1. Introduction

Most weather and climate models consist of at least two complete component models coupled together: an atmospheric general circulation model (GCM) and a land surface scheme (LSS). The coupling may be accomplished in many ways, but usually the GCM and LSS exchange information regarding the values of state variables at the interface, as well as fluxes of energy and water. Information is usually exchanged at a time interval corresponding to the integration time step of the atmospheric model, which is typically on the order of minutes. This allows the simulation of the atmosphere and land surface to evolve in tandem with feedbacks between the two systems.

All numerical weather predictions or climate simulations must have a starting point. One of the most challenging problems is accurately determining the proper initial condition from which to begin a numerical integration. Operational weather prediction models require timely analyses of the ever-evolving state of the global atmosphere, which are created from a wide array of in situ and remotely sensed observations from a number of sources around the world that are assimilated into a GCM to produce a single, consistent representation of the global atmosphere at a particular point in time. A series of these analyses represent the best historical estimate of the evolution of the global atmosphere.

Real-time analyses of the land surface state (distributions of soil wetness, temperature, snow cover, and depth, and large-scale properties of the vegetation) have historically received less emphasis. There are two main reasons. First, there is a dearth of useful data collected on the land surface, and most data that are collected are not disseminated due to the lack of a well-developed real-time global network like that which exists for atmospheric observations. Also, many land surface characteristics vary on a range of scales, down to the order of meters (Kabat et al. 1997). Hundreds of measurements of soil wetness or snow water equivalent may be necessary to arrive at an accurate, representative estimation at the resolution of the models (typically 10²–10⁵ km²), which is typically near to the characteristic spatial scale of soil-moisture variations on monthly seasonal timescales (Vinnikov et al. 1996; Entin 1998). Second, most emphasis has historically been on the needs of short-term weather forecasting, where the initial atmospheric condition is the major determinant of the evolution of weather over periods of hours to a few days. For medium-range weather forecasting and climate simulation and prediction, the land surface and ocean boundary conditions play a more important role (U.S. National Research Council 1994).

Thus, modelers have struggled to develop adequate datasets of land surface conditions to initialize LSSs in weather and climate models. In addition, some land surface parameters are often specified or updated throughout an integration to prevent the coupled land–atmosphere system from drifting to an unrealistic climate. In particular, soil wetness has been shown to be important (e.g., Shukla and Mintz 1982; Dirmeyer 1994). The seasonal variations in soil wetness are large over many areas of the globe—use of a climatological annual cycle of realistic soil wetness to furnish initial conditions is much preferred to no initialization at all. There have been attempts to produce such datasets, typically derived from the application of simplified evaporation formulations to observed data (Mintz and Serafini 1981; Mintz and Walker 1993; Liston et al. 1993) There is also interannual variability in soil wetness. Huang et al. (1996) produced a multiyear dataset over the United States based on observational data and a simple evaporation formulation. More recently, data assimilation methods have been developed where an LSS driven by observed or analyzed atmospheric conditions can be integrated to produce time series of soil wetness (Bouttier et al. 1993; Houser et al. 1998).

To address the lack of consistent, high-quality datasets of soil wetness, and to understand how different LSSs determine soil wetness, the Global Soil Wetness Project (GSWP) was created (Dirmeyer et al. 1999). GSWP is an ongoing modeling activity of the International Satellite Land-Surface Climatology Project (ISLSCP), a contributing project of the Global Energy and Water Cycle Experiment. In the pilot phase of GSWP, a two-year global dataset of soil wetness, temperature, surface fluxes, and other land surface variables was produced using a dozen different LSSs and a common set of soil and vegetation parameters and meteorological forcing: the ISLSCP Initiative I dataset (Meeson et al. 1995). The land surface datasets produced by GSWP represent the current state of the art in global quality and consistency for quantities such as soil wetness, although there is still room for improvement.

These soil wetness data were used as boundary conditions in a set of seasonal climate simulations with the Center for Ocean–Land–Atmosphere Studies (COLA), GCM, and LSS, and compared to simulations from COLA’s dynamical seasonal predictability hindcasts (DSP; Shukla et al. 2000). The results suggested that the GSWP soil wetness data improved the simulation of boreal summer climate when applied as a specified boundary condition, and that interannual variability of soil wetness was an important factor in determining regional climate variations in many locations (Dirmeyer 1999). However, that experiment used small ensembles of integrations (4 members each), and applied a GSWP soil wetness product that, although produced with the same LSS as used in the COLA climate model, used very different distributions of soil and vegetation properties than those used in the DSP hindcasts. These model inconsistencies and small sample sizes cast some doubt on the significance of the results.

In this paper, the problem addressed by Dirmeyer (1999) is reexamined with a larger and more consistently developed set of experiments. The following two questions are raised.

• Does the inclusion of GSWP soil wetness boundary conditions in a coupled GCM–LSS model significantly improve the simulation of climate?

• Does the interannual variability in a multiyear soil wetness dataset contribute to interannual variability in climate simulations?

Section 2 describes the structure of the experiments in greater detail. The impact of GSWP soil wetness boundary conditions is examined in section 3. Section 4 appraises the impact of the interannual variability in the GSWP dataset on the simulation of climate. Conclusions are presented in section 5.

2. Experiment structure

The global atmospheric and land surface models are briefly described in this section. Also, the production and application of the GSWP soil wetness data are reviewed and the framework of the numerical integrations is detailed.

3. Impact on mean climate

The difference in soil wetness as specified from the GSWP offline simulations, and the soil wetness that evolves freely in the coupled land–atmosphere model can be quite large. Figure 5 contrasts the evolution in soil wetness from May to August in the control GCM integration with the difference stipulated from the GSWP simulation. In this figure, and throughout this section, results from the 1987 and 1988 ensembles are averaged together. In both cases, there are some similar features: the signature of drying in the Northern Hemisphere mid- and high-latitudes, drying after the rainy season in the Southern Hemisphere subtropics and Tropics, and increased soil wetness over the Northern Hemisphere monsoon regions of southern and eastern Asia, the Sahel, and Mesoamerica.

However, the difference between the control and GSWP integrations shows that the free-running climate model shows excessive drying almost everywhere that drying occurs. In addition, soil wetness over much of monsoon Asia does not increase as much as in the GSWP specification. These biases are consistent with systematic errors in the COLA GCM radiation, which exhibits too little cloud cover and too much absorbed solar radiation at the ground (DeWitt and Schneider 1996). Soil wetness increases too much over most of the remainder of monsoon Asia, as well as the other parts of the Tropics where the rainy season is underway. This reflects the systematic errors in rainfall shown in Fig. 1. Thus, the character of the unregulated coupled land–atmosphere model is to simulate the correct patterns of soil wetness change during boreal summer, but with a bias toward lower soil moisture.

Figure 6 shows the spatial distribution of the June–August mean differences in precipitation, near-surface air temperature, and soil wetness between the control and GSWP cases. The differences in temperature are highly correlated with the changes in soil wetness between cases. The correlation coefficient between the two is −0.66 over land. There appears to be a strong relationship between wetter–drier soils and cooler–warmer soils. One would expect where radiation is abundant in the Tropics and summer hemisphere, that the addition of water to the soil would retard warming by increasing evapotranspiration and soil heat capacity.

The correlation over land between changes in rainfall and soil wetness is also significant, but lower, at +0.24. Nonetheless, large-scale features are well correlated. For instance, the negative differences in precipitation throughout the mid- and high-latitudes of the Northern Hemisphere correspond to lower soil wetness there. There is also a good match over southern Asia. There is inherently a great deal of small-scale structure in model precipitation fields that may mask some of the correspondence. If a 25-point smoothing is applied to the precipitation and soil wetness difference fields on the 1.8° lat × 2.8° long grid, the correlation coefficient increases to +0.34. The correlation of the large-scale fields is consistent with Vinnikov et al. (1996) and Entin (1998).

So we come to the first question posed in the introduction: Do the GSWP soil wetness data improve the simulation of climate when applied as a boundary condition? Table 2 shows the correlations and rms errors in the simulation of precipitation and near-surface air temperature for selected regions of the globe in the control and GSWP cases. As a means to establish the significance of the impact of GSWP soil wetness, we estimate the inherent noise in the climate model system. This is found by calculating the variance of the differences in the correlations or rms errors among each of the nine pairs of integrations in each ensemble. The intra-ensemble variances for the control and GSWP ensembles are then averaged together, and the square root is taken to arrive at an estimate of the noise we can expect—the standard deviation calculated from the internal variability of the coupled land–atmosphere model. The signal is simply the difference in ensemble means (GSWP minus control) in either correlation with observations, or rms error. A ratio of 2 corresponds to a significance of nearly 98% on a single tail. This approach is also used to provide the significance shading in Fig. 6.

The results suggest that there is a net improvement, but it is not universal. For precipitation, there are significant improvements in the correlations with observations globally (57°S–75°N, land grid cells only). Africa and Australia show deteriorations in correlation. There is an especially large improvement over southern Asia, and smaller notable improvements over North America and Europe. However, the rms error in precipitation is 0%–15% worse over most areas; only over South Asia is there a marked improvement.

For near-surface air temperature, reductions in rms errors are widely apparent. The rms errors are decreased globally and over every region, from 3% to 23%. Correlations between model and observed temperatures are already very high in the control case, exceeding 0.9 in every region considered. Nevertheless, there are notable improvements globally, and over North America and Africa.

The response to changes in soil wetness appear to be more robust in the temperature field than in precipitation. Soil wetness directly affects surface temperature through changes in the partitioning of net radiant energy between latent and sensible heat fluxes, and changes in the soil heat capacity. Surface temperature exerts a strong control on the near-surface air temperature. The connection from soil wetness to precipitation is more tenuous, involving impacts of both surface heat and moisture fluxes, and is modulated by the character of the atmospheric boundary layer and the general circulation. In fact, it has been shown that land surface changes may induce either increased or decreased moisture flux convergence in the overlying atmosphere, depending on the relative strengths of the perturbations in latent and sensible heat flux, and the overlying circulation (e.g., Shukla and Mintz 1982; Dirmeyer and Shukla 1994, 1996).

A clearer picture emerges when the location of the improvements is investigated, particularly when the grid cells where errors are relatively small in either case are eliminated. Table 3 shows the number of grid cells where precipitation and near-surface air temperature errors were reduced. Various limits are applied to mask out grid cells where the errors lie below a certain threshold. Thresholds are set at 0, 1, 2, 4, and 6 mm day⁻¹ for precipitation, and at 0, 2, 4, 6, and 8 K for temperature. For temperature, 63% of the land grid cells between 57°S and 75°N experience reduced errors during June–August when GSWP soil wetness is specified throughout the integration. That percentage increases to 74% when only cells where the temperature error exceeds 2 K are considered. The percentage tapers off slowly as the threshold is increased, but remains above the level when all points are considered. Figure 7b shows the spatial distribution of the points for the 2 K threshold. Large regions of Siberia, Canada, India, South America, and the central Sahara exhibit a reduction in surface temperature errors. Conterminous areas over Mongolia and Greenland show no improvement in their large errors.

Again, the precipitation response is less robust. When all land points are considered, there are actually more points experiencing an increase in precipitation errors than a decrease. However, as the minimum error threshold is raised, the number of grid cells experiencing a reduction of error quickly jumps above 50%, and peaks just under 60% for a threshold of 4 mm day⁻¹. Figure 7a shows the spatial distribution for a threshold of 2 mm day⁻¹. Most of the areas of improvement lie in Asia, particularly in the southern and eastern regions. The Tropics seem to be least remedied by the use of GSWP soil wetness.

4. Interannual variability

A comparison between the simulations of 1988 and 1987 reveals how the models respond to interannual variability in surface boundary conditions. Specifically, the simulated differences (1988 minus 1987) can be compared to observations to see how well the atmospheric response to year to year variations in SST and soil wetness is reproduced. In this section, the case where soil wetness is specified from the opposite year, that is, swapped with respect to the specified SSTs, is also examined. Thus, the second question is addressed:Does the interannual variability in a multiyear soil wetness dataset contribute to interannual variability in climate simulations?

Figure 8 shows the 1988 minus 1987 differences in precipitation during the three-month period from June to August. The upper left panel shows the observed precipitation differences with values over ocean masked out for clarity. The upper right panel shows the same difference field for the control simulation of the fully coupled land–atmosphere climate model. The lower left panel represents the difference when soil wetness is specified per the GSWP dataset. The last panel is from the case where GSWP soil wetness from the opposite year are specified over land. The first thing to notice is that all panels are similar in their representation of positive differences over much of tropical Africa, northern South America, and India, with negative differences over much of China, Southeast Asia, and Europe. The second conspicuous feature is that overall, the three model simulations are more similar to each other than any are to observations. This is a common feature of climate models.

Differences between the control and GSWP cases are small, but generally such that the GSWP case has better simulations of the differences. This is evident from Table 4, which shows the impact on the rms error and correlation with observations of the interannual difference by specification of GSWP and swapped soil wetness. Here again, signal to noise is estimated and used to determine the significance of GSWP soil wetness impacts. From the table, it is evident that GSWP soil moisture contributes to a reduction in rms error almost everywhere, including a global reduction of more than 12% for precipitation. The rms error reductions are particularly significant over North and South America. Correlation coefficients for precipitation are greatly increased over South America and southern Asia when GSWP soil wetness is used. Correlations are degraded over Africa, Europe, and northern Asia, where correlations were near zero to begin with. Global correlation increases from 0.26 to 0.33 when GSWP soil wetness boundary conditions are specified, and drops to 0.19 when specified soil wetness is swapped between the two years. The degradation is particularly strong over northern Asia, Australia, and North America. The impact over North America is especially visible in Fig. 8. Again, small-scale variations near the scale of the grid resolution degrade the correlations somewhat. If a 25-point smoothing is applied to the fields, the correlations increase to 0.36 for the control case, 0.48 for the GSWP case, and 0.31 for the swapped case.

Figure 9 presents a similar comparison for near-surface air temperature. The most striking trait is that the magnitude of the interannual variation in temperature is much larger in the observations than in the model simulations. This underrepresentation of temperature variability may be due in part to the soil temperature scheme used in SSiB: a two-layer force restore approach. Also, SSiB has a relatively shallow soil column for moisture storage. A heat-diffusion approach with collocated soil layers for heat and moisture, and a deeper total soil column would be physically more realistic, and probably provide the land surface with more realistic thermal inertia. Such a scheme will be implemented in future versions of the LSS. Ensemble averaging and GCM physics may also play a role in the weak variability seen in these simulations.

The lower half of Table 3 shows the global and regional correlation coefficients and rms errors for the model’s simulation of interannual temperature variability. Globally, the use of GSWP soil wetness improved the correlation from 0.22 to 0.33, and use of the swapped soil wetness reduces the correlation to 0.20. The increase in correlation is strongest over southern Asia, North America, South America, and Europe. As with precipitation, a 25-point smoothing serves to increase these correlations by approximately 0.1 in each case. GSWP soil wetness leads to significant improvements in correlations over all regions except Australia and northern Asia, whereas the swapped soil wetness actually causes improved simulation in those two regions while degrading correlations elsewhere, particularly over Europe and southern Asia. Use of GSWP soil wetness also reduces rms error everywhere except Australia.

In the mean (1987 and 1988 averaged together), even the swapped soil wetness case produces improvements in the simulation of precipitation and temperature comparable to those found for the GSWP case. Table 5 is like Table 2, but compares the swapped case rather than the GSWP case. The changes are remarkably similar between the two tables. It should be noted that on balance, the specified soil wetness improves rms errors in precipitation and temperature, regardless of whether data from the right or wrong year is used. This suggests that the climate of the fully coupled land–atmosphere system is drifting significantly relative to the cases where soil wetness is specified. However, interannual variation in the specified soil wetness seems to be important for the simulation of patterns of precipitation and temperature, as measured by the correlation coefficients. Correlations are improved globally when data from the right year is used, but diminish when data from the wrong year is applied. An analysis of the number of grid cells where rms error is reduced by use of the swapped soil wetness also produces similar results to those depicted in Table 3.

This apparent contradiction can be understood by comparing the difference between 1988 and 1987 GSWP soil wetness for the summer season (Fig. 10, also showing interannual differences in prescribed SST) with the soil wetness difference between GSWP and the control simulation (Fig. 6). The interannual variations are small and localized compared to the differences between the GSWP dataset and the soil wetness predicted by the coupled climate model. However, some of the local interannual variations are well positioned to affect the simulation of climate in those regions. Figure 11 shows the relative contributions of the interannual variations of ocean (SST) and land (soil wetness) to the simulated differences in precipitation and near-surface air temperature. These values are estimated by considering the two-by-two grid of ensembles as indicated in Table 1. There are only two aspects that are different in the simulations of 1988 versus 1987: SST and soil wetness. Initial atmospheric conditions are also different between the two years, specified from analyses from the appropriate year. However, the three-month period considered is a full month after the beginning of the integrations, and probably adequately remote so that the initial atmospheric conditions have no systematic effect on the later part of the simulation.

The relative signal contributed by variations in ocean and land conditions may be quantified as:

View Expanded

The above relationship is an approximation, based on the assumption that the mean effects of ocean and land can be separated in a linear fashion. Figure 11 shows these relative contributions for precipitation and temperature. The ocean signal in precipitation is stronger than the land signal, with a much greater area of statistical significance, but the two signals are largely collocated across the Tropics. The global rms (land points only) of the ocean signal is 0.99 mm day⁻¹, while the land signal rms is 0.37 mm day⁻¹. For temperature, the ocean signal is much more widespread than the land signal that is predominantly limited to India and central North America. These two locations are also located in the two regions where there was the greatest increase in correlation and largest decrease in rms error in Table 3. The global temperature rms of the ocean signal is 0.57 K, while the land signal rms is 0.30 K. So the response of the climate model to interannual ocean temperature variability is generally larger than the response to variability in soil wetness, but in key areas the variations in soil wetness can have important impacts on regional climate.

5. Conclusions

Ensembles of boreal summer integrations of a coupled land–atmosphere climate model are examined to determine whether the use of GSWP soil wetness as a specified boundary condition improves the simulation of climate (specifically precipitation and near-surface air temperature). The interannual differences between 1987 and 1988 in the GSWP soil wetness data have also been examined to determine whether they are large enough to impact the simulation of climate and its interannual variability.

GSWP soil wetness does improve the simulation of the pattern precipitation (as measured by correlation to observations) globally and regionally. Improvements are particularly large over monsoonal Asia. The simulation of surface temperature is also improved. The rms errors in temperature are reduced globally by 9% and over every region examined. The largest reductions in rms error occurred over Europe (23%) and North America (18%). Approximately two-thirds of the land grid cells between 57°S and 75°N experienced a reduction in rms error, including nearly three-quarters of points that had an error of 2 K or greater.

Even though the interannual variability in soil wetness in the GSWP dataset is small compared to the soil wetness differences between GSWP and control cases of the coupled climate model, they do have a definite impact on the climate simulations. The simulation of 1988 minus 1987 boreal summer differences in precipitation and near-surface air temperature are improved by the specification of GSWP soil wetness throughout the season. Globally and over most regions, correlations with observed 1988 minus 1987 differences is improved. Likewise, using GSWP soil wetness from the wrong year (swapped case) degrades correlations over most regions. A similar improvement is seen in the rms errors when GSWP soil wetness is used, although the swapped case usually maintains rms errors lower than the control case but higher than the GSWP case. This fact suggests that robust positive feedbacks between land and atmosphere components of the fully coupled climate model may drive precipitation and temperature away from observed values to an unrealistic climate balance. This is essentially a climate drift not unlike that often observed in coupled ocean–atmosphere model systems. The imposition of GSWP soil wetness every five days acts in the same manner as a flux correction, offsetting unrealistic trends in the LSS and GCM.

These results are consistent with those of Dirmeyer (1999), which had a smaller ensemble size (four vs nine members). The larger ensembles increases the reliability of the results by contributing to statistical stability. There is a positive impact on the simulation of climate when GSWP soil wetness calculated by the same LSS driven by observed meteorological forcings is applied as a boundary condition. In addition, there is a positive contribution to simulation from the interannually varying component of soil wetness. A longer time series of soil wetness calculated in the GSWP framework would be particularly useful for climate research.