Downloaded 10/01/21 03:37 AM UTC 1824 MONTHLY WEATHER REVIEW VOLUME 144 Western North Paciﬁc Ocean, It Is Still Necessary to Esti- GBVTD Retrievals

MAY 2016 S H I M A D A E T A L . 1823

Evaluation of the Accuracy and Utility of Tropical Cyclone Intensity Estimation Using Single Ground-Based Doppler Radar Observations

UDAI SHIMADA AND MASAHIRO SAWADA Meteorological Research Institute, Tsukuba, Ibaraki, Japan

HIROYUKI YAMADA University of the Ryukyus, Nishihara, Okinawa, Japan

(Manuscript received 18 July 2015, in ﬁnal form 21 December 2015)

ABSTRACT

Intensities (central pressures) of 28 cases of 22 tropical cyclones (TCs) that approached Japan were estimated by using single ground-based Doppler radar observations, and the accuracy and utility of the estimation method were evaluated. The method uses the ground-based velocity track display (GBVTD) technique, which retrieves tangential winds, and the gradient wind balance equation. Before application of the method to the 28 cases, a preliminary experiment was performed with pseudo-Doppler velocities obtained by numerical simulation to confirm that the method could reasonably estimate central pressures. Compared with best track data from the Regional Specialized Meteorological Center (RSMC) Tokyo, the estimated intensities of the 28 cases had a root- mean-square error of 8.37 hPa and showed a bias of 1.51 hPa. This level of accuracy is comparable to or better than the accuracies of Dvorak and satellite microwave-derived estimates. Two distance metrics are defined: 1) the distance between the TC center and the radar location and 2) the distance between the TC center and the weather station whose sea level pressure was used as an anchor for pressure measurement. In general, the accuracy of the Doppler radar estimates was higher when the distance metrics were shorter, as well as when wind retrieval accuracy was better and radar coverage was denser. For TCs with a radius of maximum wind of 20– 70 km, the estimated central pressures had a root-mean-square error of 5.55 hPa. These results confirm that Doppler radar intensity estimates have sufficient accuracy and utility for operational use.

1. Introduction and Nakazawa 2007; Sakuragi et al. 2014). The Dvorak technique is an empirical method based on past re- It is of great importance to analyze tropical cyclone connaissance observations and the subjective classifica- (TC) intensity (i.e., central pressure and maximum sus- tion of cloud patterns, although the advanced Dvorak tained wind) with high accuracy, not only from the per- technique is a fully automated method (Olander and spective of improving scientific understanding of various Velden 2007). Methods using satellite microwave phenomena associated with TCs, but also for disaster sounding and imager data are entirely objective, but in prevention and mitigation. TC intensity estimates are the western North Pacific, satellite methods refer to the generally made using infrared satellite data, for example, best track data as truth. Because these data are based by the Dvorak technique (e.g., Dvorak 1975, 1984; Koba mainly on the Dvorak technique, the uncertainty of the et al. 1990; Velden et al. 1998; Olander et al. 2004; satellite methods includes the estimation error associated Olander and Velden 2007), satellite microwave sounding with the Dvorak technique. In addition, these conven- data (Brueske and Velden 2003; Herndon and Velden tional methods estimate TC intensity from relevant 2004; Demuth et al. 2004; Oyama 2014), or satellite mi- physical values such as cloud patterns and upper-level crowave imager data (Bankert and Tag 2002; Hoshino warm-core anomalies, rather than by using a straightforward physical equation that describes pressure distributions. Therefore, conventional estimates can have a Corresponding author address: Udai Shimada, Typhoon Research Department, Meteorological Research Institute, 1-1 Nagamine, large margin of error. Although it is not realistically Tsukuba, Ibaraki 305-0052, Japan. possible at present to obtain dropsonde observations of E-mail: [email protected] minimum sea level pressures (MSLPs) of TCs in the

DOI: 10.1175/MWR-D-15-0254.1

Ó 2016 American Meteorological Society Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1824 MONTHLY WEATHER REVIEW VOLUME 144 western North Pacific Ocean, it is still necessary to esti- GBVTD retrievals. The Japan Meteorological Agency mate TC intensities with as much accuracy as possible, (JMA) upgraded all of the radars in its observation particularly for TCs that approach populated areas. network to Doppler radars between 2006 and 2013, and One method of addressing this problem is to estimate the new Doppler radars have collected data on many central pressures by using a physical equation, namely, TCs since their installation. the gradient wind equation. Lee et al. (2000) developed a The purpose of this study was to estimate TC in- method that uses data from a single ground-based tensities by using the accumulated Doppler radar data Doppler radar (DR; hereafter the DR method). In this and to evaluate the accuracy and utility of the DR method, the ground-based velocity track display method. Before the estimation of TC intensities, we (GBVTD) technique (Lee et al. 1999) is used to retrieve performed a preliminary experiment with pseudo VD axisymmetric tangential wind velocities, Vt,ofaTCfrom obtained by numerical simulation to confirm that the the Doppler radial velocities, VD, under the assumption method could reasonably estimate central pressures. that there is one primary circular vortex around the TC This paper consists of six sections. We describe the center and that the asymmetric radial wind is much data and TC cases used and the study methodology in smaller than the corresponding tangential wind. Then, an section 2. We describe the preliminary experiment and MSLP is estimated by using an axisymmetric pressure present its results in section 3.Insection 4, we present the deficit deduced by applying the retrieved Vt to the gra- DR estimation results for real TC cases and their statis- dient wind equation and using a sea level pressure (SLP) tical verification. In section 5, we discuss the results, in- observation around the TC as an anchor for pressure cluding their utility, reliability, and limitations for measurement. Because the DR method calculates operational use, and present additional statistical verifi- MSLPs by using a physical equation related to the TC cations. Section 6 includes a summary and conclusions. wind field, the estimated MSLPs are expected to have higher accuracy than those derived from cloud patterns 2. Data, cases, and methodology and upper-level warm-core anomalies, although the DR a. Data method is applicable only to TCs that approach radar locations. To apply the DR method, it is essential to re- We used VD data observed by JMA C-band opera- trieve Vt from VD with high accuracy. Because the tional Doppler radars. Figure 1 shows the radar loca- GBVTD technique has been successfully used to retrieve tions used and the observational range of 2-km, Vt for many TCs (Lee et al. 2000; Harasti et al. 2004; constant altitude plan position indicator (CAPPI) data Lee and Bell 2007; Zhao et al. 2008; Zhao et al. 2012), (;200 km) for each location. The Doppler observation the DR method is promising. parameters of the Okinawa radar, which is representa- The DR method has already been implemented by the tive of the JMA radars, are listed in Table 1. We used National Hurricane Center (NHC; Harasti et al. 2007), both high and low dual-pulse repetition frequency (dual- where it is called Vortex Objective Radar Tracking PRF) data for the lowest-elevation plan position in- and Circulation (VORTRAC), but the accuracy with dicator (PPI) in order to see out to ;200-km range at which the DR method estimates MSLPs and the oper- 2-km altitude (see section 2c). Before 2009, the radars ational utility of the DR method have not yet been observed VD at 10-min intervals, and since 2009 the clarified comprehensively, probably owing to the lack of observation interval has been 5 min. In this study, sufficient TC examples. It is of great importance that dealiasing of VD by the Hybrid Multi-PRI (HMP) the disaster prevention information have high reliability method (Yamauchi et al. 2006) and by a newly de- and consistency for the forecasts to be seen as credible veloped correction method (not shown) that assumes and to avoid confusion. Thus, it is essential to examine there is a wavenumber-1 VD pattern around a radar lo- the accuracy of the estimated MSLPs and to determine cation influenced by TC circulation was performed au- their suitability for operational use. In addition, analysts tomatically. Additionally, noise caused by dealiasing adopting the DR method for TC intensity analysis need failure and sea clutter was removed from the data au- to have a good understanding of the DR method’s ac- tomatically by various quality control (QC) procedures. curacy and reliability, as well as of its strengths, weak- Although problems associated with the dealiasing cor- nesses, and limitations. rection of VD can greatly affect the accuracy of TC in- Many strong TCs approach and make landfall in Ja- tensity estimates, describing how to cope with these pan. In particular, many TCs approach southwestern problems is beyond the scope of this study. Therefore, Japan, including the Okinawa Islands, where there is we examine the accuracy of the estimated MSLPs based no topography effect on the TC structure. Therefore, on the premise that the Doppler velocity data have been TCs approaching this region are ideally suited for dealiased and that most noise has been removed.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1825

range (;200 km) of a Doppler radar and we had to be able to subjectively identify the TC center from the VD pattern (an indication that, in general, radar coverage was sufficient). The circulation of a TC is generally de- picted by a dipole pattern of positive (away from the radar) and negative (toward the radar) VD across the TC center, with a wavenumber-1 velocity pattern surround- ing the radar location. This method was chosen under the supposition that analysts subjectively pick out a TC and estimate its duration by examining VD fields and radar reflectivity imagery. We selected 22 cases including TCs that struck southern Japan and strong TCs that passed over the Okinawa Islands (Fig. 1, Table 2). If the same TC was observed by two or more radars, then the observations of each radar were treated as a case, and the cases were distinguished by appending the letters A, B, or C to the TC name. Therefore, in total we estimated the intensity of 28 cases. The estimated duration of each case FIG. 1. Locations of the JMA operational Doppler radars and tracks of the TCs used in this study. The circles show the obser- was about 9 h on average. Although this number of TC vation ranges of all radars (;200 km) for CAPPI data at 2-km al- cases may not be a large enough sample size for statistical titude. Thick lines indicate TCs with maximum sustained (10-min verification, it enabled us to test the TC intensity esti- 21 average) surface winds of greater than 17 m s (i.e., tropical storms), mation method on a variety of TCs. thin lines indicate TCs with maximum sustained surface winds of 2 17 m s 1 or less (i.e., tropical depressions), and gray lines indicate c. Methodology extratropical cyclones. The data are the JMA best track data. The primary procedures used for the DR intensity estimates are based on the work of Lee et al. (1999, In addition to the VD data, we used SLP data observed 2000), Lee and Marks (2000), and Bell and Lee (2012). at weather stations at 10-min intervals. To verify our There are six steps in the procedures (Fig. 2). After the results, we used best track and Dvorak MSLPs from the dealiasing procedure and the removal of sea clutter and Regional Specialized Meteorological Center (RSMC) other noise (step 1), PPI data less than the elevation Tokyo. The best track data from the RSMC Tokyo are angle of 10.08 were interpolated into CAPPI data (step based on both Dvorak MSLPs and observations. These 2). After this step, the 2-km CAPPI data were used Dvorak MSLPs are obtained by using the Dvorak because at that altitude radar coverage for VD is the technique to obtain current intensity (CI) numbers, most extensive over the longest period, and the TC wind which are then converted into MSLPs by referring to the fields are generally approximately in gradient wind conversion table in Koba et al. (1990). When the radar balance (Bell and Montgomery 2008; Willoughby 1990). data were available at 5-min intervals, all other data Note that for radar ranges greater than ;130 km, VD were interpolated linearly to 5-min intervals. from the lowest-elevation PPI scan whose height is between 2 and 4 km is just projected onto the 2-km CAPPI. b. Cases Because vertical profiles of wind speed for TCs show a We used two criteria to select TC cases for estimation: general increase from 3 km down to below 1 km (e.g., the TC center had to be located within the 2-km altitude Franklin et al. 2003), there is a possibility that this

TABLE 1. Doppler observation parameters of the Okinawa operational C-band radar.

Lowest elevation angle Other elevation angles 0.48 1.88, 3.38, 5.08, 7.08, 10.08, 14.08, 19.08, 25.08 Frequency 5350 MHz Pulse repetition frequency (Hz) 940/752, 600/480 940/752 2 Unambiguous velocity (m s 1) 652.7, 633.6 652.7 Unambiguous range (km) 160, 250 160 Pulse length 1 ms Range resolution 250 m Rotation rate 4 rpm

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1826 MONTHLY WEATHER REVIEW VOLUME 144

TABLE 2. Cases used for the intensity estimation.

Case (name) Radar site Estimation period (UTC) Duration T0607 (Maria) Tokyo 0000–0300 9 Aug 2006 3 h T0709 (Fitow) Tokyo 1600–2300 6 Sep 2007 7 h T0813 (Sinlaku) Tanegashima 0200–1500 18 Sep 2008 13 h T0911 (Krovanh) Tokyo 0600–1000 31 Aug 2009 4 h T1007 (Kompasu) Okinawa 0200–1300 31 Aug 2010 11 h T1011 (Fanapi) Ishigakijima 0500–1900 18 Sep 2010 14 h T1109 (Muifa) Okinawa 1800 4 Aug–1200 5 Aug 2011 18 h T1115 (Roke) Murotomisaki 2300 20 Sep–0100 21 Sep 2011 2 h T1204 (Guchol) Okinawa 1000–1500 18 Jun 2012 5 h T1210 (Damrey) Tanegashima 0100–0900 1 Aug 2012 8 h T1215 (Bolaven) Okinawa 0400–1700 26 Aug 2012 13 h T1216 (Sanba) Okinawa 1300 15 Sep–0300 16 Sep 2012 14 h T1217 (Jelawat) Okinawa 1950 28 Sep–0645 29 Sep 2012 10 h 55 min T1307 (Soulik) Ishigakijima 0700–1600 12 Jul 2013 9 h T1312 (Trami) Ishigakijima 1800 20 Aug–0700 21 Aug 2013 13 h T1318 (Man-yi) Murotomisaki 1300–1900 15 Sep 2013 6 h T1323 (Fitow) Ishigakijima 1300 5 Oct–0200 6 Oct 2013 13 h T1324A (Danas) Okinawa 0200–1000 7 Oct 2013 8 h T1324B (Danas) Naze 0500–1200 7 Oct 2013 7 h T1408A (Neoguri) Ishigakijima 1900 7 Jul–0110 8 Jul 2014 6 h 10 min T1408B (Neoguri) Okinawa 0100–0410 8 Jul 2014 3 h 10 min T1411A (Halong) Naze 0200–1300 8 Aug 2014 11 h T1411B (Halong) Tanegashima 1500 8 Aug–0500 9 Aug 2014 14 h T1411C (Halong) Murotomisaki 1000–2345 9 Aug 2014 13 h 45 min T1418A (Phanfone) Naze 1700 4 Oct–0400 5 Oct 2014 11 h T1418B (Phanfone) Tanegashima 0300–1100 5 Oct 2014 8 h T1418C (Phanfone) Murotomisaki 1330–1930 5 Oct 2014 6 h T1419 (Vongfong) Okinawa 0300–1200 11 Oct 2014 9 h

CAPPI method contributes to positively biased MSLP density in a tropical environment with a virtual tem- estimates when the TC center is far from the radar. perature of 30.08C at 1000 hPa. At the second decimal Then, by using the GBVTD-simplex center-finding al- place, this value may be a little higher than that in the TC gorithm (Lee and Marks 2000; Bell and Lee 2012), inner environment, which would make the radial SLP where a radius of maximum wind (RMW) and its pos- gradient calculated from the gradient wind balance with sible range are subjectively given, TC center positions at Eq. (1) slightly steeper than the actual gradient. When 5-min intervals were detected from the first-guess center one or more SLP observations were available within positions interpolated linearly to 5-min intervals from radii of the deduced SLP gradient, the axisymmetric center positions subjectively determined at 1-h intervals SLP deficits of the TC were obtained by integrating the while viewing radar reflectivity imagery (step 3). Then, radial SLP gradient [the right-hand side of Eq. (1)] in- using the detected centers, the 2-km CAPPI data were ward from each observation point toward the center of interpolated to TC cylindrical coordinates with an azi- the TC (Lee et al. 2000). Finally, MSLPs were estimated muthal resolution of ;0.78 and a radial resolution of by adding each SLP deficit to the corresponding SLP 500 m (step 4). After the retrieval of Vt by the GBVTD observation (step 6). When an SLP observation point technique (step 5), the Vt data were substituted into the was located within 15 km outside the GBVTD’s outer- gradient wind equation: most ring, MSLPs were estimated by extrapolating the ! SLP gradient over the distance from the observation V2 ›p point. As many MSLPs were estimated as there were r t 1 f V 5 s , (1) s r t ›r observations. Note that the DR method assumes that asymmetric components of the pressure distribution where r is the radius from the TC center, f is the Coriolis have little impact on the estimation results. parameter corresponding to the latitude of each radar To stably estimate MSLPs for operational purposes, r location, ps is the axisymmetric SLP, and s is the envi- we modified the original DR method in four primary r ronmental air density at sea level. The value of s was ways. First, the GBVTD-simplex center-finding method 2 assumed to be 1.15 kg m 3, which corresponds to the was modified by restricting the center-finding area to

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1827

FIG. 2. The procedures used for the Doppler radar intensity estimation. Input and output data for each step are indicated in italics. Before (since) 2009, the estimate was performed at 10-min (5 min) intervals.

within a 5-km radius of the first-guess point and con- problem, VM? must be obtained from an independent fining the first-guess RMW to that associated with the source. The hurricane volume velocity processing primary eyewall (hereafter P-RMW). These modifica- (HVVP) technique, which provides VM? by using a tions enabled a reasonable center to be identified single-Doppler radar (Harasti 2014; Chen et al. 2013), around the guess point at all times; if there is no limi- is one possibility, but the results of our preliminary tation placed on the simplex, it is difficult to detect a examination showed that from an operational per- reasonable center for a dissipating TC or a TC with poor spective the use of the HVVP technique is not suitable radar coverage, as described by Murillo et al. (2011).In for the stable provision of VM?, owing to occasional addition, the latter modification enabled more accurate poor radar coverage (not shown). Hence, we decided to Vt to be retrieved around the P-RMW. This was im- use the cross-beam component of the TC translational portant because the SLP gradient calculated from Eq. speed as a proxy for VM? to resolve the aliasing prob- (1) is sensitive to the magnitude of Vt around the P- lem. The effect of using this proxy is described in sec- RMW. Second, we filled in missing values of Vt by tions 3b and 4a. Fourth, we introduced two QC using a spline function with the assumption that procedures into the GBVTD retrieval: 21 Vt 5 0ms at the TC center. However, when the 1) Retrieved Vt values were rejected when associated spline function produced a negative Vt or an artificial asymmetric tangential winds with an amplitude of 21 local Vt maximum in the eye region, we applied a more than 25 m s were obtained, because such a quadratic function instead of the spline function. In this large asymmetry is, in general, unlikely for TCs and way, the radial distribution of Vt within the eye region likely to be a retrieval failure because of, for exam- was deduced (Fig. 3). Third, we used the TC trans- ple, noise contamination. lational speed as a proxy for VM? in the GBVTD 2) To assess the accuracy of the GBVTD-retrieved technique. One weakness of the GBVTD technique is winds, we examined the root-mean-square difference that, in principle, it cannot retrieve the component of (RMSD) between the VD resampled from the the mean environmental wind perpendicular to a line GBVTD-retrieved winds and the observed VD at each connecting the TC center and the radar location radius, following Zhao et al. (2012). If the RMSD was 2 (hereafter, the cross-beam component of the mean more than 8 m s 1 inside the P-RMW and the TC 21 environmental wind, VM?). Instead, VM? is aliased into had a maximum absolute VD of more than 60 m s Vt, which leads to a decrease in the retrieval accuracy with a single, closed eyewall, the wind retrieval of Vt at outer TC radii (Lee et al. 1999; Harasti et al. accuracy was extremely poor, and the corresponding 2004; Chen et al. 2013). To resolve this aliasing Vt values were rejected.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1828 MONTHLY WEATHER REVIEW VOLUME 144

3. Simulation experiment To assess the accuracy with which the DR method could estimate MSLPs of TCs of known intensity, we used data obtained by numerical simulation. In particular, we verified two points of the procedures. First, we confirmed that the use of 2-km Vt and surface air density in the gradient wind equation would yield reasonable MSLPs. Second, we confirmed that we could improve retrieval accuracy by using TC motion as a proxy for the mean wind to correct an inherent bias in GBVTD-retrieved Vt. Here, we describe the settings of the simulation and present the estimation verification results.

a. Simulation settings We performed simulations of Typhoon Bolaven (2012), Typhoon Neoguri (2014), and Typhoon Vongfong (2014) with the JMA Nonhydrostatic Model (JMA-NHM; Saito et al. 2006, 2007). The physical parameterizations were mostly the same as those used by JMA’s operational mesoscale model [see Japan Meteorological Agency (2015) for details]. The horizontal grid spacing of the simulation was 2.5 km, and results were output at 5-min intervals.

FIG. 3. Radial profiles of retrieved Vt, interpolated Vt, and SLP For simplicity, we constructed pseudo-VD data at anomalies (with the sign reversed) for T1215. SLPs observed at the 2-km altitude corresponding to each radar location from Nago and Naha weather stations, which were used for the MSLP the simulated winds at 2-km altitude without consider- estimation, are shown along with the estimated MSLPs. MSLPs ing the effects of radar beamwidth and zenith angle. from Naha were estimated by extrapolating the SLP gradient over the distance from Naha. We regarded only grid points with a nonzero mixing ratio of rain, snow, and graupel at 1-km altitude as the grid points where VD at 2-km altitudes were observed, There are other limitations inherent in the GBVTD because a check for the presence of nonzero mixing technique. First, the GBVTD technique has a weakness ratios at 2-km altitude did not provide a sufficient VD in that the wavenumber-1 tangential wind can bias Vt via distribution. As in the real TCs, there were no pseudo- the nonlinear azimuthal coordinate system employed by VD values around the TC center. Although the grid of the technique, particularly when the radar is around the the simulation was coarser than the radar resolution radius of the GBVTD analysis ring [see Fig. 13a in Lee (Table 1), this was unlikely to be a serious problem for the et al. (1999)]. This can lead to poor MSLP estimations results because the resolution of the radar data was re- especially when the distance between the TC center and duced through the averaging and interpolation pro- the weather station whose sea level pressure is used as an cedures. The tracks of the simulated TCs deviated from anchor of the pressure measurement is close to that from the observed tracks, so we adjusted the simulated tracks the radar to the TC center. Second, the GBVTD- by displacing them parallel to the observed tracks. In simplex center-finding algorithm cannot detect centers addition, to increase the number of estimated cases, we when the radar is inside the P-RMW. In this case, we prepared two virtual cases by parallel displacement of the used the first-guess center positions as centers although simulated data. Because the simulated TCs were over the in most of the cases the GBVTD technique also failed ocean, where there is no topography effect, this dis- to retrieve Vt because of there being few scatterers in placement presents no disadvantage for the experiment. the eye region. Fortunately, in this study, this case ap- In total, we prepared six simulated TCs in this experiment plied to only T0813 and T1217. Third, the GBVTD (Fig. 4, Table 3). technique assumes that radial winds are much weaker Simulated Typhoon Bolaven (2012) had an RMW of than tangential winds. When this assumption is violated, ;46 km, about 3 times the observed RMW, and a typical the retrieval accuracy of Vt decreases. TC structure. The simulated track was almost the same

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1829

and in the S1408B case, the coverage of the Okinawa radar was limited to its eastern part. In addition, we prepared a third track for the simulated Neoguri displaced northeastward (0.53798Nand1.04158E) from the simulated track; this S1408C case (yellow line in Fig. 4) had better coverage from the Okinawa radar than that of the S1408B case. Simulated Typhoon Vongfong (2014) (S1419) had a disappearing eyewall. Because such TCs frequently approach Japan, it was important to verify how accurately the DR method could estimate MSLPs of such TCs. In this experiment, we were not concerned with how well the numerical simulation could reproduce the observed TC structures, including their vertical structures. For simplicity, we do not address the problem of the accuracy of the center-finding algorithm here. So, this FIG. 4. Tracks of the six simulated TCs used in the preliminary experiment. Black circles indicate the ranges, ;200 km, of the experiment started from step 4 in Fig. 2. TC centers used Ishigakijima and Okinawa radars (red dots). Each color represents by the GBVTD technique were centroid centers calcu- one simulated case. lated by applying a 1–2–1 smoothing filter 500 times (Nolan et al. 2009). Note that in this experiment, the true as the observed track, but the TC location was approx- MSLP was not the SLP at the centroid center but the imately 2 h ahead of the observation. The intensity of the lowest SLP within the eye region in the simulated TCs. We simulated Bolaven was estimated for two cases: along extracted pseudo-SLP observations at points correspond- the actual track of the simulated TC (S1215A) and ing to surface weather stations from each simulation. along a second track displaced southwestward from b. Results and verification S1215A (S1215B, blue line in Fig. 4). Simulated Typhoon

Neoguri (2014) had an RMW of ;60 km, which was We performed MSLP estimates using pseudo VD at larger than the observed RMW of 35–40 km. Because the 5-min intervals. Table 4 shows the estimates in the pre- simulated track of Neoguri deviated southwestward from liminary experiment compared with the true values in the observed track, the location of the simulated Neoguri the simulated TCs. The RMSE of the estimated MSLPs was displaced northeastward (0.53798N and 0.44158E). relative to the true MSLPs was 2.84 hPa, and the bias The intensity of the simulated Neoguri was estimated for was 20.77 hPa, although individual results could have a both T1408A (observed from the Ishigakijima radar, large effect because of the small number of TCs in- S1408A) and T1408B (observed from the Okinawa radar, cluded. The incorporation of the cross-beam component S1408B). In the S1408A case, the coverage of the Ishi- of TC motion as a proxy for VM? helped to improve gakijima radar was limited to the western part of Neoguri, the biases of the TCs. For example, the bias of S1408B

TABLE 3. The six TCs used in the simulation experiment.

Name (radar site) Outline Model initial time Estimation period S1215A (Okinawa) A simulation of T1215 0600 UTC 25 Aug 2012 0200–1500 UTC 26 Aug 2012 S1215B (Okinawa) Same as S1215A, but the track was 0600 UTC 25 Aug 2012 0200–1500 UTC 26 Aug 2012 displaced southwestward (0.48S and 0.88W) from S1215A S1408A (Ishigakijima) A simulation of T1408; the track was 1200 UTC 6 Jul 2014 1900 UTC 7 Jul–0110 UTC displaced northeastward 8 Jul 2014 (0.53798N and 0.44158E) from the simulated track S1408B (Okinawa) Same as S1408A, but for the 1200 UTC 6 Jul 2014 0100– 0410 UTC 8 Jul 2014 estimation period S1408C (Okinawa) A simulation of T1408; the track was 1200 UTC 6 Jul 2014 0100–0700 UTC 8 Jul 2014 displaced northeastward (0.53798N and 1.04158E) from the simulated track S1419 (Okinawa) A simulation of T1419; the track was 0000 UTC 11 Oct 2014 0300–1200 UTC 11 Oct 2014 displaced northward (0.14978N and 0.01118E) from the simulated track

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1830 MONTHLY WEATHER REVIEW VOLUME 144

TABLE 4. Estimation accuracy of the six TCs in the simulation experiment. The P-RMW was determined from the retrieved Vt values after step 5 in Fig. 2 within a possible range of the subjectively determined P-RMW. The P-RMW values and central pressures were averaged over the estimation periods shown in Table 3.

Central pressure (hPa) N RMSE (hPa) Bias (hPa) Correlation P-RMW (z 5 2 km) (km) Estimated True

Overall with VM? 931 2.84 20.77 0.95 — — — Overall without VM? 934 4.93 20.76 0.90 — — — S1215A with VM? (without VM?) 251 (253) 1.65 (2.55) 20.25 (1.78) — 46.00 933.61 933.86 S1215B with VM? (without VM?) 230 (231) 2.13 (2.84) 20.82 (21.78) — 45.54 932.55 933.37 S1408A with VM? (without VM?) 134 (134) 3.06 (3.32) 21.50 (2.14) — 62.34 929.46 930.97 S1408B with VM? (without VM?) 78 (78) 4.24 (9.32) 23.55 (28.54) — 61.33 924.96 928.51 S1408C with VM? (without VM?) 146 (146) 2.95 (7.24) 21.96 (25.70) — 60.75 926.22 928.18 S1419 with VM? (without VM?) 92 (92) 4.47 (5.94) 3.28 (5.06) — 62.57 950.10 946.81 was improved by ;5 hPa, from 28.54 to 23.55 hPa, some estimated MSLPs, most data points are concen- though the negative bias was not eliminated. Figure 5a trated in the vicinity of the 1:1 line (Fig. 6). The RMSE, shows the axisymmetric radial distributions of the re- bias, and correlation coefficient between the estimated trieved and true Vt in S1408B averaged over the esti- and best track MSLPs were 8.37 hPa, 1.51 hPa, and mation period. The difference between the retrieved 0.87, respectively. The MSLP estimate error was within and true Vt values was quite small on average, except 65(610) hPa in 60.3% (84.1%) of all estimates. In near the TC center, where the retrieved Vt were in- contrast, for the best track MSLP of 960 hPa, the esti- terpolated by a spline function, resulting in slightly weak mated MSLPs were distributed from 975 to 908 hPa. estimates, but the difference had little effect on the SLP Most of these large differences were attributed to T1007, field (not shown). Figure 5b shows the axisymmetric which is discussed in the next subsection. radial distributions of SLPs in S1408B averaged over the The use of the cross-beam component of translational estimation period. The differences between the true and speed as a proxy for VM? in the GBVTD technique was deduced SLPs show that at outer radii of more than remarkably effective. The MSLPs estimated using the 50 km, the pressure gradient was slightly steeper than VM? proxy had RMSEs smaller than 94% and positive the true gradient. In contrast, the pressure gradient was biases smaller than 55% of those of the MSLPs esti- slightly shallower than the true gradient within the mated without using the VM? proxy (Table 5). Although RMW. On the whole, a negative bias was evident. In for most of the TCs, biases were positive (not shown), S1419, the effect of the spline interpolation within the incorporation of VM? in the GBVTD technique gener- eye region led to a positive MSLP bias (not shown). The ally reduced this positive bias. This result is consistent trends of the pressure distributions were similar in with the results of the preliminary simulation experi- the other cases. Thus, the results of the preliminary ex- ment (section 3). periment confirmed that the DR method using VD data Therefore, MSLPs can be reasonably estimated from at 2-km altitude and the cross-beam component of TC VD at 2-km altitude by using TC motion as a proxy for motion as a proxy for VM? could reasonably reproduce VM?. Differences in the accuracy of the estimates under MSLPs, as reported by Harasti et al. (2004). several specific conditions are discussed in section 5b. b. Classification of the individual cases 4. Estimation results We grouped the MSLP estimates of the 28 cases into In this section, we present the estimation results for the following four groups: excellent MSLP estimates, the real TC cases. We first describe the overall statistical large differences between the DR and best track MSLPs, a verification results; then, we classify the TC cases into consistent positive bias relative to the best track, and er- four groups according to the estimation results and de- ratic fluctuations. scribe the characteristics of the individual cases in each We defined an estimate as excellent if the RMSE was of the four groups in detail. less than 4 hPa and the bias was within 64 hPa. Ten cases met these criteria: T0709, T0813, T1115, T1215, T1307, a. Overall statistical verification T1312, T1323, T1408A, T1408B, and T1411B. We show A scatter diagram of the estimated and best track the characteristics of T1312 in Fig. 7 as an example of MSLPs in the 28 cases shows that, despite large errors in an excellent estimate. The estimated MSLPs and their

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1831

FIG. 5. Radial distributions of (a) Vt and (b) axisymmetric SLP anomalies of S1408B averaged over the estimation period. In (a) the black line is the true Vt, the green line is the retrieved Vt, and the red line is the difference between them. In (b) the black line represents the true SLP anomalies, the green line the deduced SLP anomalies, and the red line the difference between them. trend correlated well with those of the best track (black line in Fig. 8a), and the Dvorak MSLP was con- (Fig. 7a). Figures 7b and 7c show the Doppler velocity sistently 947 hPa (dashed blue line in Fig. 8a) during the field and radial distribution of the retrieved axisym- estimation period. Considering the fact that the Vt at metric tangential wind and axisymmetric SLP deficit at around the RMW (;12 km) changed from ;63 to 2 2225 UTC 20 August 2013. Although radar coverage ;47 m s 1 during the estimation period and the ob- was not very dense, the MSLPs estimated from the SLPs served SLP was 965.1 hPa around the RMW during the at both Miyakojima (49.3 km from the TC center) and second half of the period, the MSLPs obtained by the Ishigakijima (107.5 km from the TC center) were almost the same as the best track MSLP at that time. The accuracies of all 10 cases are listed in Table 6. One of our most important findings here is that the DR method could provide almost the same MSLPs as the best track in real time, whereas the accuracy of the estimates made by the Dvorak method was not necessarily good. For T1007, large differences were found between the DR and best track MSLPs (Fig. 8). In the first half of the estimation period, T1007 had a negative VD minima of 21 about 270 m s and a positive VD maximum of about 2 55 m s 1 near the 12-km radius at 2-km altitude (Fig. 8b). 21 The DR method retrieved Vt greater than 60 m s , and the estimated MSLPs were around 930 hPa on average (Figs. 8a,c), although they fluctuated somewhat erratically. As the TC approached Okinawa Island, Vt weakened rapidly. At the time of the TC’s nearest approach to Nago on Okinawa Island, when the SLP was 965.1 hPa (Fig. 8a, blue star), Nago was located just within the RMW (;13 km from the TC center) and Vt 2 was 47 m s 1. The DR method estimated MSLPs that FIG. 6. Scatter diagram comparing the estimated MSLPs with the were more than 10 hPa below the SLP at Nago. By best track MSLPs. The corresponding RMSE, bias, and correlation contrast, the best track MSLP was consistently 960 hPa coefficient R are also shown.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1832 MONTHLY WEATHER REVIEW VOLUME 144

TABLE 5. Overall accuracy of the DR intensity estimates for the 28 TC cases. The results are shown both with and without incorporation of VM?. The overall accuracy excluding T1007 is given in parentheses.

N RMSE (hPa) Bias (hPa) Correlation

Overall with VM? (excluding T1007) 5574 (5317) 8.37 (7.13) 1.51 (2.41) 0.87 (0.91) Overall without VM? (excluding T1007) 5604 (5347) 8.94 (8.14) 2.77 (3.57) 0.86 (0.90)

DR method seem more plausible than the best track T1418C). T1418A showed a significant positive bias values. This suggests that the DR method can capture during the last 2 h of the estimation period (around some TC intensity changes on a short time scale (less 0300 UTC 5 October 2014) compared with the prior and than 6 h) that the conventional methods fail to capture. subsequent estimates. This bias was caused by poor ra- If T1007 is removed from the overall statistical verifi- dar coverage (not shown). Except for that period, cation described in section 4a, the RMSE improves to however, the DR estimates of Typhoon Phanfone seem 7.13 hPa (Table 5). Other cases in this group, with sim- reasonable because the retrieved Vt structure corre- ilarly large differences between the DR and best track sponds well to the VD distribution and because Vt values MSLPs, were T1204 and T1419 (not shown). were retrieved in the eye region (Figs. 9b,c). One pos- The estimates of five TC cases showed consistent sible reason for the consistent positive biases of the DR positive biases: T1109, T1217, T1318, T1418A, and MSLPs, compared with the best track, is that if the T1418B. Figure 9a shows the time evolution of the wavenumber-1 asymmetry is caused by a vortex Rossby MSLPs of Typhoon Phanfone (T1418A, T1418B, and wave and then it may be stationary near the eye region

FIG. 7. (a) Time evolutions of the DR MSLPs (color lines), best track MSLPs (black line), and Dvorak MSLPs (blue dashed line) for T1312. The vertical black line indicates the time of the radar observation shown in (b). The red line indicates MSLPs derived from SLPs at Miyakojima, the green line indicates MSLPs derived from SLPs at Ishigakijima, and the blue line indicates MSLPs derived from SLPs at Iriomotejima. (b) Doppler radial velocity of the 2-km CAPPI observed from the Ishigakijima radar (red dot) at 2225 UTC 20 Aug 2013. The circles show the radii at 10, 30, 60, 90, 120, and 150 km from the TC center. (c) As in Fig. 3, but for T1312 at 2225 UTC 20 Aug 2013.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1833

TABLE 6. Accuracy of the 10 cases with excellent estimation results. The P-RMW was determined from the retrieved Vt values after step 5inFig. 2 within its possible range used in the GBVTD-simplex center-ﬁnding algorithm. The P-RMW values and central pressures were averaged over the estimation periods shown in Table 2.

Central pressure (hPa) Case (name, alias) N Estimated Best track Dvorak RMSE (hPa) Bias (hPa) P-RMW (z 5 2 km; in km) T0709 (Fitow) 226 974.76 974.55 974.44 2.21 0.22 49.12 T0813 (Sinlaku) 106 980.72 981.12 981.92 2.17 20.41 18.83 T1115 (Roke) 49 951.05 949.12 926.86 3.77 1.94 23.19 T1215 (Bolaven) 314 931.66 929.12 948.46 3.88 2.54 14.07 T1307 (Soulik) 203 948.19 948.02 965.00 3.79 0.17 30.89 T1312 (Trami) 299 964.87 965.58 973.00 2.30 20.70 40.64 T1323 (Fitow) 325 963.32 961.33 968.13 2.84 2.00 73.78 T1408 (Neoguri, T1408A) 141 940.33 940.00 947.00 3.27 0.33 39.99 T1408 (Neoguri, T1408B) 78 939.70 940.37 947.00 3.12 20.67 35.21 T1411 (Halong, T1411B) 260 956.29 955.32 972.63 3.00 0.97 47.22

(e.g., Lamb 1932), it can potentially cause a constant bias inside the P-RMW, there is no assurance that asym- in the GBVTD-retrieved Vt (see section 2c), resulting in metric radial winds are negligible, and the wind retrieval consistent MSLP biases. However, because this possi- accuracy would not be good. The estimated MSLPs of bility can lead to both positive and negative biases and T1007 in the first half of the estimation period (Fig. 8a) because there was no case with a consistent negative and of T1324A (Fig. 10b) fluctuated with an amplitude bias, it may not be the reason for the consistent positive of 5–15 hPa for this reason, because in these two cases, biases. Another possible reason is that the radial SLP RMSDs between the VD resampled from the GBVTD- structure inside the P-RMW assumed in the best track retrieved winds and the observed VD were sometimes analysis was different from that deduced from the DR quite large inside the P-RMW. Although the QC pro- method. Table 7 shows the accuracies and the P-RMWs cedures described in section 2c remove retrieved winds of the TCs in this group. Except for T1318, the P-RMW with large RMSDs, the results can be problematic. was more than 70 km. When the P-RMW is large, the When too many retrieved winds were rejected by the effect of the difference between the SLP structures in- QC procedure, the estimated MSLPs tended to be too side the P-RMW assumed in the best track analysis and high. With insufficient rejection, the retrieved Vt tended deduced from the DR method on MSLP estimates may to be too strong in the eye region, and the MSLPs were be large. Interestingly, the DR MSLPs were better too low. As a result, the estimated MSLPs showed erratic correlated with the Dvorak MSLPs than with the best fluctuations. The third reason is illustrated by T1411A track MSLPs in this group (Fig. 9a, Table 7). Consider- (Fig. 10c), for which the estimated MSLPs fluctuated vi- ing these findings, the estimated MSLPs of these cases olently between two extremes. When the radar coverage characterized as having consistent positive biases may was sufficient for wind retrieval around the disappearing not be wrong, although we do not have true MSLPs with eyewall, the estimated MSLPs were about 950 hPa, which which to compare them. is consistent with the best track. By contrast, poor radar The estimated MSLPs fluctuated erratically in some coverage around the disappearing eyewall led to MSLPs cases, as they also did in some results from the simula- of around 965 hPa. Fourth, the aliasing of wavenumber-2 tion experiment (not shown). According to Harasti et al. radial winds into Vt in the GBVTD retrieval formula (see (2004), unreasonable TC center finding by the simplex details in Lee et al. 1999; Murillo et al. 2011) probably method is one reason for such erratic fluctuations, and caused wavy fluctuations with a period of a couple of we identified four additional reasons for erratic fluctu- hours, as seen in T1204, T1215, T1411B (Fig. 10c), and ations. First, noise contamination caused by failure of T1411C (Fig. 10c). If wavenumber-2 radial winds are the dealiasing correction sometimes resulted in violent dominant within the eye region and their distribution fluctuations with amplitudes of 5–15 hPa, for example, in moves around the eyewall cyclonically together with T1011 (Fig. 10a), T1216, and T1324B (not shown). In mesovortices (e.g., Kossin and Schubert 2004; Braun et al. those cases, noisy data in small areas were not fully re- 2006) or vortex Rossby waves (e.g., Montgomery and moved by the automatic processes. Second, in some Kallenbach 1997; Wang 2002), the radial winds could bias cases the GBVTD assumption that radial winds are the retrieved Vt with a period of a few hours. much weaker than tangential winds was apparently not Given these characteristics of our estimation results, valid inside the eyewall. If there are active mesovortices we recommend using a running mean of DR MSLPs of a

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1834 MONTHLY WEATHER REVIEW VOLUME 144

FIG. 8. (a) As in Fig. 7a, but for T1007. The blue star is the SLP (965.1 hPa) observed at Nago, 13.2 km from the TC center. [The vertical black line indicates the time of the observation shown in (b).] The green line indicates MSLPs derived from SLPs at Naha, and the red line indicates MSLPs derived from SLPs at Nago. (b) Doppler radial velocity of the 2-km CAPPI observed from the Okinawa radar (black dot) at 0245 UTC 31 Aug 2010. The circles show the radii at 10, 30, 60, 90, 120, and 150 km from the TC center. (c) As in Fig. 3, but for T1007 at 0245 UTC 31 Aug 2010. few hours. Additionally, by paying attention to the RMSE of Dvorak MSLPs is 11.7 hPa. According to characteristics of each group, analysts should be able to Oyama (2014), Advanced Microwave Sounding Unit utilize those MSLPs with higher accuracy and reliability (AMSU) MSLPs in the western North Pacific have an than is suggested by the overall statistical verification RMSE of 10.1 hPa relative to the JMA best track, and results described in section 4a. according to Velden et al. (2007), the RMSE of the AMSU method using the Cooperative Institute for Meteorological Satellite Studies (CIMSS) algorithm 5. Discussion is 7.5 hPa and that using the Cooperative Institute for We discuss the utility and feasibility of the operational Research in the Atmosphere (CIRA) algorithm is use of DR intensity estimates compared with conven- 10.3 hPa, relative to aircraft reconnaissance observational estimation methods and we examine additional tions. These error statistics indicate that the accuracy statistical verification results. of the MSLPs obtained by our DR method is comparable to or better than the accuracies of the Dvorak a. Comparison with conventional methods and AMSU methods. We stress that the DR intensity Koba et al. (1990) have reported that the error range estimates are totally independent of the past best of Dvorak MSLPs in the western North Pacific is 7– track. Therefore, we can now obtain plausible MSLPs 19 hPa, values that are almost equivalent to the RMSE; in real time by a method other than aircraft re- Martin and Gray (1993) have reported that the standard connaissance observations that is different from that deviation of Dvorak MSLPs in the western North Pacific used to obtain the best track. The use of DR estimates is 9 hPa; and Velden et al. (2007) have reported that the would be a new paradigm of TC intensity analysis in

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1835

FIG. 9. (a) As in Fig. 7a, but for T1418. The red star is the SLP (962.1 hPa) observed at Shionomisaki, which was 6.2 km from the TC center. [The vertical black line indicates the time of the observation shown in (b).] (b) The Doppler radial velocity of 2-km CAPPI observed from the Tanegashima radar (black dot) at 0540 UTC 5 Oct 2014. The circles show the radii at 10, 30, 60, 90, 120, and 150 km from the TC center. (c) As in Fig. 3, but for T1418 at 0540 UTC 5 Oct 2014. the western North Pacific, where aircraft observations uses the same weather station observations, and weather are not available. stations are near the radar locations. The RMW was determined from all of the retrieved b. Additional statistical verification Vt values after step 5 in Fig. 2. For TCs with an RMW of Aiming for the effective operational use of the DR 20–70 km, the estimated central pressures had an RMSE intensity estimates, we further examined the estimation of 5.55 hPa and showed a bias of 0.69 hPa. In contrast, for accuracy of the MSLPs with respect to several specific TCs with an RMW of 75–120 km, the estimates had a conditions (Fig. 11). Note, however, that because we large positive bias of 5.23 hPa, as indicated in section 4b. used only 22 TCs for the estimations, the statistical re- Large RMWs greater than 120 km may contain false sults for some specific conditions may be dependent on RMWs. Bell and Lee (2012) stated that missing data extreme cases. The RMSEs were, in general, smaller outside the radar range can cause a false RMW at a large when the distance between the TC center and the radar radius. Thus, large errors accompanying large RMWs location was shorter, and the estimates exhibited a may be due in part to the false RMWs and, more spe- positive bias when the distance was large. The positive cifically, the lack of radar data. bias may result from the CAPPI method, as described in Because the above results suggest that the pressure section 2c. Similarly, RMSEs were smaller when the gradients retrieved by the DR method have a systematic distance between the TC center and the weather station error, we examined the relationship between the obser- whose sea level pressure was used as an anchor for vational SLP gradient when there were multiple SLP ob- pressure measurement was shorter. One reason for this servations around a TC and the corresponding retrieved similarity is likely the fact that the best track analysis SLP gradient. Figure 12 shows the result schematically:

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1836 MONTHLY WEATHER REVIEW VOLUME 144

TABLE 7. As in Table 6, but for the ﬁve cases with consistent positive biases.

Central pressure (hPa) Case (name, alias) N Estimated Best track Dvorak RMSE (hPa) Bias (hPa) P-RMW (z 5 2 km; in km) T1109 (Muifa) 243 951.95 945.32 956.00 6.94 6.64 71.68 T1217 (Jelawat) 148 949.92 931.89 947.08 18.87 18.03 85.80 T1318 (Man-yi) 140 968.34 960.00 975.75 8.81 8.34 45.10 T1418 (Phanfone, T1418A) 189 957.21 943.18 953.68 14.99 14.03 92.22 T1418 (Phanfone, T1418B) 167 957.11 945.50 956.00 11.88 11.61 85.55

the retrieved SLP gradient was on average 0.55 hPa was denser and wind retrieval accuracy was better steeper than the observational SLP gradient over a (Figs. 11d,e). TC analysts should consider these features 44.3-km interval. This result is apparently inconsistent with when the DR method is used operationally. the positive biases of the MSLPs described above. How- c. Utility of the DR intensity estimation method ever, given that most of the weather stations were outside of the RMW, these ﬁndings suggest that the SLP gradient Estimation of MSLPs by the DR method has four retrieved from outside the RMW is steeper than the true advantages. First, the DR method can provide MSLPs gradient, and that the SLP gradient inside the RMW, where a few hours before a TC approaches a populated area. some of the Vt were interpolated by using a spline function, Thus, with reliable nowcasting, a meteorological agency is shallower than the gradient inferred from the best track. will be able to issue effective warnings to mitigate the These features are consistent with those of the average SLP TC’s impact. Second, the DR method can provide gradient in the preliminary experiment (Fig. 5b). The radial MSLPs at 5-min intervals with an accuracy compa- SLP gradient error is attributed to 1) Vt retrieval errors rable to or better than that of conventional methods. associated with the use of the TC motion as a proxy for VM? As a result, analysts will be able to nowcast a TC at and several limitations inherent in the GBVTD tech- shorterintervalsthanispossible with conventional nique, 2) the improper use of a value for surface air methods, which allow nowcasts at 6-h intervals or density that is a little too high and Vt at 2-km altitude in several times per day. Third, the DR method can, in the gradient wind equation (1), and 3) an interpolation principle, estimate the intensity of any TC, as long as error by the spline function. The development of a method radar coverage is sufﬁcient. By contrast, the Dvorak for interpolating Vt within the TC eye in a plausible way technique cannot handle rapid intensity changes, and is a topic for future work. the AMSU techniques, because of the coarse reso- In addition, the estimated MSLPs tended to be more lution of the instrument, are not suitable for esti- consistent with the best track MSLPs when radar coverage mating the MSLPs of a TC with a small eye (Oyama

FIG. 10. As in Fig. 7a, but for (a) T1011, (b) T1324A, and (c) T1411.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1837

FIG. 11. Scatter diagrams comparing the difference between the estimated MSLPs and the best track MSLPs (y axis) with (a) the distance between the radar location and the TC center, (b) the distance between the weather station and the TC center, (c) the RMW at 2-km altitude, (d) radar coverage, and (e) the GBVTD wind retrieval accuracy (x axis). The RMWs were determined from all the retrieved Vt values from step 5 in Fig. 2. The radar coverage is defined as the radial average of the maximum azimuthal gap (rad) at each radius on the GBVTD- specified coordinate system. The wind retrieval accuracy is defined as the overall average of the RMSD between the VD resampled from the GBVTD-retrieved winds and the observed VD, identical to the RMSE shown in Fig. 3 of Zhao et al. (2012). The error bars show the bias and the RMSE within the range indicated by the double- headed arrows.

2014). Fourth, in recent years the DR method could have However, DR intensity estimates also have some been applied operationally to four to ﬁve TCs each year limitations. Compared with the lifetime of a TC, the according to Table 2. In particular, TCs approaching the estimation period is extremely limited—only 9 h on av- vicinity of the Okinawa Islands, where there is no topo- erage. The DR method also uses some parameters, such graphical effect on the TC structure, are ideally suited for as initial-guess centers, the P-RMW, and its possible GBVTD retrievals. Therefore, given this TC frequency, it range, that must be subjectively determined beforehand. will be beneﬁcial in Japan to utilize the DR method op- In addition, the method cannot estimate MSLPs unless erationally to improve the accuracy of TC intensity ana- VD can be successfully dealiased. Furthermore, the lyses in real time. GBVTD technique has several inherent limitations, as

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1838 MONTHLY WEATHER REVIEW VOLUME 144

can be estimated from the Doppler velocities with high accuracy. A set of procedures developed by Lee et al. (1999, 2000), Lee and Marks (2000), and Bell and Lee (2012) for estimating TC intensity by using the GBVTD technique to retrieve axisymmetric tangential winds from Doppler velocities has shown promising results. In this study, we evaluated the accuracy of this method by estimating the MSLPs of 28 TC cases; identiﬁed some limitations, strengths, and weaknesses of the method; and evaluated its utility for operational use. Before estimating the intensity of real TCs, we performed a preliminary experiment using pseudo-Doppler velocities obtained from numerical simulations to con-

FIG. 12. Schematic diagram of retrieved and observed SLP gra- firm that the set of intensity estimation procedures could dients. The values were calculated when data from multiple SLP provide reasonable estimates of TC central pressures. observations around the TCs were available. To obtain as many In particular, we assessed the effect of using the cross- SLP gradients in the inner region of TCs as possible, when three or beam (normal to the line connecting the radar with more SLP observations were available, only SLP gradients relative the TC center) component of the mean environmental to the innermost observation point were calculated. wind,VM?, assumed from the TC translational speed, to correct for the aliasing effect of the GBVTD algorithm. described previously. However, it is possible to make The results showed that when Doppler velocities at 2-km use of bias information between the Dvorak estimates altitude and the assumed VM? were used, the estimated and the DR estimates even for the period before and MSLPs had an RMSE of 2.84 hPa and a bias of 20.77 hPa. after the DR estimation period to correct the Dvorak In addition, the incorporation of VM? assumed from the estimates. The dealiasing problem should be addressed TC motion in the GBVTD algorithm decreased the bias by improving dealiasing and noise-removal methods. and improved the accuracy of the estimate very well. We Because some other techniques for relaxing the limita- also confirmed that the DR method could reasonably tions of the GBVTD technique have been proposed estimate MSLPs by using Doppler velocities at 2-km (e.g., Jou et al. 2008; Wang et al. 2012; Chen et al. 2013), altitude. it is possible to further improve the DR method, which On the basis of these preliminary results, we used should be addressed in future studies. Doppler velocity data at 2-km altitude and the cross- beam component of the TC translational speed as VM? to obtain DR intensity estimates for 28 TC cases. 6. Summary and conclusions Compared with the best track MSLPs of RSMC Tokyo, It is of great importance to analyze TC intensity with the RMSE, bias, and correlation were 8.37 hPa, 1.51 hPa, high accuracy, particularly for TCs approaching popu- and 0.87, respectively. The MSLP estimation error was lated areas. However, conventional methods of TC in- within 65(610) hPa for 60.3% (84.1%) of the total tensity estimation, such as the Dvorak technique, which number of estimates. This estimation accuracy is com- estimates TC MSLPs and maximum sustained wind parable to or better than the accuracies of the Dvorak speeds from cloud patterns, and the AMSU technique, and AMSU methods. Additionally, we confirmed that which estimates MSLPs from the magnitude of the the accuracy was poor if the VM? was not incorporated upper-level warm-core anomaly retrieved from satellite into the GBVTD algorithm. Two distance metrics are microwave sounding data, have limitations with regard defined: 1) the distance between the TC center and the to analysis accuracy. Furthermore, observing MSLPs radar location and 2) the distance between the TC in a straightforward way, such as by using dropsondes, is, center and the surface weather station whose sea level for now, not a practical solution in the western North pressure was used as an anchor for pressure measure- Pacific. However, as a result of the JMA upgrading all of ment. The MSLPs estimated here tended to be more its operational radars to Doppler radars from 2006 to consistent with the best track MSLPs when the distance 2013, Doppler velocity observations are available for the metrics were shorter, when the wind retrieval accuracy many TCs that have passed near these radars. If axi- was better, and when the radar coverage was denser. symmetric tangential winds retrieved from the Doppler For TCs with an RMW of 20–70 km, the estimated velocities and the gradient wind equation can reason- MSLPs had an RMSE of 5.55 hPa and showed a bias of ably be used to infer a TC’s pressure field, then MSLPs 0.69 hPa relative to the best track. Examination of cases

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC MAY 2016 S H I M A D A E T A L . 1839 with large differences between the best track and DR tropical cyclone intensity and size estimation algorithm. J. Appl. method MSLPs suggested that the DR method can capture Meteor., 43, 282–296, doi:10.1175/1520-0450(2004)043,0282: . TC intensity changes over a shorter time scale (less than EOAMSU 2.0.CO;2. Dvorak, V. F., 1975: Tropical cyclone intensity analysis and fore- 6 h) than is possible with the conventional methods. This casting from satellite imagery. Mon. Wea. Rev., 103, 420–430, result suggests that the DR method can improve not only doi:10.1175/1520-0493(1975)103,0420:TCIAAF.2.0.CO;2. real-time intensity analyses, but also best track analyses. ——, 1984: Tropical cyclone intensity analysis using satellite data. Our results confirm the feasibility of using DR in- NOAA Tech. Rep. 11, 45 pp. tensity estimates operationally. TCs for which the DR Franklin, J. L., M. L. Black, and K. Valde, 2003: GPS drop- windsonde wind profiles in hurricanes and their operational method is applicable have approached Japan four to implications. Wea. Forecasting, 18, 32–44, doi:10.1175/ five times per year recently. Therefore, we believe that 1520-0434(2003)018,0032:GDWPIH.2.0.CO;2. adopting the DR method, in addition to the Dvorak Harasti, P. R., 2014: An expanded VVP technique to resolve primary technique and methods using satellite microwave and environmental circulations in hurricanes. J. Atmos. Oceanic sounder and imager data for operational use, would Technol., 31,249–271,doi:10.1175/JTECH-D-13-00030.1. ——,C.J.McAdie,P.P.Dodge,W.-C.Lee,J.Tuttle,S.T. enable more frequent and possibly more accurate in- Murillo, and F. D. Marks, 2004: Real-time implementation tensity analyses of TCs approaching Japan. of single-Doppler radar analysis methods for tropical cyclones: Algorithm improvements and use with WSR-88D Acknowledgments. Some data analyses were performed display data. Wea. Forecasting, 19, 219–239, doi:10.1175/ 1520-0434(2004)019,0219:RIOSRA.2.0.CO;2. with the ‘‘DRAFT’’ Doppler radar analysis tool of the ——, W.-C. Lee, and M. Bell, 2007: Real-time implementation of Meteorological Research Institute (MRI). US is deeply VORTRAC at the National Hurricane Center. 33rd Conf. on gratefultoMr.H.YamauchiandMr.E.Satoforhelpful Radar Meteorology, Cairns, QLD, Australia, Amer. Meteor. advice. Gratitude is also extended to colleagues at the Soc., P11A.6. [Available online at https://ams.confex.com/ams/ RSMC Tokyo and MRI. The authors thank three anony- 33Radar/webprogram/Paper123747.html.] Herndon, D., and C. Velden, 2004: Upgrades to the UW-CIMSS mous reviewers for valuable comments that have greatly AMSU-based TC intensity algorithm. 26th Conf. on Hurri- improved the manuscript. This work was partially sup- canes and Tropical Meteorology, Miami, FL, Amer. Meteor. ported by Ministry of Education, Culture, Sports, Science Soc., 118–119. and Technology (MEXT) KAKENHI Grant 25400468. Hoshino, S., and T. Nakazawa, 2007: Estimation of tropical cyclone’s The opinions in this paper are those of the authors and intensity using TRMM/TMI brightness temperature data. J. Meteor. Soc. Japan, 85, 437–454, doi:10.2151/jmsj.85.437. should not be regarded as official RSMC Tokyo views. Japan Meteorological Agency, 2015: Numerical weather prediction at the Japan Meteorological Agency. Japan Meteorological Agency. [Available online at http://www.jma.go.jp/jma/jma-eng/ REFERENCES jma-center/nwp/nwp-top.htm.] Bankert, R. L., and P. M. Tag, 2002: An automated method to Jou, B. J.-D., W. C. Lee, S. P. Liu, and Y. C. Kao, 2008: Generalized estimate tropical cyclone intensity using SSM/I imagery. J. Appl. VTD retrieval of atmospheric vortex kinematic structure. Part Meteor., 41, 461–472, doi:10.1175/1520-0450(2002)041,0461: I: Formulation and error analysis. Mon. Wea. Rev., 136, 995– AAMTET.2.0.CO;2. 1012, doi:10.1175/2007MWR2116.1. Bell, M. M., and M. T. Montgomery, 2008: Observed structure, Koba,H.,T.Hagiwara,S.Osano,andS.Akashi,1990:Re- evolution, and potential intensity of category 5 Hurricane lationship between the CI-number and central pressure and Isabel (2003) from 12 to 14 September. Mon. Wea. Rev., 136, maximum wind speed in typhoons (in Japanese). J. Meteor. 2023–2046, doi:10.1175/2007MWR1858.1. Res., 42, 59–67. ——, and W. Lee, 2012: Objective tropical cyclone center tracking Kossin, J. P., and W. H. Schubert, 2004: Mesovortices in Hurricane using single-Doppler radar. J. Appl. Meteor. Climatol., 51, Isabel. Bull. Amer. Meteor. Soc., 85, 151–153, doi:10.1175/ 878–896, doi:10.1175/JAMC-D-11-0167.1. BAMS-85-2-151. Braun, S. A., M. T. Montgomery, and Z. Pu, 2006: High-resolution Lamb, H., 1932: Hydrodynamics. Dover, 738 pp. simulation of Hurricane Bonnie (1998). Part I: The organiza- Lee, W.-C., and F. D. Marks Jr., 2000: Tropical cyclone kinematic tion of eyewall vertical motion. J. Atmos. Sci., 63, 19–42, structure retrieved from single-Doppler radar observations. Part II: doi:10.1175/JAS3598.1. The GBVTD-simplex center finding algorithm. Mon. Wea. Brueske, K. F., and C. S. Velden, 2003: Satellite-based tropical cyclone Rev., 128, 1925–1936, doi:10.1175/1520-0493(2000)128,1925: intensity estimation using the NOAA–KLM series Advanced TCKSRF.2.0.CO;2. Microwave Sounding Unit (AMSU). Mon. Wea. Rev., 131,687– ——, and M. M. Bell, 2007: Rapid intensification, eyewall contrac- 697, doi:10.1175/1520-0493(2003)131,0687:SBTCIE.2.0.CO;2. tion, and breakdown of Hurricane Charley (2004) near landfall. Chen, X., K. Zhao, W.-C. Lee, B. J.-D. Jou, M. Xue, and P. R. Geophys. Res. Lett., 34, L02802, doi:10.1029/2006GL027889. Harasti, 2013: The improvement to the environmental wind ——, B. J.-D. Jou, P.-L. Chang, and S.-M. Deng, 1999: Tropical and tropical cyclone circulation retrievals with the modified cyclone kinematic structure retrieved from single-Doppler GBVTD (MGBVTD) technique. J. Appl. Meteor. Climatol., radar observations. Part I: Interpretation of Doppler ve- 52, 2493–2508, doi:10.1175/JAMC-D-13-031.1. locity patterns and the GBVTD technique. Mon. Wea. Rev., Demuth, J. L., M. DeMaria, J. A. Knaff, and T. H. Vonder Haar, 127, 2419–2439, doi:10.1175/1520-0493(1999)127,2419: 2004: Validation of an Advanced Microwave Sounding Unit TCKSRF.2.0.CO;2.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC 1840 MONTHLY WEATHER REVIEW VOLUME 144

——, ——, ——, and F. D. Marks Jr., 2000: Tropical cyclone ki- operational development at JMA. J. Meteor. Soc. Japan, nematic structure retrieved from single-Doppler radar obser- 85B, 271–304, doi:10.2151/jmsj.85B.271. vations. Part III: Evolution and structure of Typhoon Alex Sakuragi, T., S. Hoshino, and N. Kitabatake, 2014: Development (1987). Mon. Wea. Rev., 128, 3982–4001, doi:10.1175/ and veriﬁcation of a tropical cyclone intensity estimation 1520-0493(2000)129,3982:TCKSRF.2.0.CO;2. method reﬂecting the variety of TRMM/TMI brightness tem- Martin, J. D., and W. M. Gray, 1993: Tropical cyclone observation perature distribution. RSMC Tokyo-Typhoon Center Tech. and forecasting with and without aircraft reconnaissance. Wea. Rev., No. 16, 15 pp. [Available online at http://www.jma.go.jp/ Forecasting, 8, 519–532, doi:10.1175/1520-0434(1993)008,0519: jma/jma-eng/jma-center/rsmc-hp-pub-eg/techrev/text16-2.pdf.] TCOAFW.2.0.CO;2. Velden, C. S., T. Olander, and R. M. Zehr, 1998: Development of Montgomery, M. T., and R. J. Kallenbach, 1997: A theory for an objective scheme to estimate tropical cyclone intensity vortex Rossby waves and its application to spiral bands and from digital geostationary satellite imagery. Wea. Fore- intensity changes in hurricanes. Quart. J. Roy. Meteor. Soc., casting, 13, 172–186, doi:10.1175/1520-0434(1998)013,0172: 123, 435–465, doi:10.1002/qj.49712353810. DOAOST.2.0.CO;2. Murillo, S. T., W.-C. Lee, M. M. Bell, F. D. Marks Jr., P. P. Dodge, ——, D. Herndon, J. Kossin, J. Hawkins, and M. DeMaria, 2007: and G. M. Barnes, 2011: Intercomparison of ground-based Consensus estimates of tropical cyclone (TC) intensity using velocity track display (GBVTD)-retrieved circulation centers integrated multispectral (IR and MW) satellite observations. and structures of Hurricane Danny (1997) from two coastal Joint 2007 EUMETSAT Meteorological Satellite/15th Conf. on WSR-88Ds. Mon. Wea. Rev., 139, 153–174, doi:10.1175/ Satellite Meteorology and Oceanography, Amsterdam, the 2010MWR3036.1. Netherlands, EUMETSAT–Amer. Meteor. Soc. [Available Nolan, D. S., J. A. Zhang, and D. P. Stern, 2009: Evaluation of online at http://www.ssec.wisc.edu/meetings/jointsatmet2007/ planetary boundary layer parameterizations in tropical pdf/velden_satcon.pdf.] cyclones by comparison of in situ observations and high- Wang, M., K. Zhao, W. C. Lee, B. J.-D. Jou, and M. Xue, 2012: The resolution simulations of Hurricane Isabel (2003). Part I: Ini- gradient velocity track display (GrVTD) technique for re- tialization, maximum winds, and the outer-core boundary trieving tropical cyclone primary circulation from aliased ve- layer structure. Mon. Wea. Rev., 137, 3651–3674, doi:10.1175/ locities measured by single-Doppler radar. J. Atmos. Oceanic 2009MWR2785.1. Technol., 29, 1026–1041, doi:10.1175/JTECH-D-11-00219.1. Olander, T. L., and C. S. Velden, 2007: The advanced Dvorak Wang, Y., 2002: Vortex Rossby waves in a numerically simulated technique (ADT): Continued development of an objective tropical cyclone. Part I: Overall structure, potential vorticity, scheme to estimate TC intensity using geostationary IR sat- and kinetic energy budgets. J. Atmos. Sci., 59, 1213–1238, ellite imagery. Wea. Forecasting, 22, 287–298, doi:10.1175/ doi:10.1175/1520-0469(2002)059,1213:VRWIAN.2.0.CO;2. WAF975.1. Willoughby, H. E., 1990: Gradient balance in tropical cyclones. ——, ——, and J. P. Kossin, 2004: The advanced objective Dvorak J. Atmos. Sci., 47, 265–274, doi:10.1175/1520-0469(1990)047,0265: technique (AODT)—Latest upgrades and future directions. GBITC.2.0.CO;2. 26th Conf. on Hurricanes and Tropical Meteorology, Miami, Yamauchi, H., O. Suzuki, and K. Akaeda, 2006: A hybrid multi- FL, Amer. Meteor. Soc., P1.19. [Available online at https:// PRI method to dealias Doppler velocities. SOLA, 2, 92–95, ams.confex.com/ams/pdfpapers/75417.pdf.] doi:10.2151/sola.2006-024. Oyama, R., 2014: Estimation of tropical cyclone central pressure Zhao, K., W.-C. Lee, and B. J.-D. Jou, 2008: Single Doppler radar from warm core intensity observed by the Advanced Microwave observation of the concentric eyewall in Typhoon Saomei, Sounding Unit-A (AMSU-A). Pap. Meteor. Geophys., 65, 35– 2006, near landfall. Geophys. Res. Lett., 35, L07807, 56, doi:10.2467/mripapers.65.35. doi:10.1029/2007GL032773. Saito, K., and Coauthors, 2006: The operational JMA non- ——, M. Xue, and W.-C. Lee, 2012: Assimilation of GBVTD- hydrostatic mesoscale model. Mon. Wea. Rev., 134, 1266–1298, retrieved winds from single-Doppler radar for short-term doi:10.1175/MWR3120.1. forecasting of super typhoon Saomai (0608) at landfall. ——, J. Ishida, K. Aranami, T. Hara, T. Segawa, M. Narita, and Quart. J. Roy. Meteor. Soc., 138, 1055–1071, doi:10.1002/ Y. Honda, 2007: Nonhydrostatic atmospheric models and qj.975.

Unauthenticated | Downloaded 10/01/21 03:37 AM UTC