Evaluation of Radar and Automatic Weather Station Data Assimilation for a Heavy Rainfall Event in Southern China

The Hourly Assimilation and Prediction System (HAPS) has been developed collaboratively by the Center for Analysis and Prediction of Storms (CAPS) at the University of Oklahoma, Shenzhen Meteorological Bureau (SZMB), and the Shenzhen Institute of Advanced Technology (SIAT), Chinese Academy of Sciences, since March 2010. The motivation is to improve the accuracy of short-term (0-12 h) forecasts of severe weather in the southern China region.

In (Hou et al., 2013), the HAPS was described and examined using two heavy rainfall events, with a focus on the overall impact of additional surface automatic weather station (AWS) and radiosonde data on rainfall forecasts. Results suggested that radiosonde data were useful in improving rainfall position accuracy and reducing rainfall overprediction. We extend that study by focusing on Doppler radar and surface data impacts on echo structure and 3-h accumulated precipitation. Particular emphasis is given to the impact of assimilating only radar or surface AWS data by using the same 15 July 2011 rainfall case in (Hou et al., 2013).

The purpose of the current study is to improve quantitative precipitation forecasting (QPF) skill of the forecasting system by including additional available observational data sources. Viable data assimilation from various data sources, including Doppler weather radar, surface observations, conventional radiosondes, Global Positioning System (GPS) water vapor, wind profile measurement, and satellite observations, is critical in modern numerical weather forecasting systems (Ruggiero et al., 1996; Cucurull et al., 2004; Gu et al., 2005; Zhao and Jin, 2008; Liu et al., 2012a). Information from Doppler radar data is especially useful for showing the dynamic structure of convective-scale precipitation systems (Sun and Crook, 1997). As Doppler radar, surface, and radiosonde observations are the only sources of available data for the studied case, the impact of such data assimilation on QPF skill is explored. Since radiosonde data assimilation experiments have already been conducted in (Hou et al., 2013), this study focuses only on the impact of radar and surface AWS data.

The three-dimensional variational data assimilation (3DVAR) method, although less optimal, is computationally cheaper than the four-dimensional variational data assimilation (4DVAR) and ensemble Kalman filter (EnKF) methods in real-time numerical weather forecasting. (Cucurull et al., 2004) demonstrated that cycling assimilation of GPS data and local surface meteorological observations improved the forecast of a snow storm. (Xiao et al., 2007) suggested there are positive impacts of 3DVAR radar data assimilation on short-range QPF. Many studies (Xue et al., 2003; Hu et al., 2006a; Zhao and Xue, 2009; Schenkman et al., 2011a) have shown that assimilating radar data using the Advanced Regional Prediction System (ARPS) 3DVAR package and cloud analysis procedure is successful in reproducing more accurate structures of tornadic thunderstorms.

In order to quantitatively evaluate the impacts of observational data assimilation on forecast skill, a heavy rainfall event [the same 15 July 2011 case as in (Hou et al., 2013)] that occurred in the southern China region is simulated. In section 2, a brief description of the forecasting system and observations is provided. In section 3, the synoptic setting and corresponding precipitation of the rainfall event is presented. Sections 4 and 5 illustrate the numerical experiments and examine the impacts of data assimilation on precipitation forecast quality. Finally, a discussion and conclusions are given in section 6.

The HAPS forecasting system, using the Advanced Research Weather Research and Forecasting (WRF-ARW) (V3.3.1) model as the forecast model and the ARPS 3DVAR package for data assimilation, consists of an outer mesoscale domain with 12-km horizontal grid spacing and a one-way nested convective-scale domain at 4-km grid spacing (Fig. 1). The 0.25^{°× 0.25°} European Centre for Medium-Range Weather Forecasts (ECMWF) analyses are used for the initial and boundary conditions. The ECMWF data have 20 vertical pressure levels, with the model top at 10 hPa, and are available every 12 h at 0000 and 1200 UTC.

The outer domain, which covers the southern China region, produces 48-h forecasts at 0000 and 1200 UTC every day. Since initializing from relatively coarse ECMWF data, a spinup period of 3-6 hours is needed to develop the smaller scale convective features (Kain et al., 2010). Thus, the first 12 h of forecasts from the mesoscale domain are not used and the remaining forecasts are interpolated hourly to the convective-scale domain to serve as a background and lateral boundary condition for inner domain assimilation. Convective-scale 12-h forecasts initialized from the inner domain analyses are run every hour (initialized at 0000, 0100, 0200 UTC, etc.).

Figure 1.  Model domain coverage and locations of radar (big dots) and AWS (small dots) sites. The black circles represent the 230-km-range rings of the assimilated Doppler radars.

For real-time forecast runs, no data assimilation is conducted in the mesoscale forecast, while radar reflectivity and radial velocity data from local Weather Surveillance Radar-1998 Doppler (WSR-98D) radars have been assimilated at the initialization time (without cycling) of every convective-scale forecast. The convective-scale forecast is set to start 3 min past every hour. It will check the availability of radar data, repeating every 15 sec for a total of 3 min, so the 12-h forecast in the convective-scale domain will proceed no later than 6 min past every hour with available radar data.

A complete real-time forecast procedure includes interpolation of the 6-h interval ECMWF data to ARPS format, data assimilation using the ARPS 3DVAR package, interpolation of the updated analyses to the WRF format grid using the ARPS WRF interface utility code wrf2arps, and finally running the WRF model. The programs for interpolating between the ARPS and WRF model levels have been proven efficient and reliable, as they have been used since 2007 in the CAPS Storm-Scale Ensemble Forecasts (SSEF) during the National Oceanic and Atmospheric Administration's (NOAA) Hazardous Weather Testbed Spring Experiments (Clark et al., 2012).

Both domains have 51 vertical levels. The Rapid Radiative Transfer Model (RRTM) for longwave radiation (Mlawer et al., 1997), Goddard shortwave radiation (Dudhia, 1989), Mellor-Yamada-Janji\'c (MYJ) planetary boundary layer (Janji\'c, 1990) and Noah land surface model (Chen and Dudhia, 2001) are used in both domains in the real-time forecasting system. The Eta microphysics and Kain-Fritsch cumulus parameterization (Kain, 2004) are used in the 12-km domain, while the 4-km domain uses the more complex Thompson scheme (Thompson et al., 2008) and runs without cumulus parameterization.

The ARPS 3DVAR package minimizes a cost function that includes the background, observation, and mass conservation constraint terms. In the ARPS 3DVAR, the background error covariance matrix is modeled by a one-dimensional recursive filter, which allows for accurate model isotropic Gaussian error correlations. The square root of the matrix is used for preconditioning to obtain the optimal solution of the analysis. A forward operation operator is used to project the analysis to the observation space. As described in (Hu et al., 2006a), the conventional data, including radial velocity data, are analyzed through the 3DVAR scheme, while radar reflectivity data are used in a cloud analysis procedure that retrieves the amount of hydrometeors and adjusts in-cloud temperature and moisture fields. The forward operator errors for conventional data are small and usually neglected (Hu et al., 2006b). Although uncertainties exist due to the use of empirical relationships between the hydrometeor variables and observed reflectivity, the 3DVAR method has been successful in simulating convective storms (Hu and Xue, 2007; Schenkman et al., 2011a; Xue et al., 2014). More details on the cost function and the minimization algorithm can be found in (Gao et al., 2004).

As described in (Hou et al., 2013), the ARPS 3DVAR includes a multiscale analysis procedure to allow for the assimilation of different data types with different spatial scales (Hu et al., 2006b, Schenkman et al., 2011a). The observation data can include single-level surface data, aircraft data, multiple-level soundings and wind profiles, as well as Doppler radar reflectivity and radial velocity. The analysis variables include the three wind components, potential temperature, pressure, and water vapor mixing ratio. Hydrometeors are not analyzed variationally, but through a cloud analysis module in ARPS.

As the distance between AWS sites varies, sensitivity experiments with three different horizontal scales of 10, 20 and 50 km are conducted. However, no experiment performs consistently better than any other for all the precipitation thresholds. Guided by the density of AWS observations, a horizontal decorrelation scale of 50 km is used in the study [as in (Hou et al., 2013)]. Different horizontal scales of 6, 9 and 20 km for radar data are also tested, and the general precipitation patterns are found to be similar. However, the experiment with the 6-km horizontal scale produces the highest degree of overprediction during the first nine forecast hours, while the experiment with the 20-km horizontal scale exhibits such overprediction during the last three forecast hours in terms of maxima for 3-h accumulated precipitation. Therefore, radar data are analyzed with a horizontal decorrelation scale of 9 km [20 km in (Hou et al., 2013)]. The vertical decorrelation scale is four grid points for both data types.

As discussed in (Hou et al., 2013), observation errors are important for the data assimilation technique. Due to a lack of reliable statistics on error information for Chinese instruments, observation errors are assumed to be uncorrelated and are specified according to estimated errors. The specified observation error covariance values for AWS observations are 1.5 m s^-1 for wind components, 2.0 hPa for pressure, 1.5 K for temperature and 5.0 % for relative humidity. The observation error for radar radial velocity observations is set to 2.0 m s^-1.

The available observations for this study are from WSR-98D radars, surface AWS, and rain gauge measurements [as in (Hou et al., 2013)]. The WSR-98D radars are S-band radars with similar characteristics to those of the Weather Surveillance Radar-1988 Doppler (WSR-88D) radars in the US operational Doppler radar network (Sheng et al., 2006; Liu et al., 2012b). There are 11 radars in the convective-scale domain (big dots in Fig. 1), including eight in Guangdong Province.

The surface AWS sites are densely distributed throughout China, with a total number of more than 30 000 (Shen et al., 2014). The distribution of the sites is heterogeneous. In eastern and southern provinces, the distance between AWS sites can be as low as around 10 km. However, only around 2000 sites are national-level and others have been developed in recent years by different provinces. Considering the existence of missing and suspicious records, 438 AWSs in the convective-scale domain (small dots in Fig. 1) are selected from those national-level sites. Such near-surface variables as 2-m temperature, 2-m dewpoint, and 10-m wind components are available at the top of every hour.

The total number of rain gauges in the convective-scale domain is 1353, including the same AWS sites used in data assimilation and an additional 915 sites in Guangdong. These rain gauge data are used for precipitation verification in later sections.

Figure 2.  (a) 850 hPa wind (arrows, m s^-1) and relative humidity (shading, %), and (b) 500 hPa wind and geopotential height (black contours, 10 m) at 0000 UTC 15 July 2011.

The same rainfall case as that studied in (Hou et al., 2013) is used here to investigate the impact of radar and surface AWS data. It is a heavy rainfall event that took place on 15-18 July 2011, associated with the southwest monsoon and troughs of low pressure. This event is selected because, firstly, the southwest monsoon is the typical synoptic setting producing rainfall in June and July in southern China; and secondly, heavy precipitation causing serious damage in multiple cities was observed. The synoptic features of the case from the ECMWF analyses are shown in Figure 2.

At 850 hPa (Fig. 2a), warm and moist air is transported from coastal areas near southern China to inland regions through the strong southwest winds, forming a southwest-northeast water vapor band. At 500 hPa (Fig. 2b), a favorable dynamic environment for precipitation is provided by a low to the west of the southern China provinces with the trough axis extending along the coast. During the following 12 h, southwesterly winds, combined with a deeper trough at higher levels, transport substantial amounts of water vapor to produce heavy precipitation over coastal areas of southern China.

The 15-18 July 2011 heavy rainfall is characterized by non-uniformity and extremely high rainfall rates in localized regions. According to rain gauge network observations in Guangdong, the average accumulated rainfall per site is 58.8 mm between 0000 UTC 15 July and 0000 UTC 18 July 2011, with 423 sites recording accumulated rainfall of between 100 and 250 mm, and at one site a maximum value of 374 mm. Figure 3 shows the observed 24-h accumulated precipitation during 1200 UTC 15 July and 1200 UTC 16 July. Heavy rain mainly occurs in Guangdong and its surrounding provinces, with a maximum value of 208.2 mm.

As our focus in this study is on short-term QPF skill of the forecasting system, only the 12-h period rainfall from 1200 UTC 15 July to 0000 UTC 16 July will be discussed in the following sections. Therefore, the forecast in the 12-km domain is initialized at 0000 UTC 15 July 2011 and runs for 48 h to 0000 UTC 17 July 2011. The 12-h forecast in the 4-km domain is from 1200 UTC 15 July 2011 to 0000 UTC 16 July 2011. The 12-h accumulated precipitation pattern during that period is similar to that shown in Fig. 3 and has a maximum of 157 mm (see Hou et al., 2013, Fig. 9).

A set of data assimilation experiments are conducted in both the 12-km and 4-km domains for the case study. For the experiments including data assimilation, the 3DVAR analysis is only conducted at the model initialization time without cycling. The basic run is a WRF-ARW forecast without the assimilation of any observational data (BASELINE, for short). Radar data assimilation is conducted at the initialization time in the 4-km domain (1200 UTC 15 July) in experiment RADONLY. To examine the impact of AWS observations, experiment AWSONLY with AWS data assimilation in both the 12-km and 4-km domains is also considered. An additional experiment, RADAWS2, assimilates both radar and AWS data. The difference between RADAWS2 and RADAWS in (Hou et al., 2013) is that AWS data assimilation in the 12-km domain is added in RADAWS2. Experiments with radiosonde data assimilation were performed in (Hou et al., 2013) and will not be discussed here. Other data sources such as satellite data, GPS water vapor, and wind profiler data have been demonstrated to be useful in improving model results in the literature (e.g., Powers and Gao, 2000; Liu et al., 2012a), but were unavailable for this study. Therefore, the focus in this study is on the impacts of radar and AWS data assimilation on the accuracy of short-term accumulated precipitation forecasts from the convective-scale domain.

Figure 3.  Observed 24-h accumulated precipitation (mm) valid at 1200 UTC 16 July 2011.

The quality of the 3DVAR analyses is first examined by comparing model initial fields from the background (ECMWF analyses) with those from AWS data assimilation. Figure 4 gives the temperature, water vapor mixing ratio and wind fields at the surface (first level above ground) from ECMWF analyses and AWSONLY in the mesoscale domain at 0000 UTC 15 July 2011. For clarity, observations from only 110 AWS sites are marked in Figs. 4b and d. A comparison between the surface temperature fields from the background and AWSONLY (Figs. 4a and b) suggests higher analyzed values over southeastern Guangxi and the coastal region of Guangdong. These changes are due to the observed higher 2-m temperature of around 27^°C-28^°C. Similarly, the water vapor mixing ratio field (Figs. 4c and d) over the same region is also improved, as values of 20 g kg^-1 or higher are observed. The surface wind field (Figs. 4e and f) over southern Guangxi is also modified, as the magnitude of the observed wind speed is much smaller. It can be seen that the model initial fields are modified towards the direct surface AWS observations after the 3DVAR assimilation.

To examine the impact of radar data on model simulated echo structure, Fig. 5 shows the observed composite reflectivity fields as well as the simulated fields from experiments BASELINE and RADONLY at the time of 3-h, 6-h and 12-h forecasts with the convective-scale domain initialization at 1200 UTC 15 July. The 9-h forecasts are not given due to the unavailability of radar observations at 2100 UTC 15 July. Composite reflectivity is defined as the maximum reflectivity in the vertical column. The observed reflectivity is gridded onto the convective-scale domain by the ARPS (with radar station locations shown in Fig. 1).

Radar observations at 1500 UTC (Fig. 5a) suggest that the convective clusters are over southern Guangdong and the coastal area. The 3-h forecasts valid at 1500 UTC from experiments BASELINE and RADONLY (Figs. 5b and c) suggest that RADONLY produces a radar pattern characterized by a larger convective region over Guangdong Province. In addition, RADONLY captures the intense reflectivity off the coast of Guangdong, while BASELINE does not. By 1800 UTC (Fig. 5d), the main convective region has moved northeast, but still remains over Guangdong and with a maximum value of 44 dBZ. In comparison, RADONLY (Fig. 5f) at the 6-h forecast is better than BASELINE (Fig. 5e) in producing the reflectivity region along the coastline over Guangdong. Meanwhile, it should be noted that the convection from radar data assimilation is too strong, with large swaths of greater than 35 dBZ. At the time as the 12-h forecast, the discrepancy between BASELINE and RADONLY (Figs. 5h and i) becomes smaller, both of which fail to produce the convective region off the coast.

To further demonstrate the effect of radar data assimilation, the predicted wind fields at 3 km above the surface from BASELINE and RADONLY are also plotted. From Fig. 6, we can see that the main patterns of the wind field are similar in both runs. However, the wind directions and magnitudes over the eastern half of Guangdong's coastal region show substantial differences between the two runs, with the wind circulation existing further to the southwest in RADONLY. As a result, the flow at 3 km is more perpendicular to the coastline in eastern Guangdong in RADONLY.

Figure 7 presents the 3-h accumulated precipitation amounts from 3-h to 12-h forecasts in the convective-scale domain (valid at 1500, 1800, 2100 UTC 15 July and 0000 UTC 16 July). Figure 7a shows a southwest-northeast oriented band of accumulated rainfall over southern coastal provinces, with the precipitation centers very close to Guangdong. The 3-h forecast from BASELINE (Fig. 7b) presents a very small precipitation center in Guangdong, with a much higher maximum value of up to 233 mm. Compared with that from observation, BASELINE fails in producing the rainfall region over southern Guangdong, especially at the border region between Guangxi and Guangdong. In comparison, RADONLY (Fig. 7c) produces a new cyclone-like rainfall region, with the rainfall location and orientation more closely matching the observations. It should also be noted that the maximum values from RADONLY are much higher than those from observations. Later, the maximum precipitation cores along Guangdong's coast have moved onto land from the southwest.

A comparison of composite reflectivity and rainfall distribution between RADONLY and BASELINE suggests that radar data are useful in improving the precipitation pattern forecast. Results from radar data assimilation generally capture the location of primary precipitation centers and more realistically reflect the overall precipitation system development. Nevertheless, radar data assimilation does not seem to change the overprediction of maximum rainfall values, and even increases the degree of overprediction in some cases.

Figure 4.  (a, b) Temperature (^°C), (c, d) water vapor mixing ratio (g kg^-1) and (e, f) wind fields (m s^-1) at the surface at 0000 UTC 15 July 2011 from background (left) and the 3DVAR analyses (right). Blue marked values in (b, d) and blue vectors in (f) are AWS observations.

Figure 5.  Composite reflectivity (dBZ) at 3-h, 6-h and 12-h forecasts from observations (left), BASELINE (middle) and RADONLY (right) with the model initialization from analyses at 1200 UTC 15 July.

Figure 6.  Wind field (m s^-1) at 3 km above the surface from (a) BASELINE and (b) RADONLY at 1500 UTC 15 July 2011.

Figure 7.  3-h accumulated precipitation amounts (mm) at 3-h, 6-h, 9-h and 12-h forecasts from observations (left), BASELINE (middle) and RADONLY (right) with the model initialization from analyses at 1200 UTC 15 July.

Surface observations, although they provide data at only one level, have the advantage of better temporal and spatial resolutions of both thermal and wind information. Root-mean-square errors (RMSEs) of 2-m temperature and relative humidity, as well as 10-m wind components, are used to evaluate the impact of AWS data assimilation in this section. Interpolation from model grid data to observation locations is first performed over the 438 sites before calculating the RMSE.

The RMSE is defined as \begin{equation} {\rm RMSE}=\left[\dfrac{1}{N}\sum_{i=1}^N(F_i-O_i)^2\right]^{\frac{1}{2}} , \end{equation} where F_i represents variables, including temperature, relative humidity, and wind components, from model results, and O_i is for the same variables from AWS data. N is the total number of observations, and is less than or equal to 438, depending on the availability of data.

The impact of AWS data assimilation is investigated in the convective-scale domain by comparing results between BASELINE and AWSONLY. Figure 8 shows the time series of RMSEs for the 2-m temperature, 2-m relative humidity and 10-m wind speed during the 12-h forecast with the model initialization from analyses at 1200 UTC 15 July in the convective-scale domain from experiments BASELINE and AWSONLY.

The RMSE for 2-m temperature (Fig. 8a) at 1200 UTC reduces slightly from 2.2^°C to 2.0^°C after AWS data assimilation, and the improvement from AWSONLY roughly exists during the first 6 forecast hours. When AWS data assimilation are only conducted in the convective-scale domain, a positive impact on 2-m temperature is apparent, but only lasts for 3 hours (Hou et al., 2013, Fig. 10c). The RMSE for relative humidity at 1200 UTC (Fig. 8b) reduces from 11.1% of AWSONLY to 7.1% of BASELINE. Positive effects of AWS data assimilation are also shown in the results for 10-m wind speed (Fig. 8c), with differences between the two experiments observed until the 9-h forecast. A comparison of RMSEs for the near-surface variables between BASELINE and AWSONLY suggests that the positive effects of AWS data assimilation generally last 6-9 hours.

Figure 8.  Time series of RMSEs for (a) 2-m temperature, (b) 2-m relative humidity and (c) 10-m wind speed for experiments BASELINE and AWSONLY, during the 12-h forecast valid from 1200 UTC 15 July to 0000 UTC 16 July in the convective-scale domain.

Figure 9.  FSS and bias scores for 3-h accumulated precipitation from the experiments BASELINE, AWSONLY, RADONLY and RADAWS2 with the model initialization from analyses at 1200 UTC 15 July: FSSs with the thresholds of (a) 1 mm, (b) 5 mm and (c) 10 mm, and biases with the thresholds of (d) 1 mm, (e) 5 mm and (f) 10 mm. The dashed lines at the value of 1.0 in the bias plots represent the perfect score.

(Ruggiero et al., 1996) demonstrated that the addition of surface data enhanced the low-level thermal gradient and winds. The study of (Sheng et al., 2006) demonstrated that surface and upper-air data assimilation on a 6-km domain forecast was beneficial, but only slightly. However, improved short-term forecasts could be achieved by continuing to assimilate surface data beyond the traditional synoptic cutoff times. Using the WRF 3DVAR cycling mode with incremental analysis updates, (Ha et al., 2011) showed that surface data played an important role in enhancing the thermal gradient and modulating the planetary boundary layer of the model. Assimilation of 5-min Mesonet observations has a positive impact on forecasting of low-level shear profile and gust front structure (Schenkman et al., 2011b). Data assimilation cycling experiment runs are also conducted in the present study, with mixed or even negative effects (figures not shown). AWS data quality control and improvement of the cycling process will be investigated in future work.

The neighborhood-based Fractions Skill Score (FSS; Roberts and Lean, 2008) is used for quantitative evaluation. It is defined as \begin{equation} {\rm FSS}=1-\dfrac{\frac{1}{M}\sum\nolimits_{j=1}^M[P_{F(j)}-P_{O(j)}]^2}{\frac{1}{M}[\sum\nolimits_{j=1}^MP_{F(j)}^2-\sum\nolimits_{j=1}^MP_{O(j)}^2]} , \end{equation} where M is the number of grid points within the neighborhood of grid point j, and P_F(j) and P_O(j) are the neighborhood probabilities at the ith grid box in the model forecast and observed fraction fields, respectively. A perfect forecast would have an FSS of 1, while a score of 0 means no skill.

The FSS scores for 3-h accumulated precipitation in the convective-scale domain are computed against rain gauge observations with the thresholds of 1, 5 and 10 mm. The radius of influence in the study is specified as 25 km.

The bias (B) is defined as \begin{equation} \label{eq1} B=\dfrac{f}{o} , \end{equation} where f is the number of events that are forecast and o is the number of events that occurred. For a specified threshold, a perfect forecast would have a bias of 1, while values of bias less than or greater than 1 represent rainfall underprediction and overprediction, respectively.

Figure 9 shows the FSS and bias scores for 3-h accumulated precipitation from the four experiments (BASELINE, AWSONLY, RADONLY and RADAWSD2) with the thresholds of 1, 5 and 10 mm and with the model initialization from analyses at 1200 UTC 15 July. For the 1-mm threshold (Fig. 9a), RADONLY and RADAWS2 have higher FSS scores than BASELINE and AWSONLY. For heavy rain prediction with the thresholds of 5 and 10 mm (Figs. 9b and c), RADONLY produces higher FSS scores than BASELINE until the time of the 9-h forecast, indicating improvement of rainfall position accuracy. The FSS scores for RADAWSD2 are higher than those for RADONLY at some lead times, but this positive effect of adding AWS data assimilation is not very consistent and sometimes negative.

As indicated by the bias scores (Figs. 9d-f), the performance of AWSONLY is better than BASELINE for all thresholds. RADONLY also outperforms BASELINE as bias values being closer to 1.0. The bias scores also suggest that precipitation areas are overpredicted in RADONLY and RADAWS2, while underpredicted in BASELINE and AWSONLY. Therefore, radar data assimilation also results in overforecasting with larger biases due to too much convection (Stratman et al., 2013). A comparison of bias scores between RADONLY and RADAWS2 shows that when radar data are assimilated, the additional AWS data assimilation even increases biases to larger values (Figs. 9e and f) of 1.54 with the 5-mm threshold and 1.93 with the 10-mm threshold at the 6-h forecast. As demonstrated by (Gallus and Segal., 2001), inclusion of mesonetwork surface observations resulted in improved equitable threat scores but also increased bias scores that already exceeded 1.0 for most precipitation thresholds. To improve initialization and prediction of convective-scale storms further, other available in situ observations should also be considered. (Stauffer et al., 1991) found that assimilation of surface wind and moisture data throughout the model PBL generally showed a positive effect on precipitation prediction, suggesting improvement by the reasonable combination of single-level surface and multi-level radiosonde data. Results presented in (Hou et al., 2013) showed a positive impact of radiosonde data assimilation in reducing rainfall overprediction.

The HAPS used at Shenzhen Meteorological Bureau, characterized by a combination of the WRF-ARW model and the ARPS 3DVAR and Cloud Analysis package, has been run as a real-time forecasting system for over three years since March 2010. To improve the short-term forecast skill of the system, continuous efforts have been made in terms of using fine-resolution initial data, assimilating local observations in model runs, and comparing the capability of different physics options in the model. The present study is focused on the quantitative evaluation of the impacts of Doppler radar and AWS data assimilation on QPF in the convective-scale domain by using a heavy rainfall event in southern China.

A set of data assimilation experiments including assimilating only radar or surface AWS data, as well as assimilating both radar and AWS data are conducted. Comparison of both echo structure and precipitation distribution suggests that radar data assimilation is useful in improving the forecast by capturing the primary precipitation centers and rainfall orientation. It is in fact the single most important analysis that contributes to QPF improvement in convection-allowing numerical weather prediction. Quantitative verification against rain gauge measurements in the convective-scale domain also suggests that radar data assimilation successfully improves the short-term QPF skill. Studies using the ARPS 3DVAR (Zhao and Xue, 2009; Schenkman et al., 2011a) show that radar data assimilation can produce more details of tornado and hurricane tracks, and thus better precipitation forecasts. This study further demonstrates the capability of the ARPS 3DVAR radar data assimilation in improving short-term precipitation forecasts at a convection-allowing resolution. It should also be noted that experiments with radar data assimilation produce too much convection and result in rainfall overprediction.

Assimilation of surface AWS data, including temperature, dewpoint, and wind components, from 438 stations is also carried out. The RMSE evaluation of forecasted near-surface variables suggest that the positive impact of AWS data generally lasts for 6-9 hours in the convective-scale domain. Quantitative evaluation of forecasted precipitation using FSS and bias scores suggests that assimilation of AWS data alone has a slight but generally positive effect in improving the QPF skill. When radar data are assimilated, the additional AWS data may increase the degree of rainfall overprediction at some lead times. As presented in (Hou et al., 2013), radiosonde data assimilation can reduce rainfall overprediction. Overall, radiosonde and AWS data are recommended to be taken into account in the HAPS real-time forecast.

In addition, only one case is selected in the study. To provide a more statistically meaningful verification, a much larger sample base, such as the summer-month cases in the past three years, is required for future work. To further improve the HAPS forecast skill, ensemble forecasts will also be considered. The uncertainty of the verification measures should also be quantified using such parameters as confidence intervals and hypothesis testing.