1. Introduction
General circulation models (GCMs) constrain the surface moisture flux between land and atmosphere with some representation of soil moisture. Manabe (1969) modified evaporation using a simple ratio between soil moisture and a critical soil moisture (0.75 of the field capacity). Since this rather simple approach for representing soil moisture control of evaporation, a number of models with different levels of complexity have evolved, including Milly and Shmakin (2002), Dai et al. (2003), and Essery et al. (2003).
In recent years there has been an increased interest in the representation of soil moisture in global climate models. The Global Land–Atmosphere Coupling Experiment (GLACE; Koster et al. 2004), which used a number of GCMs to create a multimodel ensemble, identified regions of the world where there was a significant feedback between soil moisture and precipitation. Koster et al. (2004) found a large variation between models in the strength of the feedback.
Guo et al. (2006) analyzed the GLACE output further to determine if the model variability was due to the atmospheric or land surface component and identified the effect of soil moisture on evaporation as the dominant factor in explaining these differences. Lawrence and Slingo (2005) further showed that the definition of the soil parameters within the third Hadley Centre Atmosphere Model (HadAM3) affected the land surface atmosphere coupling strength.
There have also been a number of studies that show the importance of soil moisture in accurately modeling atmospheric processes. Hennessy et al. (1998) found that the climate predictions for Australia were influenced by the level of complexity in the representation of soil processes. Koster and Milly (1997) analyzed the land surface schemes from the Project for the Intercomparison of Land-Surface Parameterization Schemes (PILPS; offline runs with identical forcing data) and found that, of all the processes represented within the models, the partitioning of throughfall into evaporation and runoff explained the majority of the intermodel variation in evapotranspiration. Gedney et al. (2000) extended this idea by characterizing a number of land surface schemes (coupled to their respective GCMs) and found, once again, that the surface hydrology dominated all other aspects of the land surface schemes.
The soil parameters in most GCMs—including HadAM3—are calculated using pedotransfer functions to relate the hydraulic properties to soil texture (Clapp and Hornberger 1978; van Genuchten 1980; Cosby et al. 1984). These functions are usually applied to a global distribution of soil textures based on the Food and Agriculture Organization (FAO) soil map (FAO 2003). The global application of a single transfer function has been found to poorly represent some regions. For instance, Tomasella and Hodnett (1998) found that transfer functions based on midlatitude soils poorly predicted soil water content when applied to Brazilian Amazonia.
Another problem when undertaking global land modeling experiments is assessing the results of different models and parameterization schemes, because comparable observations of hydrological variables are not available. One method used is to compare the modeled runoff with river flow data, as shown by Zhang et al. (1999), among others, who compared the Simple Biosphere Model with various river data.
The aim of this work is to identify a methodology to assess objectively the effect of different soil parameter sets and model parameterizations on global soil moisture stress patterns. Spaceborne radiometry vegetation data and a climatological dataset are used to determine the temporal evolution of soil moisture stress in vegetation. The advantage of using vegetation stress, as opposed to a more direct measure of soil moisture, is that this provides information over the depth of the root zone, whereas microwave products only represent the top few centimeters and also have issues in heavily vegetated regions. Global distributions of the observed stress are compared to the modeled soil moisture stress from a number of model integrations.
Sections 2 and 3 will outline the method developed to compare correlations of normalized difference vegetation index (NDVI) and precipitation with modeled soil moisture stress, and describe the datasets used. In section 4 the output from the new methodology will be presented and analyzed. In section 5 more in-depth analysis of the soil moisture patterns will be presented, with geographical regions of interest being identified.
2. Modeling experiment
The Joint U.K. Land Environment Simulator (JULES) is based on the second version of the Met Office Surface Exchange Scheme (MOSES2; Essery et al. 2003) used in the Met Office atmospheric model (Unified Model). JULES is a tiled or mosaic model with a grid box divided into fractions of nine surface types, each with their own surface energy balance. There are five plant functional types (PFTs): broadleaf, needleleaf, C3 grasses, C4 grasses, and shrub; the nonvegetated surface is represented by urban, lake, bare soil, or permanent ice.
All the runs for this work used JULES uncoupled from an atmospheric model and were forced with the meteorological driving data from the second Global Soil Wetness Project (GSWP2; Dirmeyer et al. 2005), which includes surface pressure, downward shortwave radiation, downward longwave radiation, wind speed, liquid precipitation, solid precipitation, near-surface temperature, and near-surface specific humidity.
All the data were provided on land points at a horizontal resolution of 1°, with a range of 59°S–89.5°N and 179.5°W–179.5°E. This corresponds to 15 238 land points. The temporal resolution was three hours, and the data spanned the period 1982–95. The years 1982–85 were used as spinup, leaving the data from 10-yr period 1986–95 to be used for the analysis.
Table 1 presents the model integrations undertaken during this investigation. The Met Office soil (MOsoil) parameter set was created by identifying the dominant FAO soil type and applying a variation of the Cosby et al. (1984) transfer function. The MOsoil parameters were used for a number of years in the Hadley Centre GCM. Cosby soil (COsoil) parameters were created by applying the function from Cosby et al. (1984) to a 1° soil texture map created by the International Satellite Land Surface Climatology Project (ISLSCP; Zobler 1999).
The van Genuchten (VanGen) runs used the equations from van Genuchten (1980) to calculate the hydraulic conductivity and soil suction; the parameters for the equations were created by identifying the dominant FAO soil type within a 1° grid box and using the parameters provided by the dataset for that soil type.
Two runs were used to assess the sensitivity of the modeled soil moisture stress to seasonally varying LAI. For the seasonal vegetation experiments, the fraction of absorbed photosynthetically active radiation (fAPAR) dataset of Los et al. (2000) was used to create a seasonal cycle of LAI. In MeanLai, rather than have a seasonally varying values for LAI, one value was used for each grid box. This was calculated by taking the average fAPAR over the whole of the 17-yr dataset. This value remained a fixed value for every month of the 10-yr model run. All the runs represented the surface as gridbox fractions of five plant functional types and four nonvegetated surface types.
In the final simulation, the sensitivity to different parameterizations of soil hydrology is assessed. In other studies, including Dirmeyer (1994), it has been observed that incorrect evolution of the soil moisture and excessive runoff during the spring thaw can lead to unrealistically low soil moisture and subsequent vegetation stress. In the standard version of JULES, the soil moisture flux is solved from the bottom up; that is, at each level supersaturation is checked for and water greater than the saturated value is pushed to the layer above. In some cases where frozen soils are thawing, this results in an unrealistic fraction of the soil moisture being lost to surface runoff.
To assess whether this was restricting the ability of the model to predict the seasonal evolution of soil moisture stress, a further run was undertaken. In this run (Pushdown), the soil moisture fluxes were resolved from the top down. Once again, each level was tested for supersaturation; however, in this instance any water over saturation was pushed to the next lower level. This method had been used internally in the Met Office (M. Best 2009, personal communication) to overcome this problem.
3. Using NDVI to identify the seasonal evolution of drought stress
Given sufficient insolation and a high enough temperature, natural vegetation will normally only be stressed where there is a prolonged soil moisture deficit in the root zone. We compare the seasonality of explicitly modeled soil moisture stress with an implicit moisture stress derived from the NDVI. Drought phenology is not simulated in the model; therefore, a direct comparison between modeled biomass and NDVI is not possible.
To identify seasonal vegetation stress related to soil moisture, we used correlations between monthly values of a global gridded index of vegetation greenness (NDVI from satellite observations) and the University of East Anglia’s Climatic Research Unit (CRU) precipitation totals (Mitchell and Jones 2005). A significant correlation between NDVI and preceding rainfall totals is taken to imply a relationship between vegetation greenness and soil moisture—either through restricted productivity or leaf loss.
A variety of vegetation indices are available from satellites, but the advantage of the Fourier-adjusted, sensor- and solar zenith angle–corrected, interpolated, and reconstructed (FASIR) 4.1 Advanced Very High Resolution Radiometer (AVHRR) NDVI dataset is that it provides one of the longest homogeneous global gridded monthly vegetation time series (extending from January 1982 to December 1999; Los et al. 2005; available online at http://www.neodc.rl.ac.uk). There was an interruption in the AVHRR data during 1993/94; to overcome this, a long-term climatology of NDVI was inserted for this period. There are also possible issues concerning the corrections required for the atmospheric aerosol loading following a volcanic eruption during 1991, and future corrections may improve the results found in this work. Overall, the length of the dataset outweighed any concerns, because it allowed an assessment of the seasonal cycle, which would not have been possible with shorter datasets.
Figure 1 presents regions with significant correlation between rainfall and NDVI (gold shading) and shows, in the case of July, that a substantial fraction of the globe has significant correlation at the 95% level. The figure also shows a global pattern of ψ averaged over the 10 years of July data from the COsoil run (marked with a “W”) and areas with strong gradients in available soil moisture (green contouring). The fact that regions of significant correlation are generally found in regions with intermediate ψ—for example, West Africa, southeastern United States, Australia—supports the idea that the correlation is due to moisture stress. Additionally, the areas of correlation are in well-defined regions, suggesting that the cause is regionally coherent rather than a series of random observations.
For each month of the year, the correlation between rainfall and NDVI is calculated from data over 17 years (i.e., 17 data points). This allows seasonal variations in significance and, by inference, vegetation stress to be assessed. However, to allow for comparison of model experiment results with the NDVI correlations, the analysis was conducted at 1° spatial resolution.
The four half-degree-square pixels contributing to each 1° grid box result in 68 data points overall for each grid box. The data from each contributing pixel had their mean subtracted and were normalized by their standard deviation prior to the correlation calculation. This helps to control the effect of vegetation cover type on the amplitude of the NDVI seasonal cycle. Sites exhibiting no variability in either or both variables over the full 18 years (e.g., permanent ice-covered regions and small parts of some deserts) and with extensive snow cover (north of 60°N) were excluded from analysis.
To improve signal-to-noise ratios, the correlations use NDVI data summed from two consecutive months. The standardized, concatenated NDVI values were correlated with first one month of preceding precipitation, then the two preceding months summed and so on, up to three preceding months summed. The number of grid boxes exhibiting significant correlation globally increases substantially when moving from one to two to three months of summed precipitation, but thereafter the number increases slowly. This observation is consistent with the idea that deep-rooted vegetation responds to soil moisture variations over seasonal time scales. Hence results for three months of summed precipitation are used here.
The seasonal evolution of significance for each grid box was examined to identify areas where there was a transition from nonsignificant to significant. Figure 2a shows the seasonally varying values of correlation coefficient r for an example pixel in West Africa. During the wet regional wet season (July–October), there is no significant correlation; however, in November the correlation becomes significant and is termed the stress onset month.
To isolate the effect of soil moisture stress from other possible sources, a number of restrictions were placed on the determination of stress onset. It is assumed that vegetation growth ceases below 5°C; therefore, using the GSWP2 driving data used in the model experiments, months with an average temperature below 5°C were excluded from being part of a seasonal cycle in soil moisture–related stress. Similarly, a threshold of 130 W m−2 was applied to filter out months with low shortwave radiation.
To exclude correlations in semiarid regions that are simply related to the onset of growth at the start of the wet season, the onset was only considered where the NDVI was above 0.1 and constant or decreasing. In very arid regions, a threshold of 0.1 NDVI (corresponding to an LAI of 0.2 m2 m−2) was used to exclude spurious correlations related to very low values of vegetation greenness.
To remove the effect of seasonal phenology, all the correlations were conducted on interannual perturbations from the seasonal average. Lastly, when correlation values became significant at multiple times within the year, the onset month was chosen as the month with the largest change in correlation coefficient from the previous month.
To identify the onset of stress in the model, a function of the soil moisture ψ̂ was calculated using Eq. (3) and the diagnosed soil moisture in the model. Rather than restricting the values to between zero and one, as in the calculation of ψ in the model, ψ̂ was allowed to assume any value. In other words, ψ̂ is the ration (θ − θw)/(θc − θw). Months where ψ̂ was ≥(≤)1 are considered unstressed (stressed). The same restrictions concerning temperature and LAI were applied during the analysis of the model output. To remove the influence of grid boxes where ψ̂ fluctuated around unity over the year, an additional requirement was that the model be unstressed for two months prior to the onset of stress. If in a year a grid box remains stressed for the entire year, then this grid box is not considered to have an onset of stress. This process was carried out across the 10 years of model output rather than on time average seasonal climatology.
Figure 2b shows the climatology (over the 10 years of the modeling experiment) of ψ̂ for the grid box in West Africa as the example. In this case, the month of onset of vegetation stress in the model experiment is November, because ψ̂ has fallen below 1.0 after it had exceeded this level for two months. Notice that at this location, the month of onset of vegetation stress, according to the observational data based on r (Fig. 2a), is identical to that in the model based on ψ̂; however, this is not always the case.
4. Global comparisons of stress onset
In this section, the global distribution of the NDVI- and model-derived soil moisture stress onsets are examined and regions are identified for further analysis. The number of pixels that showed an onset of stress in the NDVI data analysis was 1015, and the distribution of these points is shown in Fig. 3a. Comparing this figure with Fig. 1 we see that, especially in the midlatitudes, there are far fewer pixels with onsets identified than pixels with a significant correlation. The loss of pixels is due to a number of factors: first, many of the pixels have a constant state—that is, they are either always significant or not significant; second, in some cases the correlation is related to a secondary process—for example, temperature or radiation limitations that are reflected in the rainfall time series; third, in arid and semiarid regions, there is too little vegetation to make a meaningful assessment. The consequence of this reduction in pixels is that the remaining pixels are where there is a robust signature of water stress.
The timing of the onsets of stress for the two model runs with the maximum and minimum number of pixels with onset (COsoil and MOsoil) are presented in Figs. 3b and 3c. In the case of the model runs, the month plotted is the mean month from years when there was an onset during the 10-yr integration.
The majority of cases are situated within the tropics and subtropics (30°S–30°N), and there is a broad agreement between the NDVI and the model runs. This latitudinal limitation is in agreement with the findings of Churkina and Running (1998), who found that outside the tropical regions, temperature became a stronger influence than soil moisture.
There is good spatial agreement between the NDVI and COsoil run stress onsets in the tropical/subtropical region. However, the temporal agreement is less consistent. In the MOsoil run, the spatial extent of the coherent regions is diminished and the temporal agreement is less in agreement with the data-derived timing than COsoil. This is also highlighted in Table 2, which summarizes the number of onsets identified in all the model runs and the NDVI analysis. The modeled onsets are the average number of onsets for all of the 10 years in the runs.
The most notable feature is that the three runs with the three Cosby soil parameters (COsoil, MeanLai, and Pushdown) have nearly identical onset numbers and have the largest number of collocated onsets. These runs capture a mean of 55% of the NDVI points. The interannual range of model captured onsets is between 53% and 58% in the 10 years of the experiment. In contrast, VanGen and MOsoil capture an average of 38% and 39% of the NDVI-derived onsets, respectively. They have an interannual range of between 30% and 42% for the VanGen run and 33% and 42% for the MOsoil run. This spread among the various soil parameter runs with the similarity between the differing vegetation and hydrology runs suggests that with this model setup, it is the soil parameters that dominate the soil moisture stress, and that the change in the representation of LAI and soil hydrology has less effect.
The spatial distribution of moisture stress onset was shown in Fig. 3; however, only a subjective assessment of the timings in model- and data-derived month could be made. Figure 4 examines the global distribution of modeled stress onsets in terms of how many months the model lags the NDVI-derived onset. This is expressed as a percentage of the total NDVI-derived onsets and, because the model does not capture all the onsets identified in the data, the area below the graph does not equal 100%. A negative value on the x axis indicates that the model becomes stressed early.
All the distributions are nonsymmetric about zero, but they do have a maximum within three months of zero. This may result from the nature of the comparison between an instantaneous model variable and a lagged biophysical response to soil moisture stress.
There is, once again, a marked difference between the three Cosby soil runs, and the MOsoil and VanGen runs. The three Cosby soil runs have very similar distributions; in general, they become stressed slightly later than the NDVI analysis predicts. The noticeable peak in the distribution at +3 is due to the region that lies between −10°S and 10°N, and spans South America and Africa. A sample point from this region will be examined in the following section. The three runs have skewness values COsoil −0.57, MeanLai −0.68, and Pushdown −0.59. The VanGen and MOsoil distributions have notably smaller peaks—in agreement with the evidence in Table 2—and become stressed earlier than the NDVI analysis predicts and have larger values for skewness (VanGen, 0.94; MOsoil, 1.17).
One possible mechanism for the differing model biases mentioned earlier is the partitioning of evaporation and runoff in the hydrological cycle. Table 3 shows the overall values for evaporation and runoff for the grid boxes used in Fig. 4. All the runs with the three Cosby soil parameters have very similar values for runoff and evaporation. This might be expected for Pushdown because it only has an effect in the high latitudes, and the majority of these points are tropical. However, the lack of change when seasonality is added to the LAI suggests that overall vegetation has little effect on annual evaporation and runoff values, at the mainly tropical pixels, in the model.
Analyzing the global results has highlighted a number of points. First, the soil hydraulic parameters have a strong effect on the seasonal cycle of soil moisture stress in the model. On the basis of these results, the Cosby soil parameters outperform the VanGen and MOsoil in terms of simulating the onset of moisture stress. Conversely, the representation of seasonality in vegetation has little influence in simulating the onset of soil stress.
5. Regional comparisons of stress onset
The analysis of global results in the previous section identified a number of spatially coherent regions that were dominating the NDVI and modeled onset months. In this section the climatic and hydrological influences involved in the onset of moisture stress in the vegetation of these regions are analyzed in more detail. Output from the model is used in conjunction with the correlation analysis.
The modeled variables shown are 10-yr monthly averages over the years of the GSWP2 dataset (1986–95). The correlation coefficient is calculated from data covering 1982–99, as described in section 3. To aid legibility, only the COsoil, MOsoil, and VanGen model runs are presented in the figures (because the MeanLai and Pushdown results are nearly indistinguishable from COsoil). Four areas will be discussed here: the Soudan (10.5°N, 9.5°W) and Guinea coast (6.5°N, 0.5°E) in West Africa, northern India, and the eastern United States.
a. Soudan case study
One of the regions identified in the global analysis as having spatially coherent onsets in both the data and the model runs was western Africa. Figure 5 has values from two representative points in the region with distinct annual cycles in their climate: the left column (Figs. 5a–c) is in the Soudan region (10.5°N), and the right column (Figs. 5d–f) is in the Guinea coast region (6.5°N).
It can been seen while examining the evolution of the rainfall in Figs. 5a and 5d that the Soudan case is dominated by a single peak in rainfall during July–September, whereas the Guinea coast case has a bimodal rainfall pattern with maxima in June and October. These sequences are due to the northward passage of the intertropical convergence zone in early summer and its subsequent retreat in late summer.
The NDVI-derived onset of stress for the Soudan (Fig. 5a) is in September (one month after the maximum in rainfall). The correlation peaks in October and is nonsignificant by December; therefore, the interannual variation in the end of the growing season is dominated by the rainfall as the ITCZ migrates south.
The onset of stress in the model runs varies between runs with COsoil and VanGen becoming stressed two months later than the NDVI in November, whereas MOsoil becomes stressed in October after only briefly achieving a value greater than one. It is notable that the VanGen run attains a higher ψ than COsoil, which is unusual to the pixels with identifiable onsets.
These points can be explained by examining the seasonal hydrological cycle shown in Fig. 5c. It shows that there is a 0.5 mm day−1 difference in evaporation between the COsoil and VanGen runs during the peak month of June. The effect of this on the total evaporation for May–July is shown in Table 4, with the VanGen run losing 35 mm more through evaporation over the 3-month period. This stronger evaporation in both MOsoil and VanGen is entirely through bare soil evaporation losses from the top soil layer.
There are also differences in the seasonal evolution of runoff: the VanGen run has a smoother runoff curve with a weaker peak flow than the other runs and a longer period of flow in the dry period. The effect of this over the August–October period is that VanGen loses more than 100 mm less through runoff than COsoil. This weaker peak flow results in higher soil moisture in the lower model levels and is the cause of the high ψ values mentioned earlier. The combination of strong early-season evaporation and strong late-season runoff during the wet season explains the earlier onset month in MOsoil.
b. Guinea coast case study
The Guinea coast case has a less extreme rainfall regime than the previous case, with rainfall over much of the year and bimodal peaks in June and September. Consequently, the amplitude of the seasonal cycle in vegetation is weaker. In fact, the peaks in greenness in May and September occur when the vegetation is stressed. Examining the seasonal evolution of the NDVI correlation shows that there is only one month of the year when the NDVI is not significantly correlated with rainfall. This occurs in July after the early wet season rainfall during the short dry season. Only the COsoil run becomes unstressed at any time during the year; this occurs during October and the rainfall in the late wet season. This mismatch of stress onsets in regions with bimodal rainfall patterns in the COsoil run is the primary cause of the 3-month model lag identified in the discussion of Fig. 4.
The partitioning of soil moisture into evaporation and runoff follows a similar pattern to the Soudan case: with stronger evaporation from the bare soil early in the year in VanGen (VanGen has 24 mm or 10% more than COsoil) and to a lesser extent MOsoil. This has a dramatic effect on the subsequent runoff with VanGen run having only 11% of the runoff observed in COsoil.
c. Indian case study
The Indian case has a very strong seasonal cycle in precipitation, with a maximum greater than 250 mm month−1 and five months with less than 20 mm month−1. There are two periods in the year when the NDVI is strongly correlated with antecedent rainfall. The first period is during the greening-up period and occurs in August (the peak rainfall month), and lasts two months. In the second month, the peak NDVI occurs. The second period of significant correlation occurs three months after the peak rainfall (December) during the senescence period. The second period of correlation is probably due to the interannual variability in the end of the wet season. This is a very different pattern to the Soudan case, which also had a single wet season, where no correlation was detected during the greening up but exhibited stress within a month of the peak rainfall. In comparison to results from the observations, COsoil and VanGen model runs do not become stressed during the greening-up phase and predict an onset two months after the peak rainfall in October. This is one month earlier than the observations The MOsoil run is stressed throughout the year.
The COsoil and VanGen model runs have similar patterns of ψ until the month after the peak rainfall rate (Fig. 6b). The mechanism for this divergence can be seen in Table 5, and there is stronger evaporation in VanGen (compared to COsoil) in the three months prior to the divergence (25 mm over the three months). There is also much stronger runoff in the VanGen run during the three months after the divergence in ψ evolution.
d. Eastern U.S. case study
The final case study occurs in the transition zone between the tropics, where soil moisture is the main control on vegetation growth, and the high latitudes, where radiation or temperature is the main control. In an examination of Fig. 3a, the eastern United States stands out as a nontropical region with a large number of spatially coherent pixels with an identified onset. This occurs in both the observations and the COsoil model run. The Cosoil run has a strong early bias in this region and becomes stressed, on average, two months early.
The seasonality of the rainfall in this area (shown in Fig. 6d) is very weak and does not mimic the seasonal cycle of the NDVI (Fig. 6d) or the modeled soil moisture stress (Fig. 6e). Figures 6d and 6f show that the radiation and evaporation are highly correlated (in the model) in the early part of the year. This continues until the soil moisture stress weakens the evaporation from May to September; therefore, in this region the timing of the stress onset in driven by the insolation and evaporation.
In this pixel, the NDVI shows an onset of stress in July that coincides with the peak in monthly shortwave radiation and a small dip in rainfall and NDVI. As stated earlier, only the COsoil run predicts an onset in stress and this is in May, coinciding with the peak in modeled evaporation. The NDVI and COsoil become unstressed in October after the evaporation has weakened. The MOsoil or VanGen runs remain stressed throughout the year but, again, they have stronger evaporation than the COsoil run. The runoff in VenGen also follows the familiar patter of having a lower peak runoff than MOsoil and COsoil but maintaining some flow during the more stressed part of the year.
In all the case studies, the VanGen model run had a lower peak runoff rate but a slower decrease during the drying period. Examining the runoff rate compared to the lowest-level soil moisture in Fig. 7 reveals that the COsoil run is able to produce more runoff at high volumetric soil values, but that the VanGen maintains some runoff at much lower soil moisture values.
6. Discussion and summary
Using earth observation vegetation data and the CRU precipitation dataset, we have created a method to examine vegetation water stress in the tropics and subtropics. An analysis of the seasonal evolution of the correlation between NDVI and antecedent precipitation allows us to identify regions with stressed vegetation. Screening for the influence of temperature and radiation, we identified regions with a strong seasonal cycle in soil moisture stress in the vegetation.
A comparison of output from a number of model runs with the NDVI-derived stress onsets has highlighted the importance of soil hydraulic parameters in accurately predicting stress onsets with the model. Global analysis showed a marked difference in both the number of pixels with identifiable seasonal cycles in soil moisture stress and in the timing of the onset of stress. In this study, soil parameters based on the transfer functions of Cosby et al. (1984) produced the most realistic results.
In the global analysis, the ability of the model to simulate seasonal evolution of moisture stress highlighted a number of areas where the model performed to different standards. The case studies have shown that the model appears to predict stress well in regions with sharp changes in precipitation and well-defined wet seasons. In regions where the wet seasons are less apparent (Guinea coast) or radiation is the main climatic factor in drying the surface out (eastern United Stastes), the model does not do as well.
The version of the JULES model used in this analysis does not include the effect of agriculture on the seasonal cycle of soil moisture stress. In the future it would be interesting to rerun the analysis with new model runs to examine the sensitivity of the soil moisture stress to planting, harvesting, and irrigation. Some of the offsets identified in this work might be explained by the inclusion of these processes in the model.
An examination of a number of case study points showed that, in agreement with previous studies (Koster and Milly 1997; Gedney et al. 2000), it was the partitioning of throughfall into runoff and evaporation that was most important in defining the onset of soil moisture stress. The model runs with stronger evaporation (MOsoil and VanGen) generally remained stressed in pixels where both the NDVI and more weakly evaporating runs predicted a seasonal switch between stressed and unstressed. The functions from van Genuchten (1980) and Cosby et al. (1984) also changed the seasonal cycle of runoff within the model.
The additional evaporation in MOsoil and VanGen was due to bare soil evaporation drawing water from the top-most soil moisture layer. Equation (4) shows that the soil conductance is dependent solely on the parameter θcrit. The value of this parameter was considerably less for these two runs compared to the Cosby soil runs.
With a paucity of global hydrological data, particularly soil moisture stress, this tool will prove useful in evaluating future changes to land surface models. It will also be useful when comparing the hydraulic regimes of different land surface models.
Acknowledgments
This work was supported by the NERC-funded CLASSIC program and the Joint DECC and MoD Integrated Climate Programme (DECC GA01101 and MoD, CBC2B0417_Annex C5, respectively).
We would also like to thank the three reviewers for some constructive comments that helped to clarify and add focus the work.
REFERENCES
Churkina, G., and Running S. , 1998: Contrasting climatic controls on the estimated productivity of global terrestrial biomes. Ecosystems, 1 , 206–215.
Clapp, R., and Hornberger G. , 1978: Empirical equations for some soil hydraulic properties. Water Resour. Res., 14 , 601–604.
Cosby, B., Hornberger G. , Clapp R. , and Ginn T. , 1984: A statistical exploration of the relationships of soil moisture characteristics to the physical properties of soils. Water Resour. Res., 20 , 682–690.
Cox, P., Huntingford C. , and Harding R. , 1998: A canopy conductance and photosynthesis model for use in a GCM land surface scheme. J. Hydrol., 212–213 , 79–84.
Dai, Y., and Coauthors, 2003: The Common Land Model. Bull. Amer. Meteor. Soc., 84 , 1013–1023.
Dirmeyer, P., 1994: Vegetation stress as a feedback mechanism in midlatitude drought. J. Climate, 7 , 1463–1483.
Dirmeyer, P., Gao X. , Zhao M. , Guo Z. , Oki T. , and Hanasaki N. , 2005: The Second Global Soil Wetness Project (GSWP-2): Multi-model analysis and implications for our perception of the land surface. COLA Tech. Rep., 46 pp.
Essery, R., Best M. , Betts R. , Cox P. , and Taylor C. , 2003: Explicit representation of subgrid heterogeneity in a GCM land surface scheme. J. Hydrometeor., 4 , 530–543.
FAO, 2003: Digital soil map of the world and derived soil properties, Rev. 1. Food and Agriculture Organization of the United Nations, Land and Water Digital Media Series, CD-ROM.
Gedney, N., Cox P. , Douville H. , Polcher J. , and Valdes P. , 2000: Characterizing GCM land surface schemes to understand their responses to climate change. J. Climate, 13 , 3066–3079.
Guo, Z., and Coauthors, 2006: GLACE: The Global Land–Atmosphere Coupling Experiment. Part II: Analysis. J. Hydrometeor., 7 , 611–625.
Hennessy, K. J., Whetton P. H. , Katzfey J. J. , McGregor J. L. , Page C. M. , and Nguyen K. C. , 1998: Fine resolution climate change scenarios for New South Wales, summary report 1995–1998. CSIRO Atmospheric Research Tech. Rep., 20 pp.
Koster, R., and Milly P. , 1997: The interplay between transpiration and runoff formulations in land surface schemes used with atmospheric models. J. Climate, 10 , 1578–1591.
Koster, R., and Coauthors, 2004: Regions of strong coupling between soil moisture and precipitation. Science, 305 , 1138–1140.
Lawrence, D., and Slingo J. , 2005: Weak land–atmosphere coupling strength in HadAM3: The role of soil moisture variability. J. Hydrometeor., 6 , 670–680.
Los, S., and Coauthors, 2000: A global 9-yr biophysical land surface dataset from NOAA AVHRR data. J. Hydrometeor., 1 , 183–199.
Los, S., North P. , Grey W. , and Barnsley M. , 2005: A method to convert AVHRR Normalized Difference Vegetation Index time series to a standard viewing and illumination geometry. Remote Sens. Environ., 99 , 400–411.
Manabe, S., 1969: Climate and the ocean circulation: I. The atmospheric circulation and the hydrology of the earth’s surface. Mon. Wea. Rev., 97 , 739–774.
Milly, P., and Shmakin A. , 2002: Global modeling of land water and energy balances. Part 1: The land dynamics (LaD) model. J. Hydrol., 3 , 283–299.
Mitchell, T., and Jones P. , 2005: An improved method of constructing a database of monthly climate observations and associated high-resolution grids. Int. J. Climatol., 25 , 693–712.
Tomasella, J., and Hodnett G. , 1998: Estimating soil water retention characteristics from limited data in Brazilian Amazonia. Soil Sci., 163 , 190–202.
van Genuchten, M. T., 1980: A closed form equation for predicting the hydraulic conductivity of unsaturated soils. Soil Sci. Soc. Amer. J., 44 , 892–898.
Zhang, B., Dazlich D. , and Randall D. , 1999: Simulations of soil moisture and surface water balance using the Simple Biosphere Model 2. J. Meteor. Soc. Japan, 77 , 217–234.
Zobler, L., 1999: Global soil types, 1-degree grid. [Available online at http://www.daac.ornl.gov].
Description of the modeling experiments undertaken. An extended description is presented in the text.
The number of points where an onset was identified in the model only, and in the model and NDVI analysis. The values refer to the mean annual values over the 10 years of the model runs. There are 1015 onsets identified in the NDVI analysis.
Global values for the evaporation (mm day−1), runoff (evap; mm day−1), and the ratio of evap/(evap + runoff) for the different model runs. Total runoff refers to the summation of the surface and subsurface runoff components in the model. All the values are the average of monthly averages over the 10 years of the runs.
Evap and runoff for three of the model runs at in the Soudan and Guinea coast regions. Evap in the Soudan case is the total for May–July, and runoff is the total for August–October. In the Guinea coast case, evap is for April–June and runoff is for August–October. Units are in millimeters.
Evap and runoff for three of the model runs in the Indian and eastern U.S. regions. In the Indian case, evap is the total for July–September, and runoff is the total for September–November. In the U.S. case, evap is the total for April–June, and runoff is the total for January–March. Units are in millimeters.