1 Introduction

Numerical studies based on atmospheric general circulation models (GCMs) provide more and more evidence that the Earth’s climate is the result of a dynamic equilibrium in which the atmosphere affects the ocean and land surfaces and is affected by them. The land-atmosphere interaction has received a particular attention in global climate models due to the following reasons:

  1. 1.

    Whereas the ocean surface fluxes mainly depend on one ocean variable, namely sea surface temperature (SST), the land surface fluxes are influenced by a large number of parameters, including surface temperature but also soil moisture (SM) and many others.

  2. 2.

    Unlike SST, the land surface shows steep gradients on both the regional and subgrid scales, due to the presence of the orography, the variety of the physiography and the heterogeneous distribution of soil moisture and snow.

  3. 3.

    While SST evolves relatively slowly compared to the atmosphere, the land surface also shows a strong diurnal cycle and day-to-day variability.

As a result, the land surface cannot be treated as the ocean boundary conditions in atmospheric GCMs. Whereas monthly mean SSTs are commonly prescribed, the space-time variability of the land surface is usually represented through the use of land surface models (LSM) that basically compute water and energy budgets within each grid cell of the GCM. Such models are all the more necessary since various land surface parameters such as SM are not easy to measure, either with in situ instruments or with current remote sensing techniques, so that global climatologies are not available to prescribe these parameters in GCMs. This is a major obstacle in designing numerical experiments aimed at exploring the influence of SM on climate variability and predictability. Early studies have been based on extreme scenarios in which the soil was either dry or saturated (Shukla and Mintz 1982). Non-interactive experiments have been also proposed in which the land surface variables (SM but also possibly snow cover and surface albedo) are prescribed at each grid point based on the climatological values computed from a former interactive model run. Using this experiment design, Delworth and Manabe (1988, 1989) showed that interactive SM may substantially increase the variability of near-surface temperature and relative humidity, especially in summer at midlatitudes. Moreover, they found that SM anomalies may persist at a monthly to seasonal time scale, suggesting the relevance of initial SM conditions for seasonal climate simulations.

Global estimations of SM can now be derived from operational land surface data assimilation systems. Early assimilation techniques were very empirical, but nevertheless suggested a positive impact on short range weather predictions (Viterbo 1996) and were implemented in multi-year reanalyses such as the ERA15 dataset (Gibson et al. 1997). More recently, an optimum interpolation technique has been developed at Météo-France, that is based on the sequential assimilation of 2-m measurements of temperature and humidity (Douville et al. 1999; Giard and Bazile 2000). This method has been tested successfully against in situ observations (Douville et al. 2000) and is now also used at ECMWF, including in the recent 40-year ECMWF reanalysis project (ERA40). Another possibility for initializing SM in numerical weather prediction models is to drive LSMs with analyses of precipitation and solar radiation. This method is being currently tested at the National Center for Environmental Prediction (NCEP) and at Météo-France (over a limited domain). The main difficulty is to produce reliable real-time analyses of precipitation. This is why it is much easier to use this method for a reanalysis of past SM conditions when many more observations are available to specify the precipitation forcing.

Such a technique has been already used in the framework of the Global Soil Wetness Project. GSWP was launched by the Global Energy and Water Cycle Experiment (GEWEX) to provide high-resolution global SM climatologies (IGPO 1998). In a first step, the atmospheric forcing was provided on a 6-hourly basis from January 1987 to December 1988 on a 1 by 1° horizontal grid. Among other participants, the COLA (Center for Ocean-Land-Atmosphere studies) and CNRM (Centre National de Recherches Météorologiques) brought an original contribution to this project. After they produced an SM climatology with their respective LSMs, they performed ensembles of global seasonal atmospheric simulations for summers 1987 and 1988, in which SM was prescribed or relaxed based on their GSWP outputs (Dirmeyer 2000; Douville and Chauvin 2000; Douville et al. 2001; Douville 2002). These case studies have highlighted the relevance of using realistic SM boundary conditions for simulating seasonal precipitation anomalies.

With the recent development of long-range atmospheric forecasts, a more difficult question has been asked of the modelling community: what is the relevance of SM initialization for dynamical seasonal predictions? Despite the lack of reliable and/or long enough land surface analyses, several case studies were conducted to tackle this problem. Fennessy and Shukla (1999) showed that the impact of initial SM was mainly local and was largest on near-surface fields. Douville and Chauvin (2000) found that the impact was significant but weak in many areas due to the limited persistence of SM anomalies and to the possible drift of SM in the model. Kanamitsu et al. (2003) showed that the predictive skill of SM was much higher in arid/semiarid than in wet areas due to the low predictability of precipitation and found a significant initial SM impact on near-surface temperature. Conversely, Dirmeyer (2003) obtained no appreciable change in skill by using more realistic initial SM. This lack of sensitivity was attributed to a strong SM drift in their model, that was formerly discussed in Dirmeyer (2001). By leading to extremely dry or wet soil conditions, the drift was likely to break the SM feedback loop that, according to the author, would be otherwise generally positive and would help to support predictability of interannual variations.

Though not very encouraging, such predictability studies will have to be repeated over longer periods and/or with more reliable SM datasets (such as the forthcoming GSWP2 climatology) when they are available. Meanwhile, more idealized experiments remain useful to investigate the potential influence of SM at the monthly to seasonal time scales. The most straightforward experiment design is to remove the interannual SM variability in ensembles of seasonal hindcasts or in multi-year integrations. The simulations are then verified against one of the ensemble members rather than real observations. With this “perfect model” assumption, the aim is not to evaluate the real ability of the model to capture observed climate anomalies, but to assess the upper limit of the prediction skill (“potential” predictability) and its sensitivity to SM. Following the pioneering works of Delworth and Manabe (1988), Koster et al. (2000) conducted a series of multi-year GCM simulations with observed or climatological prescribed SSTs and with interactive or prescribed evaporation efficiency over land. They found that the influence of the land surface is important in the summer hemisphere midlatitudes and that the foreknowledge of SM boundary conditions contributes significantly to the potential predictability of precipitation in the transition zones between dry and humid climates. This study was recently revisited by Reale and Dirmeyer (2002), who found a stronger sensitivity to interannual variations in the land surface forcing in their climate model (especially in the tropics). These contrasting results pointed out the sensitivity of the model response to the model formulation, a problem that was also emphasized by Koster et al. (2002) using multi-model ensembles of July atmospheric simulations.

Another important issue in such idealized studies is the possible influence of the technique used to control the land surface variability. This question was recently addressed by Douville (2003) who explored the influence of SM on boreal summer climate variability and potential predictability with the ARPEGE (Action de Recherche Petite Echelle Grande Echelle) atmospheric GCM. The study was based on ensembles of seasonal hindcasts using either interactive or relaxed deep SM boundary conditions. The relaxation was made toward either the ARPEGE or the more realistic GSWP climatology (Douville 1998). When relaxed toward its own SM climatology, the model almost simulates the same mean climate as in the control experiment. This result is not trivial since it was not obtained in former studies where the interannual variability of the land surface hydrology was suppressed by prescribing either a monthly mean evaporation efficiency (Koster et al. 2000; Reale and Dirmeyer 2002) or the step-by-step evolution of all land surface prognostic variables (Koster et al. 2002). Such a neutral impact on the model climatology is however important, since mean and variance are sometimes dependent, particularly in the case of variables with a skewed probability distribution function (pdf) like evaporation and precipitation.

Besides this technical issue, the main objective of Douville (2003) was to assess the influence of interactive versus climatological SM boundary conditions on the variability and potential predictability of boreal summer climate. Although less important than SST variability, SM variability was shown to have a significant impact on the atmospheric variability. The impact on low-level temperature was found in both relaxed experiments and did not strongly depend on the prescribed SM climatology (ARPEGE or GSWP). The precipitation response was less robust, since the effect of the decreased evaporation variability could be offset by an increase in mean precipitation when the relaxation is made toward GSWP. SM was also shown to influence the potential prediction skill of the ARPEGE model. Both reduced and enhanced predictability of low-level temperature and precipitation were found. This heterogeneous response was attributed to the contrasted relevance of the SM feedback among different regions, but also to contrasted SM anomalies in the initial conditions.

The present study is the continuation of Douville (2003). The main objective is to better understand the model response found in the former experiments by running an additional ensemble in which the SM relaxation is activated only in June. Indeed, this new experiment allows us to compare the impacts of initial and boundary conditions of SM. While most previously discussed studies explored only one of the two impacts, they did not really answer a crucial question about the status of SM in seasonal climate predictions: is it mainly an initial value problem? This question was recently addressed by Schlosser and Milly (2002) who found a persistent impact of initial SM on near-surface temperature. They however used a simplified bucket LSM and forced their GCM with climatological SST, which limits the relevance of their results. The key issue is indeed to know whether SM is an additional source of predictability compared to the SST contribution. Moreover, their experiments focused on the role of SM initialization, while SM could also contribute to predictability by amplifying the atmospheric signals driven by the SST variability.

Based on a fully consistent set of experiments, the present study uses a classical analysis of variance to investigate the influence of both initial SM and SM feedback on ensembles of boreal summer simulations. Section 2 briefly describes the model, the experiment design and the statistical technique that is used to analyse the results. Section 3 describes the model climatology and shows the response of climate variability and potential predictability on both regional and global scales. Section 4 summarizes the results and gives concluding remarks.

2 Methodology

2.1 The ARPEGE climate model

The ARPEGE climate model originates from the ARPEGE/IFS (Integrated Forecast System) numerical weather prediction model developed jointly by Météo-France and ECMWF (European Center for Medium-range Weather Forecast). It is a spectral atmospheric model with a hybrid σ-pressure vertical coordinate. Since the first release of the ARPEGE climate model (Déqué et al. 1994), many developments have been included, both in the dynamics and the physics. Here we use ARPEGE version 3 (Déqué et al. 1999) whose main novelty compared to the former version is the semi-Lagrangian numerical integration scheme with a 30 min time step. The physical package includes the turbulence scheme of Louis (1982), the statistical cloud scheme of Ricard and Royer (1993), and the mass-flux convective scheme with Kuo-type closure of Bougeault (1985). The radiative scheme is derived from Morcrette (1990) and is activated every 3 h. More details about the physics can be found in Geleyn et al. (1995). In the present study, the model is used with 31 vertical levels and a T63 triangular truncation. The linear Gaussian grid associated with this truncation should be a 128 × 64 longitude–latitude grid, but for computation efficiency we use a reduced Gaussian grid with a poleward decrease in the number of longitudes outside the tropics.

At the Earth’s surface, the ISBA land surface scheme is used to provide a boundary condition to temperature and moisture (Mahfouf et al. 1995). The only change made in ISBA since version 2 of ARPEGE is the implementation of a soil freezing scheme. Compared to the original parametrization described by Noilhan and Planton (1989), the model includes a four-layer heat diffusion scheme, a more physical treatment of snow cover (Douville et al. 1995) and a deep drainage (Mahfouf and Noilhan 1996). It has three prognostic variables for liquid water: the surface volumetric water content, w s , the total volumetric water content, w p , and the reservoir of rain intercepted by the canopy, W r . They evolve according to the following equations:

$$\frac{{\partial w_{s} }}{{\partial t}} = \frac{{C_{1} }}{{\rho _{w} d_{1} }}(P_{g} - E_{g} ) - \frac{{C_{2} }}{\tau }(w_{s} - w_{{seq}} )$$
(1)
$$ \frac{{\partial w_{p} }} {{\partial t}} = \frac{1} {{\rho _{w} d_{2} }}(P_{g} - E_{g} - E_{{tr}} ) - \frac{{C_{3} }} {\tau }\max (0,w_{p} - w_{{fc}} ) - \frac{\delta } {{\tau _{{clim}} }}(w_{p} - w_{{clim}} ) $$
(2)
$$ \frac{{\partial W_{r} }} {{\partial t}} = veg\,P - (E_{v} - E_{{tr}} ) - R_{r} $$
(3)

The soil hydrology, Eqs. (1) and (2), assumes the same type of force-restore method for water as for heat conduction (Deardorff 1977). P g is the flux of liquid water reaching the soil, E g the evaporation at the soil surface, E tr the transpiration rate, ρ w the density of liquid water and τ the duration of 1 day. While w s is the surface volumetric water content (arbitrary depth d 1 of 1 cm), w p is the total (not the deep) volumetric water content (d 2 is the total soil depth that is prescribed according to the soil and vegetation types). The restore term in Eq. (1) is not directly estimated as a function of the vertical gradient of moisture, but involves an equilibrium value, w seq , which is reached when gravity balances the capillarity forces. Like the coefficients C 1 and C 2, w seq is a function of soil texture and moisture (Noilhan and Planton 1989). The last right-hand term in Eq. (2) represents a nudging toward climatological SM, w clim , that is usually not activated (δ = 0) but will be tested in the present study. The deep drainage coefficient, C 3, characterizes the velocity at which the water profile is restored to the field capacity, w fc , and also takes account of the soil texture (Mahfouf and Noilhan 1996). Equation (3) describes the evolution of the interception reservoir that increases with precipitation, P, over the vegetated fraction of the grid cell, veg, and decreases with interception loss, E v – E tr , and dripping, R r. Details about the calculation of the different water fluxes can be found in Noilhan and Planton (1989).

2.2 The experiment design

In the present study, the ARPEGE AGCM is used to perform global simulations in which SM is either interactive or relaxed toward a monthly mean climatology. Each experiment consists of J = 15 ensembles of n = 10 members, one ensemble for each JJAS (June to September) season between 1979 and 1993. All members for a specific year are forced by the AMIP I (Gates 1992) observed monthly mean SSTs and differ only in their initial conditions. In such seasonal hindcasts, the method used to produce the members is not very important since the growth of initial errors saturates after a few weeks. Here the 15-year ECMWF reanalyses (ERA15, Gibson et al. 1997) are used to initialize both the atmospheric and land surface variables. For each season, the n members are generated by adding weak perturbations to the May 27th analysis at 12 GMT. The perturbations are computed through a simple Monte Carlo technique based on the 12 GMT analyses from the nine consecutive days (around May 27th) preceding the beginning of the JJAS season. Only the atmospheric prognostic variables are perturbed so that all members of a given season share exactly the same land surface initial conditions. Note also that the vertical levels and the textural properties of the ECMWF soil hydrology are not the same as in ISBA, so that it is necessary to define the initial values of w s and w p from a simple interpolation scheme based on normalized soil wetness indices (see Douville 2003 for more details).

The three ensembles that have been achieved are summarized in Table 1. In the control experiment AAX, SM is evolving freely (δ = 0 in Eq. 2) so that the ISBA land surface model is fully coupled to the ARPEGE atmospheric model. In experiment A0X, the total soil water content is relaxed toward the ARPEGE monthly mean climatology derived from the control run (δ = 1), so that the low-frequency variability of the surface hydrology is prescribed. In the course of the model integration, the monthly mean SM climatology is interpolated linearly at a daily time step in order to avoid abrupt transitions between consecutive months. The relaxation time constant, τ clim , is uniform and fixed at 1 day. This value appeared as a reasonable compromise to avoid a drift in the monthly values while keeping a temporal coherence between SM and precipitation at the hourly to synoptic time scales. In experiment A1X, w p is relaxed toward the ARPEGE climatology only during the month of June. This additional experiment, compared to Douville (2003), is necessary to distinguish between the impacts of a damped interannual variability of SM in the boundary conditions (A0X) and in the initial conditions (A1X). Note that the results will be analyzed over the JAS (July to September) season, i.e. after a 1-month spin-up. This means that SM is perfectly climatological at the end of June in A0X and A1X, while it shows some interannual variability in AAX (but also some intra-ensemble variability due to the 1-month spin-up). Although this experiment design does not allow us to analyze the potentially strong influence of SM initialization during the first weeks of integration, the month of June has been discarded from the seasonal means to get rid of the predictability associated with the atmospheric initial conditions and thereby isolate the influence of SM.

Table 1. Summary of the seasonal hindcast experiments
Full size table

2.3 The ANOVA statistical model

The classical setting for ANOVA (analysis of variance) is agricultural experiments based on a sufficient number of plots and designed to determine the effect of different “treatments” on crop yield. In climate modelling, the experimental units are not plots but simulations, and the treatment can be, as in the present study, the use of different SST boundary conditions. The ensemble technique allows us to produce a sample for each treatment and, thereby, to isolate the seasonal climate signal related to the boundary conditions from the internal variability of the atmosphere related to the initial conditions.

In keeping with the one-way ANOVA model described by Von Storch and Zwiers (1999), we can write the total variance, S 2 of the J × n = 150 seasonal means produced in each experiment as the sum of two contributions:

$$S^{2} = S^{2}_{{bc}} + S^{2}_{{ic}} $$
(4)

where S bc 2 is the variance that is forced by the prescribed boundary conditions (estimated from the J = 15 ensemble means produced in each experiment) while S ic 2 is the variance that stems from the chaotic nature of the atmospheric dynamics (estimated as a residue). Note that the subscript “ic” here refers to atmospheric noise only, since there is no perturbation of the land surface variables within the 10-member ensembles. Conversely, the SM initial conditions can differ between two seasons of the control experiment and therefore contribute to S bc 2. Though somewhat arbitrary, this model of variance is based on the common assumption that SM shows a stronger persistency than most atmospheric variables.

The coefficient of multiple determination, R 2 = S bc 2/S 2, is the fraction of the total variance that is explained by the prescribed boundary conditions (and SM initialization in the control experiment). Though this ratio will be hereafter referred to as a measure of predictability, the reader must keep in mind that it is not the real prediction skill, but the “potential” skill, i.e. the upper limit of the model skill. The real skill, in which the forecasts are compared to real observations, is normally much lower than the potential skill, and the “perfect model” assumption used in the present study is just aimed at estimating the theoretical limit of the forecast skill. Note also that the ANOVA model is based on sums of squares rather than on actual variances. For the sake of simplicity, the symbol S 2 and the word variance will be used to refer to both quantities.

Following Von Storch and Zwiers (1999), the significance of the treatment effect can be tested simply through a Fisher test based on the following ratio:

$$F = \frac{{S^{2}_{{bc}} /(J - 1)}}{{S^{2}_{{ic}} /(J(n - 1))}}$$
(5)

The no treatment effect hypothesis can be tested at the (1 – p) significance level by comparing F computed from Eq. (5) against the p-quantile of F(J – 1, (n – 1)J). In the continuation of the present study, the significance level will be fixed at 5% and the distribution of R 2 and of changes in R 2 will be shown only in the areas where the treatment is considered as significant in the control experiment. Changes in total variance will be also tested through a Fisher test, against the q-quantile of F(n – 1, n – 1), but using a significance level of 10%.

Finally, it must be emphasized that the prescribed treatment and/or the considered climate system are different among the three experiments. In A0X and A1X, the treatment is the same (interannual variability of the SST boundary conditions), but the climate model has one more degree of freedom in A1X since SM is interactive during the JAS season. The comparison between these two experiments will therefore tell us if the SM feedback is likely to amplify the climate variability and predictability driven by the SST anomalies. In the control experiment, AAX, the treatment is slightly different since it also includes some interannual variability in the initial conditions of SM. Comparing AAX with A1X (both with interactive SM) will therefore tell us if, besides the SST forcing, SM initialization also contributes to climate variability and predictability during the last three months of a 4-month forecast.

3 Results

3.1 Overview of the JAS model climatology

We first briefly validate the ARPEGE climatology against available reanalyses or observations (Fig. 1). Figure 1a and b compare the global distribution of the mean JAS SM simulated by ARPEGE in the control experiment with the 2-year GSWP climatology. Recall that GSWP is presumably more realistic than ARPEGE since it has been obtained with the same land surface model, but with an observed atmospheric forcing (Douville 1998). Though ARPEGE captures the broad-scale patterns found in the GSWP climatology, it shows significant discrepancies on the regional scale, such as dry biases over Amazonia, Europe, Russia and the western USA. Such regional biases are not surprising owing to the strong sensitivity of the surface water budget to systematic errors in precipitation and radiative fluxes. Note however that the ARPEGE SM climatology is more consistent with GSWP when the model is integrated over consecutive annual cycles (not shown). In the present study, the SM biases are not only due to model deficiencies, but also to a poor SM initialization with the ERA15 reanalysis (Douville et al. 2000).

Fig. 1.
figure 1

Validation of the mean JAS control climate simulated in ARPEGE (control experiment AAX): a SM (kg/m2) compared with b the GSWP climatology, c 850 hPa temperature (Celsius) compared with d the ERA15 climatology, e precipitation (mm/day) compared with f the CMAP climatology

Full size image

Figure 1c and d compare the mean JAS temperature simulated at 850 hPa with the ERA15 reanalysis (Gibson et al. 1997). Although the difference between the two fields is not shown, the ARPEGE model is clearly too warm over Amazonia and the Northern Hemisphere mid-latitudes, which is consistent with the dry biases found in Fig. 1a. Douville (2003) showed that the simulated temperature is more realistic when SM is relaxed toward the GSWP climatology, thereby confirming the conclusion of Douville and Chauvin (2000) whereby the model climatology could be improved through the use of more realistic SM boundary conditions. Figure 1e and f compare the mean JAS precipitation simulated by ARPEGE with the CMAP climatology (Xie and Arkin 1996). The model is too dry over Europe, Russia and, to a lesser extent, around the Great Lakes in North America. In the tropics, the summer monsoon precipitation is well simulated over both West Africa and South Asia, but some biases are noticeable over Amazonia. Again, the climatology is generally more realistic (especially over Eurasia) when the model is integrated over consecutive annual cycles. This remark indicates that the ERA15 initialization is partly responsible for the precipitation biases found in Fig. 1e, thereby suggesting that initial SM biases can trigger a persistent precipitation bias in our seasonal integrations through a positive SM-precipitation feedback.

3.2 Control and perturbed SM variability and predictability

Before moving to the comparison of the atmospheric variability and predictability simulated in the various experiments, Fig. 2 shows the global distribution of S and R 2 for the JAS mean SM. In the control experiment, the total SM variability shows a heterogeneous distribution with very low values in the arid areas and maximum values in the rainy areas of the northern tropics (Fig. 2a). These patterns indicate that the interannual variability of SM is mainly controlled by that of precipitation, which is itself linked to seasonal mean precipitation (Fig. 1e) due to the skewed pdf of this variable. Figure 2b shows the ratio, R 2, of the total variance of the JAS mean SM that is explained by the SST boundary conditions and by the SM initial conditions. The distribution is again very heterogeneous and does not show any systematic relationship between variability and predictability.

Fig. 2.
figure 2

Standard deviation (kg/m2) and predictability of JAS SM in a, b AAX, c, d A1X and e A0X. Predictability is not shown in the relaxed experiment A0X, since variability is close to zero. No shading in b and d if the treatment effect is not significant at 5% in the control experiment

Full size image

Not surprisingly, SM variability is reduced in experiment A1X (Fig. 2c). This result demonstrates that ERA15 initial SM anomalies contribute to the total variability of the JAS mean SM in the control experiment. The decrease in R 2 is even more obvious (Fig. 2d), particularly in the northern extratropics and arid southern tropics. This is consistent with the well-known evidence that, besides the tropical rainy season, precipitation is poorly predictable at the seasonal time scale, so that the SM predictability found in AAX is partly related to the possible persistence of initial SM anomalies. SM variability is very low in A0X (Fig. 2e), so that R 2 is not really meaningful and is not shown for this last experiment (even if it can be theoretically computed since S 2 is not exactly zero when SM is relaxed rather than abruptly prescribed).

The issue is now to analyze the impact of the various SM treatments on the variability and potential predictability simulated by ARPEGE. Given the heterogeneous distribution of SM variability and predictability in the control experiment, such an analysis must be first conducted on the regional scale. It is therefore decided to focus on two specific regions, North America and South Asia, where the ARPEGE model shows: (1) a reasonable climatology (Fig. 1), (2) a significant SM predictability in the control experiment (Fig. 2b), (3) a contrasted sensitivity to SM relaxation in Douville (2003).

3.3 North American response

Figure 3a, b show the geographical distribution of S and R 2 for the JAS surface evaporation in the control experiment (AAX). Not surprisingly, the variability increases with mean evaporation and shows a northeast–southwest gradient over North America. The predictability is large over the tropical and subtropical oceans, where the divergent atmospheric circulation is strongly influenced by the prescribed SSTs. As expected, the predictability is weaker over land, but remains significant between 40° and 60°N, in the regions with strong SM predictability in Fig. 2b. Removing the interannual variability in the SM initial conditions leads to a significant decrease in evaporation variability (Fig. 3c) and to a strong reduction (by at least a factor 2) of evaporation predictability (Fig. 3d). Damping SM variability (A0X versus A1X) further and even more clearly inhibits evaporation variability (Fig. 3e), but does not lead to a clear decrease in R 2 (Fig. 3f). The evaporation predictability found over North America in the control experiment is therefore mainly due to the persistence of SM anomalies in the ERA15 initial conditions. Evaporation is poorly predictable in A1X since SM itself is poorly predictable when the ERA15 springtime anomalies are removed at the beginning of the seasonal integrations. R 2 is even increased over the central US in A0X versus A1X, thereby suggesting that the SM feedback represents a source of “noise” in this region when there is no anomaly in the initial conditions. This result must however be tempered by the fact that the background evaporation predictability found in A1X is very low, so that the statistical significance of the increase found in A0X is not guaranteed.

Fig. 3.
figure 3

ANOVA results for the mean JAS surface evaporation over North America: a total standard deviation, S, found in AAX, b coefficient of determination, R 2, found in AAX, c ratio of S 2 between A1X and AAX, d ratio of R 2 between A1X and AAX, e ratio of S 2 between A0X and A1X, f ratio of R 2 between A0X and A1X. No shading in b, d and f if the treatment effect is not significant at 5% in the control experiment

Full size image

Figure 4 shows the response of surface temperature, T s . In the control experiment, the maximum variability is found in the eastern USA where evaporation variability is also strong. In keeping with the impact on surface evaporation, A1X and mostly A0X show a decreased variability of surface temperature (Fig. 4c, e) indicating that latent heat plays a significant role in the summertime surface energy budget over North America. Also consistent with the evaporation response is the decreased predictability of T s in A1X versus AAX (Fig. 4d). However, removing SM variability (A0X versus A1X) has a clearer impact on surface temperature (general increase in R 2 in Fig. 4f) than on surface evaporation. This result will be further discussed in Sect. 4. It confirms that the use of interactive SM does not amplify the climate signals controlled by the prescribed SSTs, whereas initial SM anomalies do represent a significant source of predictability during the JAS season.

Fig. 4.
figure 4

a–f Same as Fig. 3, but for surface temperature

Full size image

Figure 5 shows the temperature response in the low troposphere (850 hPa), which is very consistent with the impact at the land surface. Superimposed onto the distribution of S 2 and R 2 are the mean JAS horizontal winds simulated in the control experiment. They are useful to understand how the large-scale circulation can shape the projection of the surface temperature signal into the lower atmosphere. Over North America, the maximum variability is found in the interior of the continent (Fig. 5a) where evaporation variability is relatively large and the large-scale circulation (either westerlies from the Pacific or southeasterlies from the Gulf of Mexico) is not likely to temper the hot spells that occur in summer. Figure 5b shows a latitudinal gradient of predictability. R 2 is maximum in the equatorial Pacific (not shown) where the El Ninõ Southern Oscillation (ENSO) is a dominant mode of variability, and is generally much less in the mid-and high-latitudes where the low-frequency modes are less important and partly masked by synoptic variability. Note the clear impact of the North Pacific subtropical high pressures that bring mid-latitude air into the tropics along the USA western coast and leads to a relative decrease in predictability in this region of the Pacific. Using climatological (A0X) instead of interactive (A1X) SM conditions leads to a clear reduction in temperature variability over North America (Fig. 5e) while temperature predictability is increased, in keeping with the land surface response. Conversely, using climatological (A1X) instead of ERA15 (AAX) initial SM conditions weakens the low-level temperature predictability, thereby confirming the relevance of SM initialization over North America.

Fig. 5.
figure 5

a–f Same as Fig. 3, but for temperature at 850 hPa. Also shown are the mean JAS wind vectors at 850 hPa in the control experiment

Full size image

The control distribution of S for JAS precipitation (Fig. 6a) illustrates the strong link between mean precipitation and precipitation variability, already emphasized by Douville (2003). Also clear is the larger predictability found in the tropics than in the extratropics (Fig. 6b), which is consistent with former dynamical studies on seasonal forecasting. Over North America, the maximum values of R 2 appear again in the interior of the continent, but barely exceed 0.2 thereby confirming the well-known result that precipitation is a difficult variable to predict at the seasonal time scale. SM has a weaker influence on precipitation variability (Fig. 6c–e) than on evaporation variability. Changes in precipitation predictability (Fig. 6d–f) are hardly significant but show a decrease in R 2 over the central USA in A1X versus AAX. In this region, SM anomalies are likely to induce anomalies not only in low-level temperature, but also in precipitation, and these anomalies are more predictable with ERA15 rather than climatological SM initial conditions.

Fig. 6.
figure 6

a–f Same as Fig. 3, but for precipitation

Full size image

3.4 South Asian response

Figure 7a, b show the geographical distribution of S and R 2 for the JAS surface evaporation simulated over South Asia in the control experiment. Note again the link between mean evaporation and evaporation variability, with a maximum variance over the Arabian Sea and the Bay of Bengal. Over land, the variability is much less, but still more than 0.5 mm/day over India. Evaporation predictability is strong over the tropical Indian Ocean (prescribed SST and relatively predictable low-level atmosphere), but hardly exceeds 0.3 over South Asia. Experiment A1X does not show important changes in variability (Fig. 7c), thereby confirming the result of Fig. 2c, i.e. the fact that JAS SM variability is mainly controlled by monsoon rainfall in this region, and is thus not much sensitive to initial SM. Conversely, A0X shows a clear reduction of evaporation variability over India (Fig. 7e). This decrease does not appear over Southeast Asia, where the soil moisture reservoir is deeper (presence of rainforests), so that the variability of SM stress is weaker than over India. Changes in R 2 do not show any clear large-scale response. Note however that Fig. 7f indicates a possible increase in R 2 over India when using climatological instead of interactive boundary conditions (A0X versus A1X). This result is consistent with the impact found in North America and suggests again that SM variability does not contribute to evaporation predictability when SM anomalies are removed from the initial conditions. Conversely, India and North America show a contrasted response in A1X versus AAX, since the ERA15 SM initialization does not clearly enhance the predictability of surface evaporation over India in AAX.

Fig. 7.
figure 7

ANOVA results for the mean JAS surface evaporation over South Asia: a total standard deviation, S, found in AAX, b coefficient of determination, R 2, found in AAX, c ratio of S 2 between A1X and AAX, d ratio of R 2 between A1X and AAX, e ratio of S 2 between A0X and A1X, f ratio of R 2 between A0X and A1X. No shading in b, d and f if the treatment effect is not significant at 5% in the control experiment

Full size image

Moving directly to the 850 hPa temperature, Fig. 8a shows regional maxima of variability over northwest and central India, while the strong monsoon flow reduces the variability over western India (along the Ghats) and over Southeast Asia. The large-scale circulation clearly affects the distribution of S, as indicated by the horse-shoe pattern of low variability over the Indian Ocean. Predictability (Fig. 8b) is high over the equatorial Indian Ocean, but limited over South Asia, except in the Ganges valley where it might be reinforced by snow mass anomalies at the end of spring on the southern slopes of the Himalayas. This hypothesis is confirmed by Fig. 8d that shows decreased R 2 in this region when removing the interannual variability in the initial conditions (A1X versus AAX). No significant change in S 2 appears in A1X (Fig. 8c) while A0X shows a clear damping of temperature variability over India (Fig. 8d), in keeping with the evaporation response. This damping is associated with a general increase in predictability over South Asia, including the north of the Indian peninsula (Fig. 8f). This result is consistent with the evaporation response and confirms that the SM feedback does not amplify R 2 over India.

Fig. 8.
figure 8

a–f Same as Fig. 7, but for temperature at 850 hPa. Also shown are the mean JAS wind vectors at 850 hPa in the control experiment

Full size image

The control distribution of S for JAS precipitation (Fig. 9a) highlights once more the strong relationship between mean and variance for this variable. Predictability (Fig. 9b) is high over the Arabian Sea, but also remains higher than 0.4 over the Indian peninsula and Southeast Asia. In keeping with former results, A1X does not show any change in variability over South Asia, while A0X indicates a reduced variability over India (Fig. 9e). Changes in R 2 are relatively noisy and indicate a weak influence of SM on precipitation predictability (Fig. 9d, f). This result is consistent with the sensitivity study of Douville et al. (2001) showing a weak soil-precipitation feedback in the ARPEGE model over India. In this region, the SM impact on surface evaporation is indeed counterbalanced by changes in moisture convergence, so that the net impact on simulated precipitation is particularly weak. The relative stability of R 2 is also consistent with the case study of Douville (2002) showing that seasonal hindcasts of the 1987 and 1988 Indian monsoon seasons are not much sensitive to the use of different SM boundary conditions.

Fig. 9.
figure 9

a–f Same as Fig. 7, but for precipitation

Full size image

3.5 Global response

The global response of our sensitivity experiments is summarized through the use of scatterplots. Figure 10 shows ratios of variance between A0X and A1X (y-axis) versus between AAX and A1X (x-axis). Both axes are in log-coordinates in order to give the same relative importance to inverse ratios of variance. Such plots are useful to compare our three experiments since the first diagonal represents equal variances between A0X and AAX. Looking first at surface evaporation and temperature (Fig. 10a, b), the dominant response is clearly a shift along the y-axis, indicating a strong damping of land surface variability when the SM feedback is removed. The impact of SM anomalies in the ERA15 initial conditions is less obvious, but still evident from the global distribution of the land grid points along the x-axis. Moving to the 850 hPa temperature (Fig. 10c), a majority of grid points still shows a reduced variance in A1X versus AAX (68%) and/or in A0X versus A1X (71%), but the amplitude of the response is generally weaker than for surface temperature. The SM impact is even less obvious for precipitation (Fig. 10d) although 70% of the land grid points still indicate a reduced variance without SM feedback, while only a short majority (54%) suggests a reduced variance with climatological initial conditions.

Fig. 10.
figure 10

Scatterplots of S 2 ratio between A0X and A1X (y-axis) versus S 2 ratio between AAX and A1X (x-axis) for all land grid points: a surface evaporation, b surface temperature, c temperature at 850 hPa, d precipitation. All axes are in log-coordinates

Full size image

Figure 11 shows similar scatterplots, but for changes in R 2 and only for land grid points with a significant treatment effect in the control experiment. Note that this precaution does not guarantee at all that the selected grid points show significant impacts on predictability, but only avoids to compute ratios between very low values of R 2. The evaporation response indicates a large sensitivity to initial and/or boundary conditions of SM (a majority of grid points are located below the first diagonal in Fig. 11a). However, both effects are highly variable among different regions and the impact of initial conditions is more obvious. Globally speaking, the damping of SM variability (A0X versus A1X) has a very heterogeneous impact on the land surface predictability. Whereas a majority (62%) of land grid points indicates a reduced predictability of surface evaporation in A0X versus A1X, Fig. 11b suggests an opposite response for surface temperature. This result suggests a contrasted impact of SM variability on the predictability of the different components of the surface energy budget. In some areas, the damping of the SM feedback could at the same time increase the predictability of the radiative and sensible heat fluxes (and thereby of T s ) and have a limited impact on the predictability of the latent heat flux. A detailed analysis of the surface energy budget would be necessary to confirm this hypothesis, but is beyond the scope of the present study.

Fig. 11.
figure 11

a–d Same as Fig. 10 but for R 2 and only for land grid points with a significant predictability in the control experiment

Full size image

Moving to the atmospheric predictability, the global response is even more difficult to assess since the scatterplots do not show a clear partitionning of the land grid points. Looking first at the 850 hPa temperature (Fig. 11c), the response of R 2 seems to be consistent with that of surface temperature. This result shows that SM is likely to affect the predictability of the low-level atmosphere, but does not demonstrate that this impact is a local effect whereby the surface latent heat flux controls the predictability of the overlying atmosphere. Moving to precipitation, both A0X and A1X show a wide range of impacts (Fig. 11d) whose statistical significance is however very difficult to assess due to the limited predictability of this variable (especially in the extratropics). Globally speaking, the damping of SM variability seems to lead more frequently (59% of the land grid points) to a decrease in predictability than for low-level temperature (42%), thereby suggesting a stronger link with evaporation predictability.

4 Summary and closing remarks

Long-range dynamical atmospheric forecasts are now produced operationally in several weather or climate prediction centres. This recent activity originates from the finding that, beyond the short-range deterministic predictability related to the initial atmospheric conditions, there is a residual probabilistic predictability, mainly due to the slow-evolving oceanic boundary conditions. In this framework, the status of the land surface remains unclear since SM is more sensitive to the high-frequency atmospheric variability than SST, but still shows some memory at the monthly to seasonal time scale. The optimal strategy for dealing with the land surface in dynamical seasonal forecasting systems is also a matter of debate. Systematic errors in the atmospheric integrations can accumulate rapidly at the land surface where a strong drift can reduce the atmospheric predictability (Dirmeyer 2003). As a consequence, several possibilities are conceivable for the land surface boundary conditions: (1) fully interactive if you are confident in your model and you think that the forecast can benefit from a realistic SM feedback and/or from realistic initial SM; (2) empirically corrected (as has been done for SST in coupled ocean–atmosphere models) or statistically predicted if you consider that SM is mainly an initial value problem and that the model drift and/or the poor predictability of precipitation exclude the use of interactive SM; and (3) climatological if you think that the SM drift is a serious problem, that the SM feedback is not relevant, and that the initialization and/or the statistical modelling of SM are not currently reliable due to the lack of observed SM time series on the global scale.

This study is a preliminary attempt to provide some guidelines for making the right decision. Nevertheless, the answer is obviously dependent on several factors, such as the quality of the model (here the ARPEGE atmospheric GCM coupled to the ISBA LSM) and the quality of the initial SM (here derived from the ERA15 reanalyses). Another important limitation of the present study is the “perfect model” assumption: the predictions are verified against a reference simulation rather than against real climate observations. In other words, it is not the real prediction skill that is explored, but the upper limit of the skill often referred to as “potential” predictability. Three ensembles of JJAS hindcasts have been conducted: a control experiment with interactive SM (AAX), a second experiment relaxed toward the ARPEGE SM climatology (A0X), and a last experiment in which the relaxation is removed after the first month of integration (A1X). The impact on variability and potential predictability of low-level temperature and precipitation has been assessed through a classical analysis of variance. As in Douville (2003), the ANOVA model has been applied to the JAS seasonal means derived from the JJAS integrations. This choice allows us to ignore the predictability related to the atmospheric initial conditions (which fully disappears after two weeks) and to focus only on the influence of SM.

Globally speaking, the results indicate that the atmospheric variability is mostly sensitive to the SM feedback, while the atmospheric predictability is more sensitive to initial SM. There are many regions where the predictability of seasonal precipitation is indeed too low for the soil-precipitation feedback to amplify the climate signals driven by the SST variability. As a consequence, the relevance of SM for dynamical seasonal forecasting appears mainly as an initial value problem. However, the positive impact of interannual initial SM is obviously limited in many areas where the initial SM anomalies are weak at the end of May or do not show any persistence from June to September. This problem might be overestimated in the present study, due to the limited confidence that we have in the ERA15 initial SM fields.

Interestingly, the climate predictability found in the ARPEGE model is sometimes enhanced by the damping of SM variability. The main reason is that evaporation predictability can increase with climatological SM. Such a response has been found over North America and India in our experiments. It can be understood if surface evaporation is expressed as the product of a SM stress factor and of potential evaporation: E = β × E p (even if this approach is derived from the bucket model and is a simplified view of the ISBA LSM). In the relaxed experiment (A0X), the SM stress is almost fixed and the variance of the actual surface evaporation can be written as S E 2 = β2 S Ep 2. In this case, R 2 can be considered as the predictability of E p rather than E. When there is no SM anomaly in the initial conditions (A1X), the JAS SM is mainly controlled by the JAS precipitation and is therefore poorly predictable in many areas (especially in the extratropics). The use of climatological SM boundary conditions (A0X) is thus likely to reveal the relatively higher predictability of E p , thereby leading to an increase in the predictability of the actual surface evaporation.

Another intriguing result is the different influence of SM variability on the predictability of surface evaporation and surface temperature. While the damping of the SM feedback leads to a dominant though not systematic decrease in evaporation predictability (Fig. 11a), it leads to an increased predictability of T s for more than half of the land grid points (Fig. 11b). It has been suggested in Sect. 3 that the reason for this could be due to a contrasting response of the predictability of the different components of the surface energy budget. Though this hypothesis cannot be properly demonstrated here, it is consistent with the scatterplots shown in Fig. 12. The aim is here to illustrate the possible relationships that exist between the impacts found in A0X versus A1X for different variables. Looking first at the link between surface evaporation and T s , Fig. 12a, b indicates that this link is clear for changes in variability, but does not appear for changes in predictability. Conversely, Fig. 12c, d shows that there is a strong relationship for changes in R 2 between T s and the temperature at 850 hPa, while the link is less clear but still apparent for changes in S 2. Damping SM variability in A0X has therefore a twofold effect: (1) a direct local impact on temperature variability, that decays with the altitude but is still obvious at 850 hPa; and (2) an indirect large-scale impact on the air temperature predictability, that also appears at the land surface but is primarily controlled by the large-scale divergence of the net energy flux within the atmosphere. This remark is also relevant for explaining the increased evaporation predictability found in A0X over the central USA and India. Besides the first explanation whereby SM is probably less predictable than E p in these regions, there is another reason whereby the damping of the SM feedback leads to a large-scale increase in the predictability of E p . Since E p is strongly controlled by net surface radiation, any increase in the predictability of the atmospheric energy convergence can lead to an increased predictability of the downward surface radiation and therefore of E p .

Fig. 12.
figure 12

Scatterplots of S 2 and R 2 ratio between A0X and A1X for all land grid points: a–c surface temperature (y-axis) versus surface evaporation (x-axis), b–d 850 hPa temperature (y-axis) versus surface temperature (x-axis), e–f precipitation (y-axis) versus surface evaporation (x-axis). R 2 ratio are displayed only when the predictability is significant in the control experiment. All axes are in log-coordinates

Full size image

A last remark can be made about the response of precipitation predictability. It has been suggested at the end of Sect. 3.5 that this response is more consistent with the evaporation response than is the temperature response. This is confirmed by Fig. 12f showing a significant relationship between changes in precipitation predictability and changes in evaporation predictability in A0X versus A1X. However, this link should not be considered as a precipitation response to the land surface forcing. It is rather due to an artefact whereby changes in precipitation predictability exert a local influence on evaporation predictability. Despite the use of relaxed total SM, the surface SM (top layer of 1 cm) remains interactive in A0X and is used to compute the bare ground evaporation. In the regions with a low vegetation cover, any change in precipitation predictability can therefore trigger a consistent change in the predictability of the total surface evaporation. This artefact explains why changes in predictability are somewhat consistent between evaporation and precipitation, but not between evaporation and surface temperature despite the contribution of the latent heat flux in the surface energy budget.

In summary, the SM impacts found on the atmospheric predictability are rather limited in our sensitivity experiments. Nevertheless, the regional impact found over North America is not negligible and provides some hope of improving dynamical seasonal forecasting in the summer mid-latitudes. Such a clear impact was not found over Europe, that could be related to the dry bias that appears in this region in our control experiment. This remark highlights again that our results are obviously dependent on both model formulation and SM initialization. Future studies should be devoted to compare the response of different GCMs and the use of different SM analyses. More efforts should be made to assess the persistence of SM anomalies and their impact on real predictability before hope becomes reality.