Impact of background error statistics on forecasting of tropical cyclones over the north Indian Ocean

Authors


Abstract

[1] While the quality of background error statistics (BES) is recognized as one of the key components of assimilation, considerable uncertainties exist in prescribing BES, especially since the prescription and impact of BES can also depend on the weather regime and geographical location. In this backdrop, it is necessary to quantify the impact of different BES for particular weather systems; this is particularly true for cyclones over the north Indian Ocean which have characteristics different from those over the Atlantic and the Pacific. The objective of this work is to assess the relative improvement in forecasting tropical cyclone track and intensity due to different BES. We have used global BES (GBES), computed from global model forecasts for 357 cases distributed over a period of one year and regional BES (RBES), generated from short-range forecasts with Weather Research and Forecasting (WRF) model for a 30 day period. From a series of assimilation experiments using the WRF three-dimensional variational (3D-Var) data assimilation system with different BES, and a number of parameters to quantify the impact of BES, it is shown that the use of RBES in WRF 3D-Var significantly improves prediction of track as compared to simulations with no assimilation or GBES. Further, the skill with RBES is comparable with, or better than, many operational skills, although a strict comparison is difficult due to differences in the events and the basins. While parallel and significant efforts are needed for the formulation and incorporation of BES in assimilation systems in general, this study quantifies relative advantages of using RBES in forecasting cyclones over the Indian Ocean.

1. Introduction

[2] While there has been significant progress in modeling and forecasting of cyclones over the past decades, the increasing socio-economic impact of these events also demands longer range and higher accuracy [Anderson, 1996; AMS Council, 2000; Bengtsson, 2001]. Improving forecast of track and intensity of tropical cyclones thus continues to be one of the primary concerns of the scientific community. However, while the forecast skill has significantly improved over the last few decades over most ocean basins, the progress over the north Indian Ocean has been relatively low [Frank, 1987; Vitart and Stockdale, 2001; Camargo and Zebiak, 2002; Vitart, 2006; Singh et al., 2008; Deb et al., 2010; Mukhopadhyay et al., 2011; Goswami et al., 2011]. While the dynamical forecasting techniques such as meso-scale model and data assimilation are quite generic, there are issues that need a regional approach for improving forecast skill; the choice of background error statistics (BES) in data assimilation is one such.

[3] BES is an integral component in assimilation of data in numerical weather prediction (NWP). It is well known that a major source of error, especially for short-range forecasting, is the initial state prescribed for model integration; to a certain extent this can be addressed through data assimilation. The effectiveness of data assimilation in NWP has been demonstrated in many studies [Harasti et al., 2004; Chou et al., 2006; Wu et al., 2006; Chen, 2007; Zhang et al., 2007; Zapotocny et al., 2007; Kelly et al., 2008; Brennan et al., 2009]. Effective assimilation, however, requires careful consideration of several issues; a major challenge is the BES which determines the amount and propagation of observation information in data assimilation system. The estimation of BES and its significance in assimilation of observations in global and regional models have been addressed in a number of studies [Parrish and Derber, 1992; Wu et al., 2002; Fisher, 2003; Guo et al., 2005; Pereira and Berre, 2006; Lindskog et al., 2006; Daget et al., 2009; Michel and Auligne, 2010]. Over the years, a number of techniques have emerged to generate BES using various approaches [Široká et al., 2003; Berre et al., 2006; Daget et al., 2009].

[4] Although the effectiveness and impact of assimilation of various observational data in forecasts over the Indian region have been addressed in a number of studies [Sandeep et al., 2006; Singh et al., 2008; Vinodkumar et al., 2008; Rakesh et al., 2009a, 2009b; Routray et al., 2010; Deb et al., 2010], very few studies have addressed the design of data assimilation system. While application of BES in itself remains a challenge that requires theoretical as well as methodological developments, a primary requirement is to quantitatively assess the impact of BES for regional weather systems. It is expected regional BES (RBES) to produce better forecast; however, a quantitative assessment of RBES against global BES (GBES) is missing, especially over the Indian region.

[5] Differences in RBES and GBES can come from a number of sources, such as the basic data (forecasts) to compute the BES. In the present study, GBES is computed from global model forecasts for 357 cases distributed over a period of one year. In contrast, RBES is domain specific and is generated from Weather Research and Forecasting (WRF) short-range forecasts for a 30 day period. The objective of this study is to quantitatively evaluate impact of the BES in forecasting cyclones over the north Indian Ocean. Though it is recognized that results from application of RBES over the Indian domain will not in general be applicable to other regions, it is expected that our study will contribute to answer the question of effectiveness of assimilation with RBES. We show, based on hindcasts of 10 tropical storms and cyclones over the north Indian Ocean, that inclusion of RBES can significantly improve skill in forecasting tropical cyclone track over the region.

[6] In section 2, we provide the details of design of the experiments along with a synoptic description of the selected events; this section also outlines the model configuration, computation of BES and the verification methodology. Results are described and discussed in section 3 while section 4 contains our major conclusions.

2. Design of the Experiments

2.1. Description of the Cyclone Events

[7] We have selected eight tropical cyclones over the Indian Ocean in addition to two special cyclones. The eight cases were selected (Table 1) to represent a range of intensities, tracks, places of origin (Bay of Bengal and Arabian Sea) and seasons (pre-monsoon, monsoon and post-monsoon periods). The satellite images (Figure 1) show the locations of the eight cyclones and the movement of cloud masses associated to them. The surface wind speed and direction from Quick Scatterometer (QSCAT) observation at the intensified stages of the cyclones show (Figure 2) the high (Figures 2a, 2b, and 2c) and low (Figures 2g and 2h) intensity cyclones while the other events have intermediate intensity. The maximum wind speed and minimum central sea level pressure (CSLP) associated with the events from Joint Typhoon Warning Center (JTWC) (Figure 2) also shows that the cyclones formed in June 2007, November 2007, and April 2006 (Figures 2a, 2b, and 2c) were of very high intensities whereas those formed in September 2006 and December 2005 (Figures 2g and 2h) were of much lower intensity. Since the satellite wind observations were not available over the land areas and due to data gap, we have shown wind streamlines and mean sea level pressure observations from National Centers for Environmental Prediction (NCEP) analysis (Figure 3). The characteristic features of strong cyclonic storms such as pressure drop (see Figure 15) and strong cyclonic circulations are clearly visible for June 2007, November 2007, April 2006, and May 2008 cyclone cases (Figures 3a, 3b, 3c, and 3d).

Figure 1.

Satellite images of the eight tropical cyclones over the north Indian Ocean selected in this study. The asterisk in each image shows the central location of the cyclone obtained from JTWC.

Figure 2.

Observed wind speed (m/s; shaded region) and vector wind at 2 m level for different cyclones considered in this study. The time refers to the intensified stage of the respective cyclone. The plots represent QSCAT passes averaged over ±3 h time window. The numbers in inset show the lowest pressure (hPa) and maximum wind speed (m/s) associated with the cyclone from JTWC. The cyclones are shown in decreasing order of their intensity.

Figure 3.

Surface pressure (hPa; shaded region) and 850 hPa streamline from NCEP analysis for different cyclones considered in this study.

Table 1. Details of Different Experiments With the Initial Central Locations of the Cyclones in the Analyses for the Eight Casesa
CycloneInitialization Time (UTC)Initial Central Latitude for CNT, GBES, RBES (°N)Initial Central Longitude for CNT, GBES, RBES (°E)Maximum Surface Wind Speed for CNT, GBES, RBES (m/s)Minimum Sea Level Pressure for CNT, GBES, RBES (hPa)
  • a

    The numbers in brackets show the observed values. The analysis value closest to the observation is shown in bold.

Super cyclonic storm – GONU (Arabian Sea)1200, 02 June 200714.5,14.8, 15.1 (15.4)66.5, 67, 67.1 (67.1)16.3, 17.9, 17 (28.05)994.7, 994.9, 994.7 (982)
Very Severe cyclonic storm- SIDR (Bay of Bengal)1200, 11 Nov 200710.65, 9.98, 10.12 (10.2)91.81, 91.5, 91.6 (91.9)16.3, 15.3, 16.7 (22.95)1000.8, 1001, 1000.2 (989)
Very Severe cyclonic storm- MALA (Bay of Bengal)0000, 26 April 200610.69, 11.11, 10.67 (10.3)89.56, 89.56, 89.66 (90.3)15.4, 16.7, 15.5 (22.9)1002.5, 1002.3, 1001.7 (991)
Severe cyclonic storm- NARGIS (Bay of Bengal)0000, 30 April 200814.2, 15.4, 14.64 (14.4)86.13, 86.71, 86.39 (86.7)14, 17.4, 16.2 (35.7)1001.7, 1002.1, 1000.2 (970)
Severe cyclonic storm (Arabian Sea)1200, 12 Nov 20035.85, 5.76, 5.93 (6.1)59.85, 59.86, 59.86 (60.0)16.9, 18.9, 15.1 (15.3)1006.8, 1005.4, 1005.6 (1000)
Severe cyclonic storm – AILA (Bay of Bengal)0000, 23 May 200916.17, 6.93,16.45 (16.7)88.23, 87.34, 88.45 (88.8)18.9, 20.5, 17.7 (12.7)998.1, 997.7, 998.2 (1004)
Severe cyclonic storm – MUKDA (Arabian Sea)1200, 21 Sep 200619.83, 20.01, 20.01 (19.7)66.18, 66.18, 66.2 (66.3)22.3, 23.9, 19.1 (17.8)998.5, 998.9, 998.8 (997)
Cyclonic storm (Bay of Bengal)0000, 07 Dec 20059.5, 9.37, 9.46 (10.5)87.53, 87.24, 87.23 (86.6)14.5, 15, 14.9 (28.05)1001.6, 1002.1, 1001.1 (984)

[8] In addition to these eight cyclone events, we selected two special cases (Table 2) for studying the impact of BES. The cyclone formed over the Bay of Bengal in May 2004 (Figure 4a) is considered as a special case due to its looping track that is generally harder to predict. The other special case considered is a cyclone formed over the Arabian Sea in November 2004 due to its origin very near to the equator (Figure 4b), a location where tropical cyclones rarely occur.

Figure 4.

Satellite images of the two special cyclone cases (a, b) over the Indian Ocean; the asterisk in each figure shows the central location of the cyclone obtained from JTWC.

Table 2. Details of Different Experiments With the Initial Central Locations of the Cyclones in the Analyses for the Two Special Casesa
Special Cyclone CasesInitialization Time (UTC)Initial Latitude for CNT, GBES, RBES (°N)Initial Longitude for CNT, GBES, RBES (°E)Maximum Surface Wind Speed for CNT, GBES, RBES (m/s)Minimum Sea Level Pressure for CNT, GBES, RBES (hPa)
  • a

    The numbers in brackets show the observed values for the corresponding time. The analysis value closest to the observation is shown in bold.

Cyclonic storm (Bay of Bengal)1200, 16 May 200417.28, 16.57, 17.28 (18)94.3, 93.56, 93.81 (91.1)14.8, 14.5, 14.4 (12.7)994.6, 994.6, 995 (1002)
Severe cyclonic storm (Arabian Sea)1200, 28 Nov 20041.8, 1.44, 1.43 (1.1)68.4, 68.5, 68.2 (67.5)14.19, 16.99, 15.89 (22.95)1007.5, 1007.5, 1007.2 (991)

2.2. Model Configuration

[9] The forecast model WRF [Skamarock et al., 2005] used here is a limited area, non-hydrostatic, primitive equation model with multiple options for various physical parameterization schemes. There are two dynamics solvers in the WRF software framework: the Advanced Research WRF (ARW) solver developed primarily at National Center for Atmospheric Research (NCAR), and the Non-hydrostatic Mesoscale Model (NMM) solver developed at NCEP. We have used ARW dynamic solver for the present study. This version employs a fully compressible system of equations on an Arakawa C-grid staggering for the horizontal grid. The terrain following hydrostatic pressure coordinate with grid stretching was employed in vertical. The time-split integration uses 3rd order Runge-Kutta scheme with a smaller time step for acoustic and gravity wave modes. The WRF physics options used in this study consisted of the WRF Single Moment 6 - class graupel scheme for microphysics (WSM6), which is similar to that used by Lin et al. [1983]; the new Kain-Fritsch [Kain, 2004] cumulus convection parameterization scheme; and the Yonsei University (YSU) planetary boundary layer scheme [Hong and Dudhia, 2003]. The Rapid Radiative Transfer Model (RRTM); [Mlawer et al., 1997] and Dudhia scheme [Dudhia, 1989] were used for long-wave and short-wave radiation, respectively. All experiments were conducted with two nested domains (Figure 5) consisting of 206 × 206 (36 km horizontal grid resolution in x and y direction) and 211 × 211 (12 km horizontal grid resolution in x and y direction) grid points. The number of vertical levels used is 34, with the top of the model atmosphere located at 50 hPa.

Figure 5.

Model domains deployed for numerical simulations of the 10 cyclones formed over (a) the Arabian Sea and (b) the Bay of Bengal.

2.3. Initial and Boundary Conditions

[10] The initial fields were extracted from NCEP three dimensional final (FNL) analyses with 1° × 1° resolution. However, instead of directly using the NCEP FNL analysis, 6-h WRF forecasts initialized using the NCEP FNL analysis were used as the First Guess (FG) for all the experiments. We have used two different BES while making the initial conditions for the assimilation experiments. The BES determines how the variational assimilation system corrects the meteorological fields from the differences between observed and forecasted variables. As mentioned earlier, prescription of BES remains a critical issue in data assimilation with considerable uncertainties. The two types of BES used are GBES and RBES. Both the GBES and RBES are generated using the NMC (named for National Meteorological Center USA, now called the NCEP) method [Parrish and Derber, 1992; Wu et al., 2002]. The generation of RBES and GBES differ among themselves as follows. The GBES used is from NCEP available in-built with the WRF three-dimensional variational (3D-Var) assimilation system [Skamarock et al., 2005] which can be used for any regional domain. The computation of GBES from NMC method can be formulated as below

equation image

where xt2 = 48 hr, xt1 = 24 hr forecasts for global models and xt2 = 24 hr, xt1 = 12 hr forecasts for regional models. Here Pf is the background error covariance matrix and A is the tuning parameter. The overbar denotes the mean forecast error, with x′T representing transpose of x′. The GBES used in this study is estimated from the differences of 24 and 48-h NCEP Global Forecasting System (GFS) forecasts with T170 resolution valid at the same time for 357 cases distributed over a period of one year.

[11] Unlike GBES, the RBES is computed from the WRF model forecasts. The WRF model perturbations are computed as the difference between 12 and 24-h forecasts valid at the same time for 30 days period for each cyclone case (±15 days from the intensified stage of the cyclone) and domain dependent RBES is computed by averaging such forecast differences for the 30 days period for each cyclone case. The initial conditions for these WRF forecasts are obtained from global GFS analysis. The BES are generated for the domain consisting of 206 × 206 points in x and y directions with 36 km horizontal grid resolution (Figure 5). All the WRF forecasts are initialized twice a day at 0000 UTC and 1200 UTC. The initial conditions for the assimilation experiments differ among themselves in the use of different BES in the assimilation system: the first assimilation experiment was carried out with GBES while the second one was carried out using RBES.

[12] The surface boundary conditions such as model topography, land use, soil types, and monthly vegetation fraction were derived for the outer domain from the 10-min (∼19 km) data sets from United States Geological Survey (USGS). The interpolated 6-hourly NCEP FNL analysis was used for the model boundary conditions for all the experiments. The initial, lower and lateral boundary conditions for the inner domain were obtained by interpolating the fields from the outer domain.

2.4. Design of the Experiments

[13] The assimilation was carried out using the WRF 3D-Var assimilation scheme. The WRF 3D-Var evolved from the MM5 3D-Var system [Barker et al. 2004], but the basic software interface and coordinate framework were fully updated for the WRF model. The integration time in all the experiments was 78 h to allow a forecast lead time comparable to operational lead time. A set of three experiments were conducted for each cyclone: one control (CNT; no assimilation) and two assimilation experiments (GBES and RBES) that differ in their prescription of BES. The same set of observations was assimilated in both GBES and RBES, consisting of radiosonde, buoy, ship, and satellite observations over the land and ocean. The conventional data include surface station reports as well as upper air observations. The University of Wyoming (http://weather.uwyo.edu) provides surface and upper air data over the land, while the surface marine observations (ships and buoys) were taken from ICOADS (International Comprehensive Ocean-Atmosphere Data Set). National Climate Data Centre provided ICOADS Data (www.ncdc.noaa.gov/oa/marine.html). The data from radiosondes and SHIPs were first subjected to quality control; observations with absolute differences more than 5 from the first guess (FG) were not included in the assimilation system. The satellite data assimilated were QSCAT surface wind speed and direction along with spectral sensor microwave imager (SSM/I) observed total precipitable water (TPW) (www.ssmi.com). A large number of studies [Xiao et al., 2000; Chen et al., 2004; Chelton et al., 2006; Chen, 2007; Zhang et al., 2007; Kelly et al., 2008; Singh et al., 2008; Rakesh et al., 2009b; Brennan et al., 2009] in recent past showed the positive impact of assimilation of these data in cyclone simulations. All satellite data were also subjected to quality-checking processes in order to reduce the possibility of assimilating bad observations. First, rain-contaminated data were excluded from QSCAT and SSM/I data. The rainfall probability parameter ‘p’ along with QSCAT wind product is used to exclude the rain-contaminated observations from the QSCAT winds. The SSM/I retrievals were already flagged for the rain-contaminated pixels by SSM/I data products generation team. Second, a gross error quality control was performed in which observations (QSCAT and SSM/I) that differed from the model FG by more than five times the observational errors were removed. We have used ±1 h window to assimilate the QSCAT and SSM/I observations for all the experiments.

2.5. Verification Methodology

[14] The simulated track, intensity and CSLP from different experiments were verified against observations from JTWC. The JTWC maintains an archive of data on tropical cyclone tracks, commonly referred to as “best-tracks.” Each best-track file provides locations of center and intensities (i.e., CSLP and the maximum 1-min mean sustained 10-m wind speed) of a cyclone at six-hour intervals. To construct this database JTWC uses all possible observational networks such as ship, buoy, radiosonde, surface observations, aircraft reconnaissance, radar network, Meteorological satellites, GPS dropsondes and techniques like Dvorak technique, McIDAS and other interactive systems etc. The cyclone's center is identified by locating the center of streamlines at 950 hPa from model outputs at every six hourly interval for getting the simulated tracks. Maximum value of wind speed at 10 m height within a radius of 300 km from cyclone's center is considered as the simulated intensity in terms of wind strength for different experiments while verifying against JTWC observations. Similarly, minimum value of CSLP at 950 hPa is considered as the intensity in terms of pressure drop.

3. Results and Discussion

3.1. Impact of BES on the Analysis

[15] The variations in the simulations in our case arise essentially due to differences in the initial conditions from different BES; we have therefore first examined the analyses from RBES and GBES. The positions of the initial vortex, maximum surface wind speed, and minimum sea level pressure from the analysis of CNT, GBES and RBES with the corresponding observations for the eight cyclones show (Table 1) clear differences among themselves; the analysis with RBES is closer to the observation for a greater number of cases. Initial wind speed at 950 hPa from the analyses using GBES and RBES for different cyclones also shows significant differences, particularly near the cyclone center, for most of the cases (Figures 6a6h). A reduction in wind speed is observed due to the use of RBES as compared to GBES in most of the domain and especially near the cyclone's center for most of the cases. It may be noted that significant increase in wind speed (∼3 m/s) takes place over some pockets over the southern part of the Bay of Bengal due to the use of RBES in some of the cyclone cases. The differences in initial relative humidity at 950 hPa due to the use of GBES and RBES for different cyclones are shown in Figure 7. The use of RBES significantly reduced the relative humidity as compared to GBES, particularly over the equatorial Indian Ocean south of the Indian subcontinent (Figures 7a7h). The vertical variations of wind speed at selected locations near the center of the respective cyclones (Figure 8) also show significant differences between the analyses using GBES and RBES. The magnitude of wind speed is significantly larger with GBES than with RBES at most vertical levels for greater number of cases (Figures 8a8h). In contrast, the magnitude of relative humidity is comparatively smaller in the lower levels in the analysis using GBES as compared to that using RBES for most of the cases (Figures 9a9h). The analyses from different experiments also differ among themselves for the two special cyclone cases (Table 2) in terms of spatial distribution and vertical variation (not shown).

Figure 6.

Difference in initial wind speed (m/s) at 950 hPa with assimilation using GBES and RBES for the eight cyclone cases. The asterisk shows the observed location of the center of the respective cyclone.

Figure 7.

Difference in initial relative humidity (%) at 950 hPa with assimilation using GBES and RBES for the eight cyclone cases. The asterisk shows the observed location of the center of the respective cyclone.

Figure 8.

Difference in initial wind speed (m/s) profiles between the assimilation experiments which used GBES (black line) and RBES (gray line with symbol) at a nearby point to the cyclone's center for the eight cyclone cases.

Figure 9.

Difference in initial relative humidity (%) profiles between the assimilation experiments which used GBES (black line) and RBES (gray line with symbol) at a nearby point to the cyclone's center for the eight cyclone cases.

3.2. Impact of BES on the Model Forecast

[16] In this section, we analyze how the use of different BES in WRF 3D-Var assimilation system impacted the WRF forecast. The control forecast (CNT; no assimilation) is used as a baseline for assessing the impact of BES in assimilation experiments.

[17] A comparison of observed (JTWC: OBS) and the simulated tracks from different experiments for the eight cyclones shows (Figure 10) the tracks simulated with assimilation (GBES and RBES) to be closer to the observation than those with no assimilation (CNT) for most of the cases (Figures 10a10h). Among the two assimilation experiments, RBES reproduced tracks closer to the observed tracks than those with GBES. For a more quantitative comparison, we have considered error in track prediction as a function of forecast hour (Figure 11). It is clear that use of RBES resulted in significant reduction of track error compared to both GBES and CNT for all the cyclone cases, particularly at longer leads (Figures 11a11h). The average errors (numbers in brackets in Figures 11a11h) over all the forecast intervals are also significantly lesser with RBES as compared to GBES and CNT; the error with GBES is even higher than CNT in some of the cases.

Figure 10.

Observed (OBS; best track from JTWC) and simulated tracks for 78 h forecast period for the eight cyclones. The positions of the cyclone center are given every 6 h.

Figure 11.

Absolute track errors (with respect to best tracks from JTWC) at different forecast intervals for the eight cyclones: (a) June 2007, (b) November 2007, (c) April 2006, (d) May 2008, (e) November 2003, and (f) May 2009. The number in brackets shows the average value over all forecast intervals for the respective case.

[18] We then examine the impact of BES on prediction of tracks for the two special cyclones. The looping nature of the track of the cyclone formed over the Bay of Bengal during May 2004 could not be captured well by any of the simulations (Figure 12a). Although the simulation with RBES succeeded in simulating the looping structure, it failed to reproduce the observed track; the failure can perhaps be attributed to large errors (more than 300 km) in the initial conditions (Table 2). Comparison of absolute errors in track prediction by different experiments (Figure 13a) at different forecast intervals shows that errors are slightly lesser with RBES as compared to CNT and GBES. For the cyclone formed over the Arabian Sea in November 2004 with its origin very close to equator, the simulations could reproduce the observed track satisfactorily (Figure 12b); however, the absolute errors are comparatively less with assimilation (GBES and RBES) compared to CNT (Figure 13b).

Figure 12.

Observed (OBS; best track from JTWC) and simulated tracks for 78 h forecast period for the two special cyclone cases: (a) the cyclone formed over the Bay of Bengal in May 2004 and (b) the cyclone formed over the Arabian Sea in November 2004. The positions of the cyclone center are given every 6 h.

Figure 13.

Absolute track errors in different experiments as compared to best track from JTWC for the two special cyclone cases: (a) the cyclone formed over the Bay of Bengal in May 2004 and (b) the cyclone formed over the Arabian Sea in November 2004. The number in brackets shows the average value over all forecast intervals for the respective case.

[19] Intensity of the simulated cyclone was computed in terms of wind strength (maximum wind speed) as well as pressure drop (minimum CSLP) for the inner domain (shaded region in Figure 5) at 6 h forecast intervals. A comparison of observed and simulated wind speeds for the eight cyclones shows (Figure 14) the predicted intensities to be lower than the observed intensity at most of the forecast intervals in six out of eight cases. However, once again, simulations with RBES are found to be closer to the observation in most of the cases (Figures 14a14h). The simulations also failed to reproduce the sharp drop in central pressure in four out of eight cases; however, use of RBES resulted in improvement, though small, in minimum CSLP prediction on majority of the occasions (Figure 15). Similar conclusions also hold for the two special cyclones (Figure 16). For the cyclone that formed over the Bay of Bengal in May 2004, the intensification is not captured well by any of the assimilation methods; besides, the control experiment (no assimilation) simulated the intensification of the cyclone 24 h ahead of observation (Figures 16a and 16b). For the second special case, all the experiments underestimated the intensity during initial hours (up to 36 h) and overestimated in the later hours (after 36 h) (Figures 16c and 16d) of simulation; however, the error is relatively smaller with RBES.

Figure 14.

Observed and predicted maximum wind speed (m/s) at 10 m level for the eight cyclone cases at different forecast intervals. (a) June 2007, (b) November 2007, (c) April 2006, (d) May 2008, (e) November 2003, (f) May 2009, (g) September 2006, and (h) December 2005.

Figure 15.

Observed and predicted minimum CSLP (hPa) for the eight cyclone cases at different forecast intervals. (a) June 2007, (b) November 2007, (c) April 2006, (d) May 2008, (e) November 2003, (f) May 2009, (g) September 2006, and (h) December 2005.

Figure 16.

Observed and predicted maximum wind speed and minimum CSLP for the two special cyclone cases at different forecast intervals: (a and b) the cyclone formed over the Bay of Bengal in May 2004 and (c and d) the cyclone formed over the Arabian Sea in November 2004.

[20] A more comprehensive assessment of skill in prediction of track and intensity is provided by absolute errors averaged over all the events at different forecast times (Figures 17a, 17b, and 17c). A comparison of errors in prediction of track at different forecast intervals averaged for all the 10 cyclone cases shows (Figure 17a) RBES to perform better; while the errors in track prediction with CNT and GBES are comparable, errors with RBES are considerably smaller for all the forecast intervals. Track errors in day1 (0–24 h prediction), day2 (24–48 h prediction) and day3 (48–72 h prediction) predictions are also found to be significantly smaller with RBES as compared to CNT and GBES for most of the cyclone cases (Table 3). Similar conclusions about superiority of RBES also hold for prediction of maximum wind speed (Figure 17b) and CSLP (Figure 17c) considered when the relative errors (defined as mean absolute error divided by absolute values of observation) averaged over all the 10 cyclones at different forecast hours. Relative errors in day1, day2, and day3 predictions of maximum wind speed and minimum CSLP for different experiments (Tables 4 and 5, respectively) also show the errors to be smallest with RBES. It is clear that RBES provides the least errors in intensity prediction at all the forecast intervals. On the other hand, the assimilation experiment using GBES showed larger error in intensity prediction than CNT experiment at some of the forecast intervals.

Figure 17.

Average errors in forecast of (a) track, (b) maximum wind speed, and (c) minimum CSLP at different forecast intervals for the 10 cyclone cases. The numbers in brackets show the values averaged over all forecast intervals.

Table 3. Absolute Track Forecast Errors (Km) From Different Experiments for the 10 Cyclone Cases
ExperimentsAbsolute Track Errors (Km) for Forecast Days
Day1Day2Day3
No AssimilationGBESRBESNo AssimilationGBESRBESNo AssimilationGBESRBES
Case 1150193120383450266390380194
Case 2144168112181209112158172101
Case 38740443543241065631
Case 4577145106907611213346
Case 51488160110484517810379
Case 65811675811167010016067
Case 7101136691049931300222141
Case 810660611283464160162107
Case 9 (Special Case 1)448407388446484400422437350
Case 10 (Special Case 2)1081045219111010612610695
Average140137102176168121205188121
Table 4. Relative Errors (%) in the Forecasts of 10 m Maximum Wind Speed From Different Experiments for the 10 Cyclones Studied
ExperimentsRelative Errors × 100 in Maximum Wind Speed
Day1Day2Day3
No AssimilationGBESRBESNo AssimilationGBESRBESNo AssimilationGBESRBES
Case 1471722401665391320
Case 2343023293326252321
Case 31871623101715410
Case 4192060192656181638
Case 5283927355637334225
Case 6273629364134266362
Case 7442136321742212019
Case 8225741256362195418
Case 9 (Special Case 1)1756381620465537
Case 10 (Special Case 1)442347481937451326
Average303025313022344224
Table 5. Relative Errors (%) in Minimum CSLP From Different Experiments for the 10 Cyclones Studied
ExperimentsRelative Errors × 1000 in Minimum CSLP
Day1Day2Day3
No AssimilationGBESRBESNo AssimilationGBESRBESNo AssimilationGBESRBES
Case 1232222191113633
Case 2171715171815141614
Case 367647310127
Case 433355412127
Case 5222220616142414124
Case 6303430364034212720
Case 722191813141410175
Case 8466272826282820
Case 9 (Special Case 1)42125245124
Case 10 (Special Case 2)2525246781496
Average161614211916161811

4. Conclusion

[21] While assimilation of observations can potentially improve forecast skill, several issues need careful consideration for effective implementation; the prescription of BES is one such critical issue. Formulation of BES is also one issue that requires region-specific consideration. Our objective here has been to evaluate comparative effectiveness of (3D-Var) assimilation with GBES and RBES in simulating tropical cyclones over the north Indian Ocean. For 10 cases of tropical cyclones over the Arabian Sea and Bay of Bengal, the simulations show that assimilation with RBES generally produced better results.

[22] While GBES is inbuilt with WRF 3D-Var system, RBES is domain specific; in the present case the RBES was generated from WRF short range forecasts by applying NMC method. The GBES is computed from global model forecasts for 357 cases distributed over a period of one year while RBES is representative of 30 day period. The results with the assimilations were tested against a control experiment with no assimilation which served as the baseline for a comparative assessment of impact of GBES and RBES. The observational data included both conventional (radiosondes and SHIP) as well as satellite (QSCAT and SSM/I) data. The predicted track and intensity were verified against JTWC observations.

[23] The analysis of initial fields from the 3D-Var analysis using GBES and RBES showed significant differences in terms of horizontal and vertical structure. A scrutiny of predicted track and intensity by different experiments showed RBES to outperform control (no assimilation) as well as GBES simulations. It is important to note that the degree of improvement due to the use of RBES is higher in track prediction than that in prediction of intensity. This could be because the track is more strongly governed by the environmental flow than the intensity.

[24] We have also compared our results for track and intensity prediction with those from several operational agencies like India Meteorological Department (IMD), National Hurricane Center (NHC), Japanese Meteorological Agency (JMA), UK Met Office (UKMO), Climatology and Persistence (CLIPER) model, Statistical Hurricane Intensity Prediction (SHIP) and Logistic Growth Equation Model (LGEM). A comparison of the mean errors in track prediction in the present study with those from these agencies showed (Table 6) the skill with RBES to be comparable or better than that from most of the forecast agencies. However, in case of intensity, even though RBES reduced the errors, the errors are still larger than those from the other operational forecasts (Table 7) in general. It needs to be emphasized, however, that a strict comparison of the results from the present study with the forecast skills of these agencies is not practically possible, as in most cases skill refers only to tropical cyclones formed over the Pacific and the Atlantic basin. The results of this study point to the conclusion that use of domain dependent RBES in 3D-Var assimilation system can provide significant improvement over GBES. However, more studies for different regions are required to further strengthen this conclusion.

Table 6. The Mean Absolute Track Forecast Errors (Km) From Different Experiments Averaged Over All the Cyclone Cases in This Study Along With Errors From Various Operational Agencies
Forecast SourceDay 1 ErrorDay 2 ErrorDay 3 Error
  • a

    Mean error for 11 year period 1998–2008 from the India Meteorological Department (IMD) operational model QLM over the north Indian Ocean basin [Kotal et al., 2009].

  • b

    Mean error in National Hurricane Center (NHC) official forecast for the five year period 2005–2009 and is the average value over the Atlantic and the North Pacific basin (P. J. Cangialosi and L. J. Franklin, National Hurricane Center Forecast Verification report, unpublished, 2011, http://www.nhc.noaa.gov/verification/pdfs/Verification_2010.pdf).

  • c

    Mean error in UK Met Office (UKMO) operational forecast for the five year period 2004–2008 over the north Indian Ocean basin (available at http://www.metoffice.gov.uk/weather/tropicalcyclone/tcerrors/ni.html).

  • d

    Mean error of the Japanese Meteorological Agencies (JMA) operational forecast for the years 2008 and 2009 over the north Indian Ocean basin [Kotal et al., 2009].

WRF -CNT140176205
WRF -GBES137168188
WRF -RBES102119121
CLIPER136270626
IMDa138214341
NHCb99179266
UKMOc150220330
JMAd156.5165200
Table 7. The Mean Errors (m/s) in Intensity Forecasts From Different Experiments Averaged Over All the Cyclone Cases in This Study Along With Errors From Various Operational Agencies
Forecast SourceDay 1 ErrorDay 2 ErrorDay 3 Error
  • a

    Mean error for period 2000–2007 from the India Meteorological Department (IMD) operational model Statistical cyclone intensity prediction (SCIP) model [Kotal et al., 2008].

  • b

    Mean error of National Hurricane Center (NHC) official forecast for the five year period 2005–2009 and is the average value over the Atlantic and the North Pacific (P. J. Cangialosi and L. J. Franklin, unpublished report, 2011).

  • c

    Mean error of the Decay Statistical Hurricane Intensity Prediction System (DSHIPS) model for a homogeneous sample from 2001 to 2003 for the Atlantic and Pacific basin [DeMaria et al., 2005].

  • d

    Mean error of the 5-day versions of the Statistical Hurricane Intensity Forecast (SHIFOR) model for a homogeneous sample from 2001 to 2003 for the Atlantic and the Pacific basin [DeMaria et al., 2005].

  • e

    Mean error of Logistic Growth Equation Model (statistical intensity forecast model [DeMaria, 2009]) for the four year period 2007–2010 and is the average values over the Atlantic and the North Pacific basin (from NHC annual report, 2007–2010).

WRF -CNT8.713.814.5
WRF -GBES8.514.316.8
WRF -RBES7.211.511.4
IMDa7.212.111.3
NHCb5.57.89.5
DHIPSc4.37.710
SHIFOR5d4.98.811.3
LGEMe68.28.6

[25] While the time window chosen for calculating RBES in this study may be the most appropriate in the sense that it represents both the growth and decay stages of the cyclone, it cannot be used in actual operational forecasting due to the inclusion of simulated data beyond the time of the initial condition. However, we believe that the present RBES, in a statistical sense, provides a measure of potential improvement in forecast. In real time application, the period for computing BES may be confined to a near future time (∼+5 days) with a higher lead time (∼20–25 days) in the event period. Thus, a separate study, with applicability in operational environment, and aimed at an optimum RBES, is required. For example, such a study can assess whether a BES representative for a particular month (which can be generated from ensemble forecasts of that month for a 10-year time period) can improve forecast skill when used for different years. The results from the present work should be thus considered as indicative of potential improvement in forecast skill due to RBES. Impact of BESs form other methods such as time lagged NMC or ensemble perturbation will be attempted as future directions of the present study.

Acknowledgments

[26] This work was supported by a project “Integrated Analysis for Impact Mitigation and Sustainability,” funded by CSIR, India. The authors gratefully acknowledge Mesoscale and Microscale Meteorology division at the National Center for Atmospheric Research (NCAR) for access and support of WRF and its 3D-Var assimilation system. The authors also acknowledge the National Centers for Environmental Prediction (NCEP) for making analysis data available at their site. The satellite images of the cyclones were obtained from IMD or NASA websites. The radiosonde data were obtained from University of Wyoming website. The ICOADS was obtained from ftp.dss.ucar.edu. The QSCAT and SSM/I data were obtained from www.ssmi.com. The TRMM data were taken from NASA website and was gratefully acknowledged. The authors thank Dale Barker and S. R. H. Rizvi of NCAR for discussions regarding the generation BES for WRF 3D-Var. The authors thank the anonymous reviewers for their critical and insightful comments and suggestions, which were helpful in substantially improving the content and quality of the manuscript.