If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
Representing midlatitude cyclones is important for numerical weather prediction and climate science. These systems are the primary means of transporting energy and moisture poleward at midlatitudes and can bring severe weather in terms of heavy rain, leading to flooding, and damaging strong winds to major population centres. Therefore, it is important to determine if the physical parametrization of clouds and cloud properties used in models are providing realistic representations of these phenomena.
Recent studies have shown a robust response of the storm tracks in climate change experiments, with the storms moving poleward (Fischer-Bruns et al, 2005; Yin, 2005; Held and Soden, 2006). This seems to be accompanied by a consistent poleward shift of midlatitude precipitation (e.g. Emori and Brown, 2005; Held and Soden, 2006). However, the change in the distribution of the intensity of cyclones under climate change is not as consistent across models, and the mechanisms causing these changes are still unclear. Lambert and Fyfe (2006) analysed the Coupled Model Intercomparison Project (CMIP3) multimodel ensemble and found a reduction in the total number of events and an increase in the number of intense events, and Geng and Sugi (2003) reported similar behaviour in climate change experiments with the atmosphere general circulation model (GCM) of the Japan Meteorological Agency. Some single-model studies find no changes in the number of intense storms in future climate on global scales, although they show regional changes (Bengtsson et al2006, 2009). Watterson (2006) analysed global warming simulations of the Commonwealth Scientific and Industrial Research Organisation (CSIRO) climate model and reported a reduction in the number of cyclones, but no changes in the average intensity of cyclones.
Cloud–radiative feedbacks may play an important role in the response of midlatitude cyclones to greenhouse gas perturbations (Weaver, 2003). However, Bony et al(2006) showed that extratropical cloud–radiative feedbacks are a minor contribution to the intermodel spread of the global cloud radiative forcing response to climate warming in the current generation of models. Nevertheless, Trenberth and Fasullo (2010) argued that a better simulation of the cloud radiative impact in present-day simulations is required to increase the credibility of the impact of cloud–radiative feedbacks on the storm track activity in future climate simulations. A good simulation of midlatitude cyclones is a prerequisite to capturing the main characteristics of extreme events in the present-day climate and to having confidence in the response of cyclones to climate change.
A number of recent studies have analysed how well midlatitude cyclones are forecast by operational centres. Initial mean errors in position, speed and minimum sea level pressure are found to be < 100 km, 10 km h−1 and ∼1 mb, respectively (Froude et al, 2007; Charles and Colle, 2009; Froude, 2010). These mean errors are typically found to triple over three days due to a combination of effects from the interaction of dynamical representation and physical parametrizations. Therefore it would be advantageous if just parts of the forecast system could be isolated to test how they perform.
In this article, the forecast system as a whole is not tested. Instead we focus on assessing how well the cloud and precipitation are predicted by the physical parametrizations, given that the thermodynamic and dynamic representation of the atmosphere is accurate. To do this, the model analysis derived through assimilation is used. For the assimilation a cost function is minimised. This function measures the weighted fit of an atmospheric state to observations spanning a 6-h window and the forecast from the previous assimilation/forecast cycle valid at the start of the assimilation window, called the background state. The analysis, which is the state that minimises the cost function, is valid at the start of the window for four-dimensional variational assimilation (4D-Var). In practice, incremental 4D-Var (Courtier et al, 1994) produces an analysis increment that is added to the background state's wind components, potential temperature, total water, density and Exner pressure. The assimilation process adjusts the background state so that the resulting forecast trajectory follows the observations more closely while retaining the dynamical balance needed to minimise spurious gravity waves in the early stages of the forecast. The state of the model atmosphere in this assimilation window represents the best approximation of the real atmosphere compatible with model physics and dynamics. In contrast, variables such as cloud fraction or rain are diagnosed rather than assimilated, so may not follow the corresponding observations as closely within the assimilation window.
We assume that the differences between observations and model cloud variables are primarily due to cloud-related physical parametrizations, but we cannot rule out the effect of analysis errors in other state variables. Choosing the evolved analysis at T + 0, three hours into the forecast starting from the T–3 analysis, allows the forecast model time to adjust to any residual dynamical and physical imbalances and therefore provides us with model-derived cloud variables that will deviate from observations primarily because of shortcomings in the representation of clouds.
Previous studies focus their attention on the geographical patterns of variables of interest for the description of cyclones (e.g. precipitation, wind strength). A different approach is to examine the main characteristics of cyclones by compositing data in a reference framework where the origin is at the centre of the cyclone (Lau and Crane, 1995; Chang and Song, 2006; Field and Wood, 2007), which is a useful way of removing cyclone-to-cyclone variability from both observations and modelled representations of cyclones. The results of this compositing approach can be linked with conceptual models such as that of a warm conveyor belt transporting moisture from the surface to upper levels in a midlatitude cyclones (Harrold, 1973; Browning and Roberts, 1994). Field and Wood (2007) used satellite data to demonstrate a relation between cyclone-wide composite mean rainrates and moisture transported along a warm conveyor belt. Boutle et al(2010) has shown that moisture convergence in the boundary layer is the primary source of moisture for the warm conveyor belt which leads to the efficient conversion of moisture to precipitation. Recently, Field et al(2008) examined a comparison of composite midlatitude cyclones made from satellite data and the National Center for Atmospheric Research Community Atmosphere Model (CAM3) which showed that a coarse (1° horizontal grid resolution) model was able to reproduce the warm conveyor belt relationship seen in the satellite observations.
Satellite data are overwhelmingly two-dimensional in nature, but the advent of CloudSat means that it is becoming possible to assess the three-dimensional structure of meteorological phenomena captured by models. We extend here the methodology developed by Field and Wood (2007) to include an assessment of the three-dimensional structure of cloud and precipitation in midlatitude cyclones. We apply this methodology to a version of the global Met Office Unified Model and show that this approach provides a comprehensive assessment of the simulation of cyclones in numerical models and suggests potential avenues for further research and parametrisation development.
2.1. Unified Model
Unified Model (UM) archived 0000 UTC analysis fields have been used for the analysis and were derived from the 4D-Var system described in Rawlins et al(2007). In the global model, observations of wind, temperature and humidity are assimilated within a 6 h window. At the end of the assimilation process, the constrained dynamical and thermodynamical system is run as a forecast from T–3 hours to the analysis time at T + 0 hours to generate the condensed water and cloud fields. Spin-up of other model fields that are not incremented will then occur. The precipitation after the increments are included at T–3 takes ∼1 h (Rawlins et al, 2007) to adjust. The model analysis can be viewed as a best fit to the observations that are consistent with the previous forecast, with the derived model increments for temperature, humidity and winds, and with the full model physics. The UM uses parametrizations to represent the cloud fraction (Smith, 1990), mixed-phase cloud microphysics (Wilson and Ballard, 1999) and convection (Gregory and Rowntree, 1990).
The model versions used during this period were cycles G40 to G46 (N320 L50). The horizontal resolution of the model in midlatitudes was approximately 40 km. Model fields for winds, temperature, pressure, condensed moisture (large-scale and convective), cloud fractions (large-scale and convective) were extracted. Model rainrates were derived from the sum of precipitation from the large-scale rain and convective parametrization schemes. For comparison with the CloudSat derived rainrate, only rainrates where the model freezing level was above a predefined height threshold were used. Results were prepared for profiles with freezing levels above 1 km.
For each cyclone, obtained from an existing database of cyclone locations, the model fields are translated and regridded using bilinear interpolation onto a 4000 × 4000 km domain (x, y are the eastward and northward coordinates respectively) with 100 km grid spacing and the cyclone is located centrally (x = 0, y = 0). For many of the figures shown, these data were regridded onto a coarser 200 km grid.
Bodas-Salcedo et al(2008) have developed a system to simulate CloudSat data in the UM that is consistent with the model microphysics. The Cloud Feedback Model Intercomparison Project (CFMIP) Observation Simulator Package (COSP; Bodas-Salcedo et al, 2011) is a flexible software tool that enables the simulation of data from several satellite-borne active and passive sensors from model variables. The flexibility of COSP and a common interface for all sensors facilitates its use in any type of numerical model, from high-resolution cloud-resolving models to the coarser-resolution GCMs. It is conceptually divided into three steps. First, the grid box-mean profiles are broken into subcolumns allowing for cloud and precipitation overlap. Then, the vertical profiles of individual subcolumns are passed to individual instrument simulators. Finally, a statistical module gathers the outputs from all the instruments and builds statistics that can be compared to similar statistics from observations. COSP version 1.2.2 has been used in this study to simulate the CloudSat radar (Haynes et al, 2007). (COSP is open-source software and can be downloaded from the CFMIP website, http://www.cfmip.net.) The main difference between what was done in this study and by Bodas-Salcedo et al(2008) was that data from the 100 km grid boxes were used for the forward modelling rather than at the native model resolution. Given that the assumptions about hydrometeor properties are the same between the radar forward model and the microphysics used in the UM, any differences are likely to arise through assumptions about cloud overlap. Sensitivity tests were carried out based on changing the cloud overlap assumption used by COSP. This was found to produce negligible bias (0.05 dBZ) and to have a standard deviation of 0.9 dBZ.
CloudSat flies in sun-synchronous orbit at 705 km altitude within the A-Train (Stephens et al, 2002). It carries the first millimetre-wavelength cloud-profiling radar (CPR) in space, which operates at a frequency of 94 GHz (Im et al, 2005). The CPR points in the nadir direction, and its pulses sample a volume of 480 m in the vertical, with a horizontal resolution of 1.4 km across track. We used two CloudSat products, the 2B-GEOPROF release 4 dataset (Marchand et al, 2008) and the 2C-PRECIP-COLUMN release 4 (Haynes et al, 2009). The 2B-GEOPROF dataset provides the radar reflectivity (in dBZ) and identifies where hydrometeors occur (‘cloud-mask’). Reflectivities with cloud-mask values less than 20 were ignored. The reflectivity bias error adopted here is ± 1 dBZ (Protat et al, 2009).
The 2C-PRECIP-COLUMN provides estimates of the precipitation over ocean using a path-integrated attenuation algorithm up to a maximum of ∼70 mm d−1 for a single profile. Stephens et al(2011) assumed that the error in this retrieval was dominated by attenuation and freezing-level height errors. For midlatitude cyclones, the rainrates retrieved are not adversely affected by attenuation. While Haynes et al(2009) showed that variations in freezing-level height of ± 250 m can lead to large retrieval errors, this is a random error and potential bias in this value has been found to be much smaller when comparing ground-based radar with model data (Mittermaier and Illingworth, 2003). Haynes et al(2009) explored potential errors in precipitation rate due to a number of factors. Of the seven factors considered, two could potentially introduce a bias for the rainrates observed at midlatitudes: cloud water to rain ratio and size distribution shape assumptions. Assuming these two errors combine in quadrature gives an estimated bias to the retrieved rainrate of ± 40% which has been applied to the retrieved rainrate for freezing levels higher than 1 km.
The Advanced Microwave Scanning Radiometer (AMSR-E), on the National Aeronatics Space Administration (NASA) Aqua sun-synchronous satellite, provides near-complete coverage each day. We use AMSR-E data provided by Remote Sensing Systems (RSS). For this study we use surface rainrate and column integrated water vapour estimates (Wentz, 1997; Wentz and Spencer, 1998). The estimates are made using microwave brightness temperatures observed over ocean. The data are provided at a resolution of 0.25° × 0.25° and averaged up to a 1° × 1° grid for analysis. Data within 6 h of an identified cyclone are used from one of the daytime (∼1330 LST) or night-time (∼0130 LST) overpasses. The water vapour path (WVP) product is estimated to have a random error of 1.2 kg m−2 and has no bias with respect to rainfall rate (Wentz, 1997).
For the retrieval of rain using passive microwave, there are three issues to consider: (i) the beam-filling effect, (ii) the partitioning of rain and cloud water, and (iii) rain layer thickness, all of which can lead to systematic biases in the rainrate estimate.
The beam-filling effect concerns the concave nature of the brightness temperature versus rainfall relation, which means that precipitation rate will be underestimated unless the rainfall is homogeneous within the footprint or the heterogeneity is accounted for.
The partitioning of the retrieved liquid water column into rain and cloud water is a diagnostic function of liquid water path and freezing-level height. In the current algorithm, liquid water paths of 0.18 mm are required before rain is produced. This minimum threshold will be sensitive to drop-size distribution changes associated with different synoptic regimes. The partitioning method is very simple and may be subject to systematic error under varying conditions and does not take scattering into account. Berg et al(2006) discussed the possibility that the different gross structures of rain systems could lead to regional bias in the retrieval of rain using passive microwave techniques.
The minimum reported value for rainrate is 0.1 mm h−1 for a 0.25° pixel, which equates to 0.15 mm d−1 for the 1° pixels used here.
Surface wind vector data are from the SeaWinds microwave scatterometer on the NASA Quikbird sun-synchronous platform. The local overpass times for Quikbird are ∼0600 and 1800 LST and data within 6 h of the identified cyclone are used. The SeaWinds instrument measures backscatter from wind-driven capillary waves at a number of viewing angles, allowing estimates to be made of wind speed and direction with approximately 25 km horizontal resolution. In this study, we used the products generated by RSS using the algorithm of Wentz and Smith (1999). The data are provided at a resolution of 0.25° × 0.25° and averaged up to the common 1° × 1° grid for analysis.
Chelton and Freilich (2005) determined that the random error present in the 25 km resolution Quikscat winds is 1.7 m s−1, but is much reduced for wind speeds between 6 and 15 m s−1. These random errors will be reduced by a factor of 4 when averaged over a 100 km grid box.
Top-of-atmosphere fluxes are the ISCCP (International Satellite Cloud Climatology Project) FD 3 h data on a 2.5° grid. The ISCCP FD dataset contains radiative fluxes at the top of the atmosphere, at the surface, and at three additional levels (680, 440, and 100 mb), computed from the ISCCP cloud properties and radiative transfer computations (Zhang et al, 2004). The potential bias for top-of-atmosphere fluxes is estimated to be ± 10 W m−2.
3. Composite midlatitude cyclones
Cyclone locations at 0000 UTC each day for June 2006 to January 2008 (provided by M. Bauer) were used to composite the data. The location procedure is based on tracking minima in sea level pressure using a method similar to that described by Bauer and Del Genio (2006).
This process has been applied to the National Centers for Environmental Prediction (NCEP-2) reanalysis to provide a long-term climatology of midlatitude cyclones called the Modeling, Analysis and Prediction (MAP) Climatology of Midlatitude Storminess (MCMS; http://gcss-dime.giss.nasa.gov/mcms/mcms.html). All cyclones were flipped into a Northern Hemisphere sense for subsequent analysis. Ideally only cyclones entirely located over oceans would be used to reduce the effects of land surface and orography and because the AMSR/Quikscat data products used here are retrieved only over ocean. However, the number of cyclones used in the analysis was doubled by including cyclones with areal ocean coverage of 70% or greater rather than using the stricter no-land threshold. As a test of sensitivity to this threshold, it was found that the mean precipitation, integrated water vapour column and wind speed averaged across the cyclones for oceanic area fractions ranging from 0.5 to 0.9 changed by % or less.
Some cyclone properties have been examined as a function of cyclone strength and atmospheric moisture metrics which are defined as follows. Cyclone strength, , is determined as the mean surface wind speed, measured using Quikscat, within a circle of radius 2000 km centred on the cyclone. The angled brackets indicate averaging over the cyclone within a 2000 km of the cyclone centre.
Typically, the central pressure anomaly, minimum in sea level pressure or some relative vorticity measure is used to define the cyclone strength. Here is used as an indicator of cyclone strength following Field and Wood (2007) who showed that is a natural indicator of cyclone strength from a cloud and precipitation perspective. Cyclone moisture, , is similarly determined as the mean integrated water vapour column (water vapour path), from AMSR-E, within a circle of radius 2000 km centred on the cyclone.
The single nadir beam from CloudSat means that the data coverage is sparse. Therefore, initially, a single composite of > 1900 cyclones was considered to maximise the amount of CloudSat data contained in the composite and to avoid the effects of extreme cyclones skewing the properties of the composite. The single-cyclone composite was constructed from cyclones within the following range of model-derived cyclone-wide mean values: : 4.9–12.3 m s−1 and : 10–33 kg m−2.
For each cyclone, CloudSat and AMSR-E overpasses within six hours of 0000 UTC were assigned to that cyclone and typically each cyclone identified had one or two overpasses. Assuming that midlatitude cyclones move at around 10 m s−1 means that over 6 h the centre of a cyclone would move by ∼200 km. This will likely lead to some potential smearing of the structure in the composite. However, the smoothing effect of inter-cyclone variability will be much greater. For CloudSat the number of cyclones sampled in each 100 km grid box increases from about 100 in the south of the Northern-Hemisphere-sense composite cyclone to in excess of 150 in the north. Only profiles with freezing levels higher than 1 km were used to form composite rainrates. This reduces the number of cyclone samples per grid box to . In addition, there is a decrease in samples towards the Pole because of the lowering in altitude of the freezing level. For the hydrometeor profiles and composite reflectivity slices that just make use of the measured reflectivity, the filtering based on freezing-level height did not need to be done. Model and AMSR-E cyclone samples per composite grid box were of the order ∼1500.
Figures 1 and 2 show cross-sections through the composite cyclone regridded to 200 km × 200 km pixels for the UM and CloudSat, respectively. This plot is an arithmetic mean of the reflectivity power (where it was greater than –20 dBZ) at different altitudes for all of the individual profiles sampled. The standard error of the mean reflectivity for CloudSat is typically ∼1 to 2 dBZ, and a factor of 5 smaller for the model-derived data (given the larger number of samples). While the CloudSat composite cross-sections are noisier than the model, both the model and CloudSat composites exhibit similar structures. For the lowest altitude (1.2 km, Figures 1(a) and 2(a)), the CloudSat 4 dBZ contour covers a slightly larger area than the model. The 2.2 km slices (Figures 1(b) and 2(b)) are very similar. Concentrating on the 4.1 km (Figures 1(c) and 2(c)) and 6.0 km (Figures 1(d) and 2(d)) slices shows a tongue of relatively higher reflectivity arcing up from the southeast and then west towards the centre of the cyclone. At 6.0 km, the CloudSat cross-section has greater reflectivities to the southeast and a 4 dBZ contour that is absent in the model.
Further horizontal cross-sections have been generated showing the fraction of the grid box exhibiting greater than –20 dBZ reflectivity (Figures 3, 4). The error in the mean for fractions of 10% are ∼1% for the CloudSat data and smaller again for the model. These figures show a maximum fraction near the centre of the composite cyclone at low levels (Figures 3(a), 4(a)) that gradually extends further eastwards and south at higher levels tracing out the clouds associated with frontal structures. At lower levels (Figures 3(a, b) and 4(a, b)), the 40% contour for CloudSat covers a greater area than the model, but the 15% contour for the model covers a larger area at 2.2 km (Figures 3(b), 4(b)) than CloudSat.
For the upper levels (Figures 3(c, d) and 4(c, d)), the 40% contour again covers a larger area for CloudSat than for the model. The 10% contour to the west of the cyclone centre appear similar for both model and CloudSat.
Following Bodas-Salcedo et al(2008), we have produced joint histogram plots of radar reflectivity versus height from the COSP simulator using UM condensed water fields (Figure 5) and from CloudSat (Figure 6). These distributions have been formed for each 100 km × 100 km grid box and then subsequently combined into quadrants of the composite cyclone within 2000 km of the centre. Above 4 km, the joint histograms generated from the model look similar to those generated from CloudSat, although the mode in the CloudSat distribution exhibits slightly greater reflectivities, indicating that the model is predicting not enough ice water content or that the snow size distributions are too narrow or a combination of both these elements. There is a rapid reduction of reflectivity with height in the joint histograms. Such behaviour in the reflectivity can be reproduced by the ice particle size distribution used in the model by assuming that the difference between liquid water saturation and ice water saturation mixing ratios gives the ice water content at that height. Such a curve is shown for a subtropical temperature profile (Fels, 1986) applied to the ice representation in the UM. Below 4 km there are larger differences between the model and CloudSat similar to those reported by Bodas-Salcedo et al(2008). The main difference is that the model has a conspicuous mode at relatively high reflectivities compared to CloudSat which instead tends to have a flatter or even bimodal distribution of reflectivities.
Integrating across the joint histograms as a function of height gives the fraction of pixels with greater than –20 dBZ, 0 dBZ and 10 dBZ reflectivities as a function of altitude for the four quadrants (Figure 7). Error bars indicating the effect of a ± 1 dBZ bias in the reflectivity are also shown. For reflectivities > − 20 dBZ, all quadrants show a maximum fraction of between 0.2 and 0.35 at 1 km altitude, a minimum at 2 km and then a slight secondary maximum between 5 and 8 km. The southwest quadrant has the lowest fractions overall and the northeast quadrant has the greatest fractions. Reflectivity thresholds of 0 and 10 dBZ generally indicate decreasing fractions with altitude and again the northeast quadrant exhibits the highest fractions while the southwest has the lowest fractions.
For all quadrants, the model underpredicts the fraction of profiles with reflectivities greater than 10 dBZ by a factor of ∼10 or more. The fraction of profiles exceeding the 0 dBZ threshold is similar for CloudSat and the model at altitudes greater than 4 km, but the model overpredicts this fraction below 4 km. For the southern quadrants, the fraction of profiles with reflectivities exceeding –20 dBZ from the model agrees well with the observations. For the northern quadrants, the model underpredicts the fraction of profiles exceeding –20 dBZ by ∼20%, with the largest discrepancies found in the northeast.
Using ISCCP 3 h instantaneous fluxes, the outgoing short-wave at the top of the atmosphere has been composited for the cyclones (Figure 8). These plots represent a diurnal average for the composite cyclones. The spatial structure is dominated by the comma shape to east of the cyclone centre, reflecting the structure of the cloud shield associated with the fronts. Difference plots show local differences can exceed –10 W m−2. A consistent feature is that the model looks brighter in the short-wave near the centre of the cyclones, but is dimmer relative to ISCCP away from the centre. Averaging over the entire domain reveals that the midlatitude composite cyclone appears –8 W m−2 dimmer from the model in the short-wave. The top-of-atmosphere outgoing long-wave signal again follows the spatial structure of the frontal cloud shield, but there is much better agreement with ISCCP; averaging over the domain reveals a difference of less than 1 W m−2.
Turning now to surface rainrates, Figure 9 shows composite mean rainrates (freezing level greater than 1 km) from the UM (Figure 9(a)), AMSR-E (Figure 9(b)), CloudSat (Figure 9(c)) and AMSR-E (Figure 9(d)), again when subsampled along the CloudSat data. These have all been regridded onto a 200 km × 200 km grid and only CloudSat tracks containing more than 150 samples within each grid box have been used. The AMSR-E data subsampled on the CloudSat track is much more noisy than the composite produced from the full swath data. The error in the mean is ∼40% (reducing to ∼30% for higher rainrates). For the model and AMSR-E full swath data, the error in the mean is reduced to less than 10%. The satellite and UM products show the peak in rainrate located in a comma shape to the east and southeast of the composite centre. The structure seen here agrees well with that seen in composites from other models and observing ship data (Chang and Song, 2006). The pattern of rainrate exhibited by model and AMSR-E is similar, but the magnitude is reduced for AMSR-E. Both CloudSat and the UM show greater intensity mean rainrates close to the centre of the composite than AMSR-E. The ratio of the AMSR-E mean to the CloudSat mean precipitation rates is approximately 0.5, in agreement with Stephens et al(2010) who found a similar difference between AMSR-E and CloudSat in midlatitudes.
It is interesting to identify what fraction of the rain in the model comes from the convection parametrization as opposed to the large-scale precipitation (overplotted on the UM rainrate in Figure 9(a)). It can be seen that in the poleward quadrants, the fraction of rain from the convection parametrization is < 0.5. For the equatorward quadrants, while the field is smoothed by cyclone variability, the fraction is greater and exceeds 0.7 in a region associated within and behind cold fronts. Comparing Figures 9(b) and (d) gives an indication of the effect of the sparse CloudSat sampling. Assuming a bias in the CloudSat precipitation rate of up to 40% combined with the error in the mean suggests that the model rainrate is not significantly different from that observed by CloudSat.
Rainrate histograms from the 100 km × 100 km mean rainrates derived from profiles with a freezing level higher than 1 km are shown in Figure 10 for CloudSat, AMSR-E and the UM. The Northern-Hemisphere-sense midlatitude cyclone composite was split into four quadrants and normalised histograms (divided by the total number of 100 km × 100 km precipitation samples, including zero) were derived. Mean rainrates for each quadrant show that the CloudSat mean is greater than the model, which is in turn greater than the AMSR-E estimate. However, bearing in mind a potential ± 40% bias in the CloudSat precipitation leads to potential agreement between the model and CloudSat means in all quadrants. The AMSR-E mean precipitation rates in the western quadrants are still too small for a 40% potential bias. Looking in more detail at the histograms reveals that the shape of the CloudSat rainrate distribution is similar to UM and AMSR-E distributions but shifted to higher rainrates. Only in the southwest quadrant does there seem to be a difference in the shapes of the distribution, where the CloudSat distribution exhibits a greater frequency of lighter precipitation events than the model or AMSR-E.
Using such a large composite includes a very wide range of environmental conditions. Following Field and Wood (2007), the cyclones have been subdivided into different categories according to cyclone-wide mean wind speed (strength) and integrated moisture column (moisture). The categories are
〈V〉: 4.9– 6.95, 6.95– 8.14, 8.14– 12.3 m s−1 and
〈WVP〉: 10– 18, 18– 21, 21– 33 kg m−2.
The categories are chosen to be orthogonal, so that the effects of changing one variable can be assessed while holding the other fixed.
Examining the fraction of pixels with reflectivity greater than –20 dBZ, 0 dBZ and 10 dBZ for the northeast quadrant only and these nine categories now shows that the best agreement between CloudSat and the model is seen for the cyclones with the lowest mean surface wind speed (Figure 11). The minimum number of profiles in one of these composite categories is ∼105. Assuming the profiles are independent means that a fraction of 0.01 would have a standard error for sample counts of 0.0003. If, for the worst case, it was assumed that all of the profiles in each quadrant were completely correlated, this would increase the error by a factor of 20, leading to a 0.006 error on fractions of 0.01 and 0.02 error for fractions of 0.1. This worst-case error is similar in magnitude to the effect of a potential measurement bias of ± 1 dBZ. This quadrant has been chosen because it exhibits the biggest deviation from the satellite observations. The negative discrepancy for the –20 dBZ threshold between CloudSat and the UM is only present above 2 km and becomes progressively more obvious with increasing strength. For the highest strength categories, the 0 dBZ fraction is underpredicted by the model above 3 km. The fraction of 10 dBZ profiles is always underpredicted by the model.
Both model and satellite data provide rich resources for extensive analysis. In this study, satellite data have been compared with operational model analysis fields to identify salient similarities or differences. Analysis fields are the basis of the forecast and provide the best model estimate of the state of the atmosphere. Because the cloud fields are not directly assimilated, the analysis fields provide a means of assessing the physical parametrizations that control the cloud without folding in the effects of dynamical and thermodynamical errors from other physical parametrizations or the dynamical evolution of the model. While this can be viewed as testing the model data assimilation system, it is felt to be the most appropriate way of attributing model–observation discrepancies to cloud-related parametrizations rather than to accumulating errors in the large-scale circulation as the forecast progresses (e.g. Bodas-Salcedo et al, 2008).
The main differences seen between the model and the satellite data are a lack of cloud above 2 km in the model associated with the main precipitation region, underpredicted fractions of reflectivities greater than 10 dBZ and underprediction of the top-of-atmosphere outgoing short-wave radiation.
In contrast, given a potential ± 40% bias in the precipitation product from CloudSat, the model precipitation rates agree with CloudSat. The exception is the southwest quadrant where the frequency of lighter rainrates appear underpredicted by the model.
It has been demonstrated that a good correlation exists between cyclone-wide composite mean rainrate (<R >) and the product of the cyclone-wide composite mean water vapour column and cyclone-wide composite mean windspeed, i.e. < R > = k < V > < WVP > , where the angled brackets represent means computed over the entire cyclone within 2000 km of the centre (Field and Wood, 2007; Field et al, 2008). This behaviour was attributed to the idea of a ‘warm conveyor belt’ that lifts moisture up and out of the boundary layer eventually forming precipitation. Figure 12 shows this correlation holds for the UM. The values of k used for the two datasets were 0.023, 0.028 for AMSR-E and UM, respectively. We attempted this analysis for the CloudSat data but the variability was too great because the relatively sparse sampling of CloudSat makes it difficult to split the data into a further nine categories.
Using the fraction of cloud greater than –20 dBZ as a proxy for cloud fraction, the lack of cirrus cloud to the northeast of the cyclone explains the negative short-wave anomaly in that region. The underprediction of midlevel cloud is consistent with the findings of Illingworth et al(2006). They used vertically pointing radar/lidar data to demonstrate that midlevel cloud between 3 and 6 km was missing from many operational weather models.
The behaviour of hydrometeor fraction (>− 20 dBZ) with cyclone strength does not produce as much of an increase with increasing strength. Naud et al(2010) suggested that moisture is not being transported to high enough altitudes in a model they studied, and that convection associated with fronts is not well represented. Similarly, Froude (2009) speculated that errors in model vertical structure or storm tilt may contribute to errors in storm evolution. The possibility of problems with vertical structure was also highlighted by Catto et al(2010) when comparing a climate model with a reanalysis product which essentially have different cloud and precipitation parametrizations.
Differences in the amount of cloud to the west of the cyclone are less apparent. Therefore, the short-wave anomaly in that region must arise from differences in the character of the cloud or the phase and size of the particles within the cloud that are present in the model.
Representing midlevel layer cloud is a challenge for global models. One significant obstacle is that the thickness of the levels is around 600 m, similar to the thickness of altostratus/cumulus layers. The second challenge is that, even if thin midlevel supercooled liquid cloud can be produced by the model, it will be efficiently transformed to ice by the ice nucleation scheme and fall out. This formation of ice reduces both the lifetime and radiative effect of these clouds. This has possible implications for the radiative effect of midlatitude cyclones on the climate system.
For the comparison of model data fields and observations, we are essentially showing results for a model that has been replaced. The current Met Office global NWP model is now running with 70 vertical levels and a horizontal grid spacing of 25 km in midlatitudes. However, it would take a further two years to accumulate enough data to carry out a similar comparison, and by that time the model will no doubt be superceded. However, because the UM is used for both NWP and climate, we are ahead of the current state of the art in climate modelling. Thus, these results will be useful for climate modellers to assess how well high-resolution climate models are likely to represent midlatitude cyclones.
In this study we have shown that the use of a compositing methodology provides a detailed picture of midlatitude cyclones and it is a valuable tool to assess the simulation of midlatitude cyclones in models. We have made use of CloudSat data to extend the two-dimensional view of previous studies to a three-dimensional perspective.
Operational global model analyses from the Met Office Unified Model have been used to construct composite midlatitude cyclones for comparison with a range of satellite data including microwave and CloudSat-derived rainrate, column integrated humidity, surface wind speeds and top-of-atmosphere radiative fluxes. Cloud and precipitation were not directly assimilated. Comparing observations with the analysis fields provides a means of assessing how well the parametrizations are able to reproduce cloud and precipitation given an accurate representation of the dynamical and thermodynamical state of the atmosphere.
CloudSat data have been used to form a 3D midlatitude composite cyclone. Using ∼1900 cyclones still resulted in a sparsely populated composite cyclone. To further subdivide the composite into different strength and moisture categories to investigate systematic structural changes would require another order of magnitude of observations.
Good qualitative agreement was seen between the reflectivity structures produced by the UM and observed by CloudSat. However, the UM underpredicted fractions of reflectivity greater than –20 dBZ and + 10 dBZ. The outgoing top-of-atmosphere short-wave flux is underpredicted in the model to the east, west and north of the composite cyclone. While precipitation rates from the UM are less than the CloudSat estimates, a potential bias of ± 40% in the retrieved rainrates means that UM precipitation rates fall within the error estimate of the CloudSat data.
We would like to thank Mike Bauer for producing the MAP Climatology of Midlatitude Storminess and making it available to us. Thanks are also due to G. Inverarity for useful discussions. A. Bodas-Salcedo was supported by the Joint DECC/Defra Met Office Hadley Centre Climate Programme (GA01101).