Impact of Doppler weather radar data on thunderstorm simulation during STORM pilot phase 2009

Transcription

1 Nat Hazards (2014) 74: DOI /s ORIGINAL PAPER Impact of Doppler weather radar data on thunderstorm simulation during STORM pilot phase 2009 S. Kiran Prasad U. C. Mohanty A. Routray Krishna K. Osuri S. S. V. S. Ramakrishna Dev Niyogi Received: 9 August 2013 / Accepted: 18 May 2014 / Published online: 29 June 2014 Ó Springer Science+Business Media Dordrecht 2014 Abstract This study assesses the impact of Doppler weather radar (DWR) data (reflectivity and radial wind) assimilation on the simulation of severe thunderstorms (STS) events over the Indian monsoon region. Two different events that occurred during the Severe Thunderstorms Observations and Regional Modeling (STORM) pilot phase in 2009 were simulated. Numerical experiments 3DV (assimilation of DWR observations) and CNTL (without data assimilation) were conducted using the three-dimensional variational data assimilation technique with the Advanced Research Weather Research and Forecasting model (WRF-ARW). The results show that consistent with prior studies the 3DV experiment, initialized by assimilation of DWR observations, performed better than the CNTL experiment over the Indian region. The enhanced performance was a result of improved representation and simulation of wind and moisture fields in the boundary layer at the initial time in the model. Assimilating DWR data caused higher moisture incursion and increased instability, which led to stronger convective activity in the simulations. Overall, the dynamic and thermodynamic features of the two thunderstorms were consistently better simulated after ingesting DWR data, as compared to the CNTL simulation. In the 3DV experiment, higher instability was observed in the analyses of thermodynamic indices and equivalent potential temperature (h e ) fields. Maximum convergence during the mature stage was also noted, consistent with maximum vertical velocities in the S. Kiran Prasad (&) U. C. Mohanty K. K. Osuri School of Earth, Ocean and Climate Sciences, Indian Institute of Technology, Satyanagar, Bhubaneswar , India skp29879@gmail.com A. Routray National Centre for Medium Range Weather Forecasting (NCMRWF), A-50, Institutional Area, Sector-62, Noida , India S. S. V. S. Ramakrishna Department of Meteorology and Oceanography, Andhra University, Visakhapatnam , India D. Niyogi Purdue University, West Lafayette, IN, USA

2 1404 Nat Hazards (2014) 74: assimilation experiment (3DV). In addition, simulated hydrometeor (water vapor mixing ratio, cloud water mixing ratio, and rain water mixing ratio) structures improved with the 3DV experiment, compared to that of CNTL. From the higher equitable threat scores, it is evident that the assimilation of DWR data enhanced the skill in rainfall prediction associated with the STS over the Indian monsoon region. These results add to the body of evidence now which provide consistent and notable improvements in the mesoscale model results over the Indian monsoon region after assimilating DWR fields. Keywords program Severe thunderstorms DWR data Variational data assimilation STORM 1 Introduction Severe thunderstorms are a dominant weather feature during the pre-monsoon season (April May) over eastern and northeastern parts of India [i.e., Gangetic West Bengal (GWB), Jharkhand, Odisha, Bihar, Assam] and parts of the northeastern states (Fig. 1). These storms generally move from the northwest to the southeast and disrupt everyday life and cause widespread loss of human and animal life and property. Improved prediction of severe thunderstorms (STS) is important but remains a challenge for the operational and research community alike. Understanding the hazard potential of the STS systems, the Department of Science and Technology (DST) and Ministry of Earth Sciences (MoES), Government of India, initiated a research program called the severe thunderstorms observations and regional modeling (STORM). The STORM program focuses on a comprehensive observational and modeling study on the genesis, evolution, and life cycle of STS through a regional network of observations, analyses, and prediction systems. Figure 1a shows the STORM domain where different observations were collected through various platforms, such as Doppler weather radar (DWR), automatic weather stations (AWS), radiosonde/rawinsonde (RS/RW), and surface synoptic observations (SYNOP). Further details about the STORM program can be found in Das et al. (2014). Understanding the thermodynamic and physical mechanisms of STS is essential and can be achieved by simulating these systems with the help of high-resolution non-hydrostatic mesoscale models with sophisticated parameterization schemes (Weiss et al. 2006). A number of studies have been conducted (Brooks and Wilhelmson 1992; Farely et al. 1992; Lopez et al. 2007) to understand the multiple processes that influence STS evolution. Litta et al. (2012) simulated the thunderstorms during the STORM field experiments in 2007, 2009, and 2010 using the WRF nonhydrostatic mesoscale model (NMM) model. The study of thunderstorm systems over the Indian region has shown that numerical weather prediction (NWP) models have modestly good capability to simulate high-impact convective events (e.g., Litta et al. 2012). The WRF model is being used in quasi-operational mode over the Indian region, and the improvements in the capabilities to simulate STS events using WRF modeling system over the Indian region are of direct relevance to the operational as well as the research community. Recently, there has been availability of DWR data for selected coastal Indian regions. It is well known that convective storms and associated mesoscale and microscale systems are efficiently studied using DWR data (Wakimoto et al. 2004; Mukhopadhyay et al. 2009; Latha and Murthy 2011). Assimilating DWR data into mesoscale models is expected to help improve the simulation and prediction of these STS events (Abhilash et al. 2007;

3 Nat Hazards (2014) 74: Srivastava et al. 2010; Roy Bhowmik et al. 2011). However, the assessment of the impact of assimilating DWR data on the performance of WRF modeling system for simulating STS events over India has not been undertaken and forms the motivation for this study. The WRF ARW modeling system is capable of assimilating radar radial velocity and reflectivity using the three-dimensional variational (3DVAR) data assimilation technique (Xiao et al. 2005, 2007). A few case studies have been reported with experiments involving assimilation of the Indian DWR data in order to simulate extreme weather events using mesoscale models (Abhilash et al. 2007; Routray et al. 2010, 2013; Srivastava et al. 2010; Roy Bhowmik et al. 2011). Abhilash et al. (2007) established that assimilation of derived Doppler radar wind data in the MM5 model using 3DVAR had a positive impact on the prediction of intensity, organization, and propagation of rainbands associated with premonsoon convective storms over northeast India. Srivastava et al. (2010) and Roy Bhowmik et al. (2011) showed that assimilating quality-controlled radar data in the advanced regional prediction system (ARPS) model showed a positive impact on model performance. Routray et al. (2010) were the first to assess the utilization of DWR radial velocity and reflectivity in a mesoscale model (WRF ARW) to predict Bay of Bengal monsoon depressions (MDs). They found that assimilation of DWR data improved the prediction of location, propagation, and development of rainbands associated with the MDs. Despite the importance of DWR data for use in extreme weather event warnings, there have been limited efforts in utilizing the same in the assimilation cycle of weather prediction models used for operational purpose in India (Das et al. 2006). The main objective of this study was to explore the impact of DWR radial velocity and reflectivity collected during the STORM pilot phase for STS simulation. This study is the first research work utilizing STORM field experiment data in an assimilation system to improve the forecast skill of STS over east India. This paper is organized as follows: Sect. 2 presents an overview of the WRF ARW and its 3DVAR analysis system. In Sect. 3, a brief description of the synoptic setup associated with two thunderstorm cases observed during the STORM field phase is presented. The details of the numerical experiments performed, and data used are provided in Sect. 4. Section 5 deals with results and discussion of the model simulations, and conclusions drawn from the study are presented in Sect WRF ARW and 3DVAR systems The WRF ARW modeling system version 3.2 was used in this study. The model physics options and parameterization details are presented in Skamarock et al. (2008). The default version as available from the WRF portal and used in the quasi-operational mode over India was adopted for this study (Table 1). The DWR data are ingested within the WRF ARW system using a 3DVAR approach following Barker et al. (2004) and Jianfeng et al. (2005). The basic target of the 3DVAR system is to produce a best estimate of the real atmospheric state at analysis time through the repetitious solution of a specified cost-function J(x) as follows. Jx ðþ¼j b þ J o ¼ 1 2 ðx xb Þ T B 1 ðx x b Þþ 1 2 ðy yo Þ T ðe þ FÞ 1 ðy y o Þ ð1þ where J b and J o represent the cost functions of the background and observations, respectively; x, x b, and B represent the state vector, first guess, and background error statistics covariances, respectively. y denotes the observation space (y = Hx), whereas H is

4 1406 Nat Hazards (2014) 74:

5 Nat Hazards (2014) 74: b Fig. 1 a Domain of the STORM experiment (Source: b Model domain at 3-km resolution chosen for the present study Table 1 WRF model configuration used in the present study Dynamics Nonhydrostatic Main prognostic variables u, v, w, p 0, h 0, U 0 Number of domains Central point of the domain Horizontal grid distance Integration timestep Number of gridpoints Single domain Central lat.: 23.5 N Central lon.: 87.5 E 3 km 15 s X-direction 370 points Y-direction 370 points Number of vertical levels 51 Map projection Mercator Horizontal grid distribution Arakawa C-grid Nesting No nesting Vertical coordinate Terrain-following hydrostatic-pressure coordinate (51 sigma levels) Time integration Third-order Runge Kutta Spatial differencing scheme Sixth-order centered differencing Initial conditions Three-dimensional real data (FNL: ) Lateral boundary conditions Specified options for real data Microphysics Ferrier Radiation scheme GFDL shortwave radiation/gfdl longwave Cumulus parameterization schemes Grell Devenyi ensemble scheme PBL parameterization MYJ scheme the forward (nonlinear) operator and y o represents the observations. E and F represent the instrumental and representivity error covariance matrix, respectively. The climatological B statistics are computed using the NMC (National Meteorological Center) method (Parrish and Derber 1992) for the experimental domain. For this purpose, the WRF model was run for a month (April to May) to produce 12- and 24-h forecasts at 00 and 12 UTC initial conditions. Domain-dependent regional background error statistics were then computed by averaging these WRF forecast differences valid at the same hour for the desired period. The methodology for Doppler radar data assimilation (mainly reflectivity and the vertical velocity component of radial velocity) is available in the WRF-Var analysis system (Xiao et al. 2005, 2007). The radial velocity holds information on vertical atmospheric motions which are vital for convective initiation and forecasting. The Richardson (1922) balance equation is introduced into the WRF-Var physical transformation to generate the vertical velocity increments (Xiao et al. 2005). The assimilation of reflectivity heavily influences moisture and hydrometeor fields. The primary effect of the assimilation of radial velocity is on the wind analysis, and further, its impact on moisture and hydrometeor analysis is secondary. Therefore, assimilation of both radial velocity and reflectivity together leads to adjustments in both the dynamical and thermodynamical fields (Xiao et al. 2007; Routray et al. 2013). Details of the assimilation procedure of the Doppler radial

6 1408 Nat Hazards (2014) 74: velocity and reflectivity in the WRF-Var system can be found in studies by Xiao et al. (2005, 2007) and Routray et al. (2010, 2013). 3 Synoptic overview of thunderstorms In this study, two STS that occurred on 11 May 2009 (Case 1) and 12 May 2009 (Case 2) are considered. The detailed synoptic situations for the two cases are given below. Case 1 Thunderstorms were reported at many stations over GWB starting at UTC (approximately h IST) over Kolkata, Alipur, Malda, Digha, Bankura, and Kharagpur. Kolkata DWR (22.57 N, o E) showed strong reflectivity (around 50 dbz) over Ranchi (RNC) near Dumka (DMK) at 0809 UTC (Fig. 2a) and between Krishnanagar (KRG) and Mymensingh (MNS) at 0839 UTC. The storms intensified further, one with a northeast southwest orientation and another with an east west orientation (Fig. 2b). Merging together by 1130 UTC, the storms moved southeast and dissipated over the Bay of Bengal around 1600 UTC. This STS event was accompanied by light-to-moderate rain (24 h accumulated) as recorded at Dum Dum airport (33.3 mm) and Kharagpur (16.8 mm). Case 2 Thunderstorms were reported at many stations over the GWB starting at 0300 UTC to 1800 UTC over Malda, Midnapur, Kharagpur, Kolkata, and Canning. DWR displayed weak reflectivity (around 20 dbz; Fig. 2c) to the north of Kolkata which are remnants of the convection seen earlier over Bihar at 0000 UTC. Further, high reflectivity values are seen at around 100 km northwest of Kolkata (KOL) at 0425 UTC which intensified and initially moved to the southeast. The storm merged with a propagating system (Fig. 2d) that developed near Midnapur (MDP) and Chaibasa (CBS) and dissipated when moving south over the Bay of Bengal. Strong storm-related reflectivity (around 50 dbz) was observed over Ranchi (Fig. 2d) at 0923 UTC, and the storm intensified and moved southward dissipating near Baripada (BPD) at 16 UTC. Light-to-heavy rain (24 h accumulated) occurred at a few places over GWB with 44.2 mm of rainfall over Midnapur and 35.3 mm over Kharagpur. 4 Numerical experiments and data used Two sets of numerical experiments are carried out using the WRF ARW model with a single domain of 3-km horizontal resolution (Fig. 1b) and 51 sigma vertical levels. The first set of experiments, namely the control simulation (CNTL), is conducted without data assimilation. The model s initial and boundary conditions are provided by the National Centers for Environmental Prediction (NCEP) FiNal Analysis (FNL) in horizontal resolution. The second experiment (3DV) experiment is carried out with a 6-hourly assimilation cycle using Kolkata DWR reflectivity and radial velocity data (cutoff ±3 h), starting 12 h before the start of model integration in each case. The 6-h forecast obtained from the previous cycle is used as the first guess in the next cycle (Routray et al. 2013; Xiao et al. 2005). So, in this process of cyclic assimilation, two assimilation cycles are performed to initialize the model before the start of the actual model integration. In this procedure, the cloud water (q c ) and rainwater (q r ) information produced by the previous cycle are passed on to the next cycle alleviating the spin-up problem when the model integrates from the analysis in the cycling run. Convective parameterization (CP) is used in the simulations keeping in mind that quantitative precipitation forecasts could improve with the use of CP in high-resolution simulations (Deng and Stauffer 2006; Lean et al. 2008).

7 Nat Hazards (2014) 74: Fig. 2 DWR (Kolkata) images at a 0809 UTC, b 1124 UTC on 11 May 2009, c DWR data distribution at 1.5-km above sea level for Case 1, and DWR images at d 0425 UTC, e 0923 UTCon 12 May 2009, and f DWR data distribution at 1.5-km above sea level for Case 2 The DWR Kolkata (22.57 N, E) data are used in the present study. The DWR data are obtained through a volume scan. The dataset is stored on spherical coordinates with a range of 250 km, azimuth of 360 (1 beam width), 10 elevation angles at a resolution of km. The resolution of the datasets is much higher than the model resolution (3 km) considered in this study. So, in order to process the data to a regular grid at a resolution that is compatible with the analysis system, a strategy had to be implemented. Hence, the radar data are reduced to the Cartesian grids using the same map

8 1410 Nat Hazards (2014) 74: Table 2 RMSE of DWR observations at model initial time of 0000 UTC for Case 1 and Case 2 Cases Radial velocity (m s 21 ) Reflectivity (dbz) O B O A O B O A 11 May May projection as the model before being introduced into the 3DVAR system. A preprocessor module and stand-alone software (Routray 2007; Routray et al. 2010) were used to process and control the quality of the voluminous Indian DWR data. From the DWR data, information on radial velocity and reflectivity is retrieved along with information like radar latitude, radar longitude, elevation, and azimuth step. Data quality control is a vital step in data assimilation because poor quality data could ruin the analyses. Thus, in order to pick and filter out poor quality data, quality control procedures (range folding) are employed, such as discarding radial velocity with absolute values out of the range 2 30 m s 21 and reflectivity out of the range dbz. The range folding procedures complied with the typical practices implemented at radar data processing centers, such as the Korean Meteorological Administration (KMA), South Korea and the Center for the Analysis and Prediction of Storms (CAPS), Oklahoma, USA. Since these datasets are voluminous, a thinning procedure is implemented for assimilation. In this procedure, a desired thinning grid (two-dimensional) is used to pick up a maximum specified number of radar observations. In each grid, starting from the left-lower corner, the nearest data location comprising of maximum number of vertical levels is identified and processed. The model is integrated for 24 h in both experiments from the initial time 0000 UTC of 11 May 2009 (Case 1) and 12 May 2009 (Case 2). The numbers of ingested observations are 2183 and 3011 for Case 1 and Case 2, respectively, at model initial time. Figure 2 presents the data distribution at around 1.5 km above sea level. A comparison is made between O B (observation first guess) and O A (observation analysis) to demonstrate the performance of the 3DVAR analysis system. The statistics are calculated at grid points where the observational data are available for the same valid time. From Table 2, it is clear that the root mean square error (RMSE) of O B is higher than O A before and after the 3DVAR analysis at the initial time for reflectivity and radial velocity in both cases. This indicates that 3DVAR assimilates the radar observations in good order and produces an analysis that is consistent with radar observations. 5 Results and discussion The impact of modified initial conditions when DWR data are assimilated to simulate STS over the Indian region is examined. The simulated thermodynamic indices and meteorological fields during the thunderstorm events are compared with available observations. The rainfall obtained from the model simulations is compared with IMD station observations as well as with the tropical rainfall measurement mission (TRMM-3B42) precipitation data. 5.1 Impact on the model initial conditions Prior to analyzing the impact of assimilating DWR observations in the model forecast, it is essential to determine how the initial analyses vary among themselves. Hence, a

9 Nat Hazards (2014) 74: Fig. 3 Initial wind fields (m s -1 ) at 850 hpa for a CNTL, b 3DV, and c observed for Case 1 (valid at 00 UTC 11 May 2009), d f are the same as a c but for Case 2 (valid at 00 UTC 12 May 2009) comparative discussion of global analysis (FNL) and modified analysis after assimilation of DWR observations is presented in this section to assess the potential improvement in initial conditions. The initial wind fields at 850 hpa valid at 0000 UTC 11 May (Case 1) from both FNL and DWR analyses along with observed winds are shown in Fig. 3a c. It is noted that the wind flow pattern is similar over the domain in both analyses. In Case 1, a circulation over Jharkhand and adjacent Bihar is consistent with the observed. Convergence aided by cyclonic circulation in the lower levels is one of the important factors for providing the initial impulse for the growth of a prominent cumulonimbus cell and its subsequent

10 1412 Nat Hazards (2014) 74: development into a STS over the domain (Rai Sircar 1957). The magnitude of wind (9 11 m s21) increased along the Odisha and Bengal coast as well as over northwestern Bangladesh after the assimilation of DWR as compared to the FNL analysis (8 10 m s21). In Case 2, it is clearly seen that the wind strength increased by 1 2 m s21 over a large area covering coastal Odisha, West Bengal, and Bangladesh. Further, wind of m s21 over Jharkhand, adjoining Odisha and Bihar, increased in the modified analysis (Fig. 3e) after assimilation of DWR observations as compared to the global analysis (Fig. 3d). The feature is consistent with surrounding RS/RW observation stations (Fig. 3f). The stronger southerly/southwesterly winds from the DWR analysis help provide moisture convergence over the domain which is a necessary condition to help trigger convective activity (Desai 1950). Assimilation of radial velocity and reflectivity simultaneously gives way to adjustments in both the dynamic and thermodynamic fields at model initial time (Xiao and Sun 2007). In order to analyze the impact of DWR data on the moisture field, analysis increments (DWR minus CNTL) of specific humidity (Fig. 4) at initial time 925 and 850 hpa are examined. The increase in moisture content is noted (around g kg21) in the lower atmosphere in several places of the domain due to the strengthening of wind after DWR Fig. 4 Analyses increments (3DV-CNTL) of specific humidity (g kg-1) at a 925 hpa and b 850 hpa for Case 1 (valid at 00 UTC 11 May 2009), c d are the same as a b but for Case 2 (valid at 00 UTC 12 May 2009)

11 Nat Hazards (2014) 74: Table 3 Root mean square errors (RMSE) for temperature, relative humidity, and wind at Kolkata (22.65 N, E) valid at initial time (0000 UTC) for Case 1 and Case 2 Cases Temperature Relative humidity Wind CNTL 3DV CNTL 3DV CNTL 3DV Case Case observations are assimilated. The exceptions can be found in the northern parts. A similar trend in the moisture increment is found at 925 hpa in Case 2, but at 850 hpa, there is a positive moisture increment over the area of thunderstorm initiation after assimilation of DWR data. The moisture increment in the lower levels could be attributed to stronger moisture-laden winds from the Bay of Bengal in 3DV. Table 3 presents the root mean square error (RMSE) calculated against 0000 UTC upper-air observations at Kolkata (22.65 N, E) using data at 12 vertical levels for both experiments (Case 1 and Case 2). It is noted that temperature improvement was modest in both cases with the RMSE reducing by 0.16 and 0.21 in Case 1 and Case 2, respectively. Further, the relative humidity improved marginally in Case 1 and Case 2 with the RMSE dropping by 1.36 in Case 1 and 1.95 in Case 2. The wind also shows marginal improvement, where the RMSE drops by 0.76 in Case 1 and 0.58 in Case 2. Thus, the DWR data assimilation did modestly improve the representation of the atmospheric state as compared to the FNL analysis (CNTL). 5.2 Impact on model simulations The RMSE in the forecast fields is viewed as a standard measurement for evaluation of forecast accuracy (Wang and Yongfu 2001). The RMSE of the 12-h forecast of temperature and RH at various pressure levels is computed at 10 vertical levels against upper-air special observations at Kharagpur (22.31 N, E) to assess the impact of 3DV data on the model forecast. Due to the non-availability of upper-air observations at Kolkata at 1200 UTC, the special observations at the nearby station Kharagpur at 1200 UTC were considered. The RMSE for the 12-h forecast calculated against 1200 UTC upper-air observations for both experiments in Case 1 and Case 2 is illustrated in Table 4. It shows that the 12-h temperature forecast of 3DV is better than the CNTL, and the RMSE reduces by 0.58 in Case 1 and by 0.20 in Case 2. Relative humidity exhibited a marginal improvement where the RMSE dropped by 2.43 in Case 1 and 2.69 in Case 2. Because of the lack of wind data in the 1200 UTC sounding at Kharagpur, the RMSE of wind could not be calculated Instability Atmospheric instability is one of the major factors for convection to develop and can be assessed by examining various thermodynamic indices. Previous studies demonstrated the efficiency of different stability indices, such as convective available potential energy (CAPE), convective inhibition energy (CIN), and Lifted Index (LI) for thunderstorm initiation/prediction (Neumann 1971; Anthes 1976; Schultz 1989). The critical values of the indices are provided in Table 5.

12 1414 Nat Hazards (2014) 74: Table 4 Root mean square error (RMSE) of 12-h forecast of temperature and relative humidity at Kharagpur (19.1 N, 72.8 E) valid at 1200 UTC for Case 1 and Case 2 RMSE Temperature Relative humidity CNTL 3DV CNTL 3DV Case Case Table 5 The different stability indices and their critical values for severe thunderstorms Stability indices Description Critical values for severe thunderstorms Lifted index T T parcel \- 3 (Tyagi et al. 2011) CAPE R zn g Ttparcel Ttenv dz C1000 zf Ttenv (Tyagi et al. 2011) Z n, Level of neutral buoyancy; Z f, Level of free convection CIN R ztop dz g Ttparcel Ttenv zbottom Ttenv B150 (Chaudhuri and Middey 2009) CAPE, CIN, and LI are analyzed in Fig. 5a c for Case 1 and Fig. 5d f for Case 2 at Kharagpur (22.31 N, E). In both experiments, the CAPE for Case 1 (Fig. 5a) is moderately high at 0000 UTC at around 2,800 J kg 21, which is favorable for thunderstorm initiation. Rasmussen and Wilhelmson (1983) suggested a CAPE in excess of 1,500 J kg 21 is needed for the initiation of supercell thunderstorms. IMD reported severe thunderstorms over the region, which is well simulated in both experiments. A steady increase in CAPE is noticed from 0000 to 0500 UTC in both simulations. The 3DV experiment simulated slightly larger values of CAPE compared to the CNTL experiment. This could be attributed to an increase in wind shear after assimilation of the DWR data (Fig. 3b) as well as a more moist environment at lower levels. The first maxima of CAPE around 4,300 J kg 21 is seen at 0700 UTC during the developing stage ( UTC as per IMD observations) of the thunderstorm after assimilation of DWR data. The CAPE values gradually increase in the 3DV simulation and attain the mature stage ( UTC) of the thunderstorm. However, this feature is not simulated in the CNTL. Further, it is evident that the thunderstorm is in its mature stage from 0800 to 1100 UTC following a gradual decrease in CAPE after 1100 UTC marking its dissipation in both experiments. The 3DV simulation shows a rapid dissipation after the mature stage in contrast to the CNTL simulation. It is clearly evident that the thunderstorm from the initiation to the dissipation stage is well characterized after assimilation of the DWR observations unlike the CNTL simulation. In Case 1 (Fig. 5b), the CIN values drop from 400 J kg 21 at 0000 UTC to \50 J kg 21 at 0600 UTC. Thereafter, the values are maintained close to 0 J kg 21 up to 1100 UTC in both experiments (CNTL and 3DV), which is very conducive for thunderstorm occurrence and also is in good agreement with the CAPE (Fig. 5a) during different stages of the thunderstorm. The CIN values are slightly less in the 3DV simulation as compared to the CNTL. The LI is useful to indicate the likelihood of severe thunderstorms (Galway 1956).

13 Nat Hazards (2014) 74: Fig. 5 Model-simulated diurnal variation of thermodynamic indices at Kharagpur for a CAPE (J Kg 21 ), b CINE (J Kg 21 ), and c LI for Case 1. d f are the same as a c but for Case 2 In both simulations, LI values (Fig. 5c) around -8.4 at 0000 UTC indicate an unstable environment, which gradually drop to\-11.5 at 0400 UTC making the environment more conducive to thunderstorm initiation. The probability of thunderstorm initiation is greater when the LI is less than or equal to -6 (Air Weather Service Technical Report 1990). Similar to that in CAPE, the LI also exhibits two low value peaks and shows an increase from 0400 to 0700 UTC, but the values are still very favorable for thunderstorm initiation and development. The simulated LI values in the 3DV experiment are comparatively lower than in the CNTL experiment. For Case 2, it is clearly seen that the thermodynamic parameters are well represented in the simulations mainly after assimilation of DWR observations (Fig. 5d f). It is also clear that the model-simulated CAPE values (Fig. 5d) gradually increase from the developing stage ( UTC) up to the mature stage ( UTC) mainly in the 3DV simulation. The system gradually dissipates after 1200 UTC as was also observed during

14 1416 Nat Hazards (2014) 74: that time period. The model simulations clearly represent the gradual reduction in CAPE values as time progresses primarily from the mature to the dissipating stage. CAPE values in the 3DV simulation are much higher during the developing and mature stages (around 4,000 J kg 21 ) and lower in the dissipating stage when compared to the CNTL simulation (around 3,500 J kg 21 at the mature stage). Similarly, the model-simulated CIN values (Fig. 5e) gradually decreased from the developing stage and arrived close to 0 J kg 21 during the mature stage of the thunderstorm with a sudden increase in values after the system started to dissipate. Consistent with this evolution, the CIN values are lower in the 3DV simulation as compared to the CNTL simulation. The pattern in the 3DV simulation is well correlated with the evolution of CAPE (Fig. 5d) and is very conducive for thunderstorm occurrence. Further, LI values in the 3DV experiment started to decrease from the developing stage and drop to a minimum (-11) at the mature stage, which is much lower than the critical value (Table 5). The LI values from the 3DV experiment are comparatively less during this period than the LI (-10 at the mature stage) simulated by the CNTL experiment. These thermodynamic parameters indicate that the atmosphere in the 3DV experiment in both cases was realistic and simulated a more unstable environment in comparison with that in the CNTL experiment, which is more conducive for thunderstorm initiation and development. The analysis showed that the assimilation of DWR data thus has a significant positive impact on stability indices and the overall model results pertaining to the dynamical and thermodynamical features associated with the STS life cycle Time evolution of relative humidity and rainfall Figure 6 shows the time evolution of surface relative humidity and rainfall at Kharagpur for Case 1 and Case 2. The simulated relative humidity (Fig. 6a) in the 3DV experiments is in good agreement with that of the observed. Also, the humidity does not reduce in the CNTL after the dissipation of the thunderstorm as compared to the 3DV simulation. The hourly accumulated rainfall (Fig. 6b) for over 24 h starting at 0000 UTC on 11 May 2009 shows that the initiation of the thunderstorm is early by 4 h in both experiments. The break in rainfall between 0900 and 1000 UTC is consistent with the drop in CAPE (Fig. 5a) at 1000 UTC in the CNTL simulation; however, the rainfall shows a continuous trend in the 3DV experiment during that time period. The rainfall amount (15.75 mm) predicted by the 3DV experiment is better than that of CNTL (13.2 mm) when compared to that of the observed (16.8 mm). Similarly, for Case 2, the relative humidity (Fig. 6c) simulated in both experiments is comparatively lower than that of the observations during the thunderstorm hours, but the pattern of diurnal variation of relative humidity in the 3DV experiment is reasonably well matched with the observed pattern when compared to CNTL. It is worth mentioning here that the relative humidity simulated in the 3DV experiment is higher compared to that in the CNTL experiment during the model-simulated thunderstorm hours. Maximum relative humidity attained in both experiments is still considerably less than the observed. In Fig. 6d, comparing the model-simulated rainfall with the observed indicates that both experiments simulated the thunderstorm occurrence 2 3 h earlier than the actual occurrence. However, in the 3DV experiment, the temporal error in thunderstorm initiation is reduced by 1 h. The maximum observed rainfall is 35.3 mm during the thunderstorm activity, and the exact amount of rainfall has not been captured by either simulation. However, the amount of rainfall (25.9 mm) improved after assimilation of DWR observations as compared to the CNTL simulation (20.45 mm). The observed rainfall that occurred on two occasions during the thunderstorm has been featured well in the 3DV

15 Nat Hazards (2014) 74: Fig. 6 Model-simulated diurnal variation of a relative humidity and b 1-h accumulated rainfall for Case 1 at Kharagpur. c d are the same as a, b but for Case 2 simulation, while the CNTL simulation shows three rainfall spells during the period. The rainfall amount predicted in both experiments is less than the observed which may be attributed to a relatively drier atmosphere simulated in both experiments at the time of peak thunderstorm activity Evolution of vertical velocity, equivalent potential temperature, vorticity, and divergence The prediction of the dominant convective mode is based on the assessment of the magnitude of vertical motion which is needed to initiate convection (May and Rajopadhyaya 1999). The convective instability of the atmosphere, in terms of equivalent potential temperature (h e ), can be defined as its decrease with height (Morgan 1992). In addition, strong low-level convergence and upper-level divergence favor the initiation and intensification of thunderstorms. The vertical cross section (time pressure) of vertical velocity and equivalent potential temperature from the CNTL and 3DV experiments during the period 0000 UTC 11 May 2009 to 0000 UTC 12 May 2009 is illustrated in Fig. 7a, b, and c, d for Case 1, respectively. It is clear that the 3DV experiment (Fig. 7b) simulated strong vertical velocity around 100 cm s 21 in the lower and upper atmosphere during the mature stage of the thunderstorm. However, the maximum vertical velocity from the CNTL simulation (Fig. 7a) is 80 cm s 21 between 750 and 500 hpa. Figure 7c and d shows the

16 1418 Nat Hazards (2014) 74: Fig. 7 Time-pressure cross section of vertical velocity (cm s 21 ) for a CNTL and b 3DV at Kharagpur for Case 1. c d are same as a, b but for equivalent potential temperature ( C) time pressure cross section of h e ( C) for Case 1, which indicates high values (100 and 105 C) in the 3DV experiment corresponding to the maximum updrafts (Fig. 7a, b) and initiation of strong vorticity (figure not shown) that is noticed with a comparatively low value of h e in the CNTL experiment. The high h e values support condensation of the available moisture, and the resultant latent heat helps to further increase instability leading to convection (Holton 1994). The high h e simulated by the 3DV experiment could well be due to high mixing ratios (Fig. 9) and stronger updrafts, which confirms the higher potential instability in the 3DV experiment present during thunderstorms days (Tyagi et al. 2013). In Case 1, the vertical cross section of vorticity (figure not shown) indicates a steady increase after 0500 UTC in both experiments. It is also noted that vorticity exhibited by CNTL ( s 21 ) in the lower levels is comparatively weaker than that in 3DV ( s 21 ) during thunderstorm hours ( UTC). Further, the negative vorticity simulated after 1200 UTC in both experiments indicates dissipation of the thunderstorm. Stronger divergence is simulated by the 3DV experiment ( s 21 ) between 400 and 200 hpa against a weaker divergence in CNTL ( s 21 )at 200 hpa. The vertical cross section of vertical velocity and equivalent potential temperature for Case 2 are presented in Fig. 8a, b and c, d. The updraft and downdraft features are well represented in the 3DV simulation. In the 3DV simulation (Fig. 8b), the updraft starts at lower levels extending up to 400 hpa with a maximum value of 100 cm s 21. The maximum downdraft starts at 400 hpa and extends to the lower atmosphere, which triggers the

17 Nat Hazards (2014) 74: Fig. 8 Same as Fig. 7 but for Case 2 convective activity (Mohanty et al. 2012; Vaidya and Kulkarni 2007). However, the strength of the updraft and downdraft is not represented well in the CNTL simulation (Fig. 8a) when compared to the 3DV simulation. Similarly, high equivalent potential temperature is simulated by the 3DV experiment (Fig. 8d) during the thunderstorm period as compared to the CNTL simulation (Fig. 8c). The vertical cross section (time pressure) of vorticity and divergence from both experiments during the period 0000 UTC 12 May 2009 to 0000 UTC 13 May 2009 is also analyzed (figure not shown). The 3DV simulation exhibits strong cyclonic vorticity structures ( s 21 ) extending up to 500 hpa, which are not found in the CNTL simulation (maximum vorticity is s 21 ). In both simulations, the positive vorticity increased gradually from 0500 UTC to a maximum at 1200 UTC on 12 May 2009, after which the vorticity turned negative indicating system dissipation. At the same time, strong convergence is found in the 3DV simulation at a lower atmosphere during the thunderstorm period (from 0500 to 1200 UTC), and strong divergence is also observed around hpa at ( UTC). These features are not well represented in the CNTL simulation Evolution of vertical structure of hydrometeors Hydrometeors play an important role at different stages of convective activities (Mohanty et al. 2012). Therefore, the time evolution of hydrometeor structures over Kharagpur is analyzed in this section. The time pressure cross section of hydrometeors such as water vapor, cloud water, and rainwater mixing ratios obtained from the CNTL and 3DV simulations for Case 1 is illustrated in Fig. 9a f.

18 1420 Nat Hazards (2014) 74: Fig. 9 Time-pressure cross section of water vapor mixing ratio (g kg 21 ) at Kharagpur for Case 1 for a CNTL and b 3DV. Similarly, c d and e f are the same as a b but for cloud water mixing ratio (10 27 gkg 21 ) and rain water mixing ratio (10 27 gkg 21 ) Both experiments simulated similar patterns of the water vapor mixing ratio with a gradual increase in moisture content at the lower levels after 0400 UTC. It is clearly noticed that the maximum water vapor mixing ratio around 20 g kg 21 is found at lower levels in the 3DV simulation (Fig. 9b) as compared to the CNTL simulation (18 g kg 21 ). The maximum water vapor mixing ratio (18 g kg 21 ) in the 3DV experiment is extended up to 850 hpa in comparison with the CNTL (Fig. 9a). The temporal and vertical extensions of the cloud water mixing ratio are greater in the 3DV simulation (Fig. 9d) and are not observed in the CNTL simulation (Fig. 9c). Although the values of the cloud water mixing ratio in both experiments are similar, the extent of the cloud water mixing ratio is much higher reaching 500 hpa compared to 600 hpa in the CNTL simulation. Similar trends of higher vertical extent are seen in the rain water mixing ratio (Fig. 9e f) with 3DV simulating higher values than that in CNTL. Similarly, for Case 2, the pattern of hydrometeors is simulated well in the 3DV experiment as compared to the CNTL (the figure is not shown for brevity). It is clearly found that hydrometeors are relatively stronger in the assimilation experiment after ingestion of DWR observations.

19 Nat Hazards (2014) 74: Rainfall and reflectivity One of the main objectives of assimilation of high-resolution DWR data that contain precipitation hydrometeor information is to improve the model performance for rainfall forecasts associated with convective activities (Abhilash et al. 2007; Xiao and Sun 2007; Routray et al. 2010; Srivastava et al. 2010). A comparison of 24-h accumulated modelsimulated rainfall and observed rainfall at different stations for Case 1 is presented in Table 6. Both experiments under predicted the maximum amount of observed rainfall; however, the rainfall amounts over the stations significantly improved in the 3DV experiment. From Table 6, it can be further seen that the CNTL experiment failed to simulate rainfall over a few stations. The RMSE of simulated rainfall is obtained by comparing the 24-h accumulated model-simulated rainfall with 24-h accumulated observed station rainfall for 9 stations. The RMSE decreased by 4.29 mm in the 3DV experiment as compared to the CNTL simulation as shown in Table 7. Figure 10 depicts the modelsimulated 3-h accumulated rainfall compared with TRMM rainfall during the modelsimulated thunderstorm hours for Case 1. The rainfall during UTC reflects the fact that the thunderstorm initiated much earlier in both experiments over Kharagpur (Fig. 10b, c), which is clearly seen from the TRMM rainfall (Fig. 10a) where the rainband observed is far to the northwest of Kolkata and Kharagpur during this period. During UTC, the maximum rainfall over Bangladesh observed in TRMM does not appear in both experiments because of early initiation of the thunderstorm. However, the rainband over Bangladesh is noticed during UTC. Furthermore, the rainfall pattern in the 3DV experiment is well simulated over the southwest part of the domain indicated by a rectangular box in Fig. 10f in contrast to the CNTL simulation (Fig. 10e). The feature is well correlated with the TRMM rainfall pattern (Fig. 10d). However, the 6-h accumulated rainfall during UTC (figure not shown) was well correlated with Table 6 Comparison between observed and model-simulated rainfall (mm) valid at 0300 UTC of 12 May 2009 Station names Lat ( ) Lon ( ) Observed (mm) CNTL (mm) 3DV (mm) Dum Dum Kharagpur Canning Uluberia Krishnanagar Digha Midnapore Bijanbari Alipore Table 7 RMSE (mm) of model-predicted station rainfall with that of the observed station rainfall CNTL 3DV Case 1 (11 May 2009) Case 2 (12 May 2009)

20 1422 Nat Hazards (2014) 74: Fig. 10 Model-simulated 3-h accumulated rainfall (mm) for Case 1 during 06 to 09 UTC for a TRMM, b CNTL, and c 3DV. d f are same as a b but during 09 to 12 UTC 11 May The location of Kolkata and Kharagpur is marked with a black dot and black square, respectively Table 8 Comparison between observed and model-simulated rainfall (mm) valid at 0300 UTC on 13 May 2009 Station names Lat ( ) Lon ( ) Observed (mm) CNTL (mm) 3DV (mm) Uluberia Midnapore Kharagpur Digha Canning Alipore Bijanbari Singlabazaar Dum Dum the TRMM rainfall. Corresponding to Case 2, the model-simulated and IMD-observed rainfall at different locales (Table 8) highlight that the rainfall amount significantly improved in the 3DV experiment as compared to the CNTL simulation. The RMSE of the rainfall is comparatively less in the 3DV simulation as compared to CNTL (Table 7), which was reduced by 3.1 mm after assimilation of the DWR observations. Figure 11 shows the comparison of 3-h accumulated ( and UTC) modelsimulated and satellite-derived TRMM observed rainfall on 12 May The spatial distribution and amount of rainfall (within the rectangular box in the figure) are well represented in the 3DV experiment (Fig. 11c and f) in contrast to the CNTL simulation (Fig. 11b and e) at UTC as well as UTC. The rainfall pattern obtained from 3DV is well correlated with TRMM (Fig. 11a, d). Further, to verify the skill of the simulated precipitation, equitable threat scores (ETS; Jankov et al. 2005) for different thresholds are calculated for 6-h accumulated rainfall for Case 1 and Case 2 (Fig. 12a, b). The skill scores show that 3DV exhibited higher skill in predicting rainfall at

21 Nat Hazards (2014) 74: Fig. 11 Same as Fig. 10 but for Case 2 Fig. 12 Equitable threat scores for 6-h accumulated rainfall for a Case 1 and b Case 2 all thresholds and that the skill dropped considerably after the 10-mm threshold in Case 1 and 20-mm threshold in Case 2. Composite radar reflectivity fields generated from model output have become a popular method to diagnose various convective scenarios. The main advantage of the model reflectivity product is that it facilitates for effortless visualization of mesoscale and nearstorm scale structures in detail, such as the structure of deep convection, movement of squall lines, and frontal precipitation bands (Koch et al. 2005). The observed and modelsimulated reflectivity is presented in Fig. 13 for Case 2. It can be observed from the figure that both experiments predicted the thunderstorm over Kharagpur a few hours prior to that actually observed. In order to identify the locales of thunderstorms over Kharagpur and Kolkata, a black square and circle have been drawn on each of the images. The initiation and passage of the thunderstorm can be seen in the observed and model-simulated reflectivity images. The observed reflectivity indicated that the thunderstorm initiated at 0600 UTC reached an intense stage at 0800 UTC and gradually dissipated. In the CNTL simulation, it is noticed that the thunderstorm initiated around 0400 UTC and intensified at 0500 UTC over Kharagpur approximately 2 3 h prior to the observations. However, in the 3DV experiment, the thunderstorm initiation took place at 0500 UTC and intensified at 0600 UTC. The spatial distribution and timing improved in the 3DV simulation over the

22 1424 Nat Hazards (2014) 74: Fig. 13 Reflectivity a observed from Kolkata DWR, b CNTL, and c 3DV at 5 UTC 12 May 2009 for Case 2. d f, g i, and j l are the same as a c but at 6 UTC, 7 UTC, and 8 UTC on the same day, respectively region as compared to the CNTL. The southward movement of the thunderstorm reflectivity was well captured in both experiments. The initiation and intensification times of the thunderstorm match well with the accumulated rainfall (Fig. 11), reducing the temporal error in thunderstorm initiation by 1 h after assimilation of the DWR observations. 6 Conclusions The mesoscale model WRF ARW and 3DVAR data assimilation system with a single domain of 3-km horizontal resolution are used to simulate two thunderstorm events that occurred during the STORM pilot phase 2009 over GWB. For this purpose, two numerical experiments (CNTL without assimilation, and 3DV with assimilation of DWR data) are conducted to evaluate the impact of DWR observations on STS simulations. From the above results and discussion, the following broad conclusions can be put forward. It is clearly noticed that the 3DV simulation after assimilating the DWR observations improves the dynamic and thermodynamic features of the thunderstorm as compared to the CNTL simulation. The improvement of wind and moisture in the boundary layer at model initial time with DWR data led to more moisture incursion and higher instability and therefore stronger convective activity. The instability indices and equivalent potential temperature are well captured by the 3DV experiment in comparison with that in the CNTL experiment. The larger CAPE values in the 3DV experiment could be attributed to higher moisture convergence aided by stronger moisture-laden winds in the lower atmosphere. The 3DV simulation well represented the different stages of thunderstorm evolution as compared to the CNTL run.

23 Nat Hazards (2014) 74: The 3DV experiment produced reasonably good simulations of the precipitation in terms of intensity, location, and spatial distribution over the GWB and adjoining areas, well supported by the simulated reflectivity. The diurnal variation of rainfall and relative humidity are also well produced in the 3DV simulations, and the statistical skill scores have significantly improved. The temporal error in thunderstorm initiation is significantly reduced by 1 h in Case 2 after assimilation of DWR observations, but remained unchanged in Case 1, which would be taken up for further study. In addition, more cases need to be studied to arrive at a stronger conclusion regarding the impact of DWR reflectivity and radial velocity on the simulation of STS. This study clearly demonstrates the positive impact of DWR data assimilation through a 3DVAR analysis system for the simulation of STS. Also, at 3-km resolution, the WRF ARW model simulates certain features of thunderstorms which are brought out much better when DWR data are assimilated because the meso-convective system (MCS) associated with thunderstorm is better represented after the assimilation of DWR observations. At the same time, it should also be noted that the assimilation experiment has fallen short in simulating some aspects of the thunderstorm such as exact time of evolution and rainfall intensity. So, the 3-km resolution may not be an optimum one for simulation of such strong convective systems as increasing the resolution to a much finer scale could resolve the features of the thunderstorm much better. In addition to this, further improvement could be obtained by proper representation of land surfaces in the model (Chang et al. 2009; Niyogi et al. 2006) and with the use of extensive quality-controlled multi-radar radial velocity and reflectivity through advanced data assimilation techniques such as EnKF and 4DVAR with flow-dependent background error covariance (Zhang et al. 2009, 2011). Acknowledgments This work has been carried out with financial aid from the Department of Science and Technology (DST), Govt. of India, and is duly acknowledged here. DN benefitted from NSF CAREER grant (AGS ; Dr. A. Bamzai). Authors acknowledge the use of WRF ARW and 3DVAR developed by the National Center for Atmospheric Research (NCAR), USA. The use of TRMM products is also duly acknowledged. The authors acknowledge IMD for providing the necessary datasets used in the assimilation and validation of the results. Authors sincerely thank Dallas Staley for her outstanding editing of the manuscript. Further the authors extend their sincere thanks to anonymous reviewers for their valuable comments and suggestions in improving the manuscript. References Abhilash S, Das S, Kalsi SR, Das Gupta M, Mohankumar K, George JP, Banerjee SK, Thampi SB, Pradhan D (2007) Impact of Doppler radar wind in simulating the intensity and propagation of rainbands associated with mesoscale convective complexes using MM5-3DVAR System. Pure appl Geophys 164: Air Weather Service (AWS) Technical Report 79/006 (1990) The use of the Skew T, Log P diagram in analysis and forecasting, air weather service. Scott AFB, Illinois Anthes RA (1976) Numerical prediction of severe storms certainty, possibility, or dream? Bull Am Meteorol Soc 57: Barker DM, Huang W, Guo YR, Bourgeois A, Xiao XN (2004) A three-dimensional variational data assimilation system for MM5: implementation and initial results. Mon Weather Rev 132: Brooks HE, Wilhelmson RB (1992) Numerical simulation of a low-precipitation supercell thunderstorm. Meteorol Atmos Phys 49:3 17 Chang H, Kumar A, Niyogi D, Mohanty UC, Chen F, Dudhia J (2009) The role of land surface processes on the mesoscale simulation of the July 26, 2005 heavy rain event over Mumbai, India. Glob Planet Change 67: Chaudhuri S, Middey A (2009) The applicability of bipartite graph model for thunderstorms forecast over Kolkata. Adv Meteorol. doi: /2009/270530