Assimilation of GNSS ZTD and radar radial velocity for the benefit of very-short-range regional weather forecasts



Wind, humidity and temperature observations from aircraft and radiosondes are generally used to find the best initial state of the atmosphere for numerical weather prediction (NWP). To be of use for very-short-range numerical weather forecasting (or numerical nowcasting), these observations need to be available within several minutes after observation time. Radiosondes have a typically observation latency of over 30 min and arrive too late for numerical nowcasting. Zenith Total Delay (ZTD) observations obtained from a ground-based network of Global Navigation Satellite System (GNSS) receivers can fill this gap of lacking rapid humidity information. ZTD contains information on the total amount of water vapour. Other rapidly available observations, such as radial wind estimates from Doppler weather radars, can also be exploited. Both observations are available with a delay of less than 5 min with adequate spatial resolution. In this article, the impact of assimilation of these humidity and wind observations in a very-short-range regional forecast model is assessed over a four-month summer period and a six-week winter period. As a reference for the impact, GNSS observations are also assimilated in a three-hourly NWP scheme with longer observation cut-off times. The quality of the forecasts is evaluated against radiosonde observations, radar radial wind and hourly precipitation observations. Assimilation of both GNSS ZTD and radar radial winds resulted in a positive impact on humidity, rainfall and wind forecasts.

1. Introduction

Good insight on the state of the atmosphere is essential for short-range mesoscale forecasting and nowcasting of extreme weather events. Short-range weather forecasting based on forecasts from a numerical weather prediction (NWP) model has become feasible with the growing number of rapidly available upper-air observations and growing computer resources (Benjamin, et al., 2010; de Haan and Stoffelen 2012). Observing small-scale structures is relevant for a good initialization of the model. Generally, the initial state is obtained by assimilation of wind, temperature and humidity observations. These observations are combined with a background field to estimate the status of the atmosphere at assimilation time. The optimal estimate of the status of the atmosphere is dependent on the quality of the background, the quality and number of observations, the prescribed boundary conditions and a priori assumptions on observation- and model background-error covariances.

For nowcasting applications based on a numerical model, prompt observation availability is essential: observations should be available within at least 10 min from the observation time, preferably even sooner. The method of observation by radiosondes inherently implies some data delay, as the profile is generally broadcast at the end of the ascent. Currently, some radiosonde observations are delivered in data packets, but the top of the profile will be too late for a fast assimilation scheme.

Upper-air humidity observations are sparse. Currently, in the operational high-resolution limited-area model (HIRLAM) at the Royal Netherlands Meteorological Institute (KNMI), only radiosonde observations provide information on the moisture content of the atmosphere. Aircaft reporting via AMDAR (Aircraft Meteorological Data Relay) and equipped with a humidity sensor may also provide a humidity profile, however the quality of these observations is not yet adequate. Humidity-related satellite observations can also be very valuable but cannot be assimilated in the current operational version of HIRLAM. These observations have potential; infrared radiances from geostationary satellites have been successfully used in 4D-Var experiments with HIRLAM (Stengel, et al., 2009), and Montmerle, et al. (2007) showed positive impact from infrared radiances from Spinning Enhanced Visible and Infrared Imager (SEVIRI) data.

Observations from the Global Navigation Satellite System (GNSS) can provide rapidly available information on upper-air humidity. These observations are operationally available within several minutes of the observation time (de Haan, et al., 2009) and can be valuable for nowcasting applications. A number of assimilation studies have shown positive impact of these observations (or derived observations such as precipitable water) on forecasts by numerical models (Vedel and Huang 2004; Seko, et al., 2004; Poli, et al., 2007; Benjamin, et al., 2010; Shoji, et al., 2011).

Aircraft wind observations are of good quality and may have good spatial and temporal resolution. However, an observation profile is restricted to the vicinity and operational hours of the airport. Wind observations from Doppler radar can provide three-dimensional wind information in unsampled areas, provided there are atmospheric scattering particles (precipitation or insects) (Lindskog, et al., 2004; Xiao, et al., 2008). Since 2008, Météo-France has been assimilating radial winds operationally; Montmerle and Faccani (2009) found positive impacts on the analyses, with three-hour cycles, and on precipitation forecasts.

The impact of GNSS humidity and radar radial winds is shown in this article, in an hourly NWP cycle and in a more general three-hourly NWP cycle (using GNSS humidity). The latter cycle is similar to the operational setting at KNMI, and exploits the radiosonde observation for initialization of upper-air humidity. The hourly cycle, on the contrary, has no radiosonde humidity initialization and will therefore be expected to benefit from GNSS humidity observations. The impact is shown by comparison of the forecast hourly rainfall rates with rainfall estimates from the two weather radars in the Netherlands. Also, wind and humidity forecasts are verified against radiosonde observations from De Bilt and radial winds as measured by the rain radars.

This article is organized as follows: first the observations used for assimilation and/or verification are described (GNSS, Mode-S, AMDAR, radiosonde, surface pressure observations and radar radial velocities and radar rainfall rates). Then a description is given of the NWP models used and the different settings applied. Next the impact of assimilation is shown by comparison of forecast fields with radiosonde humidity and radar rainfall rates, and radiosonde and radar wind observations. Finally the conclusions are presented.

2. Data

This section describes the observations used for assimilation and verification, and the NWP model and assimilation experiment set-up.

2.1. GNSS atmospheric observations

Since October 2005, KNMI has processed a Dutch network of 35 permanent GNSS stations using Bernese 5.0 software (Hugentobler, et al., 2006). The GNSS system consists of a space segment of at least 24 satellites, a ground segment controlling the satellites, and a user segment, receiving the signals from the satellites. The satellites transmit signals in two frequencies. A user can determine its position, when signals from more than three satellites are received. The signal is delayed by the ionosphere and the neutral atmosphere. The delay due to the ionosphere is frequency dependent due to the dispersive behaviour of the ionosphere. This delay can be cancelled to the first order by differencing the time observations from different frequencies. The atmospheric delay is related to the refractivity along the signal path (Bevis, et al., 1992). By processing all observed slant delays within a certain time window, errors and unknowns, such as satellite and receiver clock errors, can be estimated (Hugentobler, et al., 2006). An estimate of the Zenith Total Delay (ZTD), which is the slant delay mapped to the zenith, is determined for each GNSS receiver in the network. ZTD can be expressed as

equation image(1)

where za is antenna height (m), Rd = 287.05 J kg−1K−1 is the gas constant of dry air, ρ is the water vapour density (kg m−3) and equation image is the ratio of molar weights. The empirical constants are (Thayer 1974) k1 = 77.6 K hPa−1, k2 = 70.4 K hPa−1 and k3 = 373900 K2hPa−1.

The network of GNSS double-frequency receivers used here was initially constructed for operational geodetic applications (land surveying, levelling). A sub-network of the Dutch network is processed by KNMI every 15 min; the observations are available approximately 5 min after observation time. Figure 1 depicts the location of the receivers and the data provider. Raw GNSS data are gathered from six sites from the Ordnance Survey (UK), six sites using the NTRIP server at BKG (Germany) and the remainder from Kadaster (Netherlands). Within the EUMETNET GNSS water vapour programme (E-GVAP; Vedel, et al., 2010), a large number of other GNSS ZTD observations are available, but none within 10 min of observation time in the vicinity of the Netherlands. In general, the data delay for these sites is about 35 to 40 min.

Figure 1.

Location of the GNSS sites used for near-real-time estimation of the Zenith Total Delay.

2.2. Radial wind measurements

A Doppler radar generally determines the radial velocity of scattering particles by autocorrelation of the received signal for subsequent transmitted pulses. During pulse-pair processing, the velocity is effectively deduced from the phase jump of the received signal. The operational application of these radial velocity observations has been hampered by the small unambiguous velocity interval of the instrument, especially for C-band radars. The unambiguous velocity is (Doviak and Zrnić 1993)

equation image(2)

where λ is the wavelength of the radar and PRF is the pulse repetition frequency. The two KNMI radars are C-band with a wavelength of 5.3 cm. Different PRFs will result in different unambiguous velocities and, by combining velocities from two PRFs, the unambiguous velocity interval can be extended. Dazhang, et al. (1984) showed that the dual PRF unambiguous velocity is given by

equation image(3)

where equation image and equation image are the low and high unambiguous velocities. Staggered triple PRF strategies have been developed by Tabary, et al. (2005); here a dual PRF scheme is used. The high PRF is chosen to be 4/3 of the low PRF, resulting in an unambiguous velocity of four times the low PRF unambiguous velocity. The difference between the phase jumps observed at low and high PRF is employed to deduce the radial velocity and can be expressed in terms of high and low velocities. The PRF also determines the unambiguous range Ru of the radar through

equation image(4)

where c is the speed of light; a trade-off has to be made. Table 1 shows the operational settings used at KNMI of the PRFs, elevation angles, maximum range and unambiguous dual PRF velocity.

Table 1. Operational radar settings at KNMI.
(deg.)PRF (s−1)(m s−1)(km)

At the low elevations (0.4 to 2.0circ) the unambiguous velocity is 24 m s−1 and thus in high wind situations the radial velocity can be folded. Dealiasing Doppler radar radial winds has been studied (Haase and Landelius 2004; Montmerle and Faccani 2009; Xu, et al., 2011), but instead of using a so-called torus mapping of radar volume winds, image filtering or auxiliary wind observations, a look-up table is used built from the radar observations themselves. By exploiting the highest elevations with an unambiguous velocity of 48 m s−1, velocities at lower elevations can be dealiased. This is achieved by estimating a mean radial velocity at height h and azimuth a by

equation image(5)

with |hH(ρ,α)| < 0.5 km, where H(ρ,α) is the height of the wind observation approximated by a 4/3 Earth radius, Vα(a,ρ) at elevation α and range ρ. In total 72 azimuth bins are chosen while the vertical is divided into 1 km levels.

Note that the observations from high elevations (more than 10°) at a height of 8 km are within a horizontal distance of 50 km from the radar. The radial velocity of the lowest four used elevations at a given height and azimuth is changed by an integer value of the relevant unambiguous velocity such that the value is mapped into an interval with a width of 2equation image around the observed mean radial velocity at the specific azimuth and height. This unfolding method is valid under the assumption that the (radial) wind does not change by more than 24 m s−1 over a height difference of 1 km and a horizontal distance of 150 km. A wind shear of this magnitude is very rare in the Netherlands.

After the dealiasing, a pre-processing step is performed to filter and thin the radar radial wind observations to avoid overfitting in the assimilation. First the observations too close to the radar are omitted, and next the observations are thinned to a single observation per 20 km×20 km box. When the number of observations in a box exceeds 60 and the standard deviation within the box is lower than 2 m s−1, the median observation will be used for assimilation. Figure 2 depicts an example of radial velocity observations and the dealiased and thinned observations for use in NWP. The wind was from the southeast, with (radial) winds exceeding 28 m s−1. The folding of the radial wind is clearly visible south of the radar in Figure 2(a); the radial wind shows a sudden jump from −24 to +24 m s−1. The unambiguous velocity for this elevation is 24 m s−1; the dealiasing algorithm maps these radial wind observations to positive values. The unfolding in the other direction is observed just above the centre of the radar image.

Figure 2.

Example of raw and thinned data valid for 2355 UTC on 29 August 2010. (a) raw radial velocity observations by the radar at De Bilt at an elevation of 0.4°; the unambiguous velocity is 24 m s−1. (b) de-aliased and thinned observations used for assimilation in NWP.

The observations from the two Doppler radars operated by KNMI are used in this article. The combination of both radars result in a very good coverage in the region around the Netherlands.

2.3. Other observations

Aircraft observations are very important for initializing the upper atmosphere in NWP models. Here two types of observation are used: AMDAR and Mode-S. AMDAR and Mode-S wind observations are deduced from the relative position of the aircraft to the ground track. Temperature sensors on an aircraft are read by the on-board computer and transmitted to the ground using the AMDAR communication system. AMDAR wind and temperature observations are generated on board. Mode-S wind observations are derived in a similar way, using air traffic control information on ground track position, air speed, Mach number and flight level. de Haan (2011) gives more details. The benefit of the Mode-S observations is that all aircraft in an area of 300 nautical miles (≈550 km) are queried every 4 s and can be used to infer wind and temperature information. The accuracy of the Mode-S wind is comparable to AMDAR, while Mode-S temperatures are of lower quality (de Haan 2011).

Every hour, pressure observations from the available SYNOP stations are stored. The locations are mainly on land, with some (drifting) buoys and maritime platforms. Radiosondes measure a profile of temperature, wind and humidity. A radiosonde is launched approximately 40–50 min before the synoptic hour (0000, 0600, 1200 and 1800 UTC) to ensure that it reaches the tropopause (i.e. 500 hPa) around the synoptic main hour. Because of the long duration of the flight (approximately 2 h), this observation method implies that the actual observation time at a certain height does not correspond to the profile time stamp and can differ by about an hour. Moreover, the radiosonde observations arrive too late for assimilation into a numerical nowcast model with short cut-off times. The radiosonde launch sites have a horizontal separation of about 200 km over land. Over the ocean, upper-air observations come from autosondes launched from ships and in the North Sea from radiosondes launched from the Ekofisk oil rig. The horizontal and temporal resolution is too coarse for mesoscale modelling.

2.4. Numerical weather prediction

At KNMI, HIRLAM version 7.2 (Undén, et al., 2002) is run operationally for the short-term weather forecast. Two operational HIRLAM runs are carried out differing in cycle, region and forecast length. The D11 run is started every 6 h and calculates a 48 h forecast on a domain covering part of the North Atlantic Ocean; the H11 run is started every 3 h with a forecast length of 24 h and a smaller domain. A third NWP run (U11), which is the focus of this article, is started every hour on a domain inside H11 with a forecast length of 12 h. Figure 3 shows the D11, H11 and U11 NWP domains.

Figure 3.

NWP model areas for D11 (6 h cycle), H11 (3 h cycle), and U11 (1 h cycle).

Regional NWP models require lateral boundaries. The D11 cycle (largest domain) uses forecasts from the European Centre for Medium-range Weather Forecasts (ECMWF). The H11 run is nested within D11, and the U11 run is nested within H11. All three models have a horizontal resolution of 11 km and 60 vertical levels, ranging from the surface to the top of the model atmosphere at 0.1 hPa. The experiments were performed only on H11 and U11, with operational constraints such as observation cut-off time and the use of forecast fields as boundaries (Table 2).

Table 2. HIRLAM settings and experiment definitions.
Cycle (h)33111
Observation cut-off time (min)7070101010
Observations used in assimilation     
Surface pressureYesYesYesYesYes
Radar radial velocitiesNoNoNoNoYes

For both types of model (H11 and U11), different experiments were conducted over a summer period from 1 May 2010 to 5 September 2010 and a winter period from 13 January to 28 February 2011. The differences come from assimilation of different sets of observations. Assimilation is performed in order to find the best possible initial state of the atmosphere given the observations, an initial guess (also called background or first guess) and a priori defined constraints. Observations are used to constrain the initial state (the analysis) of the atmosphere with a state estimate of the previous model run. From this initial state analysis, a forecast can be computed through integration in time.

The assimilation technique used here is 3D-Var, very common in NWP. The analysis is determined by minimizing a cost function including the background-error covariances, B, and observation-error covariances, R, together with observations, y, and the background state, xb. The cost function is defined as

equation image(6)

where H, the observation operator, maps the model fields onto the observations. Generally, H is nonlinear and a linearised approximation is used to find the minimum for J(x). The background-error covariance matrix B is balanced as described in Berre (2000).

The reference run is called H11. The run H11+GNSS contains assimilated information of GNSS ZTD observations (and Mode-S observations). The hourly reference run is abbreviated as U11; U11+GNSS has GNSS ZTD observations assimilated and U11+GNSS+RAD has additionally radar radial winds assimilated. A summary is given in Table 2. Note that no separate impact experiment of solely radar radial velocities is performed, and there is no impact experiment for H11 without Mode-S and with GNSS ZTD. A test period of two weeks in spring 2010 did not reveal any different impact from Mode-S on rainfall and so this not discussed here.

The radar observation operator was developed in the HIRLAM program (Lindskog, et al., 2000, 2004; Salonen, et al. 2009) and is used for several studies (Xiao, et al., 2008; Järvinen, et al., 2009; Salonen, et al. 2009). The observation error used in this study was set to 6 m s−1. This value is quite large but, because 20 × 20 km boxes were used, the weight of the observations is reduced. The radar radial wind observations can still be correlated, but observation-error correlations are neglected in the assimilation. A larger size of box might be appropriate and will be investigated in further research. The observation operator projects the model horizontal wind vector along the radar line of sight using the azimuth angle. The rainfall velocity is neglected and pulse path bending is taken into account.

The GNSS ZTD observation operator has been developed within the COST 716 action (Elgered, et al., 2004) and is given by Eq. (1). A bias correction is applied to the observations based on the observed bias by H11 over a period of 30 days prior to the assimilation time. Experiments with a bias correction based on U11 were not successful due to the lack of other independent humidity information to constrain the model and to simultaneously correct the observation with respect to observed ZTD. The ZTD observation error was set to three times the formal error as estimated by the GNSS Bernese processing software (Hugentobler, et al., 2006), which is approximately 0.01 m. Undén, et al. (2002) and de Haan and Stoffelen (2012) provide a complete table with observation errors.

2.5. Radar rainfall estimates

Rainfall forecasts are verified against corrected hourly radar rainfall estimates. A correction to the radar rainfall observation is applied because beam blockage is a serious problem because of the presence of tall buildings in the vicinity of the radars in the Netherlands. This results in an underestimation of rainfall intensities for certain azimuth sectors. The azimuth sectors are divided into 200 range bins of 1 km, which contain values of 1 h rainfall accumulation. For a blocked azimuth sector, each range bin value is replaced by the linearly interpolated value of the range bins at the same range from the nearest left and right non-blocked azimuth sectors. Also, correction of biases due to overshoot of the beam at large distances are included. This correction method developed by Overeem (2009) uses daily rainfall rates from a very-high-density network together with hourly rainfall estimates from the synoptic observation network in the Netherlands. The result is a corrected rainfall estimate map with the resolution quality of the radar, but the absolute quality of the synoptic rainfall observations. The area around the Netherlands selected for verification is shown in Figure 4.

Figure 4.

Definition of the area used in the validation of rainfall forecasts.

3. Results

3.1. Impact on humidity forecasts

The humidity forecasts are compared to radiosonde humidity observations for the summer (4 months) and winter (6 weeks) periods. The result is shown in Figure 5. The validation times are 0000 and 1200 UTC, which implies that, for example, assessment of 6 h humidity forecasts includes NWP runs started at 0600 UTC and 1800 UTC. Moreover, the same set of observations are used for all comparison experiments, and no gross error checks are deployed to remove data for a specific experiment. In this way a fair comparison can be made. For the summer period the observation minus forecast is shown for the experiments H11, H11+GNSS, U11, U11+GNSS and U11+GNSS+RAD, while for the winter period the result of two experiments, U11 and U11+GNSS+RAD, are shown. At the lowest level for the summer period, the analysis and forecast of the hourly experiment without GNSS data (U11) are too moist (negative bias). The 3 h experiment without GNSS data (H11) shows the same behaviour for the 3 h and 6 h forecasts. At analysis time (0000 and 1200 UTC) the bias is zero. The 3 h and 6 h forecasts have no direct assimilation of upper-level humidity observations from radiosonde at De Bilt, because this radiosonde is launched only twice a day. When GNSS data are assimilated, the bias in the forecast is reduced for the 3 h run (H11+GNSS). For the hourly runs with GNSS data (U11+GNSS and U11+GNSS+RAD), the bias becomes slightly positive, with for the latter experiment a slightly lower bias. For higher levels, the 3 h run with GNSS shows a positive impact of GNSS data, and for hourly runs the bias is slightly degraded. The negative humidity bias is reduced to almost zero for the winter period in the lowest level when GNSS data is used in the hourly run (left panel of Figure 5(b)).

Figure 5.

Comparison of specific humidity forecasts (g kg−1) with radiosonde humidity observations for (a) the summer period and (b) winter period for four different height levels. Left columns show the biases (observations minus model forecast), and right columns the standard deviations.

The standard deviation for the summer period shows a neutral impact at the lowest three levels. At the top level, the lack of humidity information is revealed for the runs without radiosonde information. For the winter period, the reduction of standard deviation for the lowest two levels is obtained while assimilating GNSS and radial velocities. Note that the standard deviation of U11 is smaller for the winter period than for the summer period because the amount of total water vapour is generally smaller in winter than in summer. Only two hourly runs are shown for the winter period but no 3 h hourly runs were performed.

3.2. Impact on rainfall forecasts

Because the humidity biases are changed when GNSS data is assimilated, an effect on rainfall forecasts can be expected as well. The reference U11 experiment showed a too moist atmosphere at analysis time and the reference H11 at 3 h and 6 h forecasts. When GNSS data are assimilated, the total moisture in the lowest levels is reduced and lies closer to the observations. In the assimilation step humidity is removed, but during the forecast a large part of the water vapour amount can evolve during the transport processes and during different microphysical phase transitions. In the latter case, raindrops can be formed and, when large enough, will rain out, thus reducing the total amount of water vapour.

Rainfall forecasts are highly influenced by (e.g.) tuning of the model and the convection scheme applied. The HIRLAM model set up in this study used the Kain–Fritsch convection scheme.

Rainfall forecast verification is hampered by having good forecasts in the wrong position. Here three methods to estimate the impact of assimilation of GNSS and radar radial velocities is shown to assess the rainfall rate forecast quality by comparison with the observed rainfall. An impact assessment can be made by inspecting and comparing area-mean frequency distributions of rainfall forecasts and observed rainfall for the summer period. A second insight on the impact can be obtained by calculating the probability of (false) detection.

3.2.1. Climatological rainfall signal

For the summer period the hourly rainfall distribution is determined for different forecast lengths. These distributions are compared to corrected rainfall estimates (described in section 2.5) from radar in the area defined in Figure 4. By focussing on the rainfall rate over an area, the error due to a good rainfall forecast at the wrong position is reduced. Figure 6 shows the frequency distributions for the rainfall rate in the Netherlands for the corrected radar rainfall (dashed lines) and forecasts for the different experiments and different forecast times.

Figure 6.

Frequency distributions of total corrected hourly radar areal rainfall estimates (dashed lines) and the 1 h, 2 h and 6 h NWP rainfall forecasts (solid lines with symbols) in the Netherlands (Figure 4) for the summer period. Each panel shows the distribution for one of the five different experiments.

The top two panels represent the rainfall frequency distributions for the two 3 h experiments. Clearly, the experiment without GNSS data underestimates areal rainfall rates higher than 0.75 mm h−1 when compared to corrected radar rainfall estimates. When GNSS data are assimilated, the frequency distribution improves and compares better to the radar rainfall estimates. The bottom row shows the frequency for the hourly experiments. The U11 experiment shows a change in distribution in the first two hours of the forecast for areal rainfall rates larger than 0.9 mm h−1. Furthermore, this experiment underestimates the rainfall estimates of 0.5–1.1 mm h−1 as well as larger rainfall rates. When GNSS data are assimilated, no underestimation is present (lower centre panel). When additionally radar radial velocities are assimilated, the frequency distributions compare even better to the corrected radar rainfall estimates (lower right panel).

Comparing the H11+GNSS and the U11+GNSS or U11+GNSS+RAD, one can notice that the hourly runs are also better in forecasting extreme rain rate events (higher than 1.2 mm h−1). Thus the rainfall forecast improves even more when an hourly scheme is used.

3.2.2. Probability of detection

The probability of detection (POD) scores and probability of false detection (POFD) for areal rainfall >0.75 mm h−1 are plotted against forecast time in Figures 7 and 8, where

equation image

The statistics are created from the time series of rainfall forecasts and corrected radar rainfall estimates during the summer period over the area depicted in Figure 4.

Figure 7.

Probability of detection of areal rainfall rates of more than 0.75 mm h−1.

Figure 8.

Probability of false detection of areal rainfall rates of more than 0.75 mm h−1.

The POD for the runs without GNSS data have the lowest scores, while all experiments with GNSS data assimilated have better scores. The U11 experiment with GNSS and radar data performs slightly worse than U11 with only GNSS data assimilated, however this difference is small and probably not significant. The POFD shown in Figure 8 shows an increase when GNSS data are assimilated. This is not surprising because a run which is biased towards too little rain will result in low values of POFD. The high POD and low POFD for the experiments with GNSS data implies that the rainfall forecast of the mean precipitation over the Netherlands have improved.

3.3. Wind forecasts

An impact on wind forecasts is expected when radar radial winds are assimilated. The wind speed forecast is verified at radiosonde site De Bilt and with radial velocities observations around the Netherlands.

Figure 9(a, b) shows the statistics of the wind speed and wind direction against radiosondes for the summer period at validation hours 0000 and 1200 UTC. The 3 h experiments show for all levels a similar signature in wind speed bias: zero at analysis time gradually becoming more negative through the forecast. The zero bias at analysis time is related to assimilation of radiosondes. The hourly experiments have a large bias at the highest level. Smaller biases compared to H11 and H11+GNSS are observed for all U11 experiments at 400 hPa, with U11+GNSS+RAD having the smallest bias. For the levels 400 hPa and below, the biases of the hourly experiments become more negative with forecast time albeit less strong as for H11 and H11+GNSS. This decrease with forecast time is not present at 225 hPa and 75 hPa, while bias of U11+GNSS+RAD at the lowest level is similar to the H11 runs. The bias of U11+GNSS+RAD at 600 hPa is slightly worse than U11+GNSS in the first hour of the forecast. No correction is applied for the influence of the fall speed which contaminates the radial velocity for high elevations. This issue needs further investigation. The standard deviation of the speed observation minus forecast is shown Figure 9(a) (right panel). Apart from the highest level, the standard deviation of the hourly experiments is smaller than the H11 and H11+GNSS experiments. Especially, the lowest level shows an improvement for the hourly experiments, with U11+GNSS+RAD having a slightly smaller standard deviation than the other hourly experiments upto 3 h into the forecast. The higher standard deviation at analysis time when GNSS data is assimilated in the hourly experiments is surprising, while the hourly reference run shows a smaller standard deviation at analysis time. This effect is visible over the whole profile and is not well understood and also needs further investigation.

Figure 9.

Comparison of upper-air wind forecasts with radiosonde observations at De Bilt for five different height levels in the summer period: (a) wind speed (m s−1) and (b) wind direction (deg). Left columns show the biases (observations minus model forecast), and right columns the standard deviations.

Wind direction statistics of the forecast versus radiosonde are shown in Figure 9(b). Wind direction biases from the hourly experiments are generally closer to zero than the wind direction biases of the 3 h experiments. There is no significant difference in the wind direction bias when GNSS data is assimilated in an H11 experiment. A remarkably high wind direction bias is observed for 3 h forecasts for H11 and H11+GNSS at a height of 400 hPa. Note that the hourly experiments do not show this, while the observations sets are equal. The reason cannot be related to assimilation of Mode-S data since both H11 runs show this high wind direction bias. The origin of this high bias is also not related to observation bias, because the observation sets are equal and the U11 do not show this behaviour. Moreover, a gross error check did not reduced the bias of this outlier. The wind direction standard deviation for the hourly experiments is generally smaller than H11 and H11+GNSS for 3 h and 6 h forecast times. U11+GNSS+RAD has the smallest standard deviation at the lowest level for the hourly runs, and H11+GNSS has the smallest standard deviation. At higher levels, at analysis time the hourly runs with GNSS data show a higher value than the other experiments. In the next forecast hour this difference is not noticeable. Note that H11+GNSS has a slightly better standard deviation in the forecast than H11, due to the additional assimilation of Mode-S observations. In conclusion, for the radiosonde wind comparison, the 3 h and 6 h wind forecasts from the hourly runs are generally better (smaller wind speed standard deviation and smaller wind direction bias) than the H11 runs with the emphasis on wind forecast improvements for levels below 500 hPa. Clearly visible is the lack of wind and/or temperature information at the top level (75 hPa) at analysis time.

Figure 10 shows the statistics for experiments with respect to the forecast and observed radial velocities for the summer period for different levels. Note that the vertical is divided into intervals of 750 m. The biases are close for the lowest three levels. This implies that the observations are unbiased for the lowest levels and that by assimilation no bias is introduced. For the highest level, U11+GNSS+RAD is different from the other experiments. The impact of assimilation is clearly visible for the standard deviation. This is reduced over the complete forecast length of 6 h; the positive effect gradually fades becoming neutral at the 6 h forecast. The reason for this could be that the improvements in the wind forecast are transported away from the radar observations by the governing atmospheric motion, or that the introduced small scales are dissipated. Utilizing radar radial observations from upstream countries (UK, Belgium, France) will most likely improve the forecast at longer time-scales.

4. Conclusions and outlook

This article describes the impact of assimilation of GNSS ZTD and radar radial winds in a very-short-range regional NWP model. The experiments are performed over a four-month summer period in 2010 and a six-week winter period in 2011. Due to the short cut-off time of the fast update cycle of 1 h, radiosonde humidity observations arrive after the analysis starting time and thus cannot be used. Here it is shown that GNSS ZTD can almost fill the lack of upper-air humidity information when no radiosonde observations are available. Additionally, the impact of assimilation of radar radial winds is shown.

Figure 10.

Observation minus forecast statistics (m s−1) for radar radial wind observations at four height ranges for the summer period. All experiments are evaluated at 0000, 0300, ..., 2100 UTC.

Two different schemes are tested with the assimilation of GNSS data: a run with an update cycle of 1 h and a run with an update cycle of 3 h. In the hourly run, no radiosonde information is used, but in the latter run radiosondes are used in the assimilation.

The forecasts of specific humidity are verified against radiosonde humidity observations. Rainfall forecasts are verified against radar rainfall estimates which are corrected with hourly SYNOP observations. Wind forecasts are verified against radiosonde winds and radar radial winds.

The impact study showed that, in the summer and winter periods without humidity information, the hourly run is too moist compared with radiosonde humidity observations. The bias in humidity is changed from negative to positive when GNSS data are assimilated. When additional radar radial winds are assimilated, the forecast humidity bias around 875 hPa is the smallest of the hourly runs. Furthermore, for the winter case, assimilation of GNSS data results in a lower humidity standard deviation. For a 3 h cycle assimilating radiosonde data, an improvement of the bias in humidity around 875 hPa is observed when GNSS data are used.

Assimilation of GNSS data also has a positive impact on rainfall forecasts. Rainfall frequency distributions are improved. Without GNSS data, the high values of rainfall forecasts are underestimated, especially for the hourly run. Assimilation of GNSS data resulted in rainfall distributions comparable to corrected radar rainfall estimates. Additional assimilation of radar radial winds shows another small improvement in the rainfall distributions for large rainfall rates.

When radar radial winds are assimilated, a bias reduction in the forecast wind direction is observed in comparison with radiosondes. An improvement in wind speed and direction standard deviation is observed below 600 hPa for the hourly run with radar radial winds when compared to the hourly runs without this information. When compared to radar radial winds themselves, a large impact on the standard deviation is observed. This impact gradually becomes neutral after 6 h, probably due to dissipation of small scales by the forecast model.

No improvement in wind speed and direction was found at the 75 hPa level. Both GNSS ZTD and radar radial wind clearly do not affect wind at the this level. To improve the wind analysis and forecast, higher-level wind observations are necessary.

In this article the positive impact of both GNSS atmospheric observations as well as radar radial velocities was shown for observations close to the Netherlands. It is expected that utilizing information from a larger GNSS network and more radars from neighbouring countries will be beneficial for the quality of the numerical nowcasts. Moreover, a modification of the model cycling can be proposed to include observations with a longer delay such as radiosondes. Through this re-analysis, high-level winds for the hourly run will most likely improve. The use of satellite observations, such as scatterometer and ATOVS will also benefit from a longer cut-off time of re-analysis.


The author would like to thank Kadaster (the Netherlands), the Ordnance Survey (UK) and the Federal Agency for Cartography and Geodesy (BKG, Germany) for raw GNSS data, and the Air Traffic Control of the Netherlands (LVNL) for providing Mode-S observations. Part of this study has been funded by the Knowledge and Development Centre, Mainport Schiphol in The Netherlands (KDC;