Moist convection is a key aspect of the extratropical summer climate and strongly affects the delicate balance of processes that determines the surface climate in response to larger-scale forcings. Previous studies using parameterized convection have found that the feedback between soil moisture and precipitation is predominantly positive (more precipitation over wet soils) over Europe. Here this feedback is investigated for one full month (July 2006) over the Alpine region using two different model configurations. The first one employs regional climate simulations performed with the Consortium for Small-Scale Modeling Model in Climate Mode (CCLM) on a grid spacing of 25 km. The second one uses the same model but integrated on a cloud-resolving grid of 2.2 km, allowing an explicit treatment of convection. Each configuration comprises one control and two sensitivity experiments. The latter start from perturbed soil moisture initial conditions.
Comparison of the simulated soil moisture–precipitation feedback reveals significant differences between the two systems. The 25-km simulations sustain a strong positive feedback, while those at 2.2-km resolution are associated with a predominantly negative feedback. Thus the two systems yield not only different strengths of this key feedback but also different signs. This has important implications, with the cloud-resolving model exhibiting a shorter soil moisture memory and a smaller soil moisture–temperature feedback.
Analysis shows that the different feedback signs relate to the sensitivity of the simulated convective development to the presence of a stable layer sitting on top of the planetary boundary layer. In the 2.2-km integrations, dry initial soil moisture conditions yield more vigorous thermals (owing to stronger daytime heating), which can more easily break through the stable air barrier, thereby leading to deep convection and ultimately to a negative soil moisture–precipitation feedback loop. In the 25-km integrations, deep convection is much less sensitive to the stable layer because of the design of the employed convective parameterization. The authors also show that there are considerable differences in the simulated soil moisture–precipitation feedback between low-resolution modeling frameworks using different cloud convection schemes.
Interactions between land and the atmosphere play an important role in our climate system and have been the focus of modeling studies on a wide range of time and spatial scales (e.g., Pielke 2001; Betts 2004; Koster et al. 2004; Schär et al. 1999). On seasonal time scales, soil moisture anomalies may maintain/amplify subsequent extreme events such as droughts, floods, or summer heat waves (e.g., Beljaars et al. 1996; Trenberth and Guillemot 1996; Fischer et al. 2007a,b), and this has important implications for seasonal forecasting (e.g., Douville and Chauvin 2000; Ferranti and Viterbo 2006). On centennial time scales, land–atmosphere interactions have been shown to interfere with European summer climate variability and its sensitivity to climate change (e.g., Schär et al. 2004; Rowell et al. 2006; Seneviratne et al. 2006b; Vidale et al. 2007). On weather forecasting time scales, poorly specified initial soil moisture conditions have been found to have a detrimental impact on the skill of numerical weather predictions (NWP) (e.g., Trier et al. 2004; Sutton et al. 2006). There is also indirect observational evidence that soil moisture affects subsequent precipitation (e.g., Findell and Eltahir 1997; Koster et al. 2003; Taylor and Ellis 2006), although such evidence is limited in space and time owing to the lack of adequate soil moisture observations. Using accumulated precipitation anomalies as a proxy for soil moisture anomalies, observational studies have established their importance for seasonal forecasting (Della-Marta et al. 2007b; Vautard et al. 2007) and have detected trends in interannual summer temperature variability (Della-Marta et al. 2007a).
Several process studies have specifically investigated the mechanisms controlling the soil moisture–precipitation feedback. Soil moisture anomalies can affect subsequent precipitation mainly via an enhanced advection of water vapor into a region due to changes in the large-scale synoptic setting (e.g., Pal and Eltahir 2003; Cook et al. 2006), a direct recycling of soil moisture into precipitation within the same region (e.g., Entekhabi et al. 1992), or locally via a modification of the boundary layer characteristics (e.g., Betts et al. 1996; Schär et al. 1999; Findell and Eltahir 2003). This latter mechanism, called indirect soil moisture–precipitation feedback, is thought to make a dominant contribution to summertime convective precipitation under weak synoptic-scale forcing (e.g., Schär et al. 1999).
The indirect soil moisture–precipitation feedback has been mainly explained in two different ways. A first indirect feedback, as addressed by a series of studies (e.g., Betts et al. 1996; Eltahir 1998; Schär et al. 1999; Pal and Eltahir 2001; Findell and Eltahir 2003), has been linked to differences in low-level moist static energy (MSE) (or equivalently in equivalent potential temperature) existing between drier and wetter soil conditions. Drier soil moisture conditions yield smaller latent heat fluxes, larger sensible heat fluxes and thus higher Bowen ratios and deeper planetary boundary layers (PBLs). A deeper PBL implies a smaller MSE per unit of PBL air (for a same total amount of MSE). Also, a deeper PBL is associated with more vigorous entrainment of above-PBL air with low MSE, which acts to decrease the MSE of the PBL. Hence, drier soils tend to be associated with lower MSE per unit of PBL air and thus with a reduced potential for convective development. From this perspective, one expects wet soils to favor precipitation, which constitutes a positive soil moisture–precipitation feedback. In other words (Findell and Eltahir 2003), convection is triggered by bringing the level of free convection down to the PBL top through a strong increase in equivalent potential temperature (i.e., moistening). Moreover, Schär et al. (1999) and Pal and Eltahir (2001) have indicated that the reduction in net shortwave radiation over wet soils from the corresponding increase in cloud cover can be overcompensated by changes in net longwave radiation. This actually results in more net radiation at the earth’s surface over wet soils. The accompanied increase in the total heat flux (sum of latent and sensible heat) and thus in MSE further favors convective development.
Recently, Findell and Eltahir (2003) have highlighted a second indirect soil moisture–precipitation feedback. They have exploited differences in PBL height existing between drier and wetter soils. Again, drier soil moisture conditions yield smaller latent heat fluxes, larger sensible heat fluxes, and thus higher Bowen ratios and deeper PBLs. Although the lifting condensation level and the level of free convection are higher over dry soils than over wet soils, the more rapidly growing PBL can overcompensate this difference and reach the level of free convection in the dry soil case, while in the wet soil case the PBL remains too shallow. Instead of triggering convection by bringing the level of free convection down to the PBL top through a strong moistening, convection is here triggered by bringing the PBL top up to the level of free convection through a strong warming. In this case, precipitation is favored over dry soils, sustaining a negative soil moisture–precipitation feedback.
The aforementioned process studies have used either climate models of relatively coarse horizontal resolution or highly simplified models (e.g., conceptual boundary layer models). The first ones allow a full interaction of the land surface scheme with the planetary boundary layer and the atmosphere, but these interactions have to rely on uncertain parameterizations owing to the coarse mesh size of the models. In particular, convective processes, which are at the core of the soil moisture–precipitation feedback, have to be parameterized. Yet it is well known from idealized and real-case numerical experiments (e.g., Bechtold et al. 2004; Guichard et al. 2004) that convective parameterizations cannot realistically capture many aspects of convective development. Conceptual boundary layer models allow for a detailed and explicit representation of PBL processes. However, they are generally driven by external data and do not include the full range of interactions. The simulations in Findell and Eltahir (2003), for instance, stopped with the appearance of the first clouds, thus preventing cloud feedbacks. Given the complexity of the processes involved and their crude or incomplete representation in numerical models, it is not surprising that different models and/or parameterization choices have yielded distinct soil moisture sensitivities (e.g., Pan et al. 1996; Gallus and Segal 2000; Guo et al. 2006). Dirmeyer et al. (2006) have even suggested that global models cannot properly represent the feedback between land and the atmosphere as they do not realistically capture the observed relationships between surface and atmospheric state variables.
The overall goal of this study is to investigate the soil moisture–precipitation feedback in simulations with explicit and parameterized convection. More specifically, we seek first to assess the magnitude and sign of the soil moisture–precipitation feedback over the Alpine region. Second, we aim to uncover the specific mechanisms controlling the feedback loops and to explain the role of the use/nonuse of a convective parameterization.
To achieve our goals, we will employ two sets of simulations. The first one uses cloud-resolving resolution (2.2 km) to allow an explicit resolution of moist convection. The second one employs a lower-resolution grid spacing of 25 km, where convection is parameterized after Tiedtke (1989). To assess the soil moisture–precipitation feedback, the initial soil moisture content is perturbed. The integrations are carried out for one full month over the Alpine region and over Europe for the cloud-resolving and the lower-resolution simulations, respectively. The month considered, July 2006, was characterized by extremely warm temperatures over Europe, weak synoptic-scale forcing, and enhanced convective activity. Except for two frontal episodes with embedded convection (5–9 and 27–30 July), precipitation exclusively resulted from convective activity due to strong surface heating (see Météo France 2006; MeteoSwiss 2006; Hohenegger et al. 2008). Hence, this study mainly addresses processes related to the indirect soil moisture–precipitation feedback (see above).
Both the 2.2-km and 25-km simulations make use of a comprehensive regional climate model that allows full interactions between the land surface, the planetary boundary layer, and the atmosphere. The 2.2-km integrations allow a better representation of topography and surface fields, an explicit resolution of a larger portion (i.e., toward the smaller wavelengths) of the spectrum of atmospheric motions, and do not rely on a convective parameterization. It has been shown by means of idealized (e.g., Chaboureau et al. 2004; Guichard et al. 2004), short-range NWP (e.g., Mass et al. 2002) and monthlong (Hohenegger et al. 2008) integrations that such simulations are able to more realistically capture and simulate convective precipitation.
The outline is as follows. Section 2 describes the model and our numerical experiments. The presentation of the results follows in the subsequent three sections. Section 3 briefly summarizes the main characteristics of the control integrations before documenting the magnitude and sign of the soil moisture–precipitation feedback in the two model configurations. The mechanisms controlling the feedback loops for the 25-km and 2.2-km integrations are uncovered in sections 4a and 4b, respectively, while the differences are explained in section 4c. The sensitivity of these results to different aspects of the chosen model configurations is tested in section 5. Conclusions are given in section 6.
The simulations are performed with a preliminary release of version 4 of the Consortium for Small-Scale Modeling Model in Climate Mode (CCLM). The CCLM is a nonhydrostatic limited-area model that numerically solves the fully compressible equations for a moist atmosphere. Its formulation and design allow its use for both weather and climate applications, with spatial resolution and temporal scales ranging between 1 and 50 km and 1 day and 50 yr (see especially Steppeler et al. 2003; Doms and Förstner 2004; A. Will et al. 2008, unpublished manuscript). In this subsection we provide a brief description of the model and its parameterizations as used for the low-resolution (25 km) experiments. Specific aspects of the high-resolution (2.2 km) version are addressed in the next subsection.
The default convective parameterization of the CCLM is the Tiedtke mass-flux scheme with a moisture-convergence closure (see Tiedtke 1989). The triggering function employs a single surface air parcel ascent and does not account for entrainment of environmental air into the rising parcel. The scheme distinguishes between shallow and deep convection depending upon the strength of the moisture convergence. A CAPE-based mass-flux closure can be used instead of the default moisture-convergence closure, if desired. An alternative convective parameterization, the Kain–Fritsch–Bechtold scheme (see Bechtold et al. 2001, 2004), is also implemented in the CCLM. The latter features a CAPE closure for both shallow and deep convection. The depth of shallow clouds is restricted between 0.5 and 2 km, while deep clouds are thicker than 2 km. As a further contrast to the default Tiedtke scheme, the triggering function uses multiple mixed layer parcel ascents and includes the effect of larger-scale (i.e., grid scale) vertical motion on the parcel temperature.
The parameterization of land surface processes is performed with the multilayer soil model designed by Heise et al. (2006). (see also Doms et al. 2005). The scheme considers infiltration, percolation, capillary movement, and melting and freezing of snow as well as evapotranspiration (from interception store, snow store, bare soil, and plants) and runoff. The soil water content is predicted based on the Richards equations, while interception and snow storage are governed by appropriate budget relations. Temperature is predicted by implicitly solving the heat conduction equation. In the employed configuration, the soil model contains 10 soil layers with thicknesses of 0.01, 0.03, 0.06, 0.12, 0.24, 0.48, 1.06, 1.72, 4.04, and 7.48 m. The top 7 layers have a prognostic water content. The soil parameters (e.g., pore volume, field capacity, heat capacity, wilting point) are prescribed according to one of eight predefined soil types. Vegetation is specified by means of prescribed seasonally varying leaf area index, fractional plant cover, and root depth.
The other physical packages of the CCLM include a turbulence scheme using turbulent kinetic energy (Raschendorfer 2001), a two-category ice scheme (cloud ice and snow, see Kessler 1969), and a radiative transfer scheme after Ritter and Geleyn (1992). The latter deals with the radiative properties of aerosols, gases, and clouds, where the considered clouds encompass the contribution from subgrid-scale, grid-scale, and convective clouds. Subgrid-scale cloud cover and cloud water content are parameterized after Slingo (1987) as a function of the gridbox relative humidity and specific humidity at saturation, respectively. They only affect radiation and turbulence and cannot precipitate. In contrast, grid-scale clouds can precipitate. They are passed over from the cloud microphysical scheme. As soon as cloud water or cloud ice is present in a grid box, the total cloud cover amounts to 1. Finally, the convective cloud cover is diagnosed in the CCLM implementation used by the convective parameterization. In the case of the Tiedtke and Kain–Fritsch–Bechtold schemes, it is a function of cloud depth. The depth itself is computed by the triggering function through the performed parcel ascents.
b. Numerical experiments
To investigate the soil moisture–precipitation feedback in simulations with explicit and parameterized convection, two main sets of integrations are conducted. Each of them consists of one control and two sensitivity experiments. All the simulations are carried out for one full month, July 2006.
The first set comprises the low-resolution regional climate simulations performed with the CCLM, where convection is parameterized with the default Tiedtke scheme (with moisture-convergence closure). The control experiment uses a horizontal resolution of 0.22° (25 km), 32 levels in the vertical, 10 soil layers, and a time step of 120 s. The grid covers the European continent (see Fig. 1a) and contains 193 × 201 points. Initial and lateral boundary conditions are derived from the European Centre for Medium-Range Weather Forecasts operational analysis except for the initial soil moisture. The latter comes from a 48-yr (1958–2006) analysis-forced 25-km CCLM integration. This ensures that the initial soil moisture of the control experiment is approximately within the equilibrium of the CCLM and thus avoids a strong spinup of the model. The control simulation is called CTL25 (see Table 1). Its setup corresponds to the current state of the art in regional climate modeling and has been applied for extended climate simulations (see, e.g., Jaeger et al. 2008).
The second set includes the simulations integrated at cloud-resolving scales to allow an explicit representation of moist convection. The grid exhibits a mesh size of 0.02° (2.2 km) and spans the whole Alpine region (see Fig. 1b). The domain contains 501 × 301 grid points, 45 vertical levels, and 10 soil layers, while the time step is 40 s. Initial and lateral boundary conditions (including soil moisture) of the control experiment are provided by CTL25. The interpolation procedure from the 25-km to the 2.2-km grid uses bilinear interpolation in the horizontal and tension splines in the vertical. The control experiment is referred to as CTL2 (see Table 1). Both CTL25 and CTL2 are identical to the two integrations named CCLM25 and CCLM2 in Hohenegger et al. (2008).
The finer model mesh size used in CTL2 allows an explicit representation of deep moist convection (no convective parameterization is used) but also imposes some further changes in the CCLM setup (see section 2a) with respect to CTL25. The 2.2-km simulations employ Runge–Kutta in place of leapfrog time stepping, graupel as a third hydrometeor category (see Reinhardt and Seifert 2006), a prognostic treatment of the hydrometeors, and a smaller turbulent length scale. The external parameter fields (e.g., topography, soil texture) also differ as they have been derived from higher-resolution observational datasets for the cloud-resolving simulations.
For each of the two configurations, the sensitivity experiments are obtained by perturbing at the initial time (0000 UTC 1 July 2006) the soil water content of CTL25 and of CTL2, respectively. Observations of terrestrial water storage indicate a soil moisture variability of ±40% over Europe (see Hirschi et al. 2007), while the 48-yr 25-km CCLM simulation only exhibits a ±10% variability. Underestimation of soil moisture variability is a common feature of low-resolution regional climate models (see Hirschi et al. 2007). Given that July 2006 already starts from dry soil moisture initial conditions, we choose an initial anomaly of ±30%. The latter is applied uniformly over the full soil depth and to all the grid points of the 25-km and 2.2-km domain, respectively. In cases where the perturbed soil moisture exceeds the water-holding capacity, the additional soil water is instantaneously converted into runoff. The sensitivity experiments starting from the positive and negative soil moisture anomaly are named WET25, WET2 and DRY25, DRY2, respectively. Except for the initial perturbation in soil moisture, they use the same boundary conditions and setups as CTL25 and CTL2. Note that we will also utilize in the following the generic terms CTL, WET, and DRY to refer to CTL25 and CTL2, WET25 and WET2, as well as DRY25 and DRY2.
To investigate the robustness of the results obtained with the aforementioned numerical experiments, four additional sets of simulations are performed, each containing a full set of DRY, CTL, and WET simulations (see Table 1). The first two sets investigate the role of the convective parameterization on the soil moisture–precipitation feedback. In the first one, the Tiedtke moisture-convergence closure is replaced by a CAPE closure, while the second one employs the Kain–Fritsch–Bechtold scheme as convective parameterization (see section 2a). The respective simulations (see Table 1) are denoted with suffix TCAPE and KFB. In the third set of additional experiments (denoted with suffix NOSC), the parameterization of subgrid-scale clouds is discarded at 2.2-km resolution. This experimentation is motivated by Hohenegger et al. (2008), who found a strong sensitivity of the simulated precipitation and cloud amounts to the treatment of subgrid-scale clouds at 2.2 km. Finally, the last set tests the role of the orographic representation, which—besides the use/nonuse of a convective parameterization—constitutes the major difference between the 2.2- and 25-km simulations. The integrations are denoted with suffix TOP25. They employ cloud-resolving resolution but a smoothed topography roughly corresponding to that of the 25-km simulations (see Table 1). Except for the mentioned differences, all these integrations are identical to their parent 25-km or 2.2-km simulations.
a. Control integrations
As the two control integrations CTL25 and CTL2 have been discussed and validated in detail in Hohenegger et al. (2008), we here only briefly summarize their main findings. Compared to observations, CTL25 and CTL2 underestimate the monthly mean 2-m temperature by up to 1.5°C over the northern part of the domain and the Alps. This underestimation is less pronounced at 2.2-km resolution. Over the Po Valley, CTL25 and CTL2 slightly overestimate 2-m temperature, but the biases are generally smaller than 1°C. The simulations also exhibit too high specific humidity values in the planetary boundary layer and a generally too small saturation deficit.
In terms of precipitation, CTL25 and CTL2 reproduce the overall precipitation pattern and sequence of precipitation events characterizing July 2006. The location of the rainfall maxima is better captured in CTL2 than in CTL25 owing to a more realistic orographic representation (see Fig. 2 of Hohenegger et al. 2008). Compared to a high-resolution rain gauge dataset available over Switzerland, mean precipitation is overestimated by 0.012 mm h−1 and 0.03 mm h−1 in CTL25 and CTL2, respectively. However, there are considerable uncertainties in the observations, with radar data suggesting an underestimation (rather than overestimation) of 0.08 mm h−1 and 0.062 mm h−1 in CTL25 and CTL2 for the same region, respectively. The underestimation in CTL2 especially manifests itself over flat terrain and/or in cases of weak convective rainfalls (see Hohenegger et al. 2008). The main deficiency in CTL25 lies in its representation of the convective diurnal cycle, a problem common to atmospheric models using parameterized convection (see, e.g., Bechtold et al. 2004; Brockhaus et al. 2008). CTL25 exhibits a precipitation peak that is 4 h earlier than observed, accompanied by a too early onset and a too rapid decay of precipitation (see Hohenegger et al. 2008). The convective diurnal cycle is much better captured in CTL2. The onset of convective precipitation is delayed by about 2 h, the time of peak precipitation is shifted by a similar period, and the decay of convective activity in the afternoon is slowed down (see, e.g., Figs. 2b,d). Given the better timing of the precipitation diurnal cycle and the more realistic rainfall pattern in CTL2 as compared to CTL25, the higher-resolution control simulation provides a more reliable representation of precipitation over the Alpine region for the month considered.
b. Precipitation and soil moisture responses
Figure 2 illustrates the time evolution and the mean diurnal cycle of precipitation (averaged over the Alpine domain of Fig. 1b) in the six simulations WET25, CTL25, DRY25, WET2, CTL2, and DRY2. It is evident from Figs. 2a,c that all simulations qualitatively reproduce a similar precipitation evolution, in the sense that they all capture the two episodes of strong synoptically forced rainfalls from 5 to 9 and 27 to 30 July, and the two periods dominated by diurnal convection on 2–4 and 11–26 July.
Examining the differences between WET, CTL, and DRY in Figs. 2a,b versus Figs. 2c,d reveals that significant differences exist in the sign and magnitude of the simulated soil moisture–precipitation feedback. For most of the days (see Figs. 2a,c), WET25 sustains stronger rainfall than CTL25 and DRY25, while the reverse is the case at 2.2-km resolution. The implied difference in feedback sign is well evident in the monthly means (Figs. 2b,d) but is most prominent during the periods dominated by diurnal convection. In both sets of simulations, the precipitation sensitivity appears larger with respect to a dry than to a wet anomaly (asymmetric response).
Inspection of the mean diurnal cycles in Figs. 2b,d further indicates that the negative soil moisture–precipitation feedback in the cloud-resolving integrations predominantly follows from differences in the amplitude of the simulated convective afternoon rainfall peak. Both model resolutions yield smaller or identical precipitation over dry soil moisture conditions in the morning hours (0100–1200 UTC), while only the cloud-resolving resolution sustains larger rainfall over dry soils in the afternoon.
Figure 3 shows accumulated precipitation in CTL as well as precipitation differences between WET and DRY integrations for the two resolutions. All of the quantities have been computed over the period of diurnal convection (2–4 and 11–26 July), where the differences between the two resolutions in terms of feedback sign are the largest (see above). For this time period, the area-mean precipitation (averaged over the Alpine domain) amounts to 19 mm in CTL25 versus 16 mm in CTL2. As indicated by Figs. 3a–c, CTL25 and CTL2 agree with each other; that is, they both yield stronger rainfall over the Alpine ridge. However, the finescale pattern and the absolute rain amounts differ, which is not surprising given the fact that the two integrations use very distinct topographies and convective representations.
Despite the similarities in the large-scale CTL precipitation pattern, consideration of Fig. 3 confirms that significant differences exist in the magnitude and sign of the simulated soil moisture–precipitation feedback between the two resolutions. Over most of the domain (see Fig. 3d), WET25 produces much stronger rainfall than DRY25, indicating a predominantly positive feedback. In contrast (see Fig. 3e), WET2 yields less precipitation than DRY2 over most grid points, implying an overall negative feedback. The differences between the two resolutions still persist when the output of the 2.2-km simulations is downgraded onto the 25-km grid (see Fig. 3f). Furthermore, the overall precipitation sensitivity appears weaker at 2.2-km than at 25-km resolution. The mean precipitation decrease amounts to 16 mm in DRY25 (with respect to WET25 and as averaged over the Alpine domain), while DRY2 exhibits an increase of 7 mm (with respect to WET2). These findings are in full agreement with Fig. 2.
While the precipitation response of the simulations using explicit and parameterized convection strongly differs, the 25-km results are consistent with those of Schär et al. (1999). Using another model but the same convection scheme, they also found a predominantly positive feedback for two simulated July months.
Figures 4a,b display the soil moisture evolution for the six simulations. In the near surface soil layers, all simulations show a pronounced response to the two major precipitation events on 5–9 and 27–30 July (see Fig. 4a), as expected, while the lower layers exhibit a drying trend throughout the integration (see Fig. 4b). The initial soil moisture differences are maintained throughout the simulations but decay with time. Overall, the cloud-resolving integrations lead to a faster recovery of the soil moisture content toward CTL as compared to the 25-km simulations. This is true not only for the uppermost soil layers in Fig. 4a, but also for the total prognostic soil moisture content in Fig. 4b. Figure 4c shows the corresponding evolution of the soil moisture anomalies plotted on a logarithmic scale along with the associated values of the half-life time t1/2. The latter is calculated assuming an exponential decay of the anomalies and omitting the first five days because of spinup. The 2.2-km simulations result in t1/2 values of about two months, whereas WET25 yields three months and DRY25 more than five months. These differences between 25- and 2.2-km resolutions are due to the switch from a positive to a negative feedback (see Figs. 2, 3), while the shorter memory in WET25 compared to DRY25 relates to the occurrence of saturation in the upper soil layers of WET25 (and thus enhanced runoff in WET25 due to the high precipitation amounts in this experiment). Other modeling studies conducted with regional climate models have found t1/2 values of 1.7–3.9 months (Schär et al. 1999) and 3–4 months (Fischer et al. 2007b). From the autocorrelation values given in Vinnikov et al. (1996) (for an observational site near Moscow) and Seneviratne et al. (2006a) (for a range of global climate models), values of 1.8–2.4 months and 1.5–5 months can be deduced, respectively.
4. Process analysis
The previous section has documented the significant differences existing in the magnitude and sign of the soil moisture–precipitation feedback between the 25- and 2.2-km simulations, while the differences in terms of mean precipitation pattern and the temporal evolution between the two CTL simulations remained small. This section seeks to uncover the mechanisms sustaining a predominantly positive and negative feedback loop in the two systems by considering the diurnal evolution of the surface energy balance and of the boundary layer, as in previous studies on this subject (e.g., Schär et al. 1999; Findell and Eltahir 2003). For simplicity, we focus on the WET and DRY integrations during the period of strong diurnal convection (2–4 and 11–26 July) when the differences are the largest. Similar conclusions hold for the pairs WET and CTL or CTL and DRY and the full integration period.
a. Positive feedback at 25-km resolution
Figure 5 shows the mean diurnal cycle of the different components of the surface energy budget, 2-m temperature, 2-m dewpoint depression, and total cloud cover for the 25-km simulations averaged over the Alpine domain and the period of diurnal convection. WET25 exhibits a reduced sensible heat flux (by 36 W m−2 in the mean) and an enhanced latent heat flux (by 36 W m−2) as compared to DRY25. The associated Bowen ratio increases from 0.24 in WET25 up to 0.99 in DRY25. WET25 is also characterized by colder 2-m temperatures (by 4.7°C) and a smaller dewpoint depression (by 7.4°C) than DRY25, in agreement with the increased fraction of energy that is carried away from the surface by latent rather than sensible heat fluxes. Qualitatively, these changes are as to be expected from the higher soil moisture content of WET25 (see, e.g., Eltahir 1998; Betts 2004).
In terms of radiation and cloud cover (see Fig. 5), WET25 shows a larger cloud cover (by 17%), smaller net shortwave (by 37 W m−2), and larger incoming and smaller outgoing longwave (31 W m−2 difference) radiation at the earth’s surface. These changes in solar and thermal radiation are consistent with the increased cloud cover and the colder temperatures in WET25 (see, e.g., Schär et al. 1999; Findell et al. 2007). The increased cloud cover is consistent with the smaller dewpoint depression and the enhanced precipitation.
These results stand in full agreement with those of Schär et al. (1999) and others. As indicated in the introduction, the change in Bowen ratio between WET25 and DRY25 implies a change in the depth of the PBL, where the corresponding shallower PBL in WET25 favors the development of moist convection. In contrast to Schär et al. (1999), the decrease in net solar radiation in WET25 is not overpowered by the differences in net longwave radiation. Nevertheless, the slight (6 W m−2) reduction in net radiation at the earth’s surface is insufficient to break the positive feedback loop between soil moisture and rainfall, as it is not accompanied by a reduction in the total flux of heat (sensible and latent heat) from the surface into the boundary layer.
b. Negative feedback at 2.2-km resolution
Figure 6 is the analog to Fig. 5 but for the cloud-resolving simulations. WET2 exhibits lower sensible heat fluxes (by 43 W m−2), higher latent heat fluxes (by 47 W m−2), colder temperatures (by 2.5°C), and a smaller dewpoint depression (by 4.8°C) as compared to DRY2. These differences are in principle consistent with a positive feedback and with the depleted soil moisture content in DRY2. In terms of cloud–radiative effects, WET2 is associated with a larger cloud cover (by 8%) as well as smaller net shortwave (by 17 W m−2) and larger incoming and smaller outgoing longwave (17 W m−2 difference) radiation at the earth’s surface. The differences in radiation are in accordance with the changes in cloud cover and temperature. The increased cloud cover in WET2 is consistent with the smaller dewpoint depression.
Figures 5 and 6 thus indicate that the sensitivities of the 25-km and 2.2-km simulations upon an imposed soil moisture anomaly are in the same direction for radiative and energy fluxes, temperature, dewpoint depression, and cloud cover. However, these differences are generally less pronounced at 2.2-km resolution, in agreement with the predominantly negative feedback, which allows a faster recovery of the soil moisture content in WET2 and DRY2 (see section 3b). Only the precipitation response differs between the two resolutions. Despite a larger cloud cover and especially despite a smaller Bowen ratio and a slightly larger flux of total heat into the PBL in WET2 as compared to DRY2, WET2 produces smaller rainfalls. These considerations imply the existence of a negative feedback mechanism that is able to overpower the positive correlation between Bowen ratio and precipitation expected from the 25-km results.
To pin down such a mechanism, we consider first a single representative day. Figure 7 displays precipitation and cloud cover evolution on 13 July as well as a tephigram at 1200 UTC for WET2 and DRY2. All of the quantities have been averaged over an area of 66 × 44 km2 located over the Alps. The striking differences between the two simulations relate to the larger precipitation amounts in DRY2 (Fig. 7a), the deeper and denser (convective) clouds in DRY2 (Figs. 7b,c), and the existence of a more pronounced layer of stable air (with respect to a moist adiabat) aloft the boundary layer (between 720 and 650 hPa at 1200 UTC) in WET2 (Fig. 7d). In DRY2, a layer of slightly enhanced stability can also be recognized (between 650 and 600 hPa), but the effect is much less pronounced. These characteristics suggest the following feedback loop. DRY2 and WET2 both exhibit convective instabilities in the early morning of 13 July over the considered subdomain, which would allow deep convection to be triggered, at least from a parcel ascent point of view. In the course of the morning, initially shallow clouds begin to form in both simulations (see Figs. 7b,c). Their base is higher in DRY2 than in WET2, in agreement with a larger dewpoint depression and thus higher lifting condensation level in DRY2. The formation of shallow clouds is consistent with the existence of a stable layer sitting on top of the PBL (with respect to the corresponding moist adiabat, see Fig. 7d) that prevents deepening of the convective cells (see, e.g., Zhu and Albrecht 2003). In WET2 the clouds remain shallow throughout the day (see Fig. 7b), while in DRY2 thermals start to break through the stable layer at 1200 UTC, leading to the formation of deep convective clouds (Fig. 7c). The latter are responsible for the substantial amounts of rain recorded in DRY2 from 1200 UTC onward (Fig. 7a). As compared to WET2, DRY2 thus benefits from its stronger surface warming that overpowers the stable air barrier. Stronger valley wind circulations, as generally observed in DRY2 (not shown), might further contribute to this effect.
The generalization of these results to the Alpine domain and period of diurnal convection is illustrated by Fig. 8, which shows mean cloud cover and a tephigram. The computation of the cloud cover at each grid point is based on grid-scale cloud liquid water content (see section 2a) and ignores the contribution of subgrid-scale clouds and cloud ice (to provide a clear signature of the convective cells in the free troposphere). Comparison of Figs. 8a and 8b indicates that, in the lower atmospheric layers (below about 700 hPa), WET2 produces more clouds than DRY2. For a given cloud cover, the cloud base is lower in WET2 than in DRY2, in agreement with a lowering of the mean lifting condensation level from 789 hPa in DRY2 to 846 hPa in WET2. These changes are as to be expected from the changes in sensible and latent heat fluxes, in dewpoint depression, and from thermodynamic considerations (see Fig. 6 and Findell et al. 2007). In contrast, above about 600 hPa, the cloud cover is larger in DRY2 than in WET2. Figures 8a,b show that shallow clouds in DRY2 transform into deep convective cells during the afternoon, while in WET2 they remain predominantly confined to the vicinity of the PBL top. Regarding the stratification above the PBL top, Fig. 8c reveals the existence at 1200 UTC of a slightly stable layer in WET2 (with respect to the corresponding moist adiabat) that is absent in DRY2. This is consistent with both Figs. 8a,b, wherein clouds are confined below 700 hPa in WET2 but extend well into the troposphere in DRY2.
It is interesting to note that already at 0600 UTC the air aloft the PBL is more stable in WET2 than in DRY2 (not shown). This follows from the larger number of low-level clouds produced in WET2 than in DRY2 (see Figs. 8a,b), which feed back on the temperature profile through their radiative properties. In particular, longwave radiative cooling at their top strengthens the stability of the above-PBL air. The production of widespread low-level clouds in WET2 is itself tied to its lower dewpoint depression and smaller Bowen ratio. In a way, the smaller Bowen ratio in WET2, by reducing the surface warming and enhancing the formation of low-level clouds, seems to actually penalize the development of deep convective cells and the production of strong precipitation, especially in environments characterized by an increased stability above the PBL.
Hence, in the cloud-resolving simulations, increased soil moisture content in WET2 yields a moister boundary layer that tends to favor cloud formation, similar to the low-resolution simulations. The development includes a first phase characterized by the widespread production of shallow convective clouds, as also described in idealized studies (see, e.g., Chaboureau et al. 2004; Guichard et al. 2004). However, the produced clouds only rarely deepen owing to the presence of a shallow layer of stable air sitting on top of the PBL. In contrast, decreased soil moisture content in DRY2 yields less shallow but more frequent deep convective cells. The generation of deep convective clouds is linked to the presence of more vigorous thermals, which are more likely to break through the existing stable layer. Also, the production of widespread low-level clouds during the early morning in WET2 (related to its smaller dewpoint depression) tends to strengthen the stable air barrier. Ultimately, the control exerted by this stability barrier explains the predominantly negative feedback in the cloud-resolving simulations.
The proposed negative feedback mechanism differs from the one advanced by Findell and Eltahir (2003) to explain a preferential triggering over dry soil conditions (see introduction). In particular, Findell and Eltahir (2003) did not explicitly consider the effects of above-PBL stability and clouds on the development of convection, as their simulations stopped with the appearance of the first clouds. The role of above-PBL stability is, however, documented in Ek and Holtslag (2004). They investigated the development of boundary layer clouds (but not of precipitation) and their sensitivity to changes in soil moisture by means of coupled one-dimensional land surface–atmospheric boundary layer simulations conducted at Cabauw (Netherlands).
Finally, we emphasize that the proposed negative feedback mechanism can only operate under conditions where clouds may form. If the soil and/or atmosphere is too dry and thus the dewpoint depression too large, no cloud will form and this will prevent a negative feedback loop. In this sense, the soil moisture–precipitation feedback depends upon the climate and land characteristics of the region considered, an idea that is familiar from Koster et al. (2004).
c. Reasons for the different feedback signs
Both negative and positive feedback mechanisms appear physically plausible. The question thus arises why the 25-km simulations favor a positive feedback. This issue is examined here in more details.
Figure 9 is the equivalent of Fig. 8 for the 25-km simulations. It shows mean cloud cover and a tephigram for WET25 and DRY25. The cloud cover is split into grid-scale cloud cover (computed as in section 4b and Figs. 8a,b for the 2.2-km simulations, i.e., without considering cloud ice and subgrid-scale clouds) and convective cloud cover (diagnosed by the Tiedtke scheme as depending upon cloud depth; see section 2a). Comparison of the different panels in Fig. 9 and of Fig. 9 versus Fig. 8 reveals several interesting features. WET25 produces a larger cloud cover than DRY25 below about 700 hPa. This is in agreement with the cloud-resolving integrations in Fig. 8 and with the effects expected from a smaller Bowen ratio (see previous sections). The existence in WET25 of a layer of stable air can also be recognized in Fig. 9c between 850 and 650 hPa with a temperature lapse rate smaller than the corresponding moist adiabat. Such a stability barrier cannot be detected in DRY25 (see Fig. 9c). Again, this is in agreement with Fig. 8 and our previous findings. However, unlike the cloud-resolving simulations, WET25 and DRY25 exhibit a similar frequency of deep convective clouds. WET25 even produces more convective clouds at all levels as compared to DRY25 during the first simulation days. This stands in clear opposition to the cloud-resolving simulations, where WET2 nowhere exhibits a larger cover of deep clouds than DRY2.
There is a subtlety concerning the convective cloud cover in DRY25 as displayed by Fig. 9b. Although DRY25 yields a similar amount of deep clouds during the afternoon as compared to WET25 (see Figs. 9a,b), the former clouds only produce weak rainfalls (see Fig. 2b). This follows from the pronounced surface warming and dewpoint depression in DRY25 (see, e.g., Fig. 5), which reduces the amount of precipitable water and also enhances the evaporation of falling precipitation. Moreover, the amount of precipitation is not directly related to the cloud cover in the Tiedtke scheme but is linked to the mass flux and the closure. Other schemes might thus result in different cloud cover changes (see next section).
Figure 9 thus indicates that the presence of an above-PBL layer of stable air does not inhibit the positive feedback loop by restricting the formation of deep convective cells in WET25. Note that the increase in stability is even slightly more pronounced in WET25 than in WET2. The different cloud evolutions displayed by Figs. 9 and 8 may be tied to the use of the Tiedtke convective parameterization, as this constitutes the core difference between the two systems investigated. As can be seen in Figs. 9a,b, there is no gradual transition from shallow to deep convective clouds at 25-km resolution. The Tiedtke scheme tends to produce either shallow or directly deep convection. The distinction is based on moisture convergence and thus does not explicitly account for above-PBL stability. For the example considered in Fig. 7, the Tiedtke scheme would have, for instance, generated deep convective cells in the WET run. This effect also expresses itself in the different fraction of deep to shallow clouds generated at the two resolutions. This fraction amounts to only 10% in DRY2 (see Fig. 8) against about 70%–80% in WET25 (see Fig. 9).
The specific design of the Tiedtke mass-flux scheme in terms of triggering function and closure (see section 2a) may further overemphasize a positive feedback loop (see also next section). Since the triggering function does not consider entrainment of environmental air into the rising parcel, it is relatively easy to trigger moist convection, even with only moderate surface warming. Moreover, the moisture-convergence-based closure favors convection with wet soil conditions by design.
5. Sensitivity analysis
The previous section has documented the key role played by the presence of an above-PBL layer of stable air and has related the differences between the soil moisture–precipitation feedback at 2.2- and 25-km resolution to the Tiedtke mass-flux scheme. To test our ideas, we investigate here the precipitation response for the four sets of additional experiments using slightly different parameterization/setup choices (see section 2b and Table 1). Simulations with suffix TCAPE and KFB test the role of the convective parameterization by replacing the moisture–convergence closure in the Tiedtke scheme by a CAPE closure and by employing the Kain–Fritsch–Bechtold scheme instead, respectively. The integrations with suffix NOSC and TOP25 investigate the role of subgrid-scale clouds and high-resolution topography at 2.2-km resolution and will be discussed further below in more details.
Figure 10 shows the mean diurnal cycle of precipitation for the different sensitivity experiments. Figure 10a indicates that WET25_TCAPE produces stronger rainfalls than DRY25_TCAPE and generally more than CTL25_TCAPE. As compared to the standard Tiedtke mass-flux scheme in Fig. 2b, the differences between the integrations are nevertheless weaker. In contrast, WET25_KFB yields a smaller precipitation peak than CTL25_KFB and DRY25_KFB (see Fig. 10b). The Kain–Fritsch–Bechtold scheme thus exhibits a strong negative soil moisture–precipitation feedback and is as such more akin to the cloud-resolving results (see Fig. 2d). It is interesting to note in this context that Guichard et al. (2004) and Bechtold et al. (2004) both found that Kain–Fritsch-type schemes were best able to reproduce the convective development over land with respect to cloud-resolving models.
The different magnitudes of the simulated soil moisture–precipitation feedback indicate a significant sensitivity to the employed closure, triggering function, and in general design of the convective parameterization. The replacement of the moisture convergence in Tiedtke by a CAPE closure and the resulting weaker sensitivity confirm that a closure based on moisture considerations likely accentuates a positive feedback loop (see section 4c). The negative feedback sign obtained with the Kain–Fritsch–Bechtold scheme is noteworthy. Inspection of the cloud cover evolution (not shown) indicates that, similar to the cloud-resolving results, the formation of deep cells is almost suppressed in WET25_KFB, while the cloud cover below 700 hPa considerably increases. Since CTL25_KFB already exhibits a more pronounced increase in stability at the top of the PBL than both CTL2 and CTL25, the Kain–Fritsch–Bechtold scheme likely overemphasizes the negative feedback. As Bechtold et al. (2001) noted, the Kain–Fritsch–Bechtold scheme tends to produce weaker vertical fluxes across the upper part of the shallow cumulus layer compared to large-eddy simulations, which eventually leads to too humid and cloudy conditions at low levels and to reduced deep convective activity. Pan et al. (1996) as well as Gallus and Segal (2000) also obtained different precipitation responses upon identical initial soil moisture perturbations by switching from the Kuo to the Grell scheme and from the Betts–Miller–Janjic to the Kain–Fritsch scheme, respectively. Both studies investigated synoptically forced convective events over the United States.
Clouds, and especially shallow low-level clouds which do not precipitate, represent another critical factor in the feedback loop. They contribute to the stabilization of the above-PBL air and to a reduced surface warming and, in this sense, favor the development of a negative feedback loop. To test this hypothesis, we consider the simulations with suffix NOSC in which subgrid-scale clouds have been discarded. Figure 10c shows that DRY2_NOSC yields larger precipitation amounts than CTL2_NOSC and WET2_NOSC. However, as compared to Fig. 2d, the differences between the integrations are smaller, thus revealing a weaker negative soil moisture–precipitation feedback. Discarding the formation of subgrid-scale clouds reduces the simulated total cloud cover by 10% and, especially, the amount of low clouds by 15%. This allows for a larger warming at the earth’s surface and for the generation of stronger thermals that are more likely to break through the stable layer and trigger deep convection, even in the WET run. Comparison of the temperature profiles between WET2 and WET2_NOSC further confirms the presence of a weaker stable air barrier in WET2_NOSC (not shown).
Finally, to investigate the relative role of high- versus low-resolution topography, we consider the simulations with suffix TOP25. They employ a smooth orographic representation roughly corresponding to that used at 25-km resolution. Figure 10d reveals that DRY2_TOP25 produces more precipitation than CTL2_TOP25 and WET2_TOP25. The differences between the integrations are roughly akin to the ones in Fig. 2d. Hence, the use of a higher-resolution topography in the 2.2-km simulations relative to that at 25 km is not of major importance for the simulated magnitude and sign of the soil moisture–precipitation feedback, although it clearly affects the simulated rainfall amounts in each integration. In other words, specific circulation patterns generated by the 2.2-km topography play a negligible role and cannot explain the different feedback signs between the two resolutions.
Besides the aforementioned sensitivity studies, we have tested further aspects of the 25-km and 2.2-km simulations over short periods of time. In particular, the role of the different vertical resolutions and domain sizes between the two systems has been investigated. We have also tested the effect of switching on a parameterization for shallow convection at 2.2 km, as the latter resolution is not sufficient to properly resolve shallow convection. All of these supplementary experiments have sustained a positive feedback with 25-km and a negative feedback with 2.2-km grid spacing, thus confirming our previous results.
In this study, we investigated the soil moisture–precipitation feedback over the Alpine region in simulations with explicit and parameterized convection. To that aim, we conducted two sets of simulations with grid spacings of 0.02° (2.2 km) and 0.22° (25 km), each set comprising one control integration and two simulations with perturbed (±30%) soil moisture initial conditions. All of our integrations were carried out for one full month, July 2006. The main results of the study are the following:
The two model configurations exhibit significant differences in the sign and magnitude of the simulated soil moisture–precipitation feedback. The 25-km integrations sustain a strong positive feedback, where an increase in soil moisture yields an increase in precipitation. In contrast, a negative feedback loop develops at 2.2-km resolution with the drier run actually producing more precipitation than the wetter run (where dry and wet refer to the initial soil moisture content).
The mechanisms controlling the positive and negative feedback loops act as follows: In both systems, positive initial soil moisture anomalies yield an enhancement of latent (at the expense of sensible) heat flux and a moister shallower boundary layer, which by itself would tend to favor the development of moist convection. However, both systems also exhibit a shallow layer of stable air located immediately above the PBL. This layer acts as a barrier for the triggering of deep convective cells. It can be overcome through the production of vigorous thermals. This presupposes a strong surface warming and thus tends to favor the development of moist convection over dry rather than wet soil conditions. In the case of the Tiedtke parameterization using a moisture-convergence closure, the first effect (moistening) dominates and leads to enhanced precipitation over wet soils, irrespective of the presence of the stable layer. With explicit convection, the second effect dominates and thus leads to an increase of precipitation over dry soils. The different responses mainly follow from the use of a convective parameterization at 25-km resolution and from the specific design of the Tiedtke mass-flux scheme.
Besides distinct sensitivities at 2.2- and 25-km resolutions, low-resolution integrations employing different convective parameterizations also exhibit substantial differences in the simulated soil moisture–precipitation feedback. Using an instability-based closure in place of the default moisture convergence reduces the strength of the feedback signal. Replacing the Tiedtke scheme with the Kain–Fritsch–Bechtold scheme even produces a negative feedback—as obtained in the cloud-resolving simulations.
Our study has some significant limitations. First, the results may depend on the considered domain and integration period. It cannot be excluded that flatter integration domains could sustain somewhat different results. The existence of the above-PBL layer of stable air may be in this respect a peculiar property of the domain or month considered. From a climatological perspective the selected month appears to be at least well suited, as it entails long convective periods with an intense diurnal cycle. Also, the fact that the same feedback sign is reproduced over many diurnal cycles consolidates our results. The consideration of longer integration periods is hampered at the moment by the large computational costs of cloud-resolving simulations. Second, the comparison of parameterized against resolved convection involves a number of parameterization choices. We have tested some of them in detailed sensitivity studies, yet there are additional choices related, for instance, to the representation of the cloud microphysics, turbulence, and land surface processes. Third, it is generally accepted that higher resolution than 2.2 km is desirable for the explicit representation of moist convection. It is possible that cloud effects are overemphasized in the 2.2-km simulations, which—in general—have some difficulties in triggering deep convective cells and in producing precipitation (e.g., Hohenegger et al. 2008). Nevertheless, the latter integrations are associated with a better precipitation diurnal cycle than their lower-resolution counterparts. Moreover, the analysis of the underlying processes has resulted in a coherent interpretation of the key results (i.e., preference for positive versus negative soil–precipitation feedback).
Notwithstanding these limitations, our study suggests that there are considerable uncertainties regarding the representation of the soil moisture–precipitation feedback (and likely land–atmosphere feedbacks in general) in current global and regional climate models. In particular, moisture-convergence closures, such as used in the original Tiedtke scheme, may likely overemphasize soil moisture–atmosphere interactions. This result might have important implications for seasonal forecasting and climate change studies. In particular, a system with a negative feedback and thus a shorter soil moisture memory will tend to more rapidly damp the effects of an anomalous soil water content. Our critique of convective parameterizations appears also consistent with the general experience regarding the overall unsatisfactory performance of low-resolution atmospheric models in representing extratropical summer precipitation climates (e.g., Jacob et al. 2007; Lenderink and van Meijgaard 2008). Further studies addressing these difficulties and investigating in detail the different model sensitivities, also by using observationally based analyses, are thus urgently needed. Ultimately, we consider the use of cloud-resolving models in climate studies as a promising avenue into the future, but the inflicted computational costs imply that this avenue is not immediately available.
The numerical simulations have been performed on the CRAY XT-3 at the Swiss National Supercomputing Centre (CSCS) in the framework of the Swiss-ALPS Grant. Partial support for this study has been provided by the Swiss National Science Foundation through NCCR Climate and by the European Commission (ENSEMBLES project). The COSMO and CCLM communities as well as MeteoSwiss are acknowledged for providing access and support to the model. The authors would also like to thank Erich Fischer, Bjorn Stevens, and the reviewers for useful comments/discussion as well as Daniel Lüthi for technical support.
* Current affiliation: Department of Atmospheric Sciences, University of Washington, Seattle, Washington.
Corresponding author address: Cathy Hohenegger, Department of Atmospheric Sciences, University of Washington, Box 351640, Seattle, WA 98195-1640. Email: firstname.lastname@example.org