## 1. Introduction

The north Indian Ocean (NIO), including the Bay of Bengal (BoB) and the Arabian Sea (AS), experiences two tropical-cyclone (TC) seasons, with a primary maximum in TC frequency during the postmonsoon season (October–December) and a secondary maximum during the premonsoon season (April–early June). Because of growing population and development, damage from landfalling TCs over the NIO has shown a steady increase. For disaster warnings and mitigation efforts, forecasting TC tracks is a critical component; track forecast errors over the NIO are high relative to those over the Atlantic and Pacific Oceans (Mohapatra et al. 2013b). It is evident that synoptic and statistical methods have limitations in predicting track and intensity beyond 24 h over the NIO (Mohanty and Gupta 1997; Gupta 2006). With advancements in resolution, physics, and initializations, global and mesoscale models have the potential to provide forecast guidance for TC genesis, intensity, and movement for a period of 72 h for disaster mitigation and warning systems.

The Advanced Research version of the Weather Research and Forecasting system (ARW), developed at the National Center for Atmospheric Research (NCAR), is one of the two distinct dynamical cores of the Weather Research and Forecasting (WRF) model. The other core version, the Nonhydrostatic Mesoscale Model (NMM), was developed at the Environmental Modeling Center of the National Centers for Environmental Prediction (NCEP). Over the Indian monsoon region, and indeed globally, the ARW model is being widely used for the simulation of a variety of weather events, such as heavy rainfall (Niyogi et al. 2006; Routray et al. 2010; Dodla and Ratna 2010; Hong and Lee 2009) and TCs (Osuri et al. 2012a; Pattanaik and Rama Rao 2009; Davis et al. 2008). The ARW model has been used for real-time TC forecasting since 2007. The application of such mesoscale models for TC forecasting over the NIO is a recent development. Osuri et al. (2012a) demonstrated the promising ability of the ARW model for real-time prediction of TC track and intensity over the NIO during the 2008 season. This study also demonstrated that the performance of the ARW model was reasonably good in comparison with other global models. The current TC operational model in use at the India Meteorological Department (IMD), New Delhi, is the Quasi-Lagrangian Model (QLM), apart from global and regional models like the Global Forecast System (GFS) and the ARW. The QLM has shown little skill over the NIO Basin because it runs at 40-km horizontal resolution with 16 vertical levels (see online at http://www.imd.gov.in/section/nhac/dynamic/RSMC-2011.pdf; Roy Bhowmik and Kotal 2010; Kotal and Roy Bhowmik 2011). As compared with other operational centers, NCEP in the United States uses the high-resolution hurricane WRF (HWRF) model, which is based on the NMM dynamical core of WRF; operating at cloud-permitting (3 km) resolution, it has achieved significant skill in hurricane track and intensity forecasts (Gopalakrishnan et al. 2006, 2012; Tallapragada et al. 2008) over the Atlantic and Pacific Ocean basins. The recent case studies by Mohanty et al. (2013) and Pattanayak et al. (2012) reported better results using the HWRF model for BoB cyclones as well.

The prime objective of this study is to evaluate the performance of the ARW model for track predictions over the NIO on the basis of 100 forecast cases involving 17 TCs that occurred over the NIO between 2007 and 2011. Note that this 5-yr period considered is comparable to the period typically used in other operating centers such as NCEP. This is an effort to document prediction skills over the NIO, because there is no well-established baseline for mesoscale modeling capability over this region. A semioperational effort of this kind may provide a basis for improving track and intensity prediction skills in this region. This semioperational effort with ARW shall become the benchmark for evaluating any future developments over the NIO region for TC predictions.

## 2. Modeling system and configuration

The model configuration including domain and physical parameterization schemes follows the sensitivity study of Osuri et al. (2012a). The initial and lateral boundary conditions for the ARW model are obtained from the analysis and forecast fields of the NCEP GFS. The lateral boundary conditions are updated in 6-h intervals with a fixed sea surface temperature throughout the model integration, with no regional data assimilation used in this study. The land surface boundary conditions are taken from the U.S. Geological Survey with a horizontal grid spacing of 10 min. A single experimental domain is fixed between 77° and 102°E and between 3° and 28°N over the BoB, and between 48° and 78°E and between 5° and 30°N over the AS, with 51 vertical layers and 27-km horizontal grid resolution with an Arakawa C grid. This study used a time step of 90 s with the Kain–Fritsch cumulus parameterization scheme, the slab model for the land surface representation, the WRF single-moment 3-class microphysics scheme, and the Yonsei University planetary boundary layer scheme. Details of the model equations, physics, and dynamics are available in Skamarock et al. (2005; see also online at http://wrf-model.org).

## 3. Method

On the basis of criteria adopted by the IMD, a TC is designated as a cyclonic storm if the associated maximum sustained surface wind (MSW) is 34–47 kt (1 kt ≈ 0.51 m s^{−1}), as a severe cyclonic storm if it has MSW of 48–63 kt, as a very severe cyclonic storm if it has MSW of 64–119 kt, and as a supercyclonic storm if it has MSW of 120 kt or more (IMD 2011). It is considered a depression if the MSW is 17–33 kt. Seventeen TCs during 2007–11 over the NIO are considered in this study. Of these 17 TCs, 13 were formed over the BoB and 4 formed over the AS. A real-time TC forecast up to 120 h was performed 2 times per day (0000 and 1200 UTC) throughout the TC life with effects from the genesis stage (formation of depression). Table 1 shows the period of simulations as well as the time of landfall for each TC. The numerical simulations result in 100 prediction “cases.” The details of synoptic situations and best-track data of the 17 TCs were obtained from the IMD Regional Specialized Meteorological Centre (RSMC) reports from 2008, 2009, 2010, 2011, and 2012 (http://www.imd.gov.in/section/nhac/dynamic/rsmc1.htm).

Details of the model simulations and observed landfall time of each TC. For intensity, CS = cyclonic storms, SCS = severe cyclonic storms, VSCS = very severe cyclonic storms, and SuCS = supercyclonic storms. For nature, SM = straight-moving TCs and RC = recurving TCs.

The model predictions were evaluated on the basis of the calculation of different types of errors such as direct position error (DPE), longitudinal (zonal, or DX), and latitudinal (meridional, or DY) errors. These errors were calculated with respect to the best-track estimates of the IMD. The IMD best-track estimates are mainly based on satellite data and expert interpretation when the TC is in the midsea region. The satellite estimates are sometimes tweaked on the basis of the available mean sea level pressure (MSLP) and 10-m wind observations from ships and buoys and of scatterometer winds from the satellite in the region. When the TC approaches the coast and is within radar range, the best-track estimates are based on the radar observations followed by satellite data. When the system lies close to the coast, then the IMD meteorologist gives higher weight to operational synoptic observations followed by radar and satellite. On a synoptic level, the location of MSLP values and the center of the 10-m wind circulations are considered to determine the best track of a TC.

The TC positions have been identified by using an automated tracking scheme that is based on Marchok (2002). The center of the TC is determined on the basis of the spatial distribution of seven parameters including the minimum MSLP and relative vorticity maxima, geopotential height minima, and wind speed minima at 850 and 700 hPa. Many global centers, including NCEP, determine TC positions by following the same procedure as that documented in Marchok (2002).

The DPE is the great-circle distance between the model forecast position and the observed TC position at a particular time. The systematic error DX (or DY) is defined to be positive if the forecast position is right (or ahead) of the best-track position. The systematic errors can provide some information about the directionality of the errors in the zonal or meridional directions. In the case of northward- and westward-moving TCs, however, there are additional difficulties in interpreting the systematic errors on the basis of DX and DY errors. To determine whether the forecast position is left or right and slower or faster, the cross-track (CT) and along-track (AT) errors can be calculated relative to the observed track. The calculation method of these errors is shown in Fig. 1. The full details on calculation of AT and CT errors can be found in Fiorino et al. (1993). The perpendicular distance from the model track forecast position to the observed track is the CT error. It is positive (negative) when the forecast position lies to the right (left) of the observed track. CT errors give the spatial spread of the simulated tracks. At a given forecast time, AT errors are defined as the distance from the observed position to the position along the observed track where the model forecast track meets perpendicular to the observed track. If the perpendicularity meets the track ahead of (behind) the observed position, the value of AT is positive (negative) and indicates faster (slower) movement of the simulated TC relative to actual movement of the TC. The information on CT and AT errors is very helpful for TC disaster management, because the CT error will help in determining the area of evacuation needed in case of a landfalling TC and the AT error will help in determining the time of evacuation. A relatively lower CT error is desirable by disaster managers because it will lower the cost of evacuation.

Positive (negative) value represents gain (loss) in model skill. Apart from this gain in skill, the confidence intervals (CI) for mean errors are also calculated.

## 4. Results and discussion

### a. Initial vortex position errors

Figure 2 shows the “observed” best tracks of all 17 TCs over the Arabian Sea (Fig. 2a) and Bay of Bengal (Fig. 2b) during 2007–11 as provided by the IMD. Figure 3 provides the mean initial vortex position error of each TC and the ensemble mean of 100 cases over the NIO as based on 27-km horizontal resolution. The mean initial vortex position errors vary from 30 to 94 km for 17 TCs. The average error is ~61 km with 99% CI of 4 km over the NIO Basin. The mean initial position errors are ~58 and 68 km over the BoB and AS, respectively. The 99% CIs of the mean initial vortex position error for the BoB and the AS are 6 and 10 km, respectively. The initial position error may be due to poor data near and around the vortex over the NIO and also to the coarser resolution of the global analyses. Further, the NCEP considers Joint Typhoon Warning Center (JTWC) TC positions for calculation of positional errors and TC vortex relocation in the GFS model. IMD best-track positions were used for calculation of ARW positional errors because the IMD is the official agency for determination of best track for TCs over the NIO. There is approximately a 30–70-km difference between the IMD and JTWC best-track positions over the NIO (Falguni et al. 2004). Similar interagency differences also exist for other ocean basins. Ahn et al. (2002) demonstrated that, when comparing the departure between JTWC and RSMC, Tokyo is greater (~50 km) when the system is at the depression stage and less (30 km) when the system is at the severe-intensity stage. Thus the difference between the estimates of TC positions by JTWC and IMD might have contributed to this higher initial error also.

The “observed” best tracks of tropical cyclones (from IMD) over the (a) AS and (b) BoB during 2007–11.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

The “observed” best tracks of tropical cyclones (from IMD) over the (a) AS and (b) BoB during 2007–11.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

The “observed” best tracks of tropical cyclones (from IMD) over the (a) AS and (b) BoB during 2007–11.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean initial vortex position error (km) for each TC and for the overall mean, as based on 27-km resolution.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean initial vortex position error (km) for each TC and for the overall mean, as based on 27-km resolution.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean initial vortex position error (km) for each TC and for the overall mean, as based on 27-km resolution.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

### b. Mean track forecast errors over the NIO

Figure 4 shows the model-simulated tracks with 27-km horizontal resolution at different initial conditions along with the IMD best track for all TCs for 2007–11. The model could realistically predict the tracks for these TCs with most of the initial conditions. The exception is TC Gonu, where the model predicted the initial movement toward Oman and the first landfall over Oman with some southward displacement. The model could not predict the second landfall of Gonu over Iran. Hence, the lead time of the model-predicted tracks is represented with respect to the first landfall (0000 UTC 6 June 2007) in Fig. 4b. In a similar way, the model could not simulate the realistic movement in some cases of Sidr and Laila. The TC Sidr moved northward under the influence of the upper-tropospheric trough to the west of the system and the upper-tropospheric ridge located to the east. The deep-layer mean wind (850–200 hPa) obtained from the ARW model was analyzed for TC Sidr for the initial conditions at 1200 UTC 11 November–1200 UTC 13 November 2007, and it could not predict the steering circulations. With better representation of the deep-layer mean wind from 0000 UTC 14 November, however, the model could simulate a realistic track that was close to what was observed. This appears to be due to the fact that the ARW model could not pick up the correct boundary conditions because the domain was limited to 78°–103°E but the system was actually near 89°–90°E longitude. It is obvious that the domain size plays an important role in providing better boundary conditions (Landman et al. 2005) along with steering forcing. The TC Nargis moved eastward for about 3 days (from 29 April to 1200 UTC 1 May 2008) mainly because it was under the joint influence of an upper-level anticyclone lying to the southeast and midlatitude upper-tropospheric westerlies (http://www.imd.gov.in/section/nhac/dynamic/RSMC-2009.pdf). The ARW model could resolve and capture the presence of the upper-tropospheric anticyclonic circulation to the south of the system as observed and the low-level (850 hPa) vorticity advection reasonably well (not shown).

Model-predicted tracks as based on 27-km horizontal resolution (in gray) of (a) Akash, (b) Gonu, (c) Yemyin, (d) Sidr, (e) Nargis, (f) Rashmi, (g) KhaiMuk, (h) Nisha, (i) Bijli, (j) Aila, (k) Phyan, (l) Ward, (m) Laila, (n) Phet, (o) Giri, (p) Jal, and (q) Thane at different initial times (shown in Table 1) along with the IMD best track (black line with bigger cyclone symbol). The numbers in the legend represent the lead forecast hours. The dotted tracks in (n) represent tracks with 132-h (short dashes) and 144-h (long dashes) lead times. The small cyclone symbols over gray color tracks represent TC position at each 24-h interval.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Model-predicted tracks as based on 27-km horizontal resolution (in gray) of (a) Akash, (b) Gonu, (c) Yemyin, (d) Sidr, (e) Nargis, (f) Rashmi, (g) KhaiMuk, (h) Nisha, (i) Bijli, (j) Aila, (k) Phyan, (l) Ward, (m) Laila, (n) Phet, (o) Giri, (p) Jal, and (q) Thane at different initial times (shown in Table 1) along with the IMD best track (black line with bigger cyclone symbol). The numbers in the legend represent the lead forecast hours. The dotted tracks in (n) represent tracks with 132-h (short dashes) and 144-h (long dashes) lead times. The small cyclone symbols over gray color tracks represent TC position at each 24-h interval.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Model-predicted tracks as based on 27-km horizontal resolution (in gray) of (a) Akash, (b) Gonu, (c) Yemyin, (d) Sidr, (e) Nargis, (f) Rashmi, (g) KhaiMuk, (h) Nisha, (i) Bijli, (j) Aila, (k) Phyan, (l) Ward, (m) Laila, (n) Phet, (o) Giri, (p) Jal, and (q) Thane at different initial times (shown in Table 1) along with the IMD best track (black line with bigger cyclone symbol). The numbers in the legend represent the lead forecast hours. The dotted tracks in (n) represent tracks with 132-h (short dashes) and 144-h (long dashes) lead times. The small cyclone symbols over gray color tracks represent TC position at each 24-h interval.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

The mean error statistics (km) of the model at 27-km resolution, such as DX, DY, CT, and AT errors, as based on 100 cases, for the NIO systems are given in Table 2. The mean DX of the ARW model is positive for all forecast periods. The average ARW forecast track thus lies to the right of the best-track position for all of the simulations. In other words, the ARW model shows a bias to predict right or eastward movement of TCs over the study domain. In contrast to the ARW model, persistence-based TC forecast positions to the left (negative DX) of the observed positions have a left or westward bias. The mean CT error of the ARW model is slightly negative at the 12-h forecast, and it is found that the slightly negative value of CT is mainly due to higher negative values in a few cases [e.g., TCs Sidr (Fig. 4d) and Bijli (Fig. 4i)] at the 12-h forecast, leading to a higher standard deviation. The CT error becomes positive for higher forecast lengths because of rightward movement of the simulated TCs by the ARW model, however. Similar to the analyses of DY, the analyses of mean AT errors of the ARW model reveal that track positions are generally behind the observed tracks as seen by the result that AT errors are negative for all forecast lengths. In comparing the magnitudes of AT and CT errors of the ARW model, it is seen that the AT errors are larger for all forecast lengths. That is, the track error is elliptical in nature with its major axis along the track. In other words, the spread of the track relative to the observed track is less. The 99% CI of mean error from the ARW forecasts at all forecast intervals is smaller when compared with those of persistence track errors suggesting that the ARW model forecasts are in general more consistent for all forecast intervals.

Mean DX, DY, CT, and AT errors (km) of predicted tracks from the ARW model of 27-km horizontal resolution and from the persistence method for up to 72-h forecast length for NIO TCs. The 99% CI is given in parentheses.

The mean DPE (km) of TC forecast positions as based on 27-km horizontal resolution is shown in Fig. 5a over the NIO. The mean 12-, 24-, 48-, and 72-h forecast errors of the ARW (persistence) tracks are 113 (121), 140 (230), 248 (500), and 375 (770) km, respectively. The 99% CI of mean DPE of the ARW model for the 12-, 24-, 48-, and 72-h forecast is approximately 18, 23, 49, and 94, respectively. There is a gain in forecast skill that varies between 7% and 51% for the 12- to 72-h forecast lengths (Fig. 5a) over the persistence track. The gain in skill is generally greater with an increase in the forecast period.

Mean DPE (km; shaded bars) and gain in skill (%; line) of the model with 27-km resolution for the TCs over the (a) NIO, (b) BoB, and (c) AS. Open bars give mean DPE of persistence tracks.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean DPE (km; shaded bars) and gain in skill (%; line) of the model with 27-km resolution for the TCs over the (a) NIO, (b) BoB, and (c) AS. Open bars give mean DPE of persistence tracks.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean DPE (km; shaded bars) and gain in skill (%; line) of the model with 27-km resolution for the TCs over the (a) NIO, (b) BoB, and (c) AS. Open bars give mean DPE of persistence tracks.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

### c. Mean track forecast errors over the BoB

Table 3 provides the mean errors (similar to Table 2) for 13 BoB TCs as based on 75 forecast cases. The mean DX and DY values are also similar to those for all TCs taken together over the NIO. Thus, similar to the results found for the NIO, model-based TC positions are biased to the right and behind observed TC positions for all forecast lengths over the BoB. This is further noted in CT and AT errors (Table 3). The ARW-based CT error values are positive at all forecast hours except for the 12-h forecast over the NIO, and the AT errors are largely negative at all forecast times. When comparing the CT and AT errors of the model, it is seen that the shape of the errors is again elliptical with the major axis along the track for all forecast lengths for TCs over the BoB, similar to those over the NIO. The mean DPE of the ARW model for the BoB TCs (Fig. 5b) at 12-, 24-, 48-, and 72-h forecast lengths is 108, 137, 243, and 353 km, respectively. The 99% CI for the mean errors at the same forecast lengths is 22, 22, 57, and 124, respectively. The ARW model shows a similar gain in skill for BoB TCs as that for TCs over NIO for all forecast periods. For individual TCs, the error is higher for recurving TCs Bijli and Ward. This may be due to the fact that the initial vortex position error is more for TC Bijli (94 km, with a lower intensity of 996 hPa) and TC Ward (91 km, with a lower intensity of 996 hPa).

### d. Mean track forecast errors over the AS

From Table 4, the DX errors are positive for all TC forecast lengths over the AS. It shows that the ARW track positions are also biased toward the right over the AS Basin. The DY errors are negative for all forecast lengths, suggesting that the model track positions lie behind the actual positions as also observed over the NIO and BoB. The CT errors of the ARW model for the AS TCs are positive up to 60 h and become negative afterward, which suggests that the model positions are to the right of the observed track up to 60 h and shift left of the observed track, similar to that found in the persistence method. The CT and AT error values from the ARW forecast are also smaller when compared with the persistence errors. The AT errors are considerably higher in comparison with CT errors for all forecast lengths, demonstrating that the error is elliptical in shape with its major axis along the track, similar to that over the NIO and BoB. Figure 5c shows the mean DPE of the ARW model on the basis of 25 cases for the 12-, 24-, 48-, and 72-h forecasts at 125, 142, 270, and 413 km, with 99% CI of 33, 49, 99, and 166 km, respectively. For the same forecast lengths, the persistence tracks are 118, 227, 503, and 831 km, respectively. This result shows a positive gain in ARW skill of approximately 37%, 46%, and 50% at 24-, 48-, and 72-h forecast time, respectively. The model has a negative gain (loss) in skill (−6%) for AS systems at the 12-h forecast, unlike the NIO and BoB TCs. The 99% CI of mean DPEs corresponding to the ARW model is small in comparison with that of persistence track errors, suggesting that the ARW model is more consistent, similar to that over the BoB.

### e. Mean track forecast errors for straight movers and recurving TCs

All of the TCs over the NIO are classified into two categories on the basis of characteristic movement (irrespective of the location of genesis) as 1) straight-moving or 2) recurving TCs. Sidr, Nargis, Bijli, Phyan, Ward, Laila, and Phet come under the category of recurving TCs (contributing to 52 forecast cases) and the remaining TCs (Akash, Gonu, Yemyin, Rashmi, KhaiMuk, Nisha, Aila, Giri, Jal, and Thane) are classified as straight-moving TCs (contributing to 48 cases). The mean track errors and associated gain in skill of the model at 27-km resolution for the above two categories of TCs at different forecast hours are shown in Figs. 6a and 6b, respectively. The 99% CI for the mean error is also shown in Fig. 6a. It is clear that the mean errors and the 99% CI associated with straight-moving TCs are significantly less than those of recurving TCs. The mean track errors vary from 55 to 250 km from 0- to 72-h forecast lengths for straight-moving TCs and from 74 to 440 km for recurving TCs. The 99% CI of mean error at initial time, 12-, 24-, 48-, and 72-h forecast lengths are 11, 23, 33, 61, and 101 km for straight movers and 16, 27, 36, 86, and 186 km for recurving TCs, respectively. The gain in skill of the model in simulating these two categories is calculated with reference to persistence tracks. The skill of the model is increasingly positive for straight-moving TCs for all forecast lengths (30%–67% for 12–72-h forecast). The gain in skill is higher in the case of straight-moving TCs by approximately 20%–30% when compared with recurving TCs for different forecast lengths. Even the skill is negative for the 12-h forecast period for recurving TCs. Ramarao et al. (2006) also demonstrated that the forecast errors increase for recurving cyclones and that they are difficult to predict on the basis of the QLM model for NIO TCs. Mohapatra et al. (2013b) demonstrated that the straight movers along the southern periphery of a subtropical ridge have higher predictability than recurving systems, which have a low predictability. This analysis showed that numerical models such as the ARW still have difficulty in predicting recurving TCs, and these difficulties can be mainly attributed to the following factors. Large-scale steering can play a principal role in deciding storm track, especially in the case of recurving TCs. Sometimes the ARW model may not be able to capture the large-scale steering ridge and westerly trough in the upper troposphere, leading to recurvature of the system because of the limited domain (77°–102°E). It is obvious that mesoscale models are mainly driven by lateral boundary conditions available from global analysis/forecast products. In this case, the model could not pick up the correct boundary conditions because of either the limited domain size or the lack of representation of such steering features in the global model initial and/or lateral boundary conditions, resulting in poor predictability of recurving systems. Landman et al. (2005) also demonstrated the role of domain size on simulation of TCs as vortices using a regional climate model. Apart from this, model initialization is also another important factor. Osuri et al. (2012b) showed that assimilation of satellite-derived sea surface winds improved the forecast of recurving TC Nargis. Therefore, high-resolution assimilation of quality-controlled observations into initial conditions as well as coupling with ocean models for better air–sea interaction will help to improve predictions of recurving systems.

(a) Mean DPEs (km) and (b) gain in skill (%) of the model with 27-km resolution for straight movers and recurving TCs over the NIO.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

(a) Mean DPEs (km) and (b) gain in skill (%) of the model with 27-km resolution for straight movers and recurving TCs over the NIO.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

(a) Mean DPEs (km) and (b) gain in skill (%) of the model with 27-km resolution for straight movers and recurving TCs over the NIO.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

From Table 5, the analyses of mean zonal and meridional errors associated with the straight-moving and recurving TCs over the NIO reveal that the track positions are to the right of (behind) the observed TC position given that DX (DY) is positive (negative) for straight movers as well as for recurving TCs. The analyses of CT and AT errors support the elucidation for straight-moving TCs, however, whereas for recurving TCs the CT errors are positive up to the 36-h forecast only. After that, the CT errors become negative up to 72 h, which implies that the model positions are to the right of the observed position up to 36 h and shift to left of the observed position from 36 h onward. In comparing the CT and AT errors, it is seen that the 99% CI is considerably less for straight-moving TCs when compared with recurving TCs. Like NIO TCs, the error is elliptical in shape, with its major axis along the track, for 12–72-h forecast lengths for these two categories of TCs.

Mean DX, DY, CT, and AT errors of the ARW model for up to 72-h forecast lengths for straight-moving and recurving TCs as based on 27-km horizontal resolution over the NIO. The 99% CI is given in parentheses.

### f. Mean track forecast errors relative to intensity of TCs

The performance of the ARW model for track prediction is further extended by considering TC forecasts initialized at different intensity stages such as depression, cyclonic storm (CS), and severe cyclonic storm (SCS) stages. The analysis is carried out with respect to the stage of intensity at the time of model initialization. Of 17 TCs considered in the study, 8 TCs reached CS intensity and the remaining 9 reached SCS intensity (Table 1). Accordingly, there are 38, 28, and 34 forecast cases issued with depression, CS, and SCS stages at the time of model initialization, respectively. It is observed that there is no significant difference in track errors up to 36 h, when predictions are conducted from the depression and CS stages. There is an improvement of about 20–25 km in 48–72-h forecast times, however, when predictions are carried out from the CS stage in comparison with those at the depression stage (not shown). A considerable difference in track forecasts errors is noticed with respect to initializations at the CS and SCS stages (Fig. 7a). The 99% CI is also shown in Fig. 7a. The mean initial vortex position error (60 km at the CS stage and 51 km at the SCS stage with 99% CI of 23 and 10 km, respectively) and the track forecast errors for all forecast lengths are smaller in the case of forecasts initialized at the SCS stages. The 99% CI is also significantly less for the SCS initializations (12 km at 12-h forecast and 90 km for the 72-h forecast) when compared with that of CS initializations (47 and 187 km, respectively). This result can be due to better representation of the TC vortex at the SCS stage in terms of horizontal and vertical structure as a result of stronger intensity. The model performance when initialized at the SCS stage (i.e., for stronger storms) is consistent with findings over the Atlantic basin (Gopalakrishnan et al. 2012). The model shows a better gain in skill in track prediction when compared with persistence track when initialized at the SCS stage (varies between 22% and 61%) rather than the CS stage (6%–43%) for all forecast lengths (Fig. 7b). The 99% CI of the mean CT and AT errors is smaller for forecasts issued at the SCS stage in comparison with those at the CS stage. That is, the stronger the storm is, the lower is the track forecast error range. These results are comparable over the western Pacific Ocean using the ARW model (Ryerson et al. 2007). Mohapatra et al. (2013b) also demonstrated the above fact by verifying the operational TC forecasts of IMD for the CS and SCS categories.

(a) As in Fig. 6, but with respect to cyclonic storms (maximum sustained wind speed of 34–47 kt) and severe cyclonic storms (maximum sustained wind speed is 48 kt or more).

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

(a) As in Fig. 6, but with respect to cyclonic storms (maximum sustained wind speed of 34–47 kt) and severe cyclonic storms (maximum sustained wind speed is 48 kt or more).

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

(a) As in Fig. 6, but with respect to cyclonic storms (maximum sustained wind speed of 34–47 kt) and severe cyclonic storms (maximum sustained wind speed is 48 kt or more).

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

### g. Mean track forecast errors relative to TC season of occurrence

The model performance is also evaluated for the two cyclone seasons (pre- and postmonsoon) on the basis of 48 and 50 forecast cases over the NIO Basin (Fig. 8). Overall, the mean initial vortex position error is less for postmonsoon TCs (52 km with a 99% CI of 11 km) when compared with the premonsoon TCs (65 km with a 99% CI of 18 km). Note that in this comparison TC Yemyin is not included in either of the seasons because it formed during the active monsoon period (25–26 June 2007). The model shows fewer errors with a minimum 99% CI for postmonsoon TCs as compared with premonsoon TCs (Fig. 8a). The 99% CI at 12-, 24-, 48-, and 72-h forecast lengths is approximately 27, 39, 89, and 174 km and 22, 31, 71, and 123 km, respectively, for pre- and postmonsoon TCs. The gain in skill is greater for postmonsoon TCs at all forecast lengths up to 72 h (Fig. 8b) over persistence tracks. The relatively higher track errors for premonsoon TCs may be due to the fact that most of the premonsoon systems are recurving systems. Yang et al. (2011) studied the seasonal variability of a number of TCs between 1980 and 2009 and found that westward- and northward-moving TCs (that come under the straight-mover category) occur more often (over 85%) during September–January (which covers the postmonsoon season), with a peak in October. Recurving TCs are fewer in number during the postmonsoon season, however, occurring more frequently during the months of April and May (premonsoon season). From 2003–11 TCs, Mohapatra et al. (2013b) also demonstrated that 60% of the total TCs in the premonsoon season are recurving TCs. In general, a deep trough in the mid- and upper-tropospheric westerlies is predominant during the premonsoon season over the Indian region (Yang et al. 2011). If a system lies to the right of this upper-level westerly trough, it is expected to recurve to the east or northeast. This situation can be observed more frequently during the premonsoon season than in the postmonsoon season. Such recurving TCs would be expected to have higher errors relative to straight movers (Pike and Neumann 1987; Zhang et al. 2013; Mohapatra et al. 2013b).

As in Fig. 6, but for premonsoon and postmonsoon TCs.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

As in Fig. 6, but for premonsoon and postmonsoon TCs.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

As in Fig. 6, but for premonsoon and postmonsoon TCs.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

## 5. High-resolution retrospective forecast of TCs

To further investigate the potential improvement for future seasons, the study also investigated the role of horizontal resolution on the model forecast by running all of the cases listed in Table 1 again but with 18- and 9-km resolution. Except for the higher resolution, the configuration of the ARW model such as domain, model physics, and initial and boundary conditions was the same as that of the real-time setup.

### a. TC track prediction at 18-km resolution

Relative to 27-km resolution, the 18-km horizontal resolution improved track predictions. Table 6 shows the mean error statistics for the 18-km runs for all of the subcategories as discussed for the 27-km runs. For the NIO TCs, the forecast errors vary from 106 to 359 km for the 12–72-h forecast. The 99% CI for the mean error varies from 20 to 92 km for the same forecast length. The 18-km runs yield an improvement of 10% when compared with the 27-km runs for the NIO Basin. For the 75 forecast cases over the BoB, the mean DPEs (99% CI) are 122 (21), 229 (38), and 337 (103) for the 24-, 48-, and 72-h forecast, with an improvement of about 11%, 6%, and 5%, respectively, over the 27-km runs. The increased resolution provided a better track forecast over the AS with an improvement in error of about 11%, 14%, and 5% for the same forecast lengths. A similar improvement in track forecast errors was found for other categories such as straight-moving TCs (11% for the 24-h forecast, 13% for the 48-h forecast, and 10% for the 72-h forecast), recurving TCs (9%, 8%, and 6%), CS initializations (10%, 15%, and 6%), SCS initializations (6%, 7%, and 3%), premonsoon systems (9%, 10%, and 7%), and postmonsoon systems (10%, 10%, and 8%). The directionality errors (CT and AT) of the 18-km runs are very similar to those of the 27-km predictions, showing a consistent eastward bias with slow movement of the TC for all forecast periods. Note that the 99% CI of the mean error is significantly less for the 18-km runs when compared with the 27-km runs. These results demonstrate that the ARW model would be able to improve the track forecast using 18-km resolution.

Mean DPE (km) of ARW simulations at 18-km resolution for different TC categories at various forecast lead times. The values in the parentheses represent the 99% CI.

### b. TC track prediction at 9-km resolution

Figure 9 provides the track forecast positions for four representative TCs: Nargis, Phet, Jal, and Thane at 6-h intervals. Table 7 summarizes the mean DPE, CT, and AT errors up to 72 h for all cases. The gain in skill (%) of the 9-km-resolution forecast over the 27- and 18-km-resolution forecasts is also calculated. The mean DPE of the 9-km runs is 115, 204, and 329 km for the 24-, 48-, and 72-h forecast, respectively. The mean DPE for the NIO Basin is reduced considerably by 8%–24% over the 27-km-resolution runs and by 4%–10% over the 18-km-resolution runs for the 12–72-h forecast length. In other words, the impact of the higher-resolution forecast (in this case 9- vs 27-km grid spacing) is such that it provided almost 12 h of advantage in terms of the error. That is, the 12-h forecast error of the 27-km resolution is now comparable to the 24-h forecast error in the 9-km-resolution runs. Other features like DX, DY, CT, and AT are similar to those found in the 27-km runs. The mean AT errors—that is, the errors associated with the translation speed of the TC—are considerably smaller in the higher-resolution (9 km) runs, however. The mean DX and CT errors became positive for all forecast lengths, unlike in the 27-km-resolution runs. Given the difficulty in predicting recurving systems, the impact of high-resolution simulations for recurving TCs was analyzed. In particular, the benefit of 9-km-simulations can be seen in the track spread of recurving TCs (e.g., Figs. 4e,n and 9a,b for Nargis and Phet), and the track errors are significantly reduced when compared with the 27- and 18-km runs. The 12-, 24-, 48-, and 72-h track forecast errors that are based on the 9-km runs are 109, 127, 244, and 389 km, respectively, with an improvement of 12%–28% over the 27-km runs and 4%–14% over the 18-km runs. The 99% CI of mean error for the same forecast lengths is also lower. The 12-h DPE is still high (~104 km) in high-resolution (9 km) runs, however, and it is expected that it can be further reduced by vortex initialization and relocation, high-resolution regional data assimilation (for better genesis), sufficient domain size to capture the large-scale environmental forcing, ocean coupling for better analysis, improved model physics, and so on.

Model forecast tracks with 9-km horizontal resolution for (a) Nargis, (b) Phet, (c) Jal, and (d) Thane at different initial times (shown in Table 1) along with the IMD best track. The labels are the same as in Fig. 4.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Model forecast tracks with 9-km horizontal resolution for (a) Nargis, (b) Phet, (c) Jal, and (d) Thane at different initial times (shown in Table 1) along with the IMD best track. The labels are the same as in Fig. 4.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Model forecast tracks with 9-km horizontal resolution for (a) Nargis, (b) Phet, (c) Jal, and (d) Thane at different initial times (shown in Table 1) along with the IMD best track. The labels are the same as in Fig. 4.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean DX, DY, CT, and AT errors and mean DPE of the ARW model with 9-km horizontal resolution for different forecast lengths (up to 72 h) for NIO TCs. The corresponding 99% CIs are given in parentheses.

### c. TC intensity prediction

Although this study mainly focuses on track prediction, a brief description about intensity prediction with 18- and 9-km resolution is also given here. Figure 10 gives the mean intensity error (model predicted − observed) for both 18- and 9-km runs. Positive (negative) values represent overestimation (underestimation) of intensity by the model. The average intensity error of ARW-WRF for 18-km-resolution runs is −13, −10, and −6 m s^{−1} for 24-, 48-, and 72-h forecast lengths, respectively, and it is −8, −6, and −4 m s^{−1} for 9-km-resolution runs. An error that ranges from −8 to −10 m s^{−1} is noticed at initial time in the global model analyses. The higher errors in intensity forecast during the first 24-h period may be due to the spinup problem associated with the ARW model and the initial intensity error. The spinup time can be reduced by adopting finer grid resolutions and digital filter initialization (Lynch et al. 1997; Brousseau et al. 2008). Xiao (2011) also mentioned that, because of the model spinup problem, the model exhibits higher errors for MSW at 24 h over both 48 and 72 h. Osuri et al. (2012b) also showed higher TC intensity errors during the first 24–30 h of the forecast, which is again a spinup issue. The error in the initial vortex (in terms of position, strength, and structure) can be reduced with a vortex initialization and relocation procedure (Gopalakrishnan et al. 2012; Hsiao et al. 2010). Xiao (2011) demonstrated that the use of (bogus) vortex initialization can not only minimize the spinup period but also improve the structure and intensity of the initial TC vortex. This discussion highlights the need for the improvement in the initial vortex over the NIO Basin either through better observations or by vortex initialization procedures. Hsiao et al. (2010) showed better intensity and track predictions by adopting advanced vortex initialization techniques. The intensity error from the 18-km-resolution runs is greater at all forecast lengths. The 99% CI for the mean intensity error varies from 6 to 11 m s^{−1} for 6–72-h forecast intervals. At 9-km resolution, however, it varies from 4 to 7 m s^{−1} at the same forecast intervals. The ARW model shows an overall underestimation in intensity prediction with both 18- and 9-km resolutions when compared with observed intensity. The 9-km resolution did improve the intensity prediction by 15%–40% over the 18-km predictions up to the 72-h forecast. It is noticed that the model could not predict the peak intensities of the TCs and that the maximum error is observed for systems with an intensity of very severe cyclonic storm (VSCS) or higher (MSW ≥ 64 kt) such as Gonu (65 m s^{−1}), Nargis (46 m s^{−1}), and Phet (43 m s^{−1}) because of underestimation. The 72-h maximum intensity predicted by 18- and 9-km-resolution runs for Gonu is 40 and 48 m s^{−1}, for Nargis is 36 and 42 m s^{−1}, and for Phet is 34 and 45 m s^{−1}. Both resolutions could predict the intensity of CS and SCS well, however. This result implies that the higher the intensity is, the poorer is the peak intensity prediction by the model. In comparison with the operational intensity forecast errors of IMD over the NIO Basin, however, the error of the ARW model is higher at 24- and 48-h forecast lengths and is comparable at the 72-h forecast length (Mohapatra et al. 2013a).

Mean intensity errors in terms of 10-m maximum winds predicted by 18- and 9-km-resolution runs for the NIO cyclones. The error bars are at 99% CI.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean intensity errors in terms of 10-m maximum winds predicted by 18- and 9-km-resolution runs for the NIO cyclones. The error bars are at 99% CI.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Mean intensity errors in terms of 10-m maximum winds predicted by 18- and 9-km-resolution runs for the NIO cyclones. The error bars are at 99% CI.

Citation: Journal of Applied Meteorology and Climatology 52, 11; 10.1175/JAMC-D-12-0313.1

Better prediction at 9-km resolution is due to the fact that the ARW model could resolve the convection better than it could for runs at coarse resolution (27 and 18 km). These results are also supported by recent studies that demonstrated that TC predictions in terms of track and intensity can be improved by increasing the resolution (Gopalakrishnan et al. 2012; Torn and Davis 2012; Han and Pan 2011). Torn and Davis (2012) and Han and Pan (2011) reported that proper treatment of convection is an important aspect for realistic TC track predictions the achievement of which is only possible at higher resolutions.

## 6. Summary and conclusions

The overall performance of the ARW model for track and intensity prediction of TCs over the NIO and individual TC basins (BoB and AS) is assessed at different horizontal resolutions (27, 18, and 9 km). This evaluation is based on 100 forecast cases of 17 landfalling TCs that occurred during 2007–11.

The ARW model has a good overall capability to predict TCs over the NIO Basin as it exhibits a reasonable skill irrespective of the region of formation, nature of movement, intensity, and season of formation. The mean track forecast errors (skill with reference to persistence) over the NIO vary from 113 to 375 km (7%–51%) for 12–72-h forecast lengths at 27-km resolution. In a comparison of the track forecast errors with respect to TC intensity at initialization, the ARW model performed better for track prediction if the model is initialized at the SCS stage rather than at the CS/depression stage. With respect to season of occurrence, the model exhibits less error and 3%–15% more gain in skill for postmonsoon TCs than is found for premonsoon TCs.

The ARW model results showed a bias toward predicting eastward movement for the NIO TCs for both the BoB and AS. This bias is particularly true for straight-moving TCs. In the case of recurving TCs, the model showed a rightward bias up to the 36-h forecast and thereafter a westward bias. The analyses of latitudinal systematic errors as well as AT errors show that the model forecast positions are biased to the south of (behind) the observed positions. The ARW forecasts are in general slower relative to the actual translation speed of the system for all forecast lengths, and they predict a delayed landfall. The magnitude of CT errors is less in comparison with AT errors in the ARW model. Hence the ARW model is more accurate in predicting TC landfall location than landfall time.

The higher the resolution of the model is, the better are the track and intensity predictions by the model. The high-resolution predictions (18 and 9 km) reduced the track errors for NIO TCs and yielded an improvement of about 4%–10% and 8%–24% in mean DPE over 27-km runs. The 9-km predictions are found to be better for recurving TC track predictions by about 12%–28% as compared with 27-km runs and 4%–14% as compared with 18-km predictions. A comparison of 18- and 9-km predictions showed that the intensity prediction at 9-km resolution was improved by ~15%–40% over the 18-km predictions up to the 72-h forecast. The model underestimates the peak intensities of systems of VSCS or higher intensity and experiences maximum errors in the case of these systems; the intensity prediction of CS and SCS is reasonably simulated, however. Still, the ARW-based intensity errors are higher for 24 and 48 h when compared with the operational intensity forecast errors of the IMD. This can be further reduced by improving the initial intensity and structure of the TC vortex by increasing the observational network of buoy, ships, and aircraft reconnaissance over the NIO region or through advanced vortex initialization techniques and ocean–atmosphere coupling for better heat, moisture, and momentum exchanges.

## Acknowledgments

The Indian National Center for Ocean Information Services (INCOIS), Ministry of Earth Sciences, government of India, and the U.S. National Science Foundation (CAREER Grant 0847472) are gratefully acknowledged for providing financial support to carry out this research. The authors also thank the IMD for providing TC best-track positions for validation of model tracks. The authors gratefully acknowledge NCEP and NCAR for their input datasets as well the MMM division of NCAR for making the ARW model available. We also thank Mrs. Ammaji for her help during revision. Thanks are given to Dallas Staley for her usual outstanding editing. We also sincerely thank the anonymous reviewers for their constructive suggestions and comments to improve the quality of the paper.

## REFERENCES

Ahn, M.-H., Y.-K. Seo, H.-S. Park, and A.-S. Suh, 2002: Determination of tropical cyclone center using TRMM Microwave Imager data.

,*Geophys. Res. Lett.***29**, doi:10.1029/2001GL013715.Brousseau, P., and Coauthors, 2008: A prototype convective-scale data assimilation system for operation: The Arome-RUC. HIRLAM Tech. Rep. 68, 23–30.

Davis, C. A., and Coauthors, 2008: Prediction of landfalling hurricanes with the Advanced Hurricane WRF model.

,*Mon. Wea. Rev.***136**, 1990–2005.Dodla, V. B. R., and S. B. Ratna, 2010: Mesoscale characteristics and prediction of an unusual extreme heavy precipitation event over India using a high resolution mesoscale model.

,*Atmos. Res.***95**, 255–269.Falguni, P., C. M. Kishtawal, P. K. Pal, and P. C. Joshi, 2004: Geolocation of Indian Ocean tropical cyclones using 85 GHz observations from TRMM Microwave Imager.

,*Curr. Sci.***87**, 504–509.Fiorino, M., J. M. Goerss, J. J. Jensen, and E. J. Harrison Jr., 1993: An evaluation of the real-time tropical cyclone forecast skill of the Navy Operational Global Atmospheric Prediction System in the western North Pacific.

,*Wea. Forecasting***8**, 3–24.Gopalakrishnan, S. G., N. Surgi, R. Tuleya, and Z. Janjic, 2006: NCEP’s two-way-interactive-moving-nest NMM-WRF modeling system for hurricane forecasting. Preprints,

*27th Conf. on Hurricanes and Tropical Meteorology,*Monterey, CA, Amer. Meteor. Soc., 7A.3. [Available online at https://ams.confex.com/ams/pdfpapers/107899.pdf.]Gopalakrishnan, S. G., S. Goldenberg, T. Quirino, X. Zhang, F. Marks, K.-S. Yeh, R. Atlas, and V. Tallapragada, 2012: Toward improving high-resolution numerical hurricane forecasting: Influence of model horizontal grid resolution, initialization, and physics.

,*Wea. Forecasting***27**, 647–666.Gupta, A., 2006: Current status of tropical cyclone track prediction techniques and forecast errors.

,*Mausam***57**, 151–158.Han, J., and H.-L. Pan, 2011: Revision of convection and vertical diffusion schemes in the NCEP Global Forecast System.

,*Wea. Forecasting***26**, 520–533.Hong, S.-Y., and J.-W. Lee, 2009: Assessment of the WRF model in reproducing a flash-flood heavy rainfall event over Korea.

,*Atmos. Res.***93**, 818–831.Hsiao, L.-F., C.-S. Liou, T.-C. Yeh, Y.-R. Guo, D.-S. Chen, K.-N. Huang, C.-T. Terng, and J.-H. Chen, 2010: A vortex relocation scheme for tropical cyclone initialization in Advanced Research WRF.

,*Mon. Wea. Rev.***138**, 3298–3315.IMD, 2011: Tracks of cyclones and depressions over North Indian Ocean (from 1891 onwards). India Meteorological Department Tech. Note Cyclone eAtlas—IMD, version 2.0. [Available online at http://www.rmcchennaieatlas.tn.nic.in/Help/TechNote2011.pdf.]

Kotal, S. D., and S. K. Roy Bhowmik, 2011: A multimodel ensemble (MME) technique for cyclone track prediction over the North Indian Sea.

,*Geofizika***28**, 275–291.Landman, A. W., A. Seth, and S. J. Camargo, 2005: The effect of regional climate model domain choice on the simulation of tropical cyclone–like vortices in the southwestern Indian Ocean.

,*J. Climate***18**, 1263–1274.Lynch, P., D. Giard, and V. Ivanovici, 1997: Improving the efficiency of a digital filtering scheme for diabatic initialization.

,*Mon. Wea. Rev.***125**, 1976–1982.Marchok, T. P., 2002: How the NCEP tropical cyclone tracker works. Preprints,

*25th Conf. on Hurricanes and Tropical Meteorology,*San Diego, CA, Amer. Meteor. Soc., P1.12. [Available online at https://ams.confex.com/ams/pdfpapers/37628.pdf.]Mohanty, U. C., and A. Gupta, 1997: Deterministic methods for prediction of tropical cyclone tracks.

,*Mausam***48**, 257–272.Mohanty, U. C., K. K. Osuri, and S. Pattanayak, 2013: A study on high resolution mesoscale modeling systems for simulation of tropical cyclones over the Bay of Bengal.

,*Mausam***64**, 117–134.Mohapatra, M., B. K. Bandyopadhyay, and D. P. Nayak, 2013a: Evaluation of operational tropical cyclone intensity forecasts over north Indian Ocean issued by India Meteorological Department.

,*Nat. Hazards***68**, 433–451.Mohapatra, M., D. P. Nayak, R. P. Sharma, and B. K. Bandyopadhyay, 2013b: Evaluation of official tropical cyclone track forecast over north Indian Ocean issued by India Meteorological Department.

,*J. Earth Syst. Sci.***122**, 589–601.Niyogi, D., T. Holt, S. Zhong, P. C. Pyle, and J. Basara, 2006: Urban and land surface effects on the 30 July 2003 mesoscale convective system event observed in the southern Great Plains.

,*J. Geophys. Res.***111**, D19107, doi:10.1029/2005JD006746.Osuri, K. K., U. C. Mohanty, A. Routray, A. K. Makarand, and M. Mohapatra, 2012a: Sensitivity of physical parameterization schemes of WRF model for the simulation of Indian seas tropical cyclones.

,*Nat. Hazards***63**, 1337–1359.Osuri, K. K., U. C. Mohanty, A. Routray, and M. Mohapatra, 2012b: Impact of satellite derived wind data assimilation on track, intensity and structure of tropical cyclones over North Indian Ocean.

,*Int. J. Remote Sens.***33**, 1627–1652.Pattanaik, D. R., and Y. V. Rama Rao, 2009: Track prediction of very severe cyclone ‘Nargis’ using high resolution Weather Research Forecasting (WRF) model.

,*J. Earth Syst. Sci.***118**, 309–329.Pattanayak, S., U. C. Mohanty, and S. G. Gopalakrishnan, 2012: Simulation of very severe Cyclone Mala over Bay of Bengal with HWRF modeling system.

,*Nat. Hazards***63**, 1413–1437.Pike, C. A., and C. J. Neumann, 1987: The variation of track forecast difficulty among tropical cyclone basins.

,*Wea. Forecasting***2**, 237–241.Ramarao, Y. V., H. R. Hatwar, and G. Agnihotri, 2006: Tropical cyclone prediction by numerical models in India Meteorological Department.

,*Mausam***57**, 47–60.Routray, A., U. C. Mohanty, S. R. H. Rizvi, D. Niyogi, K. K. Osuri, and D. Pradhan, 2010: Impact of Doppler weather radar data on simulation of Indian monsoon depressions.

,*Quart. J. Roy. Meteor. Soc.***136**, 1836–1850.Roy Bhowmik, S. K., and S. D. Kotal, 2010: A dynamical statistical model for prediction of a tropical cyclone.

,*Mar. Geod.***33**, 412–423.Ryerson, W. R., S. R. Russel, L. Elsberry, and J. Wegiel, 2007: Evaluations of the AFWA Weather Research Forecast Model tropical cyclone predictions. Preprints,

*27th Conf. on Hurricanes and Tropical Meteorology,*Monterey, CA, Amer. Meteor. Soc., 7A.5. [Available online at https://ams.confex.com/ams/pdfpapers/108856.pdf.]Skamarock, W. C., J. B. Klemp, J. Dudhia, O. G. David, D. M. Barker, W. Wang, and J. G. Powers, 2005: A description of the Advanced Research WRF version 2. NCAR/TN–468+STR, 88 pp. [Available online at http://www.mmm.ucar.edu/wrf/users/docs/arw_v2.pdf.]

Tallapragada, V., N. Surgi, Q. Liu, Y. Kwon, R. Tuleya, and W. O’Connor, 2008: Performance of the advanced operational HWRF modeling system during pre-implementation testing and in real-time 2007 hurricane season. Recorded Presentation,

*28th Conf. on Hurricanes and Tropical Meteorology,*Orlando, FL, Amer. Meteor. Soc., 4A.5. [Available online at https://ams.confex.com/ams/28Hurricanes/webprogram/Paper138066.html.]Torn, R. D., and C. A. Davis, 2012: The influence of shallow convection on tropical cyclone track forecasts.

,*Mon. Wea. Rev.***140**, 2188–2197.WMO, 2009: Standard format for verification of TC forecast. World Meteorological Organization TCM-VI/Doc. 2.4, 6 pp. [Available online at www.wmo.int/pages/prog/www/tcp/documents/Doc2.4_Verification.doc.]

Xiao, Q., 2011: Improvements of hurricane forecast with vortex initialization using WRF Variational (WRF-Var) data assimilation.

*Recent Hurricane Research—Climate, Dynamics, and Societal Impacts,*A. Lupo, Ed., InTech, 297–318. [Available online at http://www.intechopen.com/books/recent-hurricane-research-climate-dynamics-and-societal-impacts.]Yang, L., W.-W. Li, D. Wang, and Y. Li, 2011: Analysis of tropical cyclones in the South China Sea and Bay of Bengal during monsoon season.

*Recent Hurricane Research—Climate, Dynamics, and Societal Impacts,*A. Lupo, Ed., InTech, 227–246. [Available online at http://www.intechopen.com/books/recent-hurricane-research-climate-dynamics-and-societal-impacts.]Zhang, W., Y. Leung, and J. C. L. Chan, 2013: The analysis of tropical cyclone tracks in the western North Pacific through data mining. Part I: Tropical cyclone recurvature.

,*J. Appl. Meteor. Climatol.***52**, 1394–1416.