1. Introduction

Land surface modeling is cited as one of the major causes of the uncertainties in current climate change predictions (Houghton et al. 1996). This is especially significant given the need to assess many of the impacts of climatic change, such as the effect on agriculture and changes in the occurrence of extreme hydrological events such as drought and flooding.

Some climate simulations suggest that there will be an increased likelihood of summer drought in midlatitudes in an enhanced greenhouse gas climate (Gregory et al. 1997). Successful assessment of its potential severity is dependent on adequate modeling of the land surface. However PILPS (Henderson-Sellers et al. 1996) demonstrates that there is a wide disparity between GCM land surface schemes (LSSs), with little agreement on the modeled soil moisture and evaporation.

Koster and Milly (1997) analyzed part of the PILPS 2a experiment where identical atmospheric conditions were used to force a number of LSSs offline, which in principle have identically prescribed vegetation and soil characteristics. They showed that the hydrological behavior of the LSSs is fundamentally dependent on the interplay between evaporation and runoff, and that it was possible to understand the differences in terms of a few simple effective characteristics that can be associated with the monthly mean water balance of each LSS.

We use a similar methodology to assess the role of LSSs in predicted changes in surface hydrology as a result of climate change. This is carried out with a number of experiments using different GCMs and LSSs (described in section 2). First, the behavior of each LSS is characterized in the 1 × CO₂ simulations over three specific regions: Amazonia, the Sahel, and southern Europe (section 3a). This is achieved by examining the sensitivity of the partitioning between runoff and evapotranspiration to soil moisture for each LSS. We show that this can be used to understand the regionally averaged annual mean soil moisture of each LSS in the 1 × CO₂ control runs (section 3a).

This characterization can then be used to estimate and explain the LSS response to the change in climate forcing, as defined by changes in precipitation and available energy. The methodology for this is described in section 3b and is applied to Amazonia, the Sahel, and southern Europe in section 4.

2. Experiment

Four groups participated in the GCM experiments: the Hadley Centre (HC), the Laboratoire de Météorologie Dynamique du CNRS (LMD), Météo-France (CNRM), and the University of Reading (UR). Each performed two time-slice experiments, both consisting of a 1 × CO₂ and 2 × CO₂ run. The Atmospheric Model Intercomparison Project (AMIP; Gates 1992) sea surface temperature and sea-ice fields were used in the 1 × CO₂ experiment. The temperature and sea-ice changes produced by the Hadley Centre transient climate change run (Mitchell et al. 1995) were applied to the AMIP climatological fields to produce the boundary conditions for the 2 × CO₂ time-slice simulations. All simulations were at least 10 yr long after an initial spinup period.

Each experiment pair was implemented with two different LSSs coupled to an otherwise identical GCM. The differences between each LSS pair were not controlled, but they turned out to be mainly associated with surface hydrology (see Table 1 for a brief description of the models and the shorthand notation for each simulation). Two of the groups (HC and CNRM) included the effects of CO₂-induced stomatal closure in one of their 2 × CO₂ runs, with the 2 × CO₂ Cb simulation using the stomatal conductance changes diagnosed from Hb (Douville et al. 2000). Indeed, this was the only difference between Ca and Cb, so these models share the same 1 × CO₂ control run.

Although most of the LSSs share similar underlying assumptions, there are some LSS specific points that are worth noting. First, the bucket model used in “Lb” may be considered as slightly anomalous since this is the very simplest LSS used in early GCM simulations (Manabe 1969). It is included here because it represents one end of the range of LSS complexity, and also because it is still in use in some climate and numerical weather prediction models. Offline intercomparisons have suggested that the behavior of the bucket model differs from more complex LSSs primarily because it does not include an additional surface resistance to evaporation (Henderson-Sellers et al. 1996). However, the bucket model also relies on a different runoff generation mechanism than most current day LSSs, which are often dominated by slow “Darcian” drainage. By contrast, the bucket produces runoff only when the soil moisture exceeds the specified field capacity, which makes this model more sensitive to the temporal variability of the rainfall. Of the other LSSs used here, only SECHIBA (Schématisation des Echanges Hydriques à l’Interface entre la Biosphère et l’Atmosphère; Ducoudre et al. 1993) as used in La, is likely to share this sensitivity. Also, the novel soil moisture scheme in this model means that there is no simple relationship between rootzone soil moisture and evapotranspiration, because vegetation becomes unstressed if the movable top layer of soil becomes saturated, even if the rootzone as a whole is moisture deficient.

These LSS differences are relevant to the next section where we analyze the behavior of each model in the three distinct regions of Amazonia, the Sahel, and southern Europe. Table 2 lists the latitude and longitude ranges defining these study areas.

3. Characterizing the land surface schemes

Following the general approach of Koster and Milly (1997), a “bucket-type” formulation is utilized to model how the suite of LSSs respond to limited soil moisture under snow-free conditions and ignoring interception, E_i.

In the usual bucket model actual evaporation is calculated as the product of a soil moisture stress (“β”) factor and the potential evaporation rate, E_p (Manabe 1969). Unfortunately, the potential evaporation rate was not available from all of the GCM experiments, so we replaced E_p in the usual bucket formulation with effective available energy, A_e:

View Expanded

Here A_e has been scaled to have the units of a water flux, and R_N, G, and L are the net radiation, the ground heat flux, and the latent heat of vaporization. (We also define the effective available energy as A_e = A − E_i, where A is the available energy.) In this model, evapotranspiration (i.e., bare soil evaporation plus transpiration), E_e, is therefore assumed to be dependent on the effective available energy, A_e, and the rootzone soil moisture, M, through a soil moisture–dependent evaporative fraction, f:

E_efA_e(2)

This equation is very closely related to the standard bucket model as applied by Koster and Milly (1997), because potential evaporation and effective available energy, A_e, are strongly correlated. Correlation coefficients, between monthly mean values of E_p and A_e, were found to be well in excess of 0.9 for all of the GCMs that produced both of these diagnostics.

For convenience the soil moisture is scaled by its active range, which is defined as the point at which evapotranspiration starts (M_wilt) to the point beyond which evapotranspiration is not constrained by soil moisture (M_crit):

View Expanded

The critical and wilting points were estimated from the regionally averaged values of the parameters in each LSS. In terms of this scaled variable, f takes the simple piecewise linear dependence:

View Expanded

where f_c is the evaporative fraction at the critical soil moisture concentration. Among the more complex LSSs f_c tends to vary due to differences in the parameterization of stomatal resistance.

The relatively small range of μ experienced by each LSS allows us to use a linear approximation for the stress and runoff curves:

View Expanded

where f′ and Y′ are the gradients of f and Y with respect to μ, and f_c and Y_c are the values of f and Y at the critical point (μ = 1).

For each model the monthly mean values of Y and E_e/A_e from the control runs were fitted to Eqs. (7) and (6), respectively, using a least squares regression. For Amazonia and the Sahel all 12 months were utilized in the fitting exercise, but for the southern European region only the months from April to September were used, to avoid periods when the available energy is so small that f is ill-defined. This procedure provides an estimate of the characteristics of each LSS in terms of f_c, Y_c, f′, and Y′. Values for Amazonia, the Sahel, and southern Europe are given in Tables 3, 4, and 5.

a. Estimating the surface hydrology of the control climate

In this section we test the characterization of each LSS by attempting to predict its long-term mean soil moisture status and annual mean evaporation E using information about the 1 × CO₂ control climates. On the multiannual timescale, changes in soil moisture storage are negligible, so throughfall precipitation (i.e., precipitation P minus interception) balances evapotranspiration and runoff:

E_eYP_e(8)

where the bars denote multiannual means. Substituting in Eqs. (2), (6), and (7) gives:

A_ef_cfμY_cYμP_e(9)

Thus our simple model gives an estimate of the long-term annual mean value of μ in terms of the control climate (P_e and A_e) and the surface characteristics (f_c, f′, Y_c, and Y′):

View Expanded

Substituting this form for μ into Eqs. (2) and (6) also yields an expression for E_e. Figure 2 compares the estimated values of μ and the annual mean evaporation to the long-term means taken from each of the 1 × CO₂ GCM runs. [The estimates of annual mean evaporation were calculated by adding the GCM interception rates onto the expression for E_e. The annual mean quantities for P_e and A_e were used for calculating μ and E in all the regions studied, even though the linear fit for southern Europe was based on data from April to September alone (see section 3).] The simple linear characterization of each LSS is clearly able to capture the large variation in moisture stress and evaporation between the models, for each of the three regions.

Insight into the reasons for the discrepancies between the models can be gained by examining the conditions required for the annual mean evaporation not to be moisture limited. From Eq. (9), μ ≥ 1, if

χϕ,(11)

where we have defined two new dimensionless parameters:

View Expanded

The parameter, ϕ, is akin to Budyko’s aridity index as described by Koster and Suarez (1999), while χ represents the runoff fraction when μ = 1. Figure 3a plots all regional values for each model in the χ–ϕ phase space. The dot–dashed line represents ϕ + χ = 1, with evaporation being moisture limited in all models to the right of this line. The large variation in ϕ for a given region between the schemes indicates that the control climates differ markedly, and therefore some must be unrealistic. Moreover there is no clear ranking of ϕ between the regions, with, for instance, Amazonia being more arid than the Sahel in some models.

For each model there is a range of χ values that is largely associated with the different climatic regimes (as denoted by the subscripts). The values of χ also show some clustering according to LSS. Thus, for example, the Hadley Centre models (“H”) typically have small values of runoff at the critical point, which implies low values of χ, and little or no soil moisture limitation. On the other hand, the University of Reading models (“U”) always have much higher values of Y_c, usually resulting in larger χ values, ensuring that they are moisture limited in all three regions.

Figure 3b plots values of μ against ϕ + χ, for each of the 1 × CO₂ simulations over each of the study areas. Low μ values are associated with high values of ϕ + χ, which can arise from either the aridity of the driving climate (i.e., high ϕ) or from high values of runoff at the critical point (i.e., high χ).

Figure 3a shows that in many cases there is substantially more scatter in χ than ϕ. This suggests that differences in the runoff curves (Y_c) are at least as important as differences in the GCM driving climates and f_c, and in some cases Y_c is the dominant source of discrepancy among the models. This can be demonstrated further by forcing all the LSSs with the same climate for Amazonia (i.e., the same A_e and P_e). Any major differences in the resulting LSS states must then be the result of differences in the LSSs (i.e., differences in f_c and Y_c). If we take the annual mean effective available energy (3.7 mm day⁻¹) and throughfall (4.3 mm day⁻¹) from the most unstressed model, Ha (see Fig. 1a), the resulting values of ϕ + χ are greater than unity for all the LSSs other than Ha and Hb (see Table 3 for values of Y_c and f_c). This indicates that only Ha and Hb would be unstressed under such a climate. With the exception of the bucket model, most of the differences between the LSSs are a result of Y_c rather than f_c, indicating that Y_c plays the largest role in the differences between the LSSs.

b. Estimating the hydrological response to climate change

Having demonstrated that the simple linear characterization can capture the annual mean surface hydrology predicted by each LSS, here we extend the method to analyze the changes in surface hydrology on doubling CO₂. Specifically, we consider the changes in annual mean soil moisture and evaporation arising from a prescribed climate change, as given by the changes in annual mean throughfall precipitation, ΔP_e, and effective available energy, ΔA_e. The perturbation equation for the annual mean hydrological balance [Eq. (9)] is

A_ef_cfμμA_efYP_e(14)

where Δμ is the change in the scaled annual mean soil moisture resulting from the prescribed climate perturbation (ΔP_e and ΔA_e). The expression for Δμ can be simplified using Eq. (6):

View Expanded

Here, a further dimensionless parameter, ζ, has been introduced, which represents the ratio of the sensitivities of evaporation and runoff to soil moisture:

View Expanded

The corresponding perturbation to the evaporation is derived from Eqs. (2) and (6):

E_efA_efA_eμ(17)

The soil moisture perturbation Δμ can be eliminated using Eq. (15), to give an expression for the change in the annual mean evaporation resulting from ΔA_e and ΔP_e:

View Expanded

This equation has been written in the form of the “chain rule” to allow easy identification of the coefficients representing the sensitivity of evaporation to precipitation and available energy, respectively:

View Expanded

Values for these coefficients, as diagnosed from the 1 × CO₂ control climates, are given in Tables 3, 4, and 5. The variation among them can be understood in terms of the dependencies of f and ζ on the control state. Consider the extreme cases of a very dry soil (μ ≪ 1), for which the evaporative fraction will be strongly dependent on soil moisture, and a very wet soil (μ ≫ 1), for which the evaporative fraction will be independent of soil moisture:

View Expanded

where we have assumed f(0) ∼ 0 in Eq. (6). If the Y–μ curve also passes close to the origin (such that Y′ ∼ Y_c), Eq. (20) can be substituted into Eq. (10) to give an expression for ζ in each of these two regimes:

View Expanded

Now substituting Eq. (21) into Eq. (19) and using μ from Eq. (10) yields expressions for the evaporation sensitivity coefficients explicitly in terms of the control climate (through ϕ) and the LSS characteristics (through χ and f_c):

View Expanded

Equations (11), (22), and (23) enable the LSS behavior to be summarized as follows:

• In a moist climate (i.e., where ϕ ≤ 1):

  • —An LSS may be stressed or unstressed depending on its value of χ (i.e., its runoff at the critical point, Y_c), as is illustrated by Eq. 11.

  • —In unstressed models (μ ≫ 1) evapotranspiration is independent of throughfall, but optimally dependent on effective available energy.

  • —In stressed models (μ ≪ 1) evapotranspiration is dependent on both the effective available energy and throughfall. These dependencies both tend to reduce with increasing χ (i.e., increasing Y_c).

• In a more arid climate (i.e., where ϕ > 1):

  • —All LSSs are moisture limited (for physically realistic values of χ > 0).

  • —The LSSs are mainly sensitive to throughfall, and this sensitivity reduces with increasing Y_c.

We use this framework to explain the regional responses as described in the next section.

4. Climate change prediction

In this section the GCM climate change predictions are analyzed for each of the three regions using the LSS characterization as a framework for understanding the model differences. Figures 4, 5, and 6 are used to summarize the findings for Amazonia, the Sahel, and southern Europe, respectively. In each case, plots a and b compare the GCM simulations of climate and surface hydrology for the control state, while c and d do likewise for the predicted climate change. Plots e and f test the ability of the characterization to reproduce the changes in evaporation and scaled soil moisture for each GCM. The “estimated” values were derived by applying the GCM changes in throughfall (ΔP_e) and effective available energy (ΔA_e) to Eqs. (18) and (15). The GCM changes in interception were added to the calculated value of dE_e to produce the estimated changes in total evaporation (dE), as shown in the e plots. Plots g and h of each figure make use of the linear characterization to diagnose the relative sensitivities of each model to changes in throughfall and available energy. A detailed description for each region follows, and section d provides a summary of the main points.

5. Discussion

We have analyzed results from eight climate change experiments, carried out with four separate GCMs, each of which used two distinct land surface schemes (LSSs). In the absence of a common coupling interface, such as that currently under development in PILPS phase 4 (Polcher et al. 1998), it was not feasible to use the same LSS in more than one GCM. Instead, the differences between the LSSs were chosen at the discretion of each participating group. This approach prevented the complete diagnosis of the impact of each LSS (including feedbacks), but guaranteed that a large range of parameterizations would be used.

To understand the role of the LSSs in the GCM simulations of climate and climate change, a simple characterization has been developed based on a generic “bucket” model, in which each LSS is described by curves representing runoff (Y) and evaporative fraction (f) as a function of scaled rootzone soil moisture (μ). This characterization has been applied to isolate the impact of LSS differences on the GCM-predicted changes in surface hydrology, over three very different climatic regions (Amazonia, the Sahel, and southern Europe).

A linearization of the bucket model was able to reproduce the very different soil moisture regimes experienced by each GCM–LSS pair. In each region, the range of simulated soil moistures was shown to be dependent on the variation between both the LSSs and the control climates. There is also likely to be a feedback between the two from moisture recycling. There are considerable deficiencies in some of the simulated control climates and hydrology. For example, some models simulated low precipitation and high moisture stress over Amazonia, while others produced marginal stress in the Sahel. The most important LSS differences were summarized in terms of the runoff, Y_c, which occurs when evaporation is marginally moisture-limited (i.e., at the critical soil moisture). The higher this critical point runoff is, the more moisture stressed the model tends to be.

Our analysis suggests that the relationship between climate and evaporation sensitivity is actually highly dependent on the characteristics of the LSS. Recently, Koster and Suarez (1999) have shown that the sensitivity of evaporation rate to climate variability can be understood in terms of a local aridity index. They have also shown this to be the case for the Amazonian simulations presented here (R. Koster and M. Suarez 1999, personal communication). The success of these two different methodologies, suggests that the atmosphere and land are closely coupled, with not only the atmosphere influencing the surface state, but also the LSS altering the atmospheric state. For example, over Amazonia the LSSs that had a tendency to be stressed, produced low evaporation that correlated with unrealistically low rainfall.

The variation in the LSS responses has been analyzed by partitioning the behavior into two parts; throughfall and available energy sensitivity. In very arid regions, evaporation is primarily dependent on throughfall changes, but this sensitivity decreases significantly with increased critical runoff. In moist climates, the models may range from being mainly sensitive to available energy changes, if they are unstressed, to being predominantly sensitive to throughfall, if moisture stressed.

Such a large spread was simulated over Amazonia. This explained why only four out of the six LSSs without CO₂-induced stomatal closure simulated a reduction in evaporation, even though they all had qualitatively similar forcing perturbations. The models singled out here tended not to be moisture limited and were therefore more dependent on the available energy increase than the rainfall reduction. Such gross variation in the land surface response can occur in any region that experiences perturbations in precipitation and available energy that are of opposite sign. In the other regions studied there was also a substantial scatter in the LSS sensitivities.

In some respects, the use of regional averages complicates the analysis, because the nonlinear nature of the evaporative fraction and runoff curves makes them sensitive to soil moisture heterogeneity. However, the idealized, piecewise linear f–μ curve still appears to offer a valuable conceptual model, separating the moisture-limited and non-moisture-limited regimes clearly. Also, while soil moisture heterogeneity might be expected to enhance regional-scale runoff, differences among the LSS runoff characteristics were much larger than could be explained by different subregional soil moisture distributions (not shown). Instead, the different LSS runoff characteristics are probably indicative of assuming different parameters at the gridpoint scale.

For example, the LSS used in the Hadley Centre “b” experiment shares a similar soil model structure to that used in the University of Reading simulations. Both LSSs use Darcy’s equation to simulate vertical flow between four soil layers and each describes the variation of soil water suction and hydraulic conductivity with soil moisture concentration according to Clapp and Hornberger (1978). Both schemes also have similar definitions of the wilting and critical soil moisture concentrations (defined as the soil moistures for which the suction equals −1.5 MPa and −0.0033 MPa, respectively). The major differences are in the choice of extremely variable soil parameters such as the saturated soil water suction and the saturated hydraulic conductivity. The different values chosen yield values of Y_c of 1.3 and 0.014 mm day⁻¹ for Ua (Ub) and Hb, respectively, in Amazonia. Such a discrepancy partially explains the very different moisture regimes simulated in these experiments.

Overall, our analysis supports many of the conclusions of Koster and Milly (1997). We have found that differences among the GCM predictions of changes in surface hydrology are particularly sensitive to how the runoff is parameterized over the active soil moisture range. Variations in the runoff at the critical soil moisture point (below which evaporation is moisture limited), account for much of the spread in the GCM predictions of hydrological change. Reducing the uncertainties in these predictions will require improved treatments of runoff appropriate for the GCM spatial resolution, rather than further sophistication in the treatment of evapotranspiration at the point scale.