deeper soil. The data used in this study were the ensemble mean values for 1987 and 1988, in order to reduce bias for a specific year. In this paper, W_s is the value in the first layer in JMA-SiB, and W_r is the integrated value of the first and second layers.

Figure 1 shows time series of W_s and W_r over a year. These are average values over the Indochina Peninsula (10°–20°N, 95°–110°E), east China (20°–30°N, 110°–125°E), India (10°–20°N, 75°–85°E), and Mississippi (0°–10°S, 60°–75°W) regions. When small fluctuations are ignored and only seasonal variation is considered, the W_s and W_r curves seem to be connected to each other. All sets of W_s and W_r for the four regions in Fig. 1 show a similar seasonal change, although the range and phase of each differ. For example, W_s in the region shown in Fig. 1a reaches a minimum in April and peaks in October. Like W_s, W_r in that region has seasonal minimum and maximums, but the minimum is in May and the maximum is in November. In the region in Fig. 1a, W_s ranges from approximately 0.30 to 0.90, while W_r ranges between 0.37 and 0.80.

Figure 2 shows the one-to-one relationship between W_s and W_r for the regions in Fig. 1, where the dots represent the average value for each deca-day in a year, and a black star indicates the first deca-day average of the year (from 1 to 10 January). Figure 1 shows clearly that the relationship between W_s and W_r forms a loop over the course of a year. The most obvious loop is seen in Fig. 2a, which has the lowest correlation coefficient (0.868). This annual loop is due to the phase difference in the time series between W_s and W_r seen in Fig. 1. In order to derive a transfer function for W_s and W_r, it is necessary to define and remove this phase difference and the consequent loop.

b. Maximum and minimum peak index method

In order to quantify the phase difference between W_s and W_r, we introduced a new index called the “maximum and minimum peak index.” First, a smoothing procedure was adopted for both spatial and temporal distribution. Each 1° × 1° soil moisture value was averaged with the eight surrounding grids for spatial smoothing. For temporal smoothing, a 5-deca-day moving average was applied. Second, the maximum and minimum values for each grid box were determined. Using this procedure, the variation in soil moisture was assumed to reach one maximum and one minimum after the smoothing procedure, as is the case in most grids after smoothing. Finally, the deca-day values for these two peaks were set to the maximum and minimum peak indices. The deca-day number specifies the deca-day in a year, and ranges from 1 (the first 10 days in January) to 36 (the last 11 days in December). Four global 1° × 1° sets of the peak index were then obtained, that is, the maximum and minimum peaks for W_s and W_r.

Figure 3 shows global maps of LAG between W_s and W_r for the maximum (Fig. 3a) and minimum (Fig. 3b) peaks. Globally, LAG is usually positive, indicating that W_s reaches its maximum or minimum peak before W_r. In most regions, LAG values are from zero to less than 3 deca-days (30 days), and are at most a few months (30–90 days). In some areas, however, LAG is negative or has too high a positive value. These areas are primarily located at high latitudes, such as in Canada, Scandinavia, northern China, and Siberia, or in areas with high elevations, such as the Rocky Mountains in the United States, the European Alps, and the Tibetan Plateau. The W_r value in these regions (not shown) is almost constant and fluctuates over only a small range due to the high moisture, supplemented from snow, and the lower evaporation ratio. This makes it difficult to define peaks for W_r. In addition, the curves for W_s over time in these regions (not shown) have several peaks corresponding to the rainy season, snowfall period, and snowmelt within a year. This makes it difficult to specify single maximum and minimum peaks of W_s in each year. We conclude that our simple algorithm for detecting the phase lag between W_s and W_r cannot be applied to these regions because of their specific characteristics. However, since the satellite that we will use in this study does not observe high latitudes, this limitation of the algorithm can be neglected. Moreover, it is not necessary to retrieve root-zone soil moisture for these regions, since it can be regarded as constant in these regions.

Figure 3 also shows that LAG values are larger at the minimum peak (Fig. 3b) than at the maximum peak (Fig. 3a). This difference in LAG may be explained by the differences in water movement characteristics, either upward or downward, in wet and dry periods.

The matric potential between the surface and root-zone layers governs water movement between the two layers. At the beginning of the wetting period, soil moisture in the root-zone increases only when the surface layer becomes wet enough to supply water to the deeper soil. Therefore, the minimum peak in the root zone, after which the soil moisture starts to increase, is delayed compared to the surface layer. At the beginning of the dry period, root-zone soil moisture is consumed mainly by evapotranspiration, which is independent of surface layer conditions. The beginning of the decrease in soil moisture is almost simultaneous in the two layers, as a result, the LAG at maximum peak is usually small.

c. Phase-shifting method

Using the LAG at the maximum and minimum peaks defined above, the W_s data were shifted. Figure 4 shows a schematic figure of the procedure. Since LAG differs for wet and dry periods, the shift time ranges were calculated for these two periods separately. First, the maximum and minimum peak times of the new time-shifted value (W^*_s) are set to be the same as those of the W_s. The shift time ranges for the wet and dry periods were then calculated by interpolating the LAGs linearly. Next, the W^*_s was then linearly interpolated from the vicinity value of the original W_s data according to the shifted time. As showed in the example in Fig. 4, if a LAG at the maximum peak is shorter than that of the minimum peak, the length of the dry period of W_s is expanded and the wet period is shortened in W^*_s.

Figure 5 is the same as in Fig. 2, but with W^*_s. After the phase-shift corrections, it is clear that the relationships between W_s and W_r can be approximated by linear regression lines with high correlations. The correlation coefficients of the regression lines in Fig. 5 are as high as 0.99 in all four regions.

Figure 6 shows the global distribution of W_r (Fig. 6a), transferred W_r (Fig. 6b), and their difference (Fig. 6c) for the first deca-day in August. Globally, W_r appears to be successfully transferred from W_s. For example, there is good agreement for the dry area in the middle of Africa, the wet and dry boundaries in Asia, and the dry area in western North America. Figure 7 shows scatterplots of W_r and W_r transferred from W_s. Each dot is the value for a 1° × 1° grid pixel. Although there are some large errors in December, W_r retrieved from W_s corresponds fairly well with the original W_r in these 3 months.

Both LAG and regression addressed here is derived from only one data source, averaged soil moisture in 1987 and 1988 of JMA-SiB; therefore, these relations are implicitly assumed to be stationary for different years or models. This assumption can not be validated at this stage where only 2-yr global soil moisture data are available. Longer global soil moisture data by GSWP-like offline experiments or longer observations may validate this assumption in future.

d. LAG behavior of the observed soil moisture in the Oklahoma Mesonet

In order to address whether the lag behavior demonstrated for JMA-SiB is also seen in observed soil moisture, soil moisture in the Oklahoma Mesonet (Robock et al. 2003, manuscript submitted to J. Geophys. Res.; Luo et al. 2003, manuscript submitted to J. Geophys. Res.) is analyzed. Soil moisture at four different depths (5, 25, 60, and 75 cm below the surface) are observed every 30 min at more than 72 stations around the state of Oklahoma and archived. Observation data in 1998 at 52 sites are selected in terms of the data availability, averaged in deca-day, and the SWI is then estimated from the volumetric percentage by an equation of Clapp and Hornberger (1978) based on soil types. Figure 8 shows the time series of SWI in each soil layer at Boise City (left) and Ketchum Ranch site (right). As in the case of JMA-SiB data, the surface layer varies earlier than the deeper layer and their seasonal changes are similar. Both sites show several peaks during the year and the simple maximum and minimum peaks are not able to be picked up, indicating that the peak index method suggested here is not directly applicable to point observation data.

After using a 5-deca-day moving average for temporal smoothing, the number of peaks greater than two still remain while small fluctuations disappear (not shown). This inconsistency is mainly due to the difference of the scale between the large grid area of JMA-SiB and the point observation.

Defining the largest maximum peak and the smallest minimum peak of temporal smoothed data as the seasonal maximum and minimum peaks, respectively, LAGs for maximum and minimum peaks are calculated. Figure 9 shows the LAG between the first (top) layer (5 cm) and the fourth (75 cm) layer. Both maximum (left) and minimum (right) LAGs show that the first layer varies either than the fourth layer at most of the points; however, unlike the previous results of JMA-SiB, LAG range is longer in maximum peak than in minimum peak.

In short, four results are addressed for the seasonal variation (deca-day) of soil moisture observation (comments in brackets explain the comparison to the previous conclusions):

1. Surface layer varies earlier than the deeper layer (consistent).

2. Variation curves are similar among different layers (consistent).

3. There are several peaks in a year (inconsistent).

4. LAG of maximum peak is longer than the minimum peak (inconsistent).

a. Soil moisture observed by TRMM/PR

A precipitation radar sensor (PR) was launched on board the Tropical Rainfall Measuring Mission (TRMM) satellite in November 1997. Since the major objective of TRMM/PR is to estimate the tropical rain rate by bouncing microwaves off raindrops to observe the backscatter, TRMM/PR covers most of the globe from 36°N to 36°S. The footprint of TRMM/PR is approximately 4 km at the nadir angle, and the maximum incident angle of TRMM/PR is 17° across its track, with a frequency of 13.8 GHz. Each ground point is observed approximately 30 times per month, so that each ground point is observed once or twice a month at a particular incident angle (Kummerow et al. 1998).

Using the backscatter from the land surface from TRMM/PR, Oki et al. (2000) and Seto (2003) produced a global soil moisture dataset. Their algorithm calculates the land surface soil moisture as a volumetric percentage from the backscatter coefficient (σ₀) from land by combining different values of σ₀ observed at different incident angles, and by excluding vegetation effects with leaf area index (LAI) information. Using the integrated equation model of Fung (1994) to relate volumetric soil moisture and incident angle with the reflection coefficient, the monthly mean soil moisture distribution was calculated with a spatial resolution of 1° longitude × 1° latitude (Oki et al. 2000; Seto 2003).

Since microwaves penetrate the ground to a depth of 5 cm or less (Schmugge 1983), these data reflect the physical state in only the first few centimeters. As the sampling depth of a 10.0-GHz wave is from 0.5 to 5 cm in wet to dry soil, respectively (Cihlar and Ulaby 1974), the 13.8-GHz frequency information obtained by TRMM/PR is regarded as representing the first few centimeters of the soil layer. Since the depth of the first layer of JMA-SiB is similar to the penetration depth of TRMM/PR, the surface soil moisture distribution observed by TRMM/PR can be considered as corresponding to the W_s of JMA-SiB. The TRMM-derived soil moisture data used in this paper is a version that used 2-yr (1987/88) averaged JMA-SiB surface data (W_s) to relate the seasonal anomaly of σ₀ of TRMM/PR to seasonal anomaly of SWI.

As shown in Fig. 10, ranges of W_s and TRMM-derived soil moisture in 1998 (T98) are similar. There are considerable difference between W_s and T98, however. Possible reasons of this difference are 1) the data year inconsistency (average of 1987/88 for W_s and 1998 for T98) and 2) difference of model-derived soil moisture and satellite-oriented soil moisture. Figure 11 shows scatterplots of T98 to W_s and T98 to T99 (TRMM-derived soil moisture in 1999). From Fig. 11, difference between W_s and T98 (left) is much larger than the other (right), indicating that the difference due to the data source is larger than the difference of the data year.

In order to focus on the effectiveness of the transfer method of soil moisture used in this paper, comparisons of GCM simulations with model-derived soil moisture, satellite-derived soil moisture, and without soil moisture boundary are not addressed. Only two different GCM simulations with satellite-derived soil moisture with and without the transfer procedure are discussed instead.

b. Soil moisture transferability between different land surface models

As mentioned in section 2, the differences in soil moisture determined by different LSMs are large, even in cases when they are obtained by offline simulation using the same parameters and the same forcing. Since the transfer method investigated in this study was obtained from the soil moisture estimate using JMA-SiB, that method and the obtained root-zone soil moisture cannot be directly applied to the CCSR/NIES bucket model, which is the LSM coupled to the CCSR/NIES AGCM. Ideally, it is useful to investigate the effect of inputting satellite-derived soil moisture to the AGCM with a multilayer SiB-type LSM; however, because of the limitations of available models, the transferred root-zone soil moisture of TRMM/PR was converted to the value of the CCSR/NIES bucket model.

Figure 12 compares the root-zone soil moisture determined by JMA-SiB (W_r) and the soil moisture estimate of the CCSR/NIES bucket model obtained under the GSWP (hereafter referred to as W^ccsr_r). As with W_r, W^ccsr_r is an ensemble mean of 1987 and 1988 data. Comparison of these values showed that the difference due to different LSMs could be corrected globally using linear functions in each grid. In Fig. 12, high correlation coefficients between W_r and W^ccsr_r (exceeding 0.95) are seen in all regions. These relationships between W_r and W^ccsr_r are not surprising, because these two LSMs were forced with the same climatological data under the GSWP. Koster and Milly (1997) showed that the responses of ideal LSMs to the same atmospheric forcing are the same when their responses are normalized with their active soil moisture ranges. Our comparison of W_r and W^ccsr_r confirmed their results with the finding that the deca-day-averaged soil moisture normalized as SWI shows a fairly good linear relationship between two different LSMs under the same atmospheric forcing.

In highly vegetated regions, such as the Indochina Peninsula, however, the relationship between W_r and W^ccsr_r is not perfectly linear, probably because of the different parameterizations of the vegetation effect in the two models, such as transpiration and interception loss. Assuming that the errors in these regions are small enough to neglect, simple linear functions between W_r and W^ccsr_r obtained for each grid are used as the transfer function from W_r to W^ccsr_r.

c. Model and experiment design

The AGCM used in this study is the CCSR/NIES AGCM (Numaguti et al. 1997) at a spectral resolution of T42 (approximately 2.8°) and 20 vertical levels. The LSM coupled to the GCM is the CCSR/NIES bucket model. The surface water budget of this model is computed using the bucket method (Manabe 1969).

The experimental period was the 3 months of the 1998 boreal summer (JJA). In these 3 months, the snowmelt effects in the northern continents are smallest and many regions have a rainy season. Integration was conducted in ensembles of nine runs, which were initialized at 0000 UTC on 23–31 May and integrated forward through the end of 31 August.

Two sets of ensembles with specified soil moisture were conducted. One ensemble used the original TRMM/PR monthly soil moisture (Oki et al. 2000; Seto 2003) as the boundary condition, as if the value was representative of the state of deeper soil. The volumetric soil moisture value was the SWI of the TRMM/PR data multiplied by the depth of the CCSR/NIES bucket (20 cm). This ensemble is referred to as T98 in this paper. The other ensemble specified the integrated total soil moisture for the CCSR/NIES bucket, which was transferred from T98 using the new algorithm. The latter ensemble with transferred satellite soil moisture data is referred to as N98. First, the transfer function for phase shift and linear transformation in section 2 (obtains W_r from W_s via W^*_s) was applied to T98, assuming that neither the time lag nor the regression line change interannually, and that the depth of T98 was the same as that of W_s. The root-zone value for T98 in the preliminary step was then transferred to the value that is representative of the CCSR/NIES bucket by linear transformation (obtain W^ccsr_r from W_r). All these procedures used to transfer soil moisture from T98 to N98 are shown schematically in Fig. 13, where f(x, y, t) is the phase shift and linear transformation for transferring W_s to W_r, and g(x, y) is the unique linear function relating W_r and W^ccsr_r. Figure 10 shows the region-averaged value of T98 and N98. In Indochina and India (Figs. 10a,c), variance ranges of N98 become larger than T98 and peak timings are shifted. The relatively small difference found in East China and Mississippi (Figs. 10b,d) indicate that the large difference between T98 and W_s may make the transfer transformation useless.

In both experiments, the CCSR/NIES GSWP 2-yr ensemble mean was used for boundary soil moisture in high latitudes not covered by the orbit of TRMM/PR. All the other conditions were the same between these two ensembles, where climatic conditions were initialized with National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) daily reanalysis (Kalnay et al. 1996), and the monthly SST data observed by NCEP–NCAR (Reynolds 1988) were used as the boundary condition over ocean. The ensemble names, integration periods, the SST boundary condition used, and the soil moisture boundary conditions used are summarized in Table 1.

d. Results

The simulated seasonal (JJA) total precipitation in the two experiments were compared. In general, the agreement between the observed and simulated precipitation in the GCM is not perfect because of the bias in the model; comparing the errors of the two sets of comparisons with the observations of T98 and N98 clearly shows the differences between two GCM simulations using different soil moisture boundaries.

Figure 14 compares the errors in the 3-month precipitation simulated in experiments T98 and N98. Dots indicate error values in the terrestrial GCM grid. The precipitation errors were calculated by subtracting the precipitation observed by the Global Precipitation Climatology Center (GPCP; Huffman et al. 1997). Although large precipitation errors are seen in both experiments, sometimes exceeding 1000 mm, the precipitation error is less in N98 than in T98. For example, points with a large overestimation in T98 are reduced in N98.

Figure 15 maps where the precipitation error was lower in the two experiments. Light gray indicates where the precipitation error is lower in N98, and dark gray indicates where T98 has the better result. Regions with a small precipitation error (below 50 mm) in both are masked on the map. The smaller precipitation error in N98 was obvious over land in tropical areas, such as most of South America and Africa and suggests the impact of better soil moisture information. Figure 16 shows the averaged difference in soil moisture content in the two experiments. There are large-scale differences in southern Africa, most of South America, India, and northeastern Australia. From these figures, it is obvious that the regions where the improved precipitation is seen in Fig. 15 correspond to the regions where a large difference in soil moisture is seen in Fig. 16. This suggests that the soil moisture in N98 for these areas is appropriate, as compared to that in T98. Over a wide area of Southeast Asia, however, the precipitation error in N98 was larger than in T98. One possible reason for this is that the effect of the land surface is comparatively small here because, in the main, a large circulation induced by sea surface conditions dominantly governs precipitation in this region. Moreover, since the impact of the land surface was small during the study period, due to the strong monsoon wind (Kanae et al. 2001), an experiment examining the impact of surface soil moisture in other months is needed for further discussions.