Flash floods, caused by extreme convective events, pose a major threat to public and private infrastructure in densely populated areas like Western Europe (Vanneuville et al, 2006). An ever-increasing built-up area and a larger number of dwellings built in areas prone to flash flooding urge the need for better prevention. Significant progress in weather forecasting plays a major role in taking precautious measures on short time-scales and in issuing warnings to the public. Advances in the physics parametrizations of numerical atmospheric models, integrated at increasingly smaller grid spacing down to convection-resolving scales, have brought significant improvements in the circulation patterns and variables with relatively low temporal and spatial variability such as pressure and temperature fields. Quantitative precipitation forecasting on the other hand has not progressed at the same pace and still deals with strong deficiencies (Hense et al, 2003).
The causes of these deficiencies are numerous, and a debate is going on in the scientific community concerning which of these causes prevails over others:
(1)Initial states of soil moisture fields are largely unknown and simple assumptions about the initial state hence introduce important errors. Soil moisture availability for instance strongly affects the soil heat capacity (Entekhabi et al, 1996) and the partitioning of surface latent and sensible heat fluxes, boundary-layer evolution and convective stability. By varying the initial volumetric soil moisture content from 60% drier to 30% wetter than a control simulation, Gallus and Segal (2000) found an increase in surface precipitation with increasing soil moisture in the Midwest United States. Wetter soils led to more than 50% more surface precipitation compared to the very dry soils for this strong convective case. Martin and Xue (2006) found a reverse relationship between soil moisture content and surface precipitation for a dryline case in western Texas and no effect during a cold-frontal situation. The proposed mechanism would be that increased soil moisture leads to a reduced surface heating and hence reducing the amplitude of convection. A similar conclusion of wetter soil suppressing convection was found by Cheng and Cotton (2004) for the development of a mesoscale convective system in Texas. In contrast to the near-surface potential temperature, water vapour mixing ratio and Convective Available Potential Energy (CAPE), Trier et al(2008) found precipitation to be much less sensitive to variations in soil moisture during a 12-day warm season period over the central United States. This apparent contradiction between studies of soil moisture influence on surface precipitation is further discussed in section 5 in the context of our findings.
(2)Precipitation formation also involves a large number of processes, ranging from microphysics processes, turbulence or even radiation, which all need to be parametrized in current operational numerical models. Microphysical processes play a key role not only in precipitation formation, but also in the thermodynamics and dynamics of the atmosphere through, for example, their impact on latent heating or radiative properties (Khain et al, 2000). McCumber et al(1991) and Cohen and McCaul (2006) found little sensitivity of surface precipitation to their microphysics sensitivity experiments for idealized convective storms, although the vertical distribution of hydrometeors changed significantly. Serafin and Ferretti (2007) found for both a stratiform and a convective event that all of the tested microphysics schemes produced similar surface precipitation and that none of them performed significantly better than others. Strongly reducing the snowfall speed in the Reisner et al(1998) scheme resulted only in a geographical redistribution of the rainfall volumes, rather than in a variation in their magnitude. Woods et al(2007) came up with somewhat different results testing a method of snow habit prediction in a bulk microphysical scheme. They found little change in the precipitation rates at the ground compared to the scheme lacking the habit prediction as long as the cold-frontal rainband was offshore, but they found a significantly increased precipitation sensitivity when the rain band interacted with mountainous areas. In contrast to the aforementioned studies, Gilmore et al(2004) found that weighting the hail/graupel category towards small graupel instead of large hail resulted in only a fourth of the surface precipitation for their idealized simulations of severe convection. Their sensitivity experiments included a very broad range of possible intercept parameters and graupel/hail densities.
In many of these fields, continuing research has been devoted to investigating their single influence on precipitation formation. However, it is still under debate how the surface precipitation is affected by combinations of these processes. A major gap of knowledge remains as to how the different processes contributing to uncertainty in the precipitation forecasts compare to each other. Many of the aforementioned studies have been carried out for idealized simulations initialized with a single sounding and a temperature perturbation. The main advantage of these studies is that processes can be simplified so that interpretation is less ambiguous. The question arises however of whether the same sensitivities occurring in those idealized simulations could be expected in real-case simulations. Further, a general drawback of idealized simulations is the lack of a ‘ground truth’ available to compare the model results with. Many studies take the most complex physics parametrization as such a ground truth, but there is no solid ground to assume that each further sophistication leads to a model improvement. Real-case simulations can be compared with observations and hence have a more reliable ground truth available.
The scope of the present research hence is to gain further insight into how modifications in two potential sources of errors in the quantitative precipitation forecast interrelate and if current insights from idealised model integrations can be confirmed for real-case simulations of severe convection across north-western Europe. Currently, operational numerical weather prediction models use a grid spacing up to only a few kilometres. In this study we focus on a model with such an operationally applied set-up. Hence a number of sensitivity studies were devised with the high-resolution non-hydrostatic Advanced Regional Prediction System (ARPS), which is specifically designed for storm-scale simulations both for atmospheric research and operational numerical weather prediction. The model was integrated for two convective events in Belgium, with changes in soil moisture initialization and model microphysical parametrization. The main objective of this work is to understand if and why these modifications have the potential of bringing significant improvements in the numerical representation of moist processes during extreme convective cases. As in many idealized case-studies, wind shear is included as an important degree of freedom; we selected two cases with very different environmental wind-shear profiles. An overview of the ARPS model and the selected cases is given in section 2. The experiment design is described in more detail in section 3. Results are given in section 4 and a discussion of the results and the conclusions are presented in section 5.
2. Model set-up and case description
2.1. Model description
ARPS is a non-hydrostatic mesoscale meteorological model developed at the University of Oklahoma (Xue et al, 2000, 2001). The finite-difference equations of the model are discretized on an Arakawa C-grid, employing a terrain-following coordinate in the vertical direction. Advection is solved with a fourth-order central differencing scheme and leapfrog time stepping. Land surface processes are parametrized following Noilhan and Planton (1989). The model was applied using one-way grid nesting with two successive domains. Data on a 0.25° horizontal grid spacing from the global operational model operated by the European Centre for Medium-Range Weather Forecasts (ECMWF) were used as initial conditions and as 6-hourly lateral boundary conditions for the model run with a 9 km grid spacing and a domain size of 1620 × 1620 km. Within this domain, a smaller domain, centred in Belgium and covering 540 × 540 km with a 3 km grid spacing was nested. An overview of the model domain is shown in Figure 1. In all simulations, 50 levels were used in the vertical with a spacing of 20 m near the surface, increasing to 1 km near the upper model boundary, which was located at 20 km altitude. All simulations were initialized with a 12-hour spin-up period, beginning at 1200 UTC on the previous day. All of the analysis in the following sections deal with the 0000–2400 UTC period (i.e. 12–36 hour forecast range), excluding the spin-up period, if not stated otherwise. Turbulence was represented by the 1.5-order turbulent kinetic energy (TKE) model, and Sun and Chang (1986) parametrization for the convective boundary layer. The Kain–Fritsch (Kain and Fritsch, 1993) cumulus parametrization was used in the largest domain, while convection was explicitly simulated in the smaller domain. Cloud microphysics was parametrized following Lin et al(1983) including five hydrometeor categories (cloud water, cloud ice, rain water, snow and hail); the corresponding mixing ratios will be denoted by qc, qi, qr, qs and qh. In order to suppress numerical noise a fourth-order monotonic computational mixing was applied, following Xue (2000).
2.2. Case description
Two cases of extreme convection in Belgium were selected. As Weisman and Klemp (1982) point to the important influence of both vertical wind shear and buoyancy on convective storm structure, two different environments regarding these two parameters have been selected. A first case has moderate buoyancy but strong wind shear, typical for supercell development, and the second has no wind shear but high buoyancy, favouring multi-cell development. The first case will be further referred to as the shear-driven case, while the latter will be referred to as the buoyancy-driven case. A more thorough description of the synoptic and mesoscale conditions is given in the following section.
2.2.1. Shear case: 1 October 2006
During the afternoon of 1 October 2006 several tornadic supercell thunderstorms developed over northern France and were advected over Belgium, causing severe damage. At the synoptic scale, a trough at the 500 hPa level extended over the British Isles with an Upper Level Low (ULL) across Ireland and a ridge extending from southern Europe to eastern Europe. Between the ULL and the ridge a strong gradient was obvious, leading to the development of a strong jet streak with winds up to 60 m s−1 at 200 hPa. The left exit region of the jet streak was positioned over Belgium during the afternoon. At the surface level, an occlusion, connected to a low-pressure area beneath the ULL, passed across Belgium during the forenoon preceding unstable air masses advected from northern France. While the static instability, based on sounding data at 1200 UTC in Trappes (France—Figure 1), was only moderate with surface-based CAPE values around 1000 J kg−1 and surface-based Lifted Index (LI) values around -3 K, the dynamical build-up of the atmosphere was much more exceptional. The low-level (0–1 km) vertical wind shear reached values up to 12 m s−1, while the 0–6 km shear amounted to 28 m s−1. Storm Relative Helicity (SRH) values reached up to 210 m2 s−2. According to Groenemeijer and van Delden (2007) these are typical values across north-western Europe for tornado-producing thunderstorms. Onset of the supercell development in northern France was around 1400 UTC. Several supercell storms lasted more than 5 hours and by then had reached the Netherlands and Germany. Localized precipitation accumulations up to 40 mm in 24 hours, and large hail and several tornadoes were reported during this period.
2.2.2. Buoyancy case 28 July 2006
Downstream of an intense upper-level long-wave trough over the eastern Atlantic, warm air advection occurred over large parts of western Europe on 28 July 2006. Beneath a weak upper-level trough positioned near Belgium and the Netherlands, a near-surface moisture convergence zone developed extending from central France across Belgium to the Netherlands, which led to convection initiation in the deeply unstable air masses during the afternoon. As the convergence line was almost stationary during the day, high precipitation accumulations of up to 100 mm in 24 hours were reported in this area. Around noon, widespread multi-cell convection initiated east of the moisture convergence line with localized high precipitation accumulation in the east of Belgium and the west of Germany. Surface-based CAPE reached high values of up to more than 2000 J kg−1 and surface-based LI amounted to about -6 K in De Bilt (the Netherlands—Figure 1) at 1200 UTC. Vertical wind shear and SRH showed only low values, whereas the bulk Richardson number value, defined here as the ratio of CAPE to the vertical wind shear, of 338 favoured multi-cell storm development. The first thunderstorms along the convergence line appeared at 1100 UTC and remained active until 2000 UTC. Severe convection in this case was obviously mainly driven by favourable thermodynamic instability whereas the dynamical situation was far from exceptional.
2.3. Observational data and SAL
The main advantage of integrating real-case simulations is to have observational data available to evaluate the model results. Precipitation observations are obtained from the C-band weather radar in Wideumont (Figure 1) operated by the Royal Meteorological Institute of Belgium and from a dense network of rain-gauges (1 per 135 km2) operated by the hydrological service of the Walloon region. Although a second radar exists in Belgium (Zaventem radar, operated by the air traffic safety agency Belgocontrol), its data are generally not used for quantitative precipitation estimation due to excessive clutter filtering at short range which partially removes precipitation. However, as no strong occultation effects were present in any of the simulated cases, precipitation patterns observed by both radars were very similar, so sufficient information could be obtained by using data only from the Wideumont radar. Radar-based precipitation estimates are derived from a five-elevation reflectivity scan performed every 5 minutes. The processing of the radar data and various strategies for merging radar observations with rain-gauge measurements are presented in Goudenhoofdt and Delobbe (2009). In this study, radar and gauge observations have been combined using a simple mean field bias adjustment. The 24-hour precipitation accumulations for the two episodes of interest have been calculated using this method and are shown in Figure 2.
In order to make a thorough comparison between observed and simulated surface precipitation, a novel verification score is applied that separately considers the Structure, Amplitude and Location errors of precipitation in a certain domain, referred to as SAL (Wernli et al, 2008). The amplitude component A measures the relative deviation of the domain-averaged simulated precipitation from the observations. Positive values of A indicate an overestimation of total precipitation, whereas negative values indicate an underestimation. For the S and L components, coherent observation objects are separately identified in the simulation and the observations. The location component L combines information about the displacement of the simulated precipitation field's centre of mass and about the error in the weighted-average distance of the precipitation objects from the total field's centre of mass. The structure component S is constructed in such a way that positive values occur if precipitation objects are too large and/or too flat, and negative values occur if the objects are too small and/or too peaked. The values of S and A components are within [−2, 2] and of the L component within [0, 2], a zero value yielding a perfect forecast. For a thorough description of the definition of each component, refer to Wernli et al(2008).
As the quality of the radar observations deteriorates with distance by e.g. attenuation effects, SAL analysis was performed for the region within a radius of 170 km from the radar location only, which is the area within the bold circle on Figures 2, 3 and 4. Surface precipitation outside this region is indicated by lighter shading and left out of the further quantitative analysis.
Land surface processes are parametrized following Noilhan and Planton (1989). This two-soil-layer surface scheme solves prognostic equations for the soil temperature and soil moisture based on a force–restore method. The scheme further treats vegetation processes, such as the surface canopy resistance to water transpiration, and storage and evaporation of intercepted rainfall. The two primary parameters of the model are dominant type of vegetation, which is derived from the Coordination Information Environment (CORINE) land cover data (Heymann et al, 1994), and soil texture which is assumed to be loamy soil texture, homogeneous across the domain. Among the secondary parameters of the land surface model, vegetation fraction is based on the normalized difference vegetation index (NDVI) from the Satellite Pour l'Observation de la Terre (SPOT) vegetation satellite imagery, with a resolution of 1 km.
Four experiments were designed in order to assess the influence of soil moisture initialization on the moist processes in both simulated cases. The CONTROL experiment had a homogeneous loamy soil texture and homogeneous soil moisture content, which was set to 29.0% for the deep soil layer and 26.0% for the skin soil layer, which are mean data for a surface and deep soil horizon, monitored over one year in a Belgian loam area by Nachtergaele and Poesen (2002). The loamy soil which was used across the domain has a saturated volumetric soil moisture content of 43.9%. This saturated moisture content was used as model initialization during the high soil-moisture sensitivity experiment (WETSOIL) for both the skin soil layer and the deep soil layer of the surface parametrization scheme. The low soil-moisture sensitivity experiment (DRYSOIL) was initialized using a soil moisture content of only 10.0% in the skin soil layer and 15.0% in the deep soil layer, which were the lowest values monitored by Nachtergaele and Poesen (2002). A last experiment (VARSOIL) explored the influence of initializing the model using spatially distributed soil texture and soil moisture data, instead of applying homogeneous fields. Soil texture data were obtained from the European Soil Database (ESDB: European Soil Bureau Network and the European Commission, 2004; see http://eusoils.jrc.it/ESDB_Archive/ESDBv2/index.htm) at a 1 km horizontal resolution, and actual soil moisture data for both cases were provided by the European Commission Joint Research Centre DESERT action at a 5 km horizontal resolution. These data were obtained from the daily integrated LISFLOOD hydrological model (De Roo et al, 2000). Domain-average surface soil moisture derived from this model was 25.8% during the shear-driven case and 21.6% during the buoyancy-driven case.
3.2. Microphysics sensitivity experiments
McCumber et al(1991) suggested that simulated moist processes are more strongly influenced by differences in the way hydrometeor size distributions are represented than by differences in the way microphysical processes themselves are treated in the schemes. They advise to increase the number of ice categories to at least four (cloud ice, snow, graupel and hail) and to explicitly predict hydrometeor particle size spectra in addition to the hydrometeor concentration. Although many such ‘multimoment’ microphysics schemes have been developed over the past decade (e.g. Ferrier, 1994; Milbrandt and Yau, 2005; Seifert and Beheng, 2006), most operational non-hydrostatic models still make use of the computationally less expensive one-moment schemes, which only have hydrometeor concentrations as prognostic variables. The microphysics scheme used in the CONTROL experiment of this study is the six water species (water vapour, cloud water, cloud ice, rain, snow and hail), one-moment bulk scheme developed by Lin (Lin et al, 1983). All falling hydrometeors are represented by exponential size distributions of the form
where N is the number of particles per unit volume per unit size range, D is the maximum dimension of a particle and N0 and λ are the intercept and slope of the exponential size distribution, respectively. The subscript x denotes the water species (rain, snow or hail). While the intercept parameter is assumed constant, slope parameters, assuming all hydrometeors to be constant density spheres, are determined by
where ρx is the hydrometeor density, qx is the hydrometeor mixing ratio (e.g. qh) and ρ is the air density.
Many of the assumptions made in this scheme are contradicted by observational studies and could be major sources of error in the quantitative precipitation forecast. First, intercept parameters of the rain and snow size distributions are observed to vary over several orders of magnitude in the atmosphere (Waldvogel, 1974; Houze et al, 1979). Second, schemes such as the Lin scheme inconsistently apply observed fall-speed–diameter relationships to calculate the terminal fall velocity, but mass–diameter relationships for constant density spheres for the calculation of the slope parameter. Third, hail and snow in the scheme are weighted towards large hailstones and fast-falling graupel-like snow respectively, which might be a good approximation for typical thunderstorms over the Midwest United States, but could be problematic in many other atmospheric conditions and in different regions.
In order to understand the implications of these drawbacks of the microphysics schemes for the simulation of moist processes and surface precipitation in the atmosphere, we performed a set of modifications in the scheme: (1) First, we diagnosed the intercept parameter of the rain and snow size distribution from the rain mixing ratio following Zhang et al(2008) and the temperature following Houze et al(1979) respectively, instead of using constant values. (2) Second, we applied observation-derived mass–diameter relationships for the calculation of the snow and graupel slope parameter instead of assuming a constant density sphere, following Stoelinga et al(2005):
where am and bm are empirically derived constants of the power-law relationship between mass mx and diameter Dx of the hydrometeor: . Both this mass–diameter relationship and the fall-speed–diameter power-law relationship are used in the derivation of the expression for the mass-weighted terminal fall speed Vx of the snow and graupel hydrometeors (Stoelinga et al, 2005):
Most current operational numerical weather prediction models only have two species of falling ice within their microphysics schemes: snow and hail/graupel. Some microphysics schemes, such as the Rutledge and Hobbs (1983) scheme, contain snow and small graupel. Other schemes, like the Lin et al(1983) scheme, contain snow and large hail. This is believed to have a significant impact on the microphysical processes; so (3) third, we modified the hail properties used in the original Lin equations to those typical for graupel. For the 1 October 2006 and 28 July 2006 cases, we performed two integrations weighting the snow towards graupel-like snow of lump type, following Locatelli and Hobbs (1974), and the hail towards large hail (MIRS) following Wisner et al(1972) and dense lump graupel (MIRSH) following Locatelli and Hobbs (1974), respectively. An overview of all modifications made in the Lin microphysics scheme is given in Table I.
Table I. Formulations for the intercept parameter (N0), slope parameter (λ), and terminal fall velocity (V) and density (ρ) for all precipitating hydrometeors for the CONTROL, MIRS and MIRSH experiments.
Initial soil moisture content hardly affects the general features of the 24-hour accumulated surface precipitation field during the shear-driven convective case of 1 October 2006, as can be inferred from Figure 3. This is confirmed by the S and L-components of the SAL analysis (Table II). The large positive values of the S-component indicate too-widespread storm fields in all experiments compared to the observed fine-scale storm structures. A wetter initial soil confines the storms tracks slightly as reflected in the somewhat smaller S-component. Total precipitation was very similar in the CONTROL, DRYSOIL, WETSOIL and VARSOIL experiments, but peak precipitation positively correlates with initial soil moisture (Table II), showing a 25% increase going from dry to saturated soil. The highest peak precipitation occurs in the VARSOIL experiment. Including variable soil properties also does not improve the misplacement of the storms, as can be inferred from Figure 3 and from the L-components in Table II
Table II. Structure (S), Amplitude (A) and Location (L) components of SAL analysis, 24 h domain average surface precipitation (RR), 24 h domain average surface precipitation for raining grid points only (RRGP), and 24 h domain peak surface precipitation (PR) as observed and for the CONTROL, VARSOIL, DRYSOIL, WETSOIL, MIRS and MIRSH experiments respectively.
Mean RR (mm)
Mean RRGP (mm)
Max RR (mm)
Mean RR (mm)
Mean RRGP (mm)
Max RR (mm)
Figure 4 shows line-averaged vertical cross-sections through the supercells along the mean mid-level wind vector. Perturbation potential temperature (defined as the departure of potential temperature from the average over its model level) is indicated by the grey shading, relative wind speeds (defined as the departure of the scalar wind from the domain average) by arrows, clouds by the thick contour and the different precipitation types by different hatching patterns. Averaging was done over the 20 cross-sections crossing the strongest peak updraughts and cross-sections are all centred in this maximum. Storm cross-sections are shown at their mature phase around 1700 UTC. In the CONTROL experiment the general features of a supercell thunderstorm are well reproduced (Weisman and Klemp, 1984; Doswell and Burgess, 1993), showing a strong updraught generated at the leading edge of the surface cold pool and accelerating upward, reaching its maximum value around 4000 m above ground level. Further, there is clear development of a downshear trailing anvil at mid and high levels, mainly consisting of snow. In combination with the excessively broad storm swaths in panels (b), (c) and (d) of Figure 3 it can be noticed that although the general storm features of 1 October 2006 could be nicely represented, storms grew into too-large mesocyclones, as compared to the small-scale (mini-)supercells observed. These general features remain similar in all soil moisture experiments, but cold-pool intensity is much weaker in the WETSOIL experiment, due to decreased evaporative cooling as the lifted condensation level (LCL) and hence cloud base was lower and atmospheric boundary layer (ABL) was moister
Decreasing the soil moisture leads to a significant decrease of net radiation at the surface due to a higher surface albedo (0.26 as compared to 0.20 in the CONTROL experiment) and hence a decrease of the net short-wave radiation fluxes during the convective period (1000–2000 UTC), as can be inferred from Table IV. Note that the albedo in ARPS is calculated from the soil moisture and does not take the vegetation cover into account. As the vegetation cover is about 75% during this time of the year, it is likely that the albedo effect is somewhat overestimated in ARPS. When albedo increases, the decreased available energy at the surface and the decreased evaporation result in a drier ABL. As the soil heat capacity decreases, sensible heat flux increases slightly, leading to somewhat warmer surfaces and ABLs. The drier boundary layer is the main reason for stronger cold-pool development as more evaporative cooling occurs when rain is falling through the ABL. LCL, level of free convection (LFC), CAPE and convective inhibition (CIN) are calculated based on raising pseudo-adiabatically a parcel from the lowest grid-level. Lifted air parcels reach their LCL at a higher level compared to the CONTROL, so that condensation-associated latent heat release occurs at higher altitudes, decreasing the CAPE and increasing the CIN significantly. As it takes longer to build up the necessary energy in a drier ABL to overcome the capping inversions, precipitation initiation is delayed (not shown). A lower CAPE keeps the updraught intensities and peak precipitation lower compared to the CONTROL experiment.
Increasing the soil moisture to saturation had a much smaller effect than equally decreasing the soil moisture, which is consistent with literature that soil moisture only starts to affect precipitation when soil wetness decreases to less then 30% of saturation (Koster et al, 2004). As moister soils do have a higher heat capacity, ground heat flux increased, which was compensated for by an equally slight decrease in the sensible heat flux. The effect of increasing soil moisture hence was largely an effect of cooling the surface and the ABL, without bringing much more moisture into the atmosphere. This increase in relative humidity at the surface resulted in lower LCLs and hence condensation-associated heat release occurs at lower altitudes, increasing the CAPE. This leads peak updraughts to increase as well (Figure 7). The effect of an increased CAPE is compensated by less intense cold pools, leading to equally intense storms as in the CONTROL simulations when soil moisture is high in a strong shear environment. Therefore the eventual mean precipitation is hardly affected, as compared to the CONTROL. These contrasting effects of cold pools and CAPE might explain partly the results found in other studies by e.g. Cheng and Cotton (2004) who found wetter soils in Texas to prevent the outflow boundaries of various cells from merging into a clear convective line and to reduce surface precipitation significantly despite higher CAPE. Martin and Xue (2006) found precipitation to increase as soil moisture was decreased during a case with convective initiation along a dryline, although no effect was found during a cold-frontal situation.
Introducing spatially distributed soil properties (VARSOIL) had the smallest effect on the simulated precipitation fields. Mean surface soil moisture was about the same as the homogeneously initialized value in the CONTROL experiment (as described in section 3.1), so CAPE was hardly modified too. The variable soil properties also did not affect the storm initiation as can be seen from the storm tracks in Figure 3. This finding confirms the conclusions of Cheng and Cotton (2004) that introducing spatially distributed soil moisture as compared to homogeneously initiated soil moisture hardly affects surface precipitation.
4.1.2. Buoyancy-driven case
The buoyancy-driven multi-cell case of 28 July 2006 shows relatively small influence of soil moisture on surface precipitation characteristics, similar to the shear-driven case of 1 October 2006. The surface moisture convergence line, associated with the intense observed precipitation across the central parts of Belgium, is simulated too far west in all experiments and is positioned over the north of France and the west coast of Belgium, which is outside the domain covered by observational data (indicated on Figure 5 by the lighter shading outside the radar boundary). The misplacement of the convergence line could be traced already in the 9 km grid spacing domain and hence is probably an issue of inaccuracies in the initial and boundary conditions. Multi-cell storms east of the convergence line are well captured in the 3 km grid spacing domain. The structure of the precipitation field is well represented by the model, resulting in small S-components. Despite a poor representation of the position of the moisture convergence line, L-components are very low for all soil moisture experiments in this case. This is mainly due to the fact that the centres of mass in the observed and modelled precipitation fields are quite close to one another and there appears to be some rotation around the centre of mass in the observed as compared to the modelled precipitation field, in which case the L-component should be treated with care (Wernli et al, 2008). As in the shear-driven case, a decrease in soil moisture negatively influences mean surface precipitation yielding a reduction of 13%, whereas an increase of soil moisture to saturation (WETSOIL) or introducing spatially distributed soil properties (VARSOIL) has no impact on surface precipitation (Table II, Figure 5).
Vertical cross-sections in Figure 6 are averaged as in Figure 4, but for updraughts along the convergence line in the western part of the model domain only. Storm cross-sections appear more symmetric as compared to the shear-driven case, as storms were quasi-stationary. Storms were initiated along a surface convergence line with westerly winds blowing at its west and southerly winds blowing at its east. Modifying initial soil moisture conditions did not severely affect the vertical storm structure, although cold pools in the WETSOIL experiment are weaker again, as in the shear-driven case.
As can be inferred from Table IV, the surface radiation and energy balance in the DRYSOIL experiment are altered in very much the same way as in the shear-driven case. For the same reasons as explained before there is a later onset of convection around noon and a delay in the whole convective cycle by about one hour as can be seen from Figure 7.
Increasing initial soil moisture content to saturation does not affect the surface heat balance significantly (Table IV), with again a compensating effect of a higher CAPE by weaker cold pools. The differences, though, are even smaller than for the shear-driven case, probably because the atmosphere was moister in this case.
Introducing spatially distributed soil properties again hardly affects the surface heat balance (Table IV). As sensible and latent heat fluxes are largely unmodified, CAPE is also not affected. The location and timing of storm initiation is also not significantly affected by these changes.
4.2. Sensitivity to microphysics parametrization
4.2.1. Shear-driven case
Modifying the rain and snow size distribution characteristics (MIRS) does not severely affect the patterns of the 24-hour accumulated precipitation fields (Figure 3). Two main large supercells develop in the north of France and are advected over southern Belgium. The slightly lower value of the S-component of the SAL analysis (Table II) indicates that the fine-scale structure of the observed storms is somewhat better represented by the MIRS experiment. The surface precipitation is reduced by 15% as compared to the CONTROL experiment, which mainly stems from a reduction in the area with intense precipitation (>10 mm—Table III). In the CONTROL there is an overestimation of grid cells with high precipitation intensities. This model deficiency is slightly reduced by making the microphysical assumptions more consistent. This is also reflected in the lower value of the A-component compared to the CONTROL experiment (Table II).
Table III. Number of grid points (in % of total grid points) having higher accumulated surface precipitation than the amount indicated in the left column for all experiments.
> 0.1 mm
> 5 mm
> 10 mm
> 20 mm
> 0.1 mm
> 10 mm
> 25 mm
> 50 mm
Table IV. Thermodynamical and heat balance characteristics.
CAPE = convective available potential energy, LCL = lifted condensation level height, LFC = level of free convection height, CIN = convective inhibition. RN = net all-wave radiation, SLE = surface latent heat flux, SSH = surface sensible heat flux, SGR = surface ground flux. For the shear-driven case from 1000 to 2000 UTC and for the buoyancy-driven case for the 24 h period.
CAPE (J kg−1)
CIN (J kg−1)
SRN (W m−2)
SLE (W m−2)
SSH (W m−2)
SGR (W m−2)
SRN (W m−2)
SLE (W m−2)
SSH (W m−2)
SGR (W m−2)
From Figure 8 it is clear that much more snow is present when modifying its size distribution characteristics. Average snow fall speed e.g. is increased, which also decreases the accretion of snow by hail and rain. Cloud water and cloud ice are reduced significantly. As can be seen from the vertical cross-sections in Figure 4, this has also impact on the vertical storm structure as significant amounts of cloud water and ice are limited to the updraughting cores. At the later stages in the convective development strong mid- and upper-level wind speeds lead to strong downshear anvil development as large quantities of snow are being advected out of the major updraughts. Less cloud water and ice remain and consequently hail growth by accretion of these water types is decreased which gives rise to less potential to melt and contribute to the rain, which explains the decreased surface precipitation accumulation. This results in a better correspondence with the measurements. The downshear advection of precipitating snow results in a somewhat broader area of (very) light rain although much of this falling snowmelt-associated rain evaporates before reaching the surface. As the region with significant cloud amounts is confined to the updraughting cores however, hail collection of cloud water and ice is confined to a smaller region, resulting in a more confined region of intense precipitation. As these changes in microphysical processes have no significant impact on the circulation within the storm, updraughts and downdraughts remain largely unchanged. Hence storm structure and propagation are unaffected.
Additionally, changing the size distribution of hail towards the size distribution for small graupel (MIRSH) has no significant impact on storm initiation or storm track location (Figure 3 and L-component in Table II), but increases the S-component of the surface precipitation field as compared to both CONTROL and the MIRS experiments (Table II), indicating more widespread precipitation. Accumulated surface precipitation is decreased by 12% compared to the CONTROL experiment. The decrease is due to a very strong reduction in the area with very intense precipitation (>20 mm) as can be inferred from Table III, which is more in agreement with the observations.
The much slower and lighter graupel strongly redistributes the water in the atmosphere between the precipitating hydrometeors (Figure 8). Much more graupel is being formed as the smaller particles collect more cloud ice, cloud water, snow and rain. Further, graupel fallout is reduced as its fall speed is much lower. These processes lead the domain total amount of graupel to grow five times as much as compared to the CONTROL experiment. As can be inferred from the vertical cross-sections in Figure 4, this results in cloud water and ice being limited to the updraughting cores as they are much more accreted by graupel. As graupel is hardly sublimating and falling slowly towards the surface, diffusion of graupel both upshear and downshear leads precipitating cells to grow larger. Once the graupel reaches the melting level, it melts almost instantaneously inducing a much broader area with intense surface rain (>5 mm), which leads to a stronger overestimation in this precipitation bin (Table III).
4.2.2. Buoyancy-driven case
Including more realistic assumptions for the rain and snow size distribution (MIRS) hardly affects the general features of the surface precipitation field. As precipitation becomes slightly more widespread, the S-component of the SAL analysis is somewhat increased, but the L-component remains small. Although peak precipitation is decreased, total precipitation accumulation is not affected (Table II and Figure 5).
Vertical distribution of the different water species in Figure 9 shows a strong increase in the amount of snow being formed in the MIRS experiment, at the expense of the non-precipitating cloud particles. Processes at play are similar to the shear-driven case. Anvil development is more symmetric around the main updraughting core as storms remain quasi-stationary (Figure 6). Although the processes mentioned negatively affected surface precipitation in the shear-driven case, we do not see any net effect on precipitation in this case. One reason for this might be the higher pre-storm buoyancy, which has its origin in the spin-up period of the simulation on the night before the storm outbreak. During this period, a large multi-cell system tracks much more west in the MIRS experiment (not shown) This leaves the CAPE built up during the day before in the comparison area highly unconsumed and more CAPE remains throughout the night until the storm outbreak around noon on 28 July. This is also obvious from Figure 7 as updraughts and downdraughts are accelerated in accordance with the higher buoyancy.
Modifying the hail properties towards those for small graupel (MIRSH) increases total precipitation accumulation by 20%, leading to a strongly increased A-component of the SAL analysis. This increase mainly stems from increased precipitation accumulation in the low and intermediate intensity bins, whereas the number of grid cells with very intense precipitation has actually decreased as compared to the CONTROL experiment (Table III). Peak precipitation was even decreased by 35%, bringing it closer to the observations. This was to a lesser extent also the case in the shear-driven situation. As the area with light and moderate precipitation intensities increased and the area with intense precipitation decreased, the S-component of the SAL analysis was strongly increased towards too-widespread rain (Table II). The general patterns and exact location of precipitating cells on the other hand remains similar to the CONTROL experiment, leading to almost no change in the L-component.
As in the shear-driven case, the amount of qh being formed is almost an order of magnitude larger when changing it from large hail to small graupel due to much more collection of other hydrometeors (Figure 9). Huge amounts of graupel remain aloft as can be inferred also from the vertical cross-section in Figure 6. Graupel is diffused away from the main updraughting cores leading to very broad precipitation areas. In contrast to the previous case, we do see an increased surface precipitation. As graupel growth is associated with latent heat release from fusion and freezing processes, more latent heat is released, leading to increased lifting, also in areas outside the updraughting cores. Although the domain maximum updraught is not increased (Figure 7), mean updraught velocity at the level of maximized graupel amounts is significantly increased (not shown). The mean updraught velocity is about 15% higher at mid-levels as compared to the CONTROL simulation. This additional uplift could initiate precipitation processes outside the main updraughting cores, which offers an explanation for the more widespread precipitation. Apparently, these processes are of less importance in the shear-driven case.
5. Discussion and conclusions
In this research two real-case sensitivity experiments were carried out in order to assess the potential of enhanced soil moisture initialization and microphysical assumptions for improvement in quantitative precipitation forecasting during extreme convection. This was done for one case with strong buoyancy but low vertical wind shear and one having moderate buoyancy but strong vertical wind shear, as these two parameters appear to be two main drivers of convection in midlatitudes.
The gain of spatially distributed soil moisture and texture information was small in our simulations, but it is important to have at least the mean soil moisture content right for a realistic simulation of cold-pool intensity, storm structure and surface precipitation. We propose a mechanism to explain why surface precipitation does not necessarily increase as soil moisture and hence CAPE increase, as found in our and also a number of other studies (Cheng and Cotton, 2004; Martin and Xue, 2006). Increasing soil moisture content enhances the thermodynamic conditions for storm development. However, this effect is counterbalanced by a weakening of the storms as a consequence of a weakening of the cold pools due to less evaporative cooling within the boundary layer. There are also some studies that contradict these results and where surface precipitation was found to increase with increasing soil moisture, such as Gallus and Segal (2000). These studies make use of rather coarse-scale modelling, requiring convection to be parametrized. Davis et al(2003) showed that these convective parametrization schemes (CPS) fail in many circumstances to correctly simulate storm propagation characteristics. They speculated that cold pools produced by CPSs were insufficiently cold to permit propagation. Therefore in models that use CPSs, the effect of enhancement of the storms due to thermodynamic conditions is not counterbalanced by a weakening of the storm evolution by weaker cold pools. This implies that coarse-scale models with a CPS are missing an important mechanism for storm development and are therefore less suitable for studying the effect of soil moisture anomalies on storm intensity than high-resolution models that do not depend on a CPS.
Changing the size distribution assumptions of the precipitating hydrometeors in the microphysics parametrization scheme had a larger impact on the hydrometeor distribution than on surface precipitation. Unfortunately, no polarimetric radar was present within our study area in order to determine which of the microphysics experiments led to a better representation of the hydrometeor distribution. From comparison against observed precipitation fields it is clear however that including more realistic rain and snow size distributions slightly improved the surface precipitation in a shear-driven case, while there was no change in a buoyancy-driven situation. Weighting the size distribution assumptions of the large hail to small graupel led to a poor representation of the general storm structure. Large amounts of graupel were precipitating outside the convective cores, leading to widespread and, mainly in the buoyancy-driven case, unrealistic precipitation fields. While clearly beneficial for the simulation of convection, it is not certain how sensitive a quantitative precipitation forecast would be to swapping graupel and hail properties. Hence in an operational set-up, it remains a challenge how to treat the largest ice species in the microphysics scheme. As mentioned by e.g. McCumber et al(1991) and Cohen and McCaul (2006), including both hail and graupel as prognostic variables would avoid tuning a single variable representing both hail and graupel.
It should further be stressed that our study only investigated two aspects of quantitative precipitation forecasting. Many more model components influence the simulation of deep convection, e.g. parametrization of boundary-layer processes, (shallow) convection parametrization, horizontal resolution and even the applied numerical techniques. Wisse and de Arellano (2004) and Deng and Stauffer (2006), simulating deep convection at similar horizontal grid spacing as in our study, both found the Medium-Range Forecast (MRF) boundary-layer scheme to enhance vertical mixing significantly. This mostly led to too-widespread precipitation as compared to e.g. the Eta–Mellor–Yamada (ETA) or the Gayno–Seaman (GS) boundary-layer schemes. Actually, sensitivities they found were of the same order of magnitude as sensitivities found in our experiments on soil moisture and microphysics parametrization. Deng and Stauffer (2006) found largest improvement in their simulation of a convective case over the Great Lakes area in the northern United States at a 4 km horizontal grid spacing when applying a convection parametrization, suggesting that such a grid spacing is still not able to resolve deep convection sufficiently. Weisman et al(1997), however, suggested that quantitative precipitation forecast was not much influenced by horizontal resolution for grid spacing below 4 km, although Bryan et al(2003) found that simulations at 1 km grid spacing contained an unacceptable amount of subgrid-scale turbulence kinetic energy and did not adequately resolve turbulent fluxes of total water when comparing against higher-resolution simulations. This did not have a large impact on surface precipitation for simulations with weak or strong mid-level shear, but it did increase surface rainfall by about 15% when going towards lower grid spacing in a simulation with deep shear. Probably it is very case-specific which of these components dominate over others when models fail to simulate convective storms.
In summary we can conclude that the experiments we carried out mainly affected the storm intensity, and more realistic size distribution assumptions as well as initial soil moisture assumptions can improve simulated precipitation intensity during extreme convection as compared to observations. On the other hand, we found a minor impact on the position of convective initiation and the eventual storm structure development. More research in these fields is needed but we are convinced that the set-up of a large number of sensitivity studies for real cases, which can be compared to the observed situation, provides much added value to the mostly idealized experiments carried out so far for which it is not possible to judge whether a model modification is actually a model improvement.
This research was carried out in the framework of the QUEST-B project, funded by the Flemish Fund for Scientific Research (FWO-Vlaanderen). Further we would like to acknowledge the Center for Analysis and Prediction of Storms (CAPS) of Oklahoma University for providing the ARPS source code online and Jerry Straka and Ming Xue for fruitful discussions leading to the experiments carried out. We are also grateful to the European Environment Agency for making available the CORINE land cover data, the US Geological Survey for the GTOPO30 terrain height dataset, the Deutsches Zentrum für Luft und Raumfahrt (DLR) for the processed AVHRR imagery for sea-surface temperature and the Flemish Institute for Technological Research (VITO) for the SPOT vegetation NDVI imagery. We also acknowledge the European Soil Bureau Network and the European Commission for providing spatially distributed soil texture data and the European Commission Joint Research Centre DESERT action for providing spatially distributed soil moisture data. This research is conducted utilizing high-performance computational resources provided by the University of Leuven, http://ludit.kuleuven.be/hpc.