• Bayler, G. M., Aune R. M. , and Raymond W. H. , 2000: NWP cloud initialization using GOES sounder data and improved modeling of nonprecipitating clouds. Mon. Wea. Rev., 128, 39113920, doi:10.1175/1520-0493(2001)129,3911:NCIUGS.2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • de Haan, S., 2011: High-resolution wind and temperature observations from aircraft tracked by Mode-S air traffic control radar. J. Geophys. Res.,116, D10111, doi:10.1029/2010JD015264.

  • de Haan, S., 2013: Assimilation of GNSS ZTD and radar radial velocity for the benefit of very-short-range regional weather forecasts. Quart. J. Roy. Meteor. Soc.,139, 2097–2107, doi:10.1002/qj.2087.

  • de Haan, S., and Stoffelen A. , 2012: Assimilation of high-resolution Mode-S wind and temperature observations in a regional NWP model. Wea. Forecasting,27, 918937, doi:10.1175/WAF-D-11-00088.1.

    • Search Google Scholar
    • Export Citation
  • de Haan, S., Holleman I. , and Holtslag A. A. M. , 2009: Real-time water vapor maps from a GPS surface network: Construction, validation, and applications. J. Appl. Meteor. Climatol., 48, 13021316, doi:10.1175/2008JAMC2024.1.

    • Search Google Scholar
    • Export Citation
  • de Haan, S., Marseille G.-J. , de Valk P. , and de Vries J. , 2013: Impact of ASCAT scatterometer wind observations on the High Resolution Limited Area Model (HIRLAM) within an operational context. Wea. Forecasting,28, 489–503, doi:10.1175/WAF-D-12-00056.1.

  • Derrien, M., and Le Gléau H. , 2005: MSG/SEVIRI cloud mask and type from SAFNWC. Int. J. Remote Sens., 26, 47074732, doi:10.1080/01431160500166128.

    • Search Google Scholar
    • Export Citation
  • GEPARD GmbH, 2001: CineSat software user manual. Operation Manual 120, Gepard GmbH, Vienna, Austria, 225 pp.

  • Köpken, C., Kelly G. , and Thépaut J. N. , 2004: Assimilation of Meteosat radiance data within the 4D-Var system at ECMWF: Assimilation experiments and forecast impact. Quart. J. Roy. Meteor. Soc., 130, 22772292, doi:10.1256/qj.02.230.

    • Search Google Scholar
    • Export Citation
  • Martinet, P., Fourrié N. , Guidard V. , Rabier F. , Montmerle T. , and Brunel P. , 2013: Towards the use of microphysical variables for the assimilation of cloud-affected infrared radiances. Quart. J. Roy. Meteor. Soc., 139, 14021416, doi:10.1002/qj.2046.

    • Search Google Scholar
    • Export Citation
  • Munro, R., Köpken C. , Kelly G. , Thépaut J.-N. , and Saunders R. , 2004: Assimilation of Meteosat radiance data within the 4D-Var system at ECMWF: Data quality monitoring, bias correction and single-cycle experiments. Quart. J. Roy. Meteor. Soc., 130, 22932313, doi:10.1256/qj.02.229.

    • Search Google Scholar
    • Export Citation
  • Painting, J. D., 2003: WMO AMDAR reference manual. WMO-958, WMO, Geneva, Switzerland, 80 pp.

  • Stengel, M., Lindskog M. , Undén P. , Gustafsson N. , and Bennartz R. , 2010: An extended observation operator in HIRLAM 4D-Var for the assimilation of cloud-affected satellite radiances. Quart. J. Roy. Meteor. Soc., 136, 10641074, doi:10.1002/qj.621.

    • Search Google Scholar
    • Export Citation
  • Stengel, M., Lindskog M. , Undén P. , and Gustafsson N. , 2013: The impact of cloud-affected IR radiances on forecast accuracy of a limited-area NWP model. Quart. J. Roy. Meteor. Soc., 139B, 2081–2096, doi:10.1002/qj.2102.

    • Search Google Scholar
    • Export Citation
  • Undén, P., and Coauthors, 2002: HIRLAM-5 scientific documentation. HIRLAM Project Tech. Rep., Norrköping, Sweden, 144 pp. [Available online at http://hirlam.org/index.php?option=com_content&view=article&id=64:gneral-description-of-the-hirlam-model&catid=48:synoptic-scale-model-hirlam&Itemid=101.]

  • van der Veen, S. H., 2013: Improving NWP model cloud forecasts using Meteosat Second-Generation imagery. Mon. Wea. Rev., 141, 15451557, doi:10.1175/MWR-D-12-00021.1.

    • Search Google Scholar
    • Export Citation
  • Yucel, I., Shuttleworth W. J. , Pinker R. T. , Lu L. , and Sorooshian S. , 2002: Impact of ingesting satellite-derived cloud cover into the Regional Atmospheric Modeling System. Mon. Wea. Rev., 130, 610628, doi:10.1175/1520-0493(2002)130<0610:IOISDC>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Yucel, I., Shuttleworth W. J. , Gao X. , and Sorooshian S. , 2003: Short-term performance of MM5 with cloud-cover assimilation from satellite observations. Mon. Wea. Rev., 131, 17971810, doi:10.1175/2565.1.

    • Search Google Scholar
    • Export Citation
  • View in gallery
    Fig. 1.

    HIRLAM H11 and U11 domains.

  • View in gallery
    Fig. 2.

    Changes in vertical profiles of specific humidity for two cases: (left) a cloud added by the cloud initialization using MSG and (right) a cloud removed in the initialization.

  • View in gallery
    Fig. 3.

    Mean (left) rainfall rate and (right) cloud cover over the whole U11 domain for each time step (forecast time indicates the hours into the forecast) based on cycled hourly experiments over a period of 6 days for different humidity change limits.

  • View in gallery
    Fig. 4.

    Mean difference (bottom two curves) and std dev (top two curves) between cloud cover forecast and MSG cloud cover observations of the reference run (solid) and the MSG run (dashed); the shaded area denotes uncertainty obtained by the jackknife method.

  • View in gallery
    Fig. 5.

    Mean and variation of cloud cover observations and forecasts. The two solid lines refer to the cloud forecasts from the reference run; the dashed lines refer to the cloud forecast of the MSG run, and the dotted (constant) lines refer to the cloud cover observations.

  • View in gallery
    Fig. 6.

    Time series of the mean difference and std dev of model cloud cover forecast compared to MSG cloud cover observations on a daily basis, for the whole period of the experiment.

  • View in gallery
    Fig. 7.

    Locations of the synoptic stations.

  • View in gallery
    Fig. 8.

    HSS (for different forecast lengths) of the rainfall forecast compared to the synoptic hourly rainfall observations for three threshold values: 0.5, 1, and 2 mm h−1.

  • View in gallery
    Fig. 9.

    Observed and forecast cumulative rainfall-rate distributions for the whole period for (left) the REF run and (right) the MSG run.

  • View in gallery
    Fig. 10.

    Mean and std dev of the difference between forecast upper-air temperature and observed AMDAR and Mode-S EHS temperature (01, 02, 06); the shaded area denotes the uncertainty in the mean difference and the std dev obtained by the jackknife method.

  • View in gallery
    Fig. 11.

    The 2-m temperature forecast errors; the shaded area denotes the uncertainty in the bias and the std dev obtained by the jackknife method.

  • View in gallery
    Fig. 12.

    Mean difference and std dev between forecast and observed model pressure for forecast times 1–6 h. The shaded area denotes the uncertainty in the bias and the std dev obtained by the jackknife method.

  • View in gallery
    Fig. 13.

    (left) Pressure reduced to mean sea level (contoured) and observed cloud cover from NWC SAF, (middle) 6-h cloud cover forecast from the reference run (REF), and (right) 6-h forecast from the MSG run. All model runs and observations are valid at 1200 UTC 11 Nov 2011. Locations of radiosondes are indicated by the corresponding (synoptic) stations.

  • View in gallery
    Fig. 14.

    Cloud cover forecast for (left) REF, (middle) MSG run, and (right) their difference (REF − MSG) for (top) high-, (middle) medium-, and (bottom) low-level clouds.

  • View in gallery
    Fig. 15.

    Temperature T and dewpoint temperature Td profiles from different radiosonde observations at 1200 UTC 11 Nov 2011. Locations of radiosondes are shown in Fig. 13.

  • View in gallery
    Fig. 16.

    Cloud cover observations (circles) and forecasts (curves) for the De Bilt synoptic station (06260). (top) Reference and (bottom) MSG initialization cloud cover forecasts. Each curve denotes the cloud cover forecast from +0 to +6 h.

  • View in gallery
    Fig. 17.

    Cloud cover and mean sea level pressure pmsl valid at 1300 UTC 12 Nov 2011: (left) cloud cover observations (pmsl), (middle) reference cloud cover forecast (1 h), and (right) MSG cloud cover forecast (1 h).

  • View in gallery
    Fig. 18.

    As in Fig. 17, but for observations and forecasts (6 h) valid at 1800 UTC 12 Nov 2011.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 146 57 10
PDF Downloads 86 21 2

Cloud Initialization in the Rapid Update Cycle of HIRLAM

Siebren de HaanKNMI, De Bilt, Netherlands

Search for other papers by Siebren de Haan in
Current site
Google Scholar
PubMed
Close
and
Siebe H. van der VeenKNMI, De Bilt, Netherlands

Search for other papers by Siebe H. van der Veen in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

The Nowcasting Satellite Application Facility (NWC SAF) cloud mask from the Meteosat Second Generation (MSG) satellite is introduced in the initialization step of an hourly Rapid Update Cycle (RUC) of the High Resolution Limited Area Model (HIRLAM). MSG cloud-top temperatures and synoptic cloud-base height information are combined at analysis time. This cloud initialization scheme is applied to an experimental run, which is a copy of the operational Royal Netherlands Meteorological Institute [Koninklijk Nederlands Meteorologisch Instituut (KNMI)] RUC model. The experimental run was employed during the period June–December 2011. The RUC model has a forecast length of 6 h. Cloud cover forecasts are verified against MSG cloud cover information and synoptic observations. Forecasts of precipitation, surface pressure, 2-m temperature, and upper-air temperature are verified against synoptic observations and aircraft temperature observations. It is shown that including MSG cloud information in the RUC considerably improves the forecasts of most of these model fields, when compared to the operational control RUC. Both the bias and standard deviation of the errors of the cloud cover forecast are reduced substantially. Forecasts of light precipitation show a slight negative impact, but forecasts of heavy precipitation become better. The bias in 3D temperature fields disappears nearly completely. The error bias of 2-m temperatures has become larger. Two case studies are presented. The first case study had very good forecast performance with respect to low clouds when compared to the reference run. The second case study shows an ambiguous impact; there are still some deficiencies in the cloud initialization and cloud forecast when focusing on a single location.

Corresponding author address: Siebren de Haan, Royal Netherlands Meteorological Institute, Utrechtseweg 297, NL-3731 GA De Bilt, Netherlands. E-mail: siebren.de.haan@knmi.nl

Abstract

The Nowcasting Satellite Application Facility (NWC SAF) cloud mask from the Meteosat Second Generation (MSG) satellite is introduced in the initialization step of an hourly Rapid Update Cycle (RUC) of the High Resolution Limited Area Model (HIRLAM). MSG cloud-top temperatures and synoptic cloud-base height information are combined at analysis time. This cloud initialization scheme is applied to an experimental run, which is a copy of the operational Royal Netherlands Meteorological Institute [Koninklijk Nederlands Meteorologisch Instituut (KNMI)] RUC model. The experimental run was employed during the period June–December 2011. The RUC model has a forecast length of 6 h. Cloud cover forecasts are verified against MSG cloud cover information and synoptic observations. Forecasts of precipitation, surface pressure, 2-m temperature, and upper-air temperature are verified against synoptic observations and aircraft temperature observations. It is shown that including MSG cloud information in the RUC considerably improves the forecasts of most of these model fields, when compared to the operational control RUC. Both the bias and standard deviation of the errors of the cloud cover forecast are reduced substantially. Forecasts of light precipitation show a slight negative impact, but forecasts of heavy precipitation become better. The bias in 3D temperature fields disappears nearly completely. The error bias of 2-m temperatures has become larger. Two case studies are presented. The first case study had very good forecast performance with respect to low clouds when compared to the reference run. The second case study shows an ambiguous impact; there are still some deficiencies in the cloud initialization and cloud forecast when focusing on a single location.

Corresponding author address: Siebren de Haan, Royal Netherlands Meteorological Institute, Utrechtseweg 297, NL-3731 GA De Bilt, Netherlands. E-mail: siebren.de.haan@knmi.nl

1. Introduction

Weather forecasts that look farther than a few hours into the future can only be made with numerical weather prediction (NWP) models, because the nonlinear partial differential equations that describe the atmosphere’s evolution cannot be solved analytically. Realistic initial conditions (analyses) in the NWP model based on real-time observations are crucial for making meaningful forecasts. The role of data assimilation (DA) is to compute an optimized analysis, that is, the best possible initial model state. This analysis is determined by minimizing a cost function including the background error covariances and observation error covariances together with observations and the background state or first guess. The quality of the background error covariance matrix and the quality and coverage of the observations strongly determines the quality of the analysis. In our experiments, the DA is performed using three-dimensional variational data assimilation (3DVAR).

Nowcasting can be defined as making forecasts for the very short range (0–6 h), and it is in particular relevant for warnings of dangerous weather events. For example, fog and low-visibility forecasts are crucial in aviation meteorology. Numerical nowcasting exploits the real-time atmospheric information from high-temporal-resolution and high-spatial-resolution observation datasets. But NWP models often perform poorly when used for nowcasting of clouds and precipitation development, showing unrealistic spinup and/or spindown during the first few hours of the forecast. Because of these problems alternative models are sometimes used, for example an advection model for satellite clouds like CineSat (see GEPARD GmbH 2001), which can predict satellite clouds until a few hours ahead. Also for manual nowcasting, geostationary satellite images provide essential information on clouds and cloud types.

In this article, we will focus on the nowcasting and forecasting of clouds. In the past, nowcasting of clouds using NWP was notoriously difficult. Various attempts were made to use cloud information in NWP models. Changing the upper-air cloud water mixing ratio in order to match Geostationary Operational Environmental Satellite (GOES) cloud observations resulted in a positive impact on cloud and precipitation forecasts (Bayler et al. 2000). Other studies showed an impact on cloud cover forecasts in the first 3–4 h (Yucel et al. 2002, 2003) when observed cloud fields were introduced. Experiments with direct assimilation of Meteosat water vapor (WV) clear-sky radiance data in the European Centre for Medium-Range Weather Forecasts (ECMWF) model showed a slightly positive to neutral impact on forecast quality for different areas of the globe (Munro et al. 2004; Köpken et al. 2004).

In more recent attempts, radiances of cloudy areas were introduced; see, for example, Martinet et al. (2013), who showed the possibility of improving ice and liquid water profiles in the model analysis, using artificial observations (observing system simulation experiments) for validation. Furthermore, radiances from Meteosat Second Generation (MSG) have been assimilated in cloudy scenes; see Stengel et al. (2010) who applied four-dimensional variational data assimilation (4DVAR) in the High Resolution Limited Area Model (HIRLAM). See also Stengel et al. (2013), where the impact of cloudy radiances using 4DVAR is shown to be more or less beneficial, though not in the lower troposphere.

We may conclude that although clouds play a key role in the atmosphere, operational NWP models do not (yet) assimilate clouds. This is also true for the current operational numerical nowcasting model [Rapid Update Cycle (RUC)] at the Royal Netherlands Meteorological Institute [Koninklijk Nederlands Meteorologisch Instituut (KNMI)], which has no cloud initialization or cloud assimilation in its analysis step. Only upper-air-temperature- and specific-humidity-related observations are assimilated, the origin of these data being aircraft temperature and Global Navigation Satellite System Zenith Total Delay (GNSS ZTD) observations.

In this article, a novel technique developed by van der Veen (2013) to ingest cloud information is applied in a semioperational RUC setting. Cloud mask information from the Nowcasting Satellite Application Facility (NWC SAF; Derrien and Le Gléau 2005) and cloud-base height surface observations are combined to adjust the model clouds at the initial forecast step.

The impact of the cloud initialization is shown in this paper, which is organized as follows. First, a description of the NWP model used in this study is presented, followed by an explanation of the setup of the numerical experiments. A discussion of the verification results follows in section 3.

2. Observations and the HIRLAM Rapid Update Cycle

a. Observations

To be of use for a numerical nowcasting application, latency of the observations is crucial. The NWC SAF observations are quickly available and thus very suitable for these kinds of applications. Radar radial velocity observations are available within 5 min after observation, and the same is true for (most) Aircraft Meteorological Data Relay (AMDAR; Painting 2003) and aircraft observations gathered using the Selective Enhanced Surveillance Mode (Mode-S EHS) of an air traffic control radar (de Haan 2011). The latter observations, which have a short latency, have shown to be beneficial for short-term forecasts (de Haan and Stoffelen 2012; de Haan 2013). GNSS ZTD observations require a processing method; de Haan et al. (2009) showed that these observations can be produced within 10 min every 15 min. Radiosonde observations have an inherent latency of generally more than 40 min due to the time a balloon takes to reach the top of the profile. Satellite-based observations, such as sea surface winds from scatterometers measured on board polar-orbiting satellites, may have a too long latency due to data collection at only a limited number of ground stations (de Haan et al. 2013). For other satellite observations, datasets may not be available completely within the cutoff time, which is 10 min [e.g., Advanced Television Infrared Observation Satellite (TIROS) Operational Vertical Sounder (ATOVS) brightness temperatures]. Furthermore, these raw observations need processing, which adds to the latency, implying that these observations cannot be used in a RUC scheme with a short observation cutoff time. The observations used in this study are available within 10 min; the latency of the NWC SAF product is 7 min and for the cloud-base observations it is 10 min.

b. Rapid Update Cycle and initialization of HIRLAM

KNMI runs an operational hourly RUC using HIRLAM, version 7.3 (HIRLAM 7.3; Undén et al. 2002; de Haan and Stoffelen 2012; de Haan 2013; de Haan et al. 2013), with a horizontal resolution of 11 km; the domain area is shown as U11 in Fig. 1. The lateral boundary conditions are retrieved from HIRLAM forecasts on a larger domain (H11) with the same horizontal and vertical resolution, but with a 3-h assimilation cycle. Both models (U11 and H11) use a forecast of a previous run valid at assimilation time as the first guess.

Fig. 1.
Fig. 1.

HIRLAM H11 and U11 domains.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

The operational RUC is considered to be the reference experiment. An experimental configuration, being identical to the operational, was set up to test new model versions or the use of additional observations. This experimental configuration has a separate assimilation and forecast cycle. The experiment discussed in this paper differs from the operational RUC in how humidity (and temperature) is changed after the model initialization. Both cycles used basically the same datasets of observations and used the HIRLAM 3DVAR analysis system to determine an analysis based on observations and background information. Model-state upper-air winds and temperatures are constrained using observations from AMDAR and Mode-S EHS. Pressure is initialized using the pressure observations from land stations, buoys, and ships. The humidity at analysis time is constrained by GNSS ZTD observations (see de Haan 2013). Besides GNSS observations, radar radial velocities are also assimilated in both runs; see de Haan (2013) for more details. Table 1 shows the model configuration and settings used in this study. Note that no cloud information is used in the assimilation. The experiment was performed in a semioperational setting during the period June–December 2011. The experimental run, called the MSG run in the following, was not operational and had suffered from some failures due to a lack of computer resources as well as system crashes. On a number of days during the period, the synoptic cloud-base observations were not present; for these cases the cloud initialization was not performed properly and these runs are removed from the results. The dates on which this occurred are shown in Table 2.

Table 1.

HIRLAM model settings used and assimilated observations.

Table 1.
Table 2.

Failed and successful runs.

Table 2.

The upper-air humidity and temperature changes applied to the experimental run aim at a better match between the model and the NWC SAF cloud mask, MSG cloud-top temperatures, and synoptic observations of cloud-base heights. The amount of cloud cover derived from the cloud mask is computed from a 3 × 5 pixel rectangular area, with the center pixel closest to the model grid point. It is defined as the number of cloudy pixels divided by 15. The size of the 3 × 5 area of MSG pixels is a little less than 20 km × 20 km over the Netherlands, covering nearly four HIRLAM grid boxes. Cloud-top heights are assigned to the highest HIRLAM level where the temperature equals the temperature measured by MSG with the 10.8-μm channel. Cloud-base heights are derived from synoptic observations, with a representative spatial scale of approximately 50 km. All data are resampled to the HIRLAM grid using the nearest neighbor.

The cloud initialization method as developed by van der Veen (2013) was applied in this experiment. Humidity profiles were altered in order to match the cloud mask. For a cloud-free grid box (according to the cloud mask) the profile is changed to sufficiently unsaturated, in case it was too humid. For a cloudy grid box the profile becomes more or less saturated, depending on the amount of subgrid cloud cover that MSG prescribes. In the latter case, the humidity profile is changed between the observed bottom and the assigned cloud top. Apart from humidity, temperature is changed slightly such that the stability properties of the profile are maintained (i.e., no change in virtual temperatures). Figure 2 shows an example of an initial profile and an adjusted profile, for both “cloud creation” and “cloud removal.” The appendix shows more detailed information on the cloud initialization scheme applied.

Fig. 2.
Fig. 2.

Changes in vertical profiles of specific humidity for two cases: (left) a cloud added by the cloud initialization using MSG and (right) a cloud removed in the initialization.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

In this study, a limit of 10% was imposed on changes in initial specific humidity profiles. This limit for the change in specific humidity was chosen because it appeared that in an hourly update cycle spinup effects of humidity-related model parameters occurred when no limit was applied. The influence of limiting the specific humidity change on the cloud cover and rainfall forecast was tested for hourly runs performed over a 6-day period. In Fig. 3 (left), the mean rainfall rate over the whole domain for each time step is shown for the reference (red) together with runs with humidity limits from 5% to 50%, for 144 hourly cycles. For a limit on the humidity change of 50%, spindown of both cloud cover and rainfall rate is observed. The reference run (red solid line) shows a spinup in rainfall rate and a spindown in cloud cover after 3 h in the forecast. By limiting the humidity change to 10%, model behavior exhibited the smallest variation for forecast times of 30 min and longer. Spindown in cloud cover is also reduced, but only later in the forecast. Note that the wiggles at the whole hours in the cloud cover forecast are due to the lateral boundaries where the forecast field is prescribed. The choice of 10% deviates from an earlier study (30%; see van der Veen 2013), because here we apply an hourly update cycle while in the previous study it was applied in a 6-hourly cycling scheme with cloud initialization occurring only twice a day at 0000 and 1200 UTC.

Fig. 3.
Fig. 3.

Mean (left) rainfall rate and (right) cloud cover over the whole U11 domain for each time step (forecast time indicates the hours into the forecast) based on cycled hourly experiments over a period of 6 days for different humidity change limits.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

3. Verification of forecasts

In this section, the verification results are discussed. For the entire period, cloud cover, surface and upper-air temperatures, and rainfall forecasts are evaluated by comparison with observations.

a. Total cloud cover

Figure 4 shows the bias and standard deviation of the errors of forecast total cloud cover (octas) as well as an indication of the accuracy of these statistics. It should be noted that in this paper bias is defined as “forecast − observation.” The jackknife method was applied to the dataset, in order to show the statistical significance of the observed bias and standard deviation. The dataset was divided into 50 subsets from which 50 biases and standard deviations were calculated; the mean and uncertainty of these biases and standard deviations for different forecast times are shown in Fig. 4. From this figure, it is clear that the bias and standard deviation in the cloud cover forecasts are reduced over the whole forecast range. A model spinup is still present, because the standard deviation is first reduced (in both model runs) and then increased after 3 h of forecast. In Fig. 5 the mean and variation of the cloud cover forecasts are shown for both runs, together with the observed mean and variation according to the MSG cloud mask (denoted as OBS; dashed line). The standard deviation of both model runs is higher than in the observations. The standard deviation of the reference run shows an increase, while the MSG run is more constant and initially equal to the reference run. Note also that the mean cloud cover of the reference run is lower than in the observations, while for the MSG run the mean cloud cover is higher. This is in line with the chosen value for the change limit of 10% (see Fig. 3).

Fig. 4.
Fig. 4.

Mean difference (bottom two curves) and std dev (top two curves) between cloud cover forecast and MSG cloud cover observations of the reference run (solid) and the MSG run (dashed); the shaded area denotes uncertainty obtained by the jackknife method.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

Fig. 5.
Fig. 5.

Mean and variation of cloud cover observations and forecasts. The two solid lines refer to the cloud forecasts from the reference run; the dashed lines refer to the cloud forecast of the MSG run, and the dotted (constant) lines refer to the cloud cover observations.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

Figure 6 shows the time series of the bias and standard deviation of the difference between cloud cover forecasts and observations on a daily basis. There appears to be a period for which the reference run has a large daily mean bias, up to nearly 4 octas, while the MSG run has a strongly improved bias. For this period, the standard deviation is also smaller; note that this period is the focus of two case studies presented in section 4.

Fig. 6.
Fig. 6.

Time series of the mean difference and std dev of model cloud cover forecast compared to MSG cloud cover observations on a daily basis, for the whole period of the experiment.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

b. Precipitation verification against synoptic observations

The precipitation forecasts are verified against hourly rainfall observations obtained from the synoptic network in the Netherlands (see Fig. 7). This network consists of 32 automatic weather stations that measure basic meteorological parameters, such as wind, temperature, and humidity. The automatic precipitation sensor used at these locations was developed by KNMI.

Fig. 7.
Fig. 7.

Locations of the synoptic stations.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

For the period under consideration, the Heidke skill score (HSS) is determined for both forecast runs. The HSS is a measure of the fractional improvement of the forecast over a forecast performed by chance. The range of HSS can vary from −∞ to 1, with negative values indicating that chance is better, an HSS value of 0 represents “no skill,” and an HSS equal to 1 denotes a perfect forecast. Based on the computed contingency table related to a given threshold, the HSS is defined as
e1
where a, b, c, and d are defined by a contingency table as shown in Table 3.
Table 3.

Contingency table of rainfall r for a threshold R.

Table 3.

The Heidke skill scores for rainfall are determined for threshold values of 0.5, 1, and 2 mm h−1 for both RUC runs. The results are shown in Fig. 8. For the lowest threshold (0.5 mm h−1) the impact of MSG initialization is negative, as its HSS is lower than the HSS of the reference run. However, for higher thresholds a positive impact is observed, and a more or less neutral impact for these thresholds is seen later in the forecast. The HSS based on the reference run with a threshold of 1 mm h−1 shows a clear spinup in the first 4 h. Almost no spinup is observed for the run with MSG cloud initialization. The same signal is also present for the 2 mm h−1 threshold. The MSG run shows a higher HSS in the first hours of the forecast for the 1 and 2 mm h−1 thresholds. Note that this score may suffer from a double penalty: a forecast of rainfall at a wrong location and/or time is penalized twice.

Fig. 8.
Fig. 8.

HSS (for different forecast lengths) of the rainfall forecast compared to the synoptic hourly rainfall observations for three threshold values: 0.5, 1, and 2 mm h−1.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

To obtain some notion of the general rainfall forecast performance of both runs, the cumulative distribution of hourly rainfall for the synoptic stations in the Netherlands is calculated and compared to the cumulative distribution of the rainfall-rate forecasts. The focus here is on observed rainfall events at synoptic stations; only forecasts are used for which at target time at least one synoptic station has rainfall reported. By calculating the cumulative distributions over an area (i.e., the Netherlands), instead of using single-point observations, the climatological rainfall signal of forecasts from both model runs can be obtained and compared. These distributions are shown in Fig. 9 together with the cumulative distributions of synoptic hourly rainfall observations. The reference run clearly shows an underestimation of the high rainfall forecasts. The cumulative distribution is higher than that of the observations, which implies that the number of forecasts with lower rainfall-rate values than, for example 1 mm h−1, is higher than the observed number of cases for this threshold. The cumulative rainfall-rate distribution from the MSG initialization run matches the observed distribution better, especially for high rainfall rates. For rainfall rates smaller than 0.2 mm h−1, no difference between model runs and observations can be detected. The rainfall distribution in the 1-h forecast has a closer match with the observed one for lower rainfall rates. For rainfall rates higher than 0.2 mm h−1, the two runs behave differently. The rainfall forecast for the first hour is overestimated by the MSG run, up to 0.5 mm h−1. For this rainfall-rate value, the 6-h forecast of the MSG run matches the observation distribution very well. Above 0.5 mm h−1, the 6-h forecast underestimates the total rainfall, while the 1-h forecast is close to the observations. For high rainfall rates, the cumulative rainfall distribution from the MSG run matches the observations better than the reference run. For rainfall rates between 0.2 and 0.5 mm h−1, a more or less neutral impact is observed when applying MSG initialization. This is in agreement with the HSS statistics presented earlier.

Fig. 9.
Fig. 9.

Observed and forecast cumulative rainfall-rate distributions for the whole period for (left) the REF run and (right) the MSG run.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

c. Upper-air temperature fields

Figure 10 shows the mean and standard deviation of the error of the temperature forecast compared to observed AMDAR and Mode-S EHS temperatures. Again, the shaded area denotes the uncertainty in the estimated bias and standard deviation. The bias in temperature forecasts at heights above the 925-hPa level is significantly closer to zero for MSG cloud initialization runs; this is true for all forecast times. The standard deviation of the temperature error is equal. At lower model levels, around 925 hPa and below, the bias has changed very little, while the standard deviations have increased very slightly, though not significantly.

Fig. 10.
Fig. 10.

Mean and std dev of the difference between forecast upper-air temperature and observed AMDAR and Mode-S EHS temperature (01, 02, 06); the shaded area denotes the uncertainty in the mean difference and the std dev obtained by the jackknife method.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

d. 2-m temperatures

The 2-m temperature forecasts are compared to the synoptic 2-m temperature observations including estimates of the uncertainty. These observations are not assimilated in any of the model runs. The 2-m temperature forecasts are derived from the model using a postprocessing step (L. Rontu 2013, personal communication). This step calculates the temperature at 2 m using the stability of the atmospheric profile, the turbulent latent and sensible heat fluxes, and the surface properties at the grid point. Figure 11 shows the statistics of the comparison of the 2-m temperature forecast with the observed temperature. A negative impact on the temperature bias is observed. A solution for this negative bias might lie in the revision of the tuning of the postprocessing scheme to derive the 2-m temperature. The bias becomes more negative when MSG information is assimilated. A neutral impact is observed for the 2-m temperature standard deviation.

Fig. 11.
Fig. 11.

The 2-m temperature forecast errors; the shaded area denotes the uncertainty in the bias and the std dev obtained by the jackknife method.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

e. Surface pressure

In Fig. 12, it is shown that the positive bias of the surface pressure (0.1 hPa) completely disappears when MSG clouds are introduced. At first sight, this may appear strange because the humidity initialization procedure leaves pressure fields unchanged. But, on average, humidity is increased during initialization, and this implies more rainfall, especially during the beginning stages of the forecast (Fig. 3, left). This extra rainfall removes mass (water) from the atmosphere, implying a surface pressure drop in the MSG run, averaged over a long period. The decrease of ≈0.1 hPa in Fig. 12 corresponds to a liquid water layer of approximately 1 mm. Clearly, the bias of the forecast with MSG initialization is very close to zero, unlike the reference run. The standard deviation of the MSG run and the reference do not differ significantly from each other.

Fig. 12.
Fig. 12.

Mean difference and std dev between forecast and observed model pressure for forecast times 1–6 h. The shaded area denotes the uncertainty in the bias and the std dev obtained by the jackknife method.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

4. Case studies

a. 11 November 2011

The focus of the first case study is on 11 November 2011. On this day, a high pressure region was present over eastern Europe with a strong ridge stretching over the North Sea (see Fig. 13). Due to this blocking, low pressure systems were not able to reach the continent. On this date low clouds and fog were present over the Netherlands.

Fig. 13.
Fig. 13.

(left) Pressure reduced to mean sea level (contoured) and observed cloud cover from NWC SAF, (middle) 6-h cloud cover forecast from the reference run (REF), and (right) 6-h forecast from the MSG run. All model runs and observations are valid at 1200 UTC 11 Nov 2011. Locations of radiosondes are indicated by the corresponding (synoptic) stations.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

Figure 14 shows the low- (below 2 km), medium- (between 2 and 6 km), and high-level (above 6 km) cloud cover for both runs (left shows the reference, middle shows the MSG, and right shows their difference). The major difference in cloud cover is observed over the Netherlands, Belgium, northern France, and southern Germany. The differences in cloud cover are largest for low cloud cover over land. Differences in medium-level cloud cover are largest over the North Sea, near the east coasts of England and Scotland. The radiosonde observations of temperature and dewpoint temperature at 1200 UTC are shown in Fig. 15. The locations of the radiosondes are depicted in Fig. 13.

Fig. 14.
Fig. 14.

Cloud cover forecast for (left) REF, (middle) MSG run, and (right) their difference (REF − MSG) for (top) high-, (middle) medium-, and (bottom) low-level clouds.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

Fig. 15.
Fig. 15.

Temperature T and dewpoint temperature Td profiles from different radiosonde observations at 1200 UTC 11 Nov 2011. Locations of radiosondes are shown in Fig. 13.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

The radiosondes shown in Fig. 15 are located on a more or less straight line, roughly from west to east: from Brest in France (radiosonde 07110; 48.45°N, 4.42°W) to Meiningen in Germany (radiosonde 10548; 50.57°N, 10.38°E). The most western profile is almost completely saturated up to 4-km height. Radiosonde 07145 launched from Trappes in France (48.76°N, 2.02°E) shows a saturated layer in the lowest 500 m of the profile, with a moderate temperature inversion directly above this layer. Radiosonde 06260 (De Bilt in the Netherlands; 52.06°N, 5.11°E) shows a thicker low-cloud layer and a stronger inversion. This low-cloud layer is also visible in the profile from radiosonde 10618 (Idar-Oberstein in Germany; 49.70°N, 7.33°E). The inversion above the cloud layer is weaker here. The most eastern radiosonde (Meiningen, 10548) is not saturated. The cloud cover at Brest is forecast in both runs. The observed low cloud cover at the other three radiosonde sites matches the cloud cover forecasts from the MSG run, while the reference run has no clouds forecast at these radiosonde locations.

Figure 16 shows the total cloud cover forecasts for De Bilt (for forecast times from +0 to +6 h) starting at 0000 UTC 10 November and ending at 0600 UTC 13 November. All forecasts for each hourly run are shown; each separate run is depicted by a separate curve. The differences between cloud cover at 1000 UTC 10 November and later are clearly visible. It is interesting to see that some MSG runs forecast fractured clouds (forecasts started around 1200 UTC 10 November) and they show some persistence in consecutive forecasts. Furthermore, the observed (almost) cloud-free period from 1500 UTC 11 November until 1200 UTC 12 November is, at least partially, predicted by the MSG run. The difference between the broken cloud cover in the model and observations may be influenced by the fact that cloud cover in the model is actually a parameterization based on temperature and humidity. Moreover, model representativeness will differ from the observations. An observation of 1 octa should not necessarily be equal to a model forecast of cloud cover fraction of ⅛, since these parameters are not equal. Nevertheless, it is clear from Fig. 16 that the cloud fraction forecasts of the MSG run lie closer to the observed synoptic cloud cover than the reference run. We may conclude that the case study presented here shows a positive impact on the low-level cloud (and fog) forecasts when cloud forcing is applied.

Fig. 16.
Fig. 16.

Cloud cover observations (circles) and forecasts (curves) for the De Bilt synoptic station (06260). (top) Reference and (bottom) MSG initialization cloud cover forecasts. Each curve denotes the cloud cover forecast from +0 to +6 h.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

b. Afternoon of 12 November 2011

The second case focuses on a similar weather situation occurring during the afternoon of 12 November. In Fig. 16, it is seen that at this time the MSG and reference runs are both different from the observed cloud cover in De Bilt. Figure 17 shows the total cloud cover as observed by MSG and the 1-h forecasts from the reference and MSG runs, which are valid at 1300 UTC 12 November 2011.

Fig. 17.
Fig. 17.

Cloud cover and mean sea level pressure pmsl valid at 1300 UTC 12 Nov 2011: (left) cloud cover observations (pmsl), (middle) reference cloud cover forecast (1 h), and (right) MSG cloud cover forecast (1 h).

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

Apart from a region close to the Scottish east coast, the North Sea is completely covered by clouds. These clouds partly cover the western part of the Netherlands, Belgium, and the north of France. The edge of the cloud sheet is located near De Bilt in the center of the Netherlands. Another cloud field is visible over the south of Germany. The south of England is also covered by clouds. The 1-h forecast of cloud cover from the reference run is shown in Fig. 17 (middle). Clearly, this forecast missed the clouds over the southern North Sea and Belgium, and underestimates the clouds over southern Germany, southern England, and northern France. The cloud cover forecast from the MSG run is shown in Fig. 17 (right). This cloud cover matches the MSG cloud mask better (North Sea, southern England, and southern Germany) but it also missed the cloud cover over parts of northern France. The forecast broken cloud cover west of the Netherlands is seen to be fully cloudy in the MSG cloud mask.

The 6-h forecasts of the reference and MSG runs are shown in Fig. 18, together with the MSG cloud cover valid at 1800 UTC 12 November 2011. The cloud sheet over the North Sea has moved eastward, covering the Netherlands almost completely. Near the Belgian coast, a clearing appeared. The cloud sheet over southern Germany was almost stationary for 5 h. The cloud cover forecast from the reference run shows clear skies over the Netherlands and the southern North Sea, but cloud cover over Belgium. Some small clouds in southern Germany are visible; however, they are underestimated. The 6-h forecast of the MSG run shows furthermore that the cloud sheet has moved eastward but it appears that the clouds over the Netherlands have disappeared partly. The clouds over southern Germany are still present in this forecast though.

Fig. 18.
Fig. 18.

As in Fig. 17, but for observations and forecasts (6 h) valid at 1800 UTC 12 Nov 2011.

Citation: Weather and Forecasting 29, 5; 10.1175/WAF-D-13-00071.1

5. Conclusions and outlook

In this paper, it is demonstrated that inserting MSG clouds and cloud-base information into a Rapid Update Cycle simulation of HIRLAM has a significant positive impact on quite a number of forecast model fields, such as 3D temperature fields, clouds, precipitation (except small amounts), and surface pressure. The 2-m temperature forecast when using MSG cloud initialization was detrimental, although the reason for this is not fully understood and needs further investigation. These worse forecasts may be due to the postprocessing method, which was tuned for the standard HIRLAM configuration and may need retuning when MSG information is introduced into the initialization. The control RUC run showed a spinup of the rainfall rate in the first 6 h of the forecast, which is absent in the MSG run when using the limit to specific humidity change of 10%. Both the MSG RUC and the control RUC exhibited short-lasting spinup behavior for cloud cover. Such a spinup was absent in all other model fields.

A case study showing forecast failures of low cloud and fog by the operational RUC model illustrated the importance of cloud initialization for improving cloud forecasts. Comparison with synoptic cloud cover observations showed very good agreement with MSG cloud forecasts, while the reference run clearly showed (low level) cloud forecast failures over a period of several days. A second case study, 1 day later, showed that cloud initialization and forecasts had a lower quality when compared to the NWC SAF cloud cover. Nevertheless, also on this day the run with MSG cloud initialization performed better than did the reference run in most locations.

For the whole verification period of 6 months, the overall impact of initializing clouds is clearly positive; the presented cloud initialization method has therefore been implemented in the operational RUC model at KNMI.

The method outlined in this article can also be used to initialize model liquid water content simultaneously with water vapor, provided that quantitative observations are available. Possible sources of information are radar reflectivity, ceilometer backscatter profiles, and/or satellite products like vertically integrated liquid water content. Apart from this, the “latency problem” of valuable observations can be overcome by performing an extra analysis valid for the previous hour. The 1-h forecast from this delayed analysis can then be used as the first guess for the “real time” analysis. In this way, all “late arriving” information (from, e.g., radiosonde and satellite data) is incorporated into the operational run by the assimilation using this improved first guess.

Acknowledgments

The authors thank EUMETSAT for allowing the use of the Nowcasting Satellite Application Facility cloud mask product and the MSG data. The Air Traffic Control, the Netherlands (LVNL), is kindly acknowledged for providing Mode-S EHS observations. Part of this study has been funded by the Knowledge & Development Centre Mainport Schiphol in the Netherlands (KDC; http://www.kdc-mainport.nl).

APPENDIX

Humidity Profile Adjustment

This appendix shows some details of the humidity profile adjustment as described in van der Veen (2013).

Assume we have in a HIRLAM grid box observations from MSG cloud fraction NMSG, cloud-base height hc, and cloud-top temperature Tc. The specific humidity q (and T) will be adjusted at all levels in the grid box. Consider level j. The pressure at this level is given by pj, while the height hj can be calculated by integration of the temperature and humidity profile from the surface to level j, given the surface pressure. Let be the critical relative humidity value at level j, from which, when exceeded, clouds can start to form. This critical value has different values for different heights in the profile: lowest in the midtroposphere RHmin and highest near the earth surface RHmax and at the tropopause. We define as
ea1
where ps is the surface pressure. In HIRLAM 7.3, the cloud cover N is primarily a function of the relative humidity. The relation is simplified by discarding the weak dependence on T; that is,
ea2
This relation approximates the Kain–Fritsch cloud cover definition in HIRLAM 7.3.
When the HIRLAM grid box is cloudy according to the MSG cloud product (NMSG > 0), the height of the level is larger than the cloud base (hj > hc), and the temperature is higher than the cloud-top temperature (Tj > Tc), then the specific humidity is changed according to Eq. (A2):
ea3
where qs is the saturation specific humidity.
When the HIRLAM grid box is cloud free according to the MSG cloud product and always when (Tj < Tc) and/or (hj < hc), the specific humidity is decreased such that the relative humidity will never be higher than the critical relative humidity ; that is,
ea4

After this procedure, the virtual temperature Tυ = T(1 + 0.6077qql) will have slightly changed, where ql is the liquid water content. We now adjust T such that Tυ remains as it was. But now qs has slightly changed, so the whole initialization procedure must be repeated. This repetition is performed a number of times in an (converging) iterative loop.

REFERENCES

  • Bayler, G. M., Aune R. M. , and Raymond W. H. , 2000: NWP cloud initialization using GOES sounder data and improved modeling of nonprecipitating clouds. Mon. Wea. Rev., 128, 39113920, doi:10.1175/1520-0493(2001)129,3911:NCIUGS.2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • de Haan, S., 2011: High-resolution wind and temperature observations from aircraft tracked by Mode-S air traffic control radar. J. Geophys. Res.,116, D10111, doi:10.1029/2010JD015264.

  • de Haan, S., 2013: Assimilation of GNSS ZTD and radar radial velocity for the benefit of very-short-range regional weather forecasts. Quart. J. Roy. Meteor. Soc.,139, 2097–2107, doi:10.1002/qj.2087.

  • de Haan, S., and Stoffelen A. , 2012: Assimilation of high-resolution Mode-S wind and temperature observations in a regional NWP model. Wea. Forecasting,27, 918937, doi:10.1175/WAF-D-11-00088.1.

    • Search Google Scholar
    • Export Citation
  • de Haan, S., Holleman I. , and Holtslag A. A. M. , 2009: Real-time water vapor maps from a GPS surface network: Construction, validation, and applications. J. Appl. Meteor. Climatol., 48, 13021316, doi:10.1175/2008JAMC2024.1.

    • Search Google Scholar
    • Export Citation
  • de Haan, S., Marseille G.-J. , de Valk P. , and de Vries J. , 2013: Impact of ASCAT scatterometer wind observations on the High Resolution Limited Area Model (HIRLAM) within an operational context. Wea. Forecasting,28, 489–503, doi:10.1175/WAF-D-12-00056.1.

  • Derrien, M., and Le Gléau H. , 2005: MSG/SEVIRI cloud mask and type from SAFNWC. Int. J. Remote Sens., 26, 47074732, doi:10.1080/01431160500166128.

    • Search Google Scholar
    • Export Citation
  • GEPARD GmbH, 2001: CineSat software user manual. Operation Manual 120, Gepard GmbH, Vienna, Austria, 225 pp.

  • Köpken, C., Kelly G. , and Thépaut J. N. , 2004: Assimilation of Meteosat radiance data within the 4D-Var system at ECMWF: Assimilation experiments and forecast impact. Quart. J. Roy. Meteor. Soc., 130, 22772292, doi:10.1256/qj.02.230.

    • Search Google Scholar
    • Export Citation
  • Martinet, P., Fourrié N. , Guidard V. , Rabier F. , Montmerle T. , and Brunel P. , 2013: Towards the use of microphysical variables for the assimilation of cloud-affected infrared radiances. Quart. J. Roy. Meteor. Soc., 139, 14021416, doi:10.1002/qj.2046.

    • Search Google Scholar
    • Export Citation
  • Munro, R., Köpken C. , Kelly G. , Thépaut J.-N. , and Saunders R. , 2004: Assimilation of Meteosat radiance data within the 4D-Var system at ECMWF: Data quality monitoring, bias correction and single-cycle experiments. Quart. J. Roy. Meteor. Soc., 130, 22932313, doi:10.1256/qj.02.229.

    • Search Google Scholar
    • Export Citation
  • Painting, J. D., 2003: WMO AMDAR reference manual. WMO-958, WMO, Geneva, Switzerland, 80 pp.

  • Stengel, M., Lindskog M. , Undén P. , Gustafsson N. , and Bennartz R. , 2010: An extended observation operator in HIRLAM 4D-Var for the assimilation of cloud-affected satellite radiances. Quart. J. Roy. Meteor. Soc., 136, 10641074, doi:10.1002/qj.621.

    • Search Google Scholar
    • Export Citation
  • Stengel, M., Lindskog M. , Undén P. , and Gustafsson N. , 2013: The impact of cloud-affected IR radiances on forecast accuracy of a limited-area NWP model. Quart. J. Roy. Meteor. Soc., 139B, 2081–2096, doi:10.1002/qj.2102.

    • Search Google Scholar
    • Export Citation
  • Undén, P., and Coauthors, 2002: HIRLAM-5 scientific documentation. HIRLAM Project Tech. Rep., Norrköping, Sweden, 144 pp. [Available online at http://hirlam.org/index.php?option=com_content&view=article&id=64:gneral-description-of-the-hirlam-model&catid=48:synoptic-scale-model-hirlam&Itemid=101.]

  • van der Veen, S. H., 2013: Improving NWP model cloud forecasts using Meteosat Second-Generation imagery. Mon. Wea. Rev., 141, 15451557, doi:10.1175/MWR-D-12-00021.1.

    • Search Google Scholar
    • Export Citation
  • Yucel, I., Shuttleworth W. J. , Pinker R. T. , Lu L. , and Sorooshian S. , 2002: Impact of ingesting satellite-derived cloud cover into the Regional Atmospheric Modeling System. Mon. Wea. Rev., 130, 610628, doi:10.1175/1520-0493(2002)130<0610:IOISDC>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Yucel, I., Shuttleworth W. J. , Gao X. , and Sorooshian S. , 2003: Short-term performance of MM5 with cloud-cover assimilation from satellite observations. Mon. Wea. Rev., 131, 17971810, doi:10.1175/2565.1.

    • Search Google Scholar
    • Export Citation
Save