Analysis and numerical study of a hybrid BGM-3DVAR data assimilation scheme using satellite radiance data for heavy rain forecasts*

XIONG Chun-hui (熊春晖), ZHANG Li-feng (张立凤), GUAN Ji-ping (关吉平), PENG Jun (彭军), ZHANG Bin (张斌)

College of Meteorology and Oceanography, PLA University of Science and Technology, Nanjing 211101, E-mail:chunhui0603@sina.com

Analysis and numerical study of a hybrid BGM-3DVAR data assimilation scheme using satellite radiance data for heavy rain forecasts*

XIONG Chun-hui (熊春晖), ZHANG Li-feng (张立凤), GUAN Ji-ping (关吉平), PENG Jun (彭军), ZHANG Bin (张斌)

College of Meteorology and Oceanography, PLA University of Science and Technology, Nanjing 211101, E-mail:chunhui0603@sina.com

(Received November 20, 2012, Revised April 20, 2013)

A fine heavy rain forecast plays an important role in the accurate flood forecast, the urban rainstorm waterlogging and the secondary hydrological disaster preventions. To improve the heavy rain forecast skills, a hybrid Breeding Growing Mode (BGM)-three-dimensional variational (3DVAR) Data Assimilation (DA) scheme is designed on running the Advanced Research WRF (ARW WRF) model using the Advanced Microwave Sounder Unit A (AMSU-A) satellite radiance data. Results show that: the BGM ensemble prediction method can provide an effective background field and a flow dependent background error covariance for the BGM-3DVAR scheme. The BGM-3DVAR scheme adds some effective mesoscale information with similar scales as the heavy rain clusters to the initial field in the heavy rain area, which improves the heavy rain forecast significantly, while the 3DVAR scheme adds information with relatively larger scales than the heavy rain clusters to the initial field outside of the heavy rain area, which does not help the heavy rain forecast improvement. Sensitive experiments demonstrate that the flow dependent background error covariance and the ensemble mean background field are both the key factors for adding effective mesoscale information to the heavy rain area, and they are both essential for improving the heavy rain forecasts.

heavy rain forecast, hybrid data assimilation, satellite radiance data, ensemble prediction, flood forecast

Introduction

The heavy rain is a weather event that occurs frequently in China. The persistent heavy rain not only brings about meteorological disasters but also causes floods, urban rainstorm waterlogging and other secondary hydrological disasters. An accurate heavy rain forecast plays an important role in the accurate flood forecast[1], the urban rainstorm waterlogging prevention, the canal network control, and the water regulation. The external forcing of the precipitation for the current hydrological forecast is either from the assumed precipitation models or the observational precipitation after the heavy rain events. However, the forecasts of the time-varying hydrological elements require a synchronous accurate precipitation forecast for adjustment and correction. And the hydrological forecast ability is severely restricted by the heavy rain forecast level. Without an accurate precipitation forecasts, the hydrological forecast results are not reliable. Moreover, the conventional meteorological observations are inhomogeneous. So, for a local heavy rain forecast, whether a statistical forecast based on historical data or a numerical prediction based on the dynamic theory, the forecast quality will always be subject to limitations of the resolution and the accuracy. All these problems are not only difficult in weather forecast but also in hydrodynamics. In recent years, the use of the high temporal and spatial resolution satellite data is expected to improve the precipitation forecast. And the effectiveness of the satellite data highly depends on the progress of the numerical prediction. The numerical prediction is shown to be an effective method to improve the precipitation forecast, and it relies strongly on the numerical models and the initialconditions[2]. With improved numerical models, the updated data assimilation theory and technique become more and more important for obtaining good initial conditions.

Currently, the variational and ensemble data assimilation methods represent two mainstream directions, and there were many related studies in the atmospheric and oceanic fields[3]. It was found that the background error covariance, which should be varied substantially with the flow of the day, is an important parameter in the three-dimensional variational (3DVAR) Data Assimilation (DA) framework. However, the background error covariance widely used in the 3DVAR is assumed to be static, nearly homogeneous and isotropic[4]. Although with the ensemble data assimilation methods, such as the Ensemble Kalman Filters (EnKF), the flow dependent background error covariance can easily be obtained, the heavy computation induced by the increased amount of observations restricts their further applications.

Therefore, to construct a new flow dependent background error covariance for the 3DVAR, the hybrid data assimilation schemes (the hybrid schemes for short) were proposed[4]. The main idea is to improve the analysis field quality by incorporating a weighted ensemble covariance, generated by the ensemble prediction, and a weighted static background error covariance into the present 3DVAR framework[5]. In recent years, many related researches were reported[5-10]. But applications with real data were limited to hurricane forecasts with conventional insitu data, satellite derived wind and temperature data, and radar radial velocity data[8-10]. There were no studies that directly assimilate satellite radiance data by hybrid schemes for heavy rain forecasts. Moreover, the ensemble predictions are usually generated by the EnKF and the Ensemble Transform Kalman Filter (ETKF)[5,6,8-10]. However, in the current operational ensemble prediction systems, the Breeding Growing Mode (BGM)[11]method is used as a preferred method for advantages of clear physical meaning and low computational cost. And there were no studies incorporating BGM ensemble forecast products into the hybrid scheme. Meanwhile, there are two major differences between the hybrid scheme and the 3DVAR. One is that the former adopts a flow dependent ensemble covariance as the background error covariance whereas the 3DVAR uses a static background error covariance. The other is that the former uses the mean of the ensemble prediction as background field whereas the 3DVAR uses a single deterministic forecast. The importance of the flow dependent background error covariance was noticed, but the role of the ensemble mean background field in the heavy rain forecast has not received enough attention.

In this paper, based on the WRF model, the BGM-3DVAR scheme using AMSU-A radiance data is applied to a heavy rain occurred in Hefeng, Hubei Province from June 29 to June 30, 2009. And the ensemble prediction is made by using the BGM method as in the operational ensemble forecast. The results may be used for both atmosphere and ocean numerical forecasts.

1. Hybrid BGM-3DVAR DA scheme using satellite radiance data

1.1 Hybrid BGM-3DVAR DA scheme

The atmosphere or ocean numerical models can be expressed as an initial value problem of nonlinear evolution equations as follows[12].

The solution of Eq.(1) is uniquely determined by the initial valueu. Since Eq.(1) is for the nonlinear atmosphere or ocean system, the exact solution is sensitive to the initial value. In addition, the data assimilation is an effective way to obtain an accurate initial value. Although the four-dimensional variational (4DVAR) and the ensemble DA dominate the mainstream direction, the 3DVAR still retains in many operational numerical weather prediction systems. With the 3DVAR, the data assimilation is reduced to the minimizing of the quadratic functional between the analysis and the background field, and the cost function is

where xand xbdenote the analysis and the background field variables, respectively,yois the observational variable,H represents the nonlinear observational operator,Ris the observational error covariance matrix, andB is the background error covariance matrix.

Bis one of the most crucial parameters in the 3DVAR systems, and its ability of extracting information included in the background field directly affects the data assimilation. In the real atmosphere or ocean, Bshould be varied substantially with the flow of the day. However, in nearly all the 3DVAR systems,Bis assumed to be static, homogeneous and isotropic, which is con sidered as a m ajor shortcoming of t he 3DVAR.Tocompensateforthisshortcomingandtoreveal the flow dependent characteristics ofB , in the hybrid ensemble-variational data assimilation scheme, B is decomposed by a linear combination of B1and B[6]2

Table 1 The scheme and the purpose of the control and data assimilation experiments

where B1is the static background error covariance matrix,B2is the ensemble covariance matrix,α1and α2are the weighting coefficients, and α1+ α2=1. Generally speaking,B1is obtained via the NMC method,B2is calculated based on the ensemble prediction products, and the formula is

1.2 Model and data

The WRF is a common mode for the quantitative precipitation forecast, and it is often incorporated into a limited area of coupled ocean-atmosphere models, such as the WRF-ROMS[13]. Generally speaking, it is used to study tropical cyclones, coastal storm surges, sea fogs and so on, and it has become a mainstream model for the mesoscale air-sea interaction study. The WRF 3.3.1 is configured to have a 30 km horizontal grid spacing, 180×120 horizontal grid points, 28 vertical levels, and 120 s time step. The physical bases include the New Thompson microphysics scheme, the Betts-Miller-Janjic cumulus parameterization scheme, the Rapid Radiative Transfer Model (RRTM) longwave scheme, and the Dudhia shortwave scheme. The initial and lateral boundary conditions are defined by the NCEP 1o×1oglobal reanalysis data every 6 h. The AMSU-A satellite radiance data from NOAA-15/16/ 18 are used via CRTM2.0.2. The static background error covariance is re-calculated by the NMC method, and the ensemble covariance is generated by using the BGM ensemble prediction products.

The BGM ensemble prediction is designed as follows, 5 groups of random perturbations are introduced, the perturbation form is the uniformly distributed random number in the range of (-1,1), the modes of the perturbations are the 6 h forecast root mean squares errors, the breeding cycle is 6 h, and the breeding lasts for 48 h. The initial perturbations of the ensemble forecast, including the 5 groups of breeding modes, are added and subtracted, respectively, in the initial analysis field to generate 10 new initial analysis fields. After that, the model is integrated with the new initial analysis fields to generate 10 forecast members. The breeding starts from 1800 UTC June 26, ends at 1800 UTC June 28, and the integration of the forecast ends at 0000 UTC June 29.

2. The impact of hybrid DA on heavy rain forecasts

2.1 Event

A strong rainstorm occurred in the middle and lower reaches of the Yangtze River from June 28 to July 1, 2009. In most areas, the amount of 3 d rainfall was between 0.10 m and 0.15 m. Three heavy rain centers above 0.10 m were located in southern Anhui Province, northeast and southwest of Hubei Province, respectively. Two heavy rain centers above 0.20 m were Huaining, Anhui Province and Hefeng, Hubei Province, and the maximum rainfall reached 0.2415 mand 0.3132 m, respectively. The 24 h accumulated observational rainfall from 0000 UTC June 29 to 0000 UTC June 30 in Hefeng exceeded 0.30 m, which broke the history record. It also led to flash floods and other hydrological disasters. The 24 h accumulated precipitation above 0.025 m is shown in Fig.(1a). Considering that the heavy rain was concentrated in 29oN-32oN, 108oE-120oE, this area is set as the verification region.

Fig.1 The 24 h accumulated observational precipitation from 0000 UTC June 29 to 0000 UTC June 30 (Unit: m/24 h, the contour interval is 0.025 m)

2.2 Experiment design

To reveal the impact of the hybrid scheme on the heavy rain forecasts, three experiments of the CON, the 3DVAR and the BGM-3DVAR are designed. The hybrid options are the same as those in Ref.[8]. The scheme and the purpose are listed in Table 1.

To compare the difference of the BGM-3DVAR and the 3DVAR schemes with respect to the heavy rain forecast, results are analyzed from three aspects as the 24 h accumulated precipitation distribution, the score and the initial field.

2.3 The impact on the precipitation forecast

As can be seen in Fig.1(b), the WRF model has some abilities to forecast the heavy rain band. But it is a little to the south than the observation. The CON experiment does not have the ability to forecast the maximum precipitation centers. The forecast precipitation in the southwest of Hubei Province has two centers, and the intensity is obviously weaker than that in Fig.1(a), while the forecast precipitation in the south of Anhui Province involves a larger area, and the intensity is greater. After the assimilation of the satellite radiance data with the 3DVAR, the forecast precipitation intensity is significantly enhanced (Fig.1(c)). However, in the southwest of Hubei Province, it still has two centers, and the maximum precipitation is still much weaker than that in Fig.1(a), the forecast precipitation area in the southern of Anhui Province is smaller, while the maximum precipitation is obviously greater. Fig.1(d) shows the forecast precipitation of the BGM-3DVAR. It is easy to see that the forecast precipitation in the southwest of Hubei Province has only one center, in addition, the maximum precipitation is significantly enhanced, then it is very close to that in Fig.1(a). The forecast precipitation area in the south of Anhui Province is reduced, and the intensity is weakened. Compared with the 3DVAR, the hybrid scheme has enhanced 24 h accumulated precipitation from 0.175 m to 0.275 m with an increase of 57%.

Fig.2 The SAL verification of 24 h forecast precipitation of CON, 3DVAR and BGM-3DVAR

In order to objectively and quantitatively verify the precipitation forecast, a Structure, Amplitude andLocation (SAL) quantitative assessment method proposed by Wernli et al.[14]is used. The forecast precipitation is verified according to the structure, the amplitude and the location. The smaller the absolute values of S, A and L are, the better the forecast is. The SAL verification results of the CON, the 3DVAR and the BGM-3DVAR are shown in Fig.2.

Fig.3 The difference of the initial geopotential height and the relative humidity between the data assimilation experiments and the control experiment at 700 hPa, 0000 UTC June 29. The shading represents the 24 h accumulated observational rainfall from 0000 UTC June 29 to 0000 UTC June 30 (Unit: m)

The Lrepresents the effect of the precipitation location forecast, and it is easy to see that three experiments give almost an identical value of the absolute value |A|decreases from 0.16 to 0.01. This implies that the effect of the satellite radiance on the forecast is closely related to the data assimilation schemes. That is to say, with the 3DVAR, the forecast for the precipitation intensity can not be significantly improved, while with the BGM-3DVAR, a great improvement is observed. Moreover, there are also significant differences among the three experiments for the precipitation structure forecast. With the CON, we obtain the maximum|S|while with BGM-3DVAR, we obtain the minimum one, the absolute value|S|de-L, with that of the CON being slightly larger than the other two. It indicates that the ability to forecast the location of the rain band can be improved after the assimilation of satellite radiance data with whatever data assimilation scheme. There is no significant difference between those obtained with the 3DVAR and the BGM-3DVAR. The A values of three experiments are very different. The 3DVAR gives the maximum value while the BGM-3DVAR gives the minimum one, and creases from 0.09 to 0.01. This suggests that with the 3DVAR, the precipitation structure forecast can be improved to some extent, but the BGM-3DVAR is the best. In summary, with the BGM-3DVAR, the precipitation forecast can be improved from all aspects, including the precipitation structure, the amplitude and the location, while with the 3DVAR, three indexes can not all be improved.

2.4 The impact on the initial field

From the above analysis, it is clear that the BGM-3DVAR scheme can significantly improve the precipitation forecast. The purpose of the data assimilation is to acquire accurate initial fields for numerical models, so more useful mesoscale information[15]can be described.

To reveal the impact of the DA schemes on the initial field, the difference of the initial field between the data assimilation experiments and the control experiments is analyzed. The geopotential height and the relative humidity at 700 hPa are selected for the analysis. Their distributions at 0000 UTC June 29 are shown in Fig.3.

There are significant geopotential height differences between those obtained with the 3DVAR and the CON (Fig.3(a)). The differences in the south is larger than that in the north, and the positive and negative large value centers appear around the north and south boundaries. However, the rain band lies in the small absolute value area among the large value centers. And the value in the heavy rain area is between –10 and 20. This suggests that there is no evident change of the geopotential height in the heavy rain area afterthe assimilation of the satellite radiance data with the 3DVAR, and there is no obvious improvement in the heavy rain area. This means that no mesoscale information with the same scales as the heavy rain clusters is added in the heavy rain area by the use of the 3DVAR. Figure 3(b) shows the difference of the geopotential heights obtained with the BGM-3DVAR and the CON. Similar to the Fig.3(a), there are positive and negative large value centers around the south and north boundaries (Fig.3(b)). But there are also large value centers in the heavy rain area. And the positive and negative value centers are along the northeastsouthwest directions alternatively. Especially, the negative value centers are consistent with three heavy rain centers. This suggests that there are significant changes of the geopotential height in the heavy rain area after the use of the satellite radiance data assimilation, with the BGM-3DVAR scheme, and the scales of the geopotential height change are consistent with heavy rain clusters. As can be seen in Fig.3(c), the large value centers of the relative humidity differences obtained with the 3DVAR and the CON are also not in the heavy rain area. In the heavy rain area, the value is between 0 and 2.5, and most of them are close to zero. This indicates that there are evident change of the relative humidity outside of the heavy rain area by using the 3DVAR. Figure 3(d) shows the relative humidity differences obtained with the BGM-3DVAR and the CON. Obviously, there are large value centers not only outside of but also in the heavy rain area. Moreover, the scales of the area of the centers are similar to the heavy rain clusters. This suggests that there are evident change of the relative humidity initial field in the heavy rain area by using the BGM-3DVAR, and the mesoscale information consistent with the heavy rain clusters is added to the heavy rain area.

Table 2 The scheme and the purpose of sensitive experiments

From the initial field analysis, it is easy to see that the direct satellite radiance data assimilation does not always improve the forecast, and the improvement is closely related to the data assimilation schemes. The BGM-3DVAR can play a better role in the assimilation of the satellite radiance data. The mesoscale information consistent with heavy rain clusters is added to the initial field especially in the heavy rain area, and there is a significant improvement in the precipitation structure and the intensity forecast, especially for the precipitation centers in Hefeng, Hubei Province. So developing a new data assimilation scheme is essential for improving the efficiency of the satellite data assimilation.

Fig.4 The 24 h accumulated precipitation forecast of sensitive experiments from 0000 UTC June 29 to 0000 UTC June 30 (Unit: m, the contour interval is 0.025 m)

3. Effect of background field and background error covariance in the hybrid DA scheme

It was shown that 85 percent of information comes from the background field[16]in the analysis field generated by the 3DVAR. The way of transferring the observational information from one analysis cycle to the next depends on the background field, and the statistical description of the background error covariance is critical for a successful transfer. Therefore, the background field and the background error covariance directly affect the results of data assimilation schemes. There are two main elements in the BGM-3DVAR, one is the ensemble mean background field and the other is the ensemble covariance. Furtherunderstanding of the impact of the two key factors is important for the construction of data assimilation schemes.

Fig.5 The difference of the initial fields in the data assimilation experiments and the sensitive experiments at 700 hPa, 0000 UTC June 29. The shading represents the 24 h accumulated observational rainfall from 0000 UTC June 29 to 0000 UTC June 30 (Unit: m)

In order to reveal the impact of the background field and the background error covariance on the initial field and the precipitation forecast, another two sensitive experiments are designed based on Table 1. The scheme and the purpose are listed in Table 2.

3.1 The impact on the precipitation forecast

Figure 4 shows the 24 h accumulated forecast precipitation of experiments 3DVAR-MEAN and BGM-3DVAR-CON. Compared with Fig.1, it is easy to see that the precipitation forecasts made with the 3DVAR-MEAN (Fig.4(a)) and the BGM-3DVARCON (Fig.4(b)) are close to those made with the 3DVAR and the CON. All three experiments with the 3DVAR, the 3DVAR-MEAN and the BGM-3DVARCON can be used to forecast the heavy rain band, but not the maximum precipitation centers, especially for the centers in Hefeng, Hubei province. On the other hand, the experiment with the BGM-3DVAR (Fig.1(d)) shows a very good performance of the precipitation structure and the intensity. So, it can be concluded that changing only one factor of the background and the background error covariance can not significantly improve the precipitation structure and the intensity forecast, and the significant improvement is depended on the joint action of the two factors.

3.2 The impact on the initial field

In order to further understand the impact of different data assimilation schemes on the initial field, the differences between the data assimilation experiments at 0000 UTC June 29 are researched. The effect of the background field in the hybrid scheme can be revealed by the initial field difference between the experiments of the BGM-3DVAR and the BGM-3DVAR-CON (the same as the experiments of the 3DVAR-MEAN and the 3DVAR). Similarly, to analyze the initial field difference between the experiments of the BGM-3DVAR and the 3DVAR-MEAN (the same as the experiments of the BGM-3DVAR-CON and the 3DVAR), the effect of the background error covariance can be revealed.

Figures 5(a) and 5(c) show the initial field difference of the geopotential height and the relative humidity obtained with the BGM-3DVAR and the BGM-3DVAR-CON, respectively. As can be seen in Figs.5(b) and 5(d), the same distributions as Figs.5(a) and 5(c) obtained with the 3DVAR-MEAN and the 3DVAR are obtained. It is clear that the positive and negative large value centers are not only around the south and north boundaries but also in the heavy rain area, and the scale is consistent with the heavy rain clusters (Figs.5(a) and 5(b)). Moreover, the same distributions are also shown for the relative humidity (Figs.5(c) and 5(d)). This indicates that the mesoscale information consistent with the heavy rain clusters is added in the heavy rain area after the use of the ensemble mean background field in the 3DVAR, and it does not depend on the background error covariance. In other words, whether it is the static backgrounderror covariance based on the climatic statistics or the flow dependent one based on the ensemble forecast, after the use of the ensemble mean background field in the 3DVAR, the mesoscale information consistent with the heavy rain clusters is included in the initial field.

Fig.6 The difference of the initial field in the data assimilation experiments and the sensitive experiments at 700 hPa, 0000 UTC June 29. The shading represents the 24 h accumulated observational rainfall from 0000 UTC June 29 to 0000 UTC June 30 (Unit: m)

Figures 6(a) and 6(c) show the initial field difference of the geopotential height and the relative humidity obtained with the BGM-3DVAR and the 3DVAR-MEAN, respectively. As can be seen in Figs.6(b) and 6(d), the same distributions as in Figs.6(a) and 6(c) obtained with the BGM-3DVARCON and the 3DVAR are shown. In Figs.6(a) and 6(b), it is clear that the positive and negative centers are only around the south and north boundaries and the value is relatively small in the heavy rain area. In the meantime, in Figs.6(c) and 6(d), the positive and negative centers are not only around the south and north boundaries but also in the heavy rain area, and the scale is consistent with the heavy rain clusters. Therefore, based on the different background error covariance, the change of the geopotential height in the heavy rain area is relatively small, whereas, the change of the relative humidity is significant. So it can be seen that the main effect of the background error covariance is to make the change of the low-level water vapor.

As can be seen in Figs.5 and 6, there contains the mesoscale information consistent with the heavy rain clusters added to the initial field in the heavy rain area based on the ensemble mean background field, using the 3DVAR scheme. And the changes of the geopotential height and the relative humidity are all significant. On the other hand, based on the flow dependent background covariance, there is a significant change of the relative humidity, but a relatively small change of the geopotential height. Therefore, it can be concluded that the main effect of the ensemble mean background field is to add the mesoscale information, while the main effect of the flow dependent background covariance is to transfer the information, especially for the low-level water vapor. The joint action of the two key factors contributes to the improvement of the precipitation forecast. The ensemble mean background fields include the mesoscale information with nearly the same scales as the heavy rain clusters. However, to make an efficient use of the information, the flow dependent background error covariance must be introduced, which can properly transfer the information to the heavy rain area. So the interaction of the background field and the ensemble covariance is critical for the precipitation forecast.

4. Conclusions and discussions

This study aims to improve the forecast of the heavy rain intensity and the distribution structure, a hybrid BGM-3DVAR data assimilation scheme is designed using AMSU-A satellite radiance data, the 3DVAR DA technique and the BGM ensemble prediction method. The experiments demonstrate that the hybrid data assimilation scheme improves the heavy rain forecast significantly. Main conclusions are asfollows:

(1) Compared with the traditional 3DVAR scheme, the hybrid data assimilation scheme enjoys obvious advantages. The 24 h accumulated precipitation forecast from the hybrid scheme gives an increase of 57%, in addition, the SAL score indicates that the precipitation intensity and structure are improved significantly, which is more closer to the observation.

(2) It is feasible to incorporate the ensemble forecast based on the BGM method into the framework of the hybrid data assimilation scheme. The ensemble forecast with the BGM method can provide an effective background field and the flow dependent background error covariance.

(3) Different from the 3DVAR scheme, the hybrid data assimilation scheme adds some fine mesoscale information with similar scales as the heavy rain clusters to the initial field, especially in the heavy rain area, therefore, the distribution structure and the intensity forecast can be improved significantly. The information added by the 3DVAR scheme to the heavy rain area, has a relatively large scale and the evident change is outside of the heavy rain area, which contributes little to the improvement of the heavy rain forecast.

(4) The ensemble mean background field and the ensemble covariance constructed by the BGM ensemble forecast have two important factors for the improvement of the heavy rain forecast. Both of them are beneficial for adding some fine mesoscale information with similar scales as the heavy rain clusters to the heavy rain area. The ensemble mean background field has a significant impact on all variables from low levels to the high, while the ensemble covariance mainly reflects the improvement of the low-level moisture.

The hybrid data assimilation scheme proposed in this study is applied in only one heavy rain case based on the AMSU-A satellite radiance data, more heavy rain cases should be tested before it is put into operational applications. In fact, the advantages of the hybrid data assimilation scheme are demonstrated in the heavy rain forecast, meanwhile, it should be mentioned that the hybrid scheme also has other broad application prospects. Firstly, it will provide a reference for the ocean data assimilation. By improving the background error covariance in ocean data assimilation systems, the dynamic characteristics varied with the ocean flow of the day may be enhanced, and some fine mesoscale information may be enriched in the initial field, therefore, a better understanding of the ocean mesoscale dynamic processes may be possible. Secondly, it will also directly provide more accurate initial values for coupled ocean-atmosphere models, then a better understanding of the air-sea interaction dynamic process may be achieved. Finally, the finer heavy rain forecast obtained from the hybrid scheme is a crucial input parameter for the hydrological forecast model, and it has a decisive impact on the successful flood forecast and the urban rainstorm waterlogging prevention. So, the fine heavy rain forecast obtained in this study will be of great significance for coupled hydrometeorological models in enhancing the forecast capability for floods and other disasters.

[1] BAO Hong-jun, ZHAO Lin-na. Flood forecast of Huaihe River based on TIGGE Ensemble Predictions[J]. Journal of Hydraulic Engineering, 2012, 43(2): 216-244(in Chinese).

[2] WANG Shu-chang, HUANG Si-xun and LI Yi. Sensitive numerical simulation and analysis of rainstorm using nested WRF model[J]. Journal of Hydrodyna- mics Ser. B, 2006, 18(5): 578-586.

[3] ZHU J, FEI Z. and LI X. A new localization implementation scheme for ensemble data assimilation of non- local observations[J]. Tellus A, 2011, 63(2): 244-255.

[4] LORENC A. C. The potential of the ensemble Kalman filter for NWP–A comparison with 4D-VAR[J]. Quart Journal of the Royal Meteorological Society, 2003, 129(595): 3183-3203.

[5] WANG X., BARKER D. M. and SNYDER C. et al. A hybrid ETKF-3DVAR data assimilation scheme for the WRF model. Part I: observing system simulation experiment[J]. Monthly Weather Review, 2008, 136(12): 5116-5131.

[6] WANG X., SNYDER C. and. HAMILL T. M. On the theoretical equivalence of differently proposed ensemble/3D-VAR hybrid analysis schemes[J]. Monthly Weather Review, 2007, 135(1): 222-227.

[7] LIU C., XIAO Q. and WANG B. An ensemble-based four-dimensional variational data assimilation scheme. Part Ⅰ: Technical formulation and preliminary test[J]. Monthly Weather Review, 2008, 136(9): 3363-3373.

[8] WANG X. Application of the WRF hybrid ETKF-3DVAR data assimilation system for hurricane track forecasts[J]. Wea Forecasting, 2011, 26(6): 868-884.

[9] LI Y., WANG X. and Xue M. Assimilation of radar radial velocity data with the WRF ensemble-3DVAR hybrid system for the prediction of hurricane IKE (2008)[J]. Monthly Weather Review, 2012, 140(11): 3507-3524.

[10] ZHANG F., ZHANG M. and POTERJOY J. E3DVAR: Coupling an ensemble Kalman filter with three-dimensional variational data assimilation in a limited-area weather prediction model and comparison to E4DVAR[J]. Monthly Weather Review, 2013, 141(3): 900-917.

[11] TOTH Z., KALNAY E. Ensemble forecasting at NCEP: The breeding method[J]. Monthly Weather Review, 1997, 125(12): 3297-3318.

[12] TALAGRAND O., CURTIER P. Variational assimilation of meteorological observations with the adjoint vorticity equation. Part I: Theory[J]. Quarterly Journal of the Royal Meteorological Society, 1987, 113(478): 1311-1328.

[13] NOLAN D. S., MONTGOMERY M. T. Nonhydrostatic, three-dimensional perturbations to balanced, hurricanelike vortices. Part I: linearized formulation, stability, and evolution[J]. Journal of the Atmospheric Sciences,2002, 59(21): 2989-3020.

[14] WERNLI H., PAULAT M. and HAGEN M. et a1. SAL-a novel quality measure for the verification of quantitative precipitation forecasts[J]. Monthly Wea- ther Review, 2008, 136(11): 4470-4487.

[15] LUO Yu, ZHANG Li-feng. Effect of instabilities of flow on mesoscale predictability of weather systems[J]. Journal of Hydrodynamics, 2011, 23(2): 193-203.

[16] CARDINALI C., PEZZULLI S. and ANDERSON E., Influence-matrix diagnostic of a data assimilation system[J]. Quarterly Journal of the Royal Meteorological Society, 2004, 130(603): 2767-2786.

10.1016/S1001-6058(11)60382-0

* Project supported by the National Natural Science Foundation of China (Grant No. 40975031), the National Science Foundation for Young Scientists of China (Grant No. 41205074).

Biography: XIONG Chun-hui (1984-), Male, Ph. D.

ZHANG Li-feng, E-mail: zhanglif@yeah.net