Reducing TC Position Uncertainty in an Ensemble Data Assimilation and Prediction System: a Case Study of Typhoon Fanapi (2010)

APRIL 2018 L I N E T A L . 561

Reducing TC Position Uncertainty in an Ensemble Data Assimilation and Prediction System: A Case Study of Typhoon Fanapi (2010)

KUAN-JEN LIN AND SHU-CHIH YANG Department of Atmospheric Sciences, National Central University, Taoyuan, Taiwan

SHUYI S. CHEN Department of Atmospheric Sciences, University of Washington, Seattle, Washington

(Manuscript received 17 October 2017, in ﬁnal form 1 February 2018)

ABSTRACT

Ensemble-based data assimilation (EDA) has been used for tropical cyclone (TC) analysis and prediction with some success. However, the TC position spread determines the structure of the TC-related background error covariance and affects the performance of EDA. With an idealized experiment and a real TC case study, it is demonstrated that observations in the core region cannot be optimally assimilated when the TC position spread is large. To minimize the negative impact from large position uncertainty, a TC-centered EDA approach is implemented in the Weather Research and Forecasting (WRF) Model–local ensemble transform Kalman filter (WRF-LETKF) assimilation system. The impact of TC-centered EDA on TC analysis and prediction of Typhoon Fanapi (2010) is evaluated. Using WRF Model nested grids with 4-km grid spacing in the innermost domain, the focus is on EDA using dropsonde data from the Impact of Typhoons on the Ocean in the Pacific field campaign. The results show that the TC structure in the background mean state is improved and that unrealistically large ensemble spread can be alleviated. The characteristic horizontal scale in the background error covariance is smaller and narrower compared to those derived from the conventional EDA approach. Storm-scale corrections are improved using dropsonde data, which is more favorable for TC development. The analysis using the TC-centered EDA is in better agreement with independent observations. The improved analysis ameliorates model shock and improves the track forecast during the first 12 h and landfall at 72 h. The impact on intensity prediction is mixed with a better minimum sea level pressure and overestimated peak winds.

1. Introduction limited forecasting skill in TC intensity. Recent studies have shown that intrinsic predictability is more of a Tropical cyclones (TCs) can bring catastrophic di- limiting factor for TC intensity than for TC track (Judt sasters and lead to great losses in terms of human life et al. 2016; Judt and Chen 2016; Kieu and Moon 2016). and economic interests, which is a major concern and The lack of coupling to the ocean is a major shortcoming challenge for regional numerical weather prediction (NWP). Although TC track prediction has improved in many current prediction models (Chen et al. 2007, steadily owing to advancements in numerical models 2013). Another limiting factor in the current TC pre- and data assimilation of increased satellite observations diction system may be due to the relatively poor repre- over the last few decades, the progress in TC intensity sentation of the TC structure and intensity in the initial forecasting has been modest (Rappaport et al. 2009). conditions for TC forecasting (Kurihara et al. 1993). To There are a number of factors that may contribute to the tackle these difﬁcult problems, efforts have been made to improve the model physics and mimic the real TC structure using vortex initialization techniques. Denotes content that is immediately available upon publica- Generally speaking, TC vortex initialization methods tion as open access. can be categorized into four groups. They are vortex bogusing (Ueno 1989; Leslie and Holland 1995; Pu and Corresponding author: Shu-Chih Yang, [email protected]. Braun 2001), dynamical initialization (Kurihara et al. 1993, edu.tw 1995, 1998; NguyenandChen2011), vortex relocation

DOI: 10.1175/WAF-D-17-0152.1 Ó 2018 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses). 562 WEATHER AND FORECASTING VOLUME 33

(Liu et al. 2000; Hsiao et al. 2010), and data assimilation surface wind data are assimilated, a reasonable vertical (DA) (Zou and Xiao 2000; Chen and Snyder 2007; Torn TC structure can be successfully established in some ca- and Hakim 2009; Zhang et al. 2011). These methods are ses. More specifically, such data can help generate the used alone or in combination to generate a TC vortex in classic in–up–out secondary circulation of TCs. More- the model initial conditions and have led to some degree of over, regional EnKFs implemented in cloud-resolving success. Among these methods, DA is the only method models can assimilate data at high spatial and temporal that uses observations obtained in TCs and/or TC envi- resolutions, which also have a major impact on TC ronments with the proper description about the accuracy structure and intensity forecasts. Studies suggest that the of these observations. key to improving the intensity prediction relies on the Previous studies have used either a synthetically ability to represent both the TC inner-core structure generated TC vortex or real observations to represent (Zhang et al. 2009; Weng and Zhang 2012)andenviron- TC structures that can be assimilated into a DA system. mental conditions (Wu et al. 2014). The former is also referred to as bogus data assimilation One of the advantages of the ensemble-based method (BDA) and is usually adopted within a variational DA on TC assimilation is that it takes into account the col- framework. BDA implemented in three- or four- lective flow-dependent uncertainties associated with TC dimensional variational analysis systems has yielded center location and intensity, as well as the environment, positive results in representing idealized (‘‘bogus’’) TC which determine the TC evolution. However, the TC structures and improving TC predictions (Zou and Xiao position uncertainty in EnKF remains a major chal- 2000; Pu and Braun 2001; Chou and Wu 2008; Davidson lenge. It can dominate the performance and affect the et al. 2014). However, in the early work on variational accuracy of the analysis. For the mean state, averaging methods, the background error covariance used is the TC in all ensemble members with different center usually static and not able to represent the TC-related locations yields an overly smoothed and weak TC, with uncertainty; hence, the observations cannot be used ef- an overestimated size and ill-defined inner-core struc- fectively to correct errors related to TC circulation. In ture. If the position spread is not well constrained and contrast, the ensemble Kalman filter (EnKF; Evensen becomes too large, unrealistic features, such as double 1994), which uses flow-dependent background error vortices, can occur near the core region (Chen and statistics to consider the dynamical uncertainty in TC Snyder 2007). For the ensemble-estimated background circulation naturally, has the advantage of assimilating error covariance, the position uncertainty dominates the the observations more effectively (Yang et al. 2013) and background error statistics and can mask the un- is expected to fit better with TC DA and prediction certainties associated with the TC structure and in- (Hamill et al. 2011a,b). Building on the advantages of tensity (Torn and Hakim 2009; Poterjoy and Zhang EnKF and variational methods, significant work with 2011). Excessive variance in the TC position of the en- the hybrid variational–EnKF system has also shown semble results in asymmetric corrections that can distort positive results for TC assimilation and prediction the vortex structure. Even small changes in the modeled (Poterjoy and Zhang 2014; Lu et al. 2017). TC position may cause a large unrealistic covariance Within the EnKF framework, strategies have been near the TC eyewall where large gradients are located. developed to assimilate observations to improve TC Poterjoy et al. (2014) also reported that the vortex-scale structures in the analysis. Chen and Snyder (2007) were analysis increment is sensitive to the ensemble variance the first to propose an idea for assimilating the TC advi- at the inner core. Therefore, large uncertainties in storm sory information issued by operational centers [e.g., position can significantly degrade the performance of center location, minimum sea level pressure (MSLP), and the EnKF analysis in TC initialization. peak wind] in addition to other observations to improve To address this issue, methods have been proposed to the vortex initialization. Using a simple two-dimensional tackle the TC location uncertainty in EnKF. Chen and barotropic model, they showed that the track forecast Snyder (2007) suggest that assimilating the TC position initialized from such TC-targeted EnKF analysis is im- prior to the assimilation of other observations can re- proved with a reduced spurious transient evolution of the duce the TC position uncertainty and produce a more initial vortex. Wu et al. (2010) further advanced the as- accurate TC analysis. Some studies suggest that assimi- similation of TC advisory data by frequently assimilating lating MSLP from TC advisory data as a standard SLP parameters related to the TC track and structure, in- observation helps to constrain the TC position and im- cluding the center position, velocity of TC motion, and prove the intensity (Hamill et al. 2011a; Kleist. 2011). axisymmetric surface wind. They showed that after the However, if the TC center in the background state is too initialization period, a realistic initial TC vortex can be far away from the MSLP observations, an issue of produced. With their EnKF system, even when only an artificially created double eye will be problematic APRIL 2018 L I N E T A L . 563

(Kawabata et al. 2012). To avoid such an issue, Kunii TABLE 1. Summary of the idealized experiment setup. (2015) adopted the concept from Chen and Snyder P5 P30 I2 P5I2 P30I2 (2007) to use the MSLP of the model state as the background field, instead of the simulated SLP value at Position uncertainty (km) 5 30 5 30 Intensity uncertainty (MSLP; hPa) 2 2 2 the observed TC location when assimilating the MSLP observations. A recent study by Navarro and Hakim (2014, hereafter NH14) proposed a storm-relative ensemble assimilation method for EnKF to relocate the framework, we examine the relationship among TC TC ensemble to the observed TC position before per- position error, spread, and corrections to TC structure forming the storm-scale assimilation. They showed that from the single-observation assimilation experiments in by using the storm-relative approach, the vortex has a idealized scenarios. Five experiments with different finer inner core with a more realistic asymmetric struc- types of uncertainties related to TC position and in- ture compared to the truth state, whereas the conven- tensity are presented here (Table 1). The uncertainty, tional EnKF analysis can produce errors greater than an represented by ensemble spread, is defined as the dis- order of magnitude in some cases. Although NH14 persedness of the ensemble and is quantified by the demonstrated the potential of the storm-centered ap- standard deviation. In each experiment, 1000 symmetric proach for TC assimilation in an idealized observing vortices with their position and/or intensity perturbed system simulation experiment (OSSE) using a dynami- are used as the background TC ensemble, and we ex- cally simple shallow-water model and uniformly dis- amine the impact of EDA on TC structure. The large tributed observations, the impact on TC prediction with ensemble size is used to avoid the issue of sampling error full-physics NWP models and realistic observations in and to maintain the symmetry in the mean TC structure TCs has not yet been investigated. before assimilation. The symmetric vortex is generated In this study, we first demonstrate the importance of based on the sea level pressure and tangential wind representing the TC position uncertainty in the TC profile in Holland (1980), with parameters of MSLP, ensemble-based DA (EDA) within an idealized frame- pressure of TC environment Penv, radius of maximum work and examine the impact of constraining the position wind (RMW), and shape. We note that the shape pa- uncertainty in a regional EnKF system with in situ field rameter is set to 1.0 for all experiments in order to focus observations of a real typhoon case. The regional EnKF on the TC intensity uncertainty. system consists of the Weather Research and Forecasting The experiments are as follows: 1) the TC ensemble (WRF) Model and a local ensemble transform Kalman has a position spread of 5 km but no differences in in- filter (LETKF). In addition to the conventional WRF- tensity (P5); 2) the same as in the first experiment but LETKF system, the TC-centered (TCC) DA approach with a larger position spread of 30 km (P30); 3) the TC following NH14 is implemented. The capability of the ensemble has differences in intensity, but no position TCC WRF-LETKF system is investigated based on a spread (I2); 4) the same as in the first experiment but case study of Typhoon Fanapi (2010), which was a well- with intensity spread (P5I2); and 5) the same as in the observed TC during the Impact of Typhoons on the second experiment but with intensity spread (P30I2). Ocean in the Pacific (ITOP) field campaign (D’Asaro For experiments P5 and P30, the same TC vortex is et al. 2014). obtained using an MSLP of 980 hPa, Penv of 1010 hPa, This paper is organized as follows. In section 2, the and RMW of 50 km, but arranged at different locations impact of position uncertainty on TC assimilation is il- to have TC position spreads of 5 or 30 km, respectively. lustrated under ideal scenarios. Section 3 introduces the Note that the same vortex is used in the TC ensemble of framework of the TC-centered ensemble data assimila- P5 and P30; thus, the TC uncertainty is contributed by tion method. Typhoon Fanapi (2010) is introduced in the position uncertainty only. To have the intensity section 4. The model and experimental setup are given spread in I2, P5I2, and P30I2, we further perturb the in section 5. The results of the experiments are pre- MSLP with a standard deviation of 2 hPa (leading to a 2 sented in section 6. Finally, section 7 provides a sum- 1ms 1 standard deviation in the maximum of the tan- mary and some discussion. gential wind). We should note that these parameters, including the TC intensity and spread, are chosen based on the real case of Typhoon Fanapi in 2010 (section 4). 2. Impact of TC position uncertainty in an idealized Finally, the vortex is arranged in Cartesian coordinates framework with a grid spacing of 5 km. With these TC ensembles, To illustrate the importance of TC position un- the EnKF assimilation is conducted to assimilate an certainty within the ensemble-based data assimilation MSLP observation (976 hPa; observation error is set to 564 WEATHER AND FORECASTING VOLUME 33

FIG. 1. The TC vortex structure from scenario 1: the mean TC centers and MSLP observations (denoted by the exes) are collocated for all ﬁve experiments. (top) Mean surface wind speeds (background in color shading and analysis in contours), (middle) analysis increments of SLP (color shading) and wind (arrows) from assimilating the MSLP observation, and (bottom) azimuthal-averaged surface wind speed of the reference TC (black), background (blue), and analysis (red) mean for (a),(f),(k) P5; (b),(g),(l) P30; (c),(h),(m) I2, (d),(i),(n) P512; and (e),(j),(o) P30I2. The green contour lines in (f)–(j) indicate the ensemble spread of the zonal wind, and the innovation values are the difference between the observations and the background SLP.

1 hPa) under two scenarios: the MSLP observation is 1) analysis corrections. With a large position uncertainty, located at the center of the mean TC or 2) located 40 km the background mean TC structure is overly smoothed away from the mean TC center. These setups and sce- out due to the position displacements among the en- narios are used to mimic the model ensemble states with semble, resulting in a larger eye and RMW (Fig. 1l, blue and without TC relocation and with different degrees of line). Also, the horizontal scale and magnitude of the dispersiveness for the position spread. In the following, zonal wind ensemble spread increase with the un- we focus on how the analysis increment may change the certainty in the TC position (Figs. 1f,g, green contours), TC intensity and structure. However, we should em- leading to strong, broader-scale analysis corrections in phasize that this MSLP could correspond to TC vortices both the wind and SLP fields. Although the intensity, with different characteristics and thus the true TC especially near the RMW, can be greatly increased structure is not specified. The unperturbed TC in the given a large position uncertainty and innovation (e.g., background ensemble (with MSLP of 980 hPa, Penv of Fig. 1l), the correction for the inner core leads to a larger 1010 hPa, RMW of 50 km, and shape parameter of 1.0) is RMW and change in the TC structure compared with used as a reference state to distinguish the structure of the reference TC. When the intensity uncertainty is in- the analysis increment. cluded, the results show that the ensemble spread and Figure 1 shows the wind speed of the ensemble mean analysis correction are dominated by the large position TC (Figs. 1a–e), analysis increment (Figs. 1f–j), and the uncertainty. More specifically, the correction of the TC azimuthal averaged wind (Figs. 1k–o) under the first structure is determined by position uncertainty instead scenario. The MSLP observations (gray exes) and the of the intensity uncertainty itself. In contrast, the TC centers of the mean TC are collocated. By comparing ensemble with intensity uncertainty only or an addi- the first two columns in Fig. 1, it is clear that the different tional small position uncertainty can have the storm-scale degrees of position uncertainty can result in different correction to enhance the TC circulation, as shown in APRIL 2018 L I N E T A L . 565

FIG.2.AsinFig. 1, but with the observation location 40 km away from the center of the mean TC.

Figs. 1m and 1n. The correction is now associated with position does not move toward the observations, and the intensity uncertainty, rather than being over- the intensity becomes much enhanced compared to the whelmed by a large position spread. I2 experiment in scenario 1. This implies that without In the second scenario (Fig. 2), there is position error relocation, the position error in the mean state can in the mean state (i.e., a situation without TC re- result in overcorrected intensity since the MSLP is location), and the MSLP observation is 40 km away mistakenly regarded as an SLP observation near the from the center of the background mean TC. The re- eyewall, and the corrections enhance the whole TC sults from experiments with position uncertainty in- circulation. When both intensity uncertainty and small dicate that the TC vortex can be shifted toward the position uncertainty (P5I2) are present, the correction observed TC location after assimilating the MSLP can also successfully enhance the TC for both the inner observation. Comparing P5 and P30, when the TC structure (100 km) and the outer circulation (.100 km) position spread is large enough, the TC vortex can shift without changing the TC structure too much, despite very close to the observation location, as shown in the fact that the correction for the position error is Figs. 2b and 2e. However, although the TC position has limited. As expected, the TC intensity development been corrected, an asymmetry component is in- based on the TC structure of the P5I2 and P30I2 ana- troduced at the northeastern sector of the TC, along the lyses can be very different. line between the observed and mean TC positions. The The negative impact of TC position uncertainty on larger the position spread, the larger the asymmetric ensemble TC assimilation results through the averaging component. As pointed out earlier in Fig. 1, the strong of the mean state and structure of the background error correction in P30 and P30I2 brought about by the large covariance. We should note that in other DA systems, ensemble spread corrects the intensity, and the eye and such as deterministic EnKF or hybrid approaches, the RMW become larger than the reference TC structure. issue of averaging is not a concern. However, the Also, the large position uncertainty dominates the ensemble-based background error covariance can still analysis increment, and the difference between P30 carry the information about position uncertainty, and, and P30I2 in intensity correction is less significant thus, large position uncertainty can still dominate (Figs. 2l,o). With no position uncertainty (I2), the TC the background error covariance and still induce 566 WEATHER AND FORECASTING VOLUME 33 unrealistically large corrections to the TC structure, In Eq. (1), r^5 (r 2 Rin)/(Rout 2 Rin), where r is the ra- especially in the inner-core region. dial distance from the TC center, and Rin and Rout are, As a conclusion from the experiments of the ide- respectively, the inner and outer radii of the concentric alized scenarios, the position uncertainty should ring where we merge the two analyses. The outer ra- be first taken care of in the ensemble-based TC dius Rout is 280 km, close to the size from the typhoon assimilation, in order to use the observations in the advisory dataset, and the inner radius Rin is three- inner-core regions of the TC to correct the intensity. quarters of the outer radius Rout, which is 210 km in Otherwise, the structure of the intensity correction this case. will be overwhelmed by the position spread. For such a purpose, a vortex relocation scheme can be applied to correct the position error and to eliminate 4. Typhoon Fanapi of 2010 the position uncertainty in the ensemble. We should Typhoon Fanapi developed from a tropical denote that if the relocation is only performed on the pression east of the Philippines on 14 September 2010. mean state, the TC position uncertainty can still After its formation, Fanapi moved northwestward and dominate the ensemble-based error statistics and gradually intensified into a tropical storm (TS) on degrade the analysis performance. 15 September. During its early stages, Fanapi moved slowly and turned northeastward on 16 September. Starting at 0000 UTC 17 September, Fanapi curved 3. TC-centered ensemble data assimilation toward the northwest again and then moved west- framework ward after 1200 UTC 17 September. While moving In the second part of this study, we focus on the in- westward, Fanapi reached its peak intensity with an vestigation of the impact of TC position uncertainty on MSLP of 944 hPa and maximum wind speed (MWS) of 2 the analysis and forecasting of a real observed typhoon 54 m s 1 at 0600 UTC 18 September. After making using a regional EnKF assimilation system implemented landfall in Hualien County in Taiwan at 0000 UTC 19 in a full-physics numerical model. As demonstrated in September, Fanapi weakened drastically as a result of an idealized experiment, TC position uncertainty and destruction by topographic effects, and the storm position error in the mean state during the ensemble- dissipated a few hours later after its second landfall over based TC assimilation can bring negative impacts to the China on 20 September. TC structure when the position spread becomes large. Typhoon Fanapi is selected in this study because it is To reduce such effects, the TC-centered approach used the best observed TC during the ITOP1 (D’Asaro et al. by NH14 is adapted in this study. We made a number of 2014) field campaign in 2010. During the life cycle of adjustments for implementation in the nested domains Fanapi, numerous dropsonde, airborne, and satellite of the WRF Model with the EnKF assimilation system. observations were made in the inner-core regions, and Our experiment is carried out in the following three we used many for our data assimilation experiments and steps (similar to the process described in NH14). 1) A verification. conventional data assimilation experiment is conducted to provide an analysis of the TC environment over the WRF’s outer domain. 2) Because the TC center of each 5. Model, assimilation, and experimental setup ensemble member is located at the center of the vortex- a. Model and experimental setup following inner domain, we relocate the center of the inner domain of all ensemble members to the observed The TC-centered ensemble data assimilation method TC location. The TC-centered data assimilation process is implemented within the WRF-LETKF framework. is then performed to generate an analysis in the TC area WRF-LETKF (Yang et al. 2013, 2014) is a regional data within the inner domain. 3) The results from steps 1 and assimilation and prediction system that consists of 2 are merged together with the same method used in the Advanced Research version of the WRF Model, NH14. The merging process is performed by linearly combining two analyses over a finite area, defined by a concentric ring around the TC. The weight on the con- 1 The ITOP program was a multinational field campaign de- ventional analysis is given by signed to study the oceanic response to typhoons in the western Pacific Ocean. During the lifetime of Fanapi, the U.S. Air Force pr^ 2 C-130 penetrated Fanapi and deployed dropsondes to measure the w 5 1 2 cos . (1) properties of the typhoon, and then floats and drifters were used to 2 measure the ocean’s response. APRIL 2018 L I N E T A L . 567 version 3.6.1 (Skamarock et al. 2008), with a vortex- b. Assimilated data following nested grid capability (Tenerelli and Chen 2001) Sources of the observations assimilated here include and LETKF (Hunt et al. 2007) data assimilation systems. surface stations, rawinsondes, aircraft reports, mid-to- The performance of this new TC-centered EnKF sys- upper-level atmosphere motion vectors (AMVs), the tem is evaluated based on a case study of Typhoon MSLP of the TC from the Joint Typhoon Warning Fanapi (2010), and the impact is investigated by Center (JTWC), and dropsondes from the ITOP field comparing a conventional DA experiment, which uses campaign in 2010 (D’Asaro et al. 2014). For quality the conventional WRF-LETKF system, with the ex- control, observations are rejected when the innovations periment using the new TC-centered WRF-LETKF (i.e., the difference between the background and ob- framework. The conventional DA experiment is servation) are greater than 5 times the observation er- hereafter referred to as CTL and the new TC-centered experiment as TCC. rors. In addition, the AMV data are averaged within a In CTL, the WRF Model is executed with two-way cylinder of 100-km radius and a layer of 25-hPa depth, nested domains. The horizontal grid spacings of the similar to a strategy of superobservation adopted by Wu outer and inner domains are 12 and 4 km, with di- et al. (2014). mensions of 600 3 445 and 403 3 403 grid points, re- In ensemble-based TC assimilation systems, the as- spectively. There are 36 vertical layers in both domains similation of the TC’s position is often used to constrain with a model top up to 50 hPa. The physical parame- that position (Chen and Snyder 2007; Torn and Hakim terizations include the WSM 5-class microphysics 2009; Wu et al. 2010). Torn and Hakim (2009) have scheme, the RRTM scheme for longwave and shortwave suggested that assimilating the TC position frequently radiation, the MM5 surface-layer scheme, the YSU PBL may partially reduce the large TC position spread. scheme, and the Kain–Fritsch scheme for cumulus pa- However, Wu et al. (2010) showed that although as- rameterization (outer domain only). The configuration similation of the TC position can reduce the TC position for the TCC experiment is mostly the same as that of error, it has limited impact on the near-surface wind CTL, except the inner domain is vortex following with structure. This may indicate that some inconsistency 151 3 151 grid points in order to implement the between the mass and dynamic fields can be introduced TC-centered DA. For a fair comparison, the merging by assimilating TC position information. Within an process in TCC is also performed in CTL so that the idealized framework, we found that assimilating the TC TC vortex analyses are from the inner domain and position can reduce the TC position error, but at the the analyses of the TC environment are from the same time, it also generated an asymmetric structure in outer domain. the wind fields similar to that found during the assimi- Both experiments are cold started at 1200 UTC lation of MSLP observations (Fig. 2). The larger the 12 September 2010, in which 36 ensemble members position spread, the larger the correction to the TC po- are initialized by adding perturbations, randomly sition and also the larger the asymmetries in the wind drawn from the three-dimensional variational data as- field. These results demonstrate the potential issues similation (3DVAR) background error covariance (Torn from assimilating the TC position observations during et al. 2006), to the NCEP GDAS 18318 analysis high-resolution TC assimilation. Based on these con- dataset. (Available online at https://doi.org/10.5065/ siderations, we did not assimilate the TC position in our D6M043C6.) After a 12-h ensemble forecast to spin up CTL experiment. We note that alternative strategies for the model, the first LETKF analysis is conducted at correcting the TC position error are possible. For ex- 0000 UTC 13 September followed by 6-h forecast– ample, Nehrkorn et al. (2015) use the feature calibration analysis cycles until 0000 UTC 16 September. and alignment technique to correct the TC position er- Afterward, a 3-day forecast is initialized from the ror in background fields. analysis ensemble mean at 0000 UTC 16 September. During the lifetime of Fanapi, dropsondes from A covariance localization scale of 150 km is adopted penetrating U.S. Air Force C-130 flights were avail- for the outer-domain assimilation and 50 km for the able almost every day. There were also three flights of inner-domain assimilation. The TC-centered DA is the Dropwindsonde Observations for Typhoon Sur- conducted at 1800 UTC 14 September, after the vortex veillance near the TaiwanRegion(DOTSTAR;Wu in the CTL experiment has been spun up by assimi- et al. 2005) Astra jet to measure the environment of lating the synthetic vortex winds that will be described Fanapi on 15–17 September. As an example, the lo- in the next subsection. In other words, before 1800 cations of the dropsondes at 0000 UTC 15 and UTC 14 September, both experiments are identical 16 September are shown in Fig. 4. The importance of (Fig. 3). these data in providing corrections for improving the 568 WEATHER AND FORECASTING VOLUME 33

FIG. 3. Timeline for the data assimilation and forecast experiments. The letter B shows the time when the bogus wind observation was assimilated.

TC environment and structure through data assimi- 6. Results from the real case lation will be shown in the next section. We also note a. Analysis that instead of treating dropsondes as sounding profiles with a location, each dropsonde data point at During the early stage of the assimilation, the assim- each level has its own location to take into account ilation of observations, including the bogus vortex as- the location drifting as a result of the strong wind similation, helps to form the circulation of Fanapi. As speeds in order to represent the inner-core structure shown in Fig. 5, the TC position and intensity (i.e., correctly. MSLP and maximum 10-m wind speed) in the analysis During the early developing stage of Fanapi, the MSLP mean state approached the observed values at 1800 of Fanapi is 1004 hPa at 1200 UTC 14 September from UTC 14 September. However, in the CTL analysis JTWC and a small observation increment is obtained, mean, the TC position and intensity have large fluctua- which led to ineffectiveness in establishing a reason- tions with poor performance at 1200 and 1800 able vortex structure. To spin up the TC structure, addi- UTC 15 September. At these times, a poor TC structure tional synthetic winds according to an axisymmetric in the background mean is obtained as a result of the vortex consistent with the JTWC best-track (i.e., large position spread. In comparison, TC parameters are MSLP and MWS) data are assimilated at 0000, 0600, well fit with the observed values in the TCC analysis and 1200 UTC 14 September. The generation of the mean. Although at 0000 UTC 16 September, the TC synthetic TC data is adopted from a procedure of position errors in both the CTL and TCC analysis means bogus data assimilation used in the Coupled Ocean– are comparable (8.7 km for CTL and 10.5 km for TCC), Atmosphere Mesoscale Prediction System-Tropical their MSLP values are very different, indicating that the Cyclone (COAMPS-TC) for the purposes of TC pre- TC inner-core structures derived from the TC-centered diction (Liou and Sashegyi 2012). Note that this type of and conventional assimilations are different. As will be synthetic TC data cannot fully represent the asymmetric discussed in further detail later, the large improvement structure in TCs. at 0000 UTC 16 September in the CTL analysis is

FIG. 4. Locations of the ITOP dropsondes assimilated in this study (dropsondes within an analysis time of a 63-h assimilation window; the red circles are from the C-130, and the black circles are from the DOTSTAR Astra) and the storm center location (black cross) at (a) 0000 UTC 15 Sep and (b) 0000 UTC 16 Sep. The black line is the JTWC best track of Typhoon Fanapi. APRIL 2018 L I N E T A L . 569

west component of the wind from both experiments at 1800 UTC 14 September, where the TC-centered approach is first applied. At this time, the difference between the CTL and TCC experiments is small, because the position spread is still small (i.e., 14 km in CTL and 6 km in TCC). As indicated by the black dots in the top panels in Fig. 6, the TC position spread in CTL increased steadily with time, from 14 km at 1800 UTC 14 September (Fig. 6a) to 24 km at 0000 UTC 15 September (Fig. 6b). The difference in TC structure between the two experiments becomes more noticeable as the position spread increases. At 0000 UTC 15 September, the TC in the background mean of CTL is much weaker and has a larger spread around the TC center than those shown for TCC (cf. Figs. 6b and 6d). With the assimilation of the dropsonde data at 0000 UTC 15 September (Fig. 4a), the position errors (spread) can be reduced in both experiments, from 54.2 to 28 km (from 23.8 to 11.63 km) and from 20.0 to 15.3 km (from 7.6 to 4.6 km), respectively. However, without the dropsonde data at 0600, 1200, and 1800 UTC 15 September, assimilating other observations has a limited impact on constraining the position or shrinking the spread effectively. In particular, the TC structure is too weak in the background mean of CTL, and thus the simulated MSLP is far from the observations, so that the MSLP observations at 1200 UTC 15 September are rejected by the procedure of quality control (QC). We note that if we relax the QC criterion to assimilate MSLP, significant spindown can occur (Tallapragada et al. 2014). In comparison, the MSLP observations can be successfully as- FIG. 5. (a) Model TC position error, where the simulated TC similated into the TCC experiment with a background center is defined as the location of the MSLP, (b) MSLP, and (c) MWS compared to the JTWC best-track data at 0000 UTC closer to the observations. This again confirms that the 14–16 Sep 2010. position uncertainties in the ensemble can affect the ef- fectiveness of using observations in the inner-core region of the TC. attributed to the assimilation of the ITOP dropsondes. As a result of the less constrained position un- We note that the presence of the position error in TCC is certainty, the position spread in the CTL ensemble because the TC center here is defined by MSLP, whereas grows even faster. The position spread in CTL reaches in our model setup the center of the vortex-following 56 km (Figs. 7a,b) at 0000 UTC 16 September. Except inner domain is determined by the geopotential height for the large position spread, there is also a significant at 850 hPa. When the TC is vertically tilted and the westward bias in the ensemble TC position. The large tilting extent could be different for each member, these uncertainty and bias of the TC position reflect the fact centers will not be the same. that, during early analysis cycles, the conventional as- As pointed out in previous studies and in section 2, the similation has not yet well depicted the conditions of the TC position uncertainty in the ensemble dominates the TC environment (e.g., variations in the subtropical ensemble-based background error covariance and may high). The spinup period of assimilation can be longer hinder the performance of the ensemble-based DA. To with limited observations over the open ocean and ini- illustrate the impact of the position uncertainty on tial ensemble perturbations that are less optimal to ensemble-based DA, we examined the background en- represent flow-dependent errors (Yang et al. 2013). As semble mean and the spread from both experiments. noted in section 2 and in previous studies (Chen and Figures 6a and 6c show the mean and spread of the east– Snyder 2007; Poterjoy and Zhang 2011; Navarro and 570 WEATHER AND FORECASTING VOLUME 33

FIG. 6. Background ensemble mean (color) and spread (contours) of the east–west component of the wind at the lowest model level for the (top) CTL and (bottom) TCC experiments at (a),(c) 1800 UTC 14 Sep and (b),(d) 0000 UTC 15 Sep. The black dots indicate the center locations of TCs for each member, where the red ex is the center of the ensemble mean. The black dashed line is the track of Fanapi from JTWC with the black circle representing the TC center at the corresponding analysis time.

Hakim 2014), the large TC position bias and spread lead TC area, such that the spread of the east–west compo- to a weak mean TC structure and an unrealistically large nent of the wind at the lowest model level can reach 2 ensemble spread around the TC center (the same as the 17 m s 1. In contrast, within the TC-centered DA P30I2 experiment in scenario 2). The background en- framework, the TCC background mean state (Figs. 7c,d) semble of CTL at 0000 UTC 16 September provides shows a more compact and intense TC structure with 2 evidence for such an issue. As shown in Figs. 7a and 7b, much stronger MWS (31 m s 1), lower MSLP (981 hPa), the TC in the background mean state has a very weak and smaller size. Moreover, the spread of the east–west 2 and broad structure due to the averaging. The MWS is wind is now reduced to about 8 m s 1 in maximum and is 2 only 16 m s 1 and the MSLP is 998 hPa, compared to the conﬁned in the eyewall of the TC. We should note that 2 28 m s 1 and 982 hPa values reported by JTWC. In ad- within the TCC method framework, although all the dition, the ensemble spread is unrealistically large in the TCs in the ensemble are relocated to the observed TC APRIL 2018 L I N E T A L . 571

FIG. 7. (a),(c) As in Fig. 6, but at 0000 UTC 16 Sep. (b),(d) Sea level pressure field for CTL and TCC at 0000 UTC 16 Sep. location, there is still a small TC position spread of about the TC center and the domain-wise sea level pressure, as 6–8 km. Such a position spread comes from the un- well as the east–west component of the wind. Within the certainty in the TC structure and the methodology used range of 50 km away from the center, the features of the to define the TC center. Most importantly, this allows us point covariance of CTL and TCC are similar, indicating to represent the uncertainty in the TC intensity and that a negative innovation not only decreases the central structure, without being overwhelmed by position sea level pressure, but also enhances the cyclonic cir- uncertainty. culation of the TC. However, the CTL background error Having a large position spread in the background covariance has a larger amplitude and a much wider ensemble not only leads to a large innovation, it also horizontal scale than that of the TCC background error greatly influences the structure of the background error covariance. This suggests that the CTL background er- covariance and analysis correction. Figure 8 shows the ror covariance will result in a strong and wider correc- point error covariance between the sea level pressure at tion on the background, while the TCC has the smaller 572 WEATHER AND FORECASTING VOLUME 33

FIG. 8. Background point covariance between the SLP at the TC center (black cross) and (a),(c) domain-wise points and (b),(d) the east–west component of the wind in CTL and TCC at 0000 UTC 16 Sep. correction confined in the inner core. For example, at structure of the TC is still not well constructed. In this time, the mean position error in the CTL experi- comparison, with the TC-centered DA, the TCC ment decreased from 187 to 8.6 km. background mean state (Figs. 7c,d)atthistimehas Although the position error at 0000 UTC 16 September already exhibited a stronger and more symmetric TC is comparable in both analyses, the characteristics of the structure. The analysis correction (Fig. 9d)allowsfur- TC structure, especially the inner core, are very different. ther strengthening of the inner core and shrinking of Figure 9 shows the analysis mean and analysis increment the eye. Also, the outer (100 km away from TC center) of wind speed at 0000 UTC 16 September, at the time wind speed along the western side of TC has increased. when the dropsonde observations from the ITOP field With the same observations, it is clear that the obser- campaign are available. In the CTL experiment (Fig. 9a), vational influences can be very different given the CTL the analysis increment is well collocated with the TC and TCC background ensembles. We also observed structure in the analysis mean state. Such a result in- that the structure of the corrections varies greatly dicates that, even with a very poor background state among the CTL analysis ensemble members. For the (i.e., a smooth and mislocated TC structure), the EnKF background ensemble member with a TC center al- successfully adjusts the position error and attempts to ready close to the observed TC location (Fig. 9b), the build a vortex structure with the help of the dropsonde TC looks similar to that of in the TCC analysis mean information, and the analysis increment determines the (Fig. 9d), which has a stronger wind speed in the TC structure in the analysis mean state. However, al- southeast sector of the inner core. For the background though the TC circulation is significantly enhanced, the ensemble member with a TC center away from the APRIL 2018 L I N E T A L . 573

FIG. 9. Analysis (color) and increment (black contours) of surface wind speed in TC-centered coordinates at 0000 UTC 16 Sep. (a) Analysis ensemble mean, (b) member 25, and (c) member 34 of CTL. (d)–(f) As in (a)–(c), but for TCC. The ocean surface winds from OSCAT at about 0300 UTC 16 Sep (g) before and (h) after rain correction. In (h) the rain-contaminated data are removed (white area). The black line in (a), (d), (g), and (h) shows the ﬂight path of the SFMR observations in TC-relative coordinates. The red ex (open circle) is the dropsonde location from the C-130 (DOTSTAR). observed location (Fig. 9c), the analysis correction The results from Figs. 8 and 9 also imply that the behaves like that shown in Fig. 9a, which is wider and conventional assimilation results in strong corrections collocated with the analysis TC circulation. In contrast, that may not be consistent with the model dynamics, the main characteristics of the TC in the TCC analysis while the TC-centered assimilation builds a more reli- ensemble are relatively similar. For example, the able background structure, and observations play a role analysis corrections in the TCC ensemble members in ﬁne-tuning the TC structure. With the unrealistically (Figs. 9e,f) and mean state (Fig. 9d) all show small- large ensemble spread shown in Fig. 7, we also note that scale features near the eye and a broader impact at the the strong correction in the CTL analysis can lead to outer region (.100 km). model shock, a dramatic adjustment for the model to 574 WEATHER AND FORECASTING VOLUME 33

FIG. 10. Azimuthal-averaged tangential (color) and radial (contours) wind of the (left) background and (right) analysis mean at 0000 UTC 16 Sep for (a),(b) CTL and (c),(d) TCC. The solid and dashed contours denote the positive (inward wind) and negative (outward wind) values, respectively. restore the balance. As will be demonstrated in the next correction is limited to the mid- to low levels, and the TC subsection, forecasts initialized from the CTL analysis structure in the CTL analysis is shallow. In comparison, may require a longer model spinup period to adjust to a the tangential wind, low-level inflow, and upper-level dynamically balanced state. outflow are much stronger in the TC of the TCC analysis The use of the TC-centered framework also brings a mean (Fig. 10d), accompanied by less tilting and a more great positive impact on the vertical development of compact eyewall structure. This indicates that the ob- the TC. As in Fig. 10a, the TC in the CTL background servations effectively enhance the secondary circulation mean has very weak tangential wind with a large radius of the TC in the TCC experiment, resulting in conditions of maximum wind (;140 km) and weak radial inflow that favor TC development. appearing near the surface. After assimilating the To verify the impact discussed above quantitatively, dropsonde observations, both the tangential wind and the TC wind speed from the CTL and TCC simulations radial inflow are increased (Fig. 10b). However, since analyzed at 0000 UTC 16 September are compared to the dropsondes were deployed near 700 hPa, analysis surface wind observations from the Stepped Frequency APRIL 2018 L I N E T A L . 575

FIG. 11. Surface wind speed in TC-relative coordinates. The blue and red dots show results from CTL for the (a) background and (b) analysis at 0000 UTC 16 Sep. (c),(d) As in (a) and (b), but for TCC. The black lines are the surface wind speeds observed by the SFMR on board the C-130 during 2230–0400 UTC 15–16 Sep. The gray dots are the ocean surface wind retrievals from OSCAT observed at 0300 UTC 16 September. The green exes are the dropsonde-observed surface wind speeds; the dropsondes were deployed from the C-130.

Microwave Radiometer (SFMR) on board the C-130 the TC center indicates that the OSCAT cannot cap- and the Indian Oceansat-2 Scatterometer (OSCAT) in- ture the inner-core structure of Fanapi; on the other strument. The C-130 reconnaissance carrying the SFMR hand, the SFMR and OSCAT are more consistent at the ﬂew in from the northeastern quarter of the typhoon and outer circulation (.50 km) except that the wind speed ended southeast of the typhoon, as denoted by the thick from SFMR is higher in the southeastern quadrant. For black solid line in Fig. 9a. The OSCAT surface winds the model results, as shown in Figs. 11a and 11b, there shown in Figs. 9g and 9h were observed at about 0300 is a strong correction for the CTL background to en- UTC 16 September. From Fig. 9g, the OSCAT without hance the wind speed in the inner core, while a relatively rain correction shows that Fanapi has an asymmetric small adjustment is needed for the TCC background to structure, which is stronger on the east side and even further shrink the eye. The TC in the TCC analysis stronger in the southeastern quadrant. This pattern is (Fig. 11d) shows in its intensity a much closer re- better represented in the TCC analysis (Fig. 9d) com- semblance to SFMR than to that of the CTL analysis, pare to the CTL analysis (Fig. 9a). However, we note and its wind speed in the outer circulation generally that the OSCAT observations in the southeastern agrees better with OSCAT with smaller spread. Such a quadrant have larger uncertainty due to rain contami- difference between the CTL and TCC corrections can be nation, as shown by the white area in Fig. 9h. Figure 11 more clearly demonstrated with the azimuthal average. compares the model to SFMR and OSCAT surface wind As shown in Fig. 12, the TC in the TCC analysis has a 2 speeds (SWSs) in TC-relative coordinates. The large radius of maximum wind and 34-kt (1 kt 5 0.51 m s 1) difference between SFMR and OSCAT within 50 km of wind closer to the observed values, while the RMW in 576 WEATHER AND FORECASTING VOLUME 33

forecasts initialized from the CTL and TCC analyses. Discussion focuses on the forecasts initialized at 0000 UTC 16 September (solid lines in Figs. 14 and 15). We note that forecast initialized at the later time (dashed lines in Figs. 14 and 15) is very consistent with the first 3-day forecast, which reinforces that our conclusions remain the same. Figure 14a shows the 3-day track prediction initialized from the analysis at 0000 UTC 16 September, when Fanapi is about to transition from a tropical storm into a category 1 hurricane. Overall, the track difference between the CTL and TCC forecasts is small, since the corrections for the TC environment from the conventional and TC-centered DA schemes should be similar. Both analyses lead to good track predictions with errors FIG. 12. Azimuthal-averaged surface wind speed in TC-relative smaller than 100 km for the 3-day forecast. However, the coordinates at 0000 UTC 16 Sep. The blue dashed (solid) line position errors at the initial time are reduced by differ- shows the results from the CTL background (analysis), and the ent means in both analyses. In the CTL analysis, the red dashed (solid) line shows results from the TCC background (analysis). The blue and red open circles (exes) represent the position error is mainly reduced by the assimilation of MWS and its radial location in the background (analysis) of the dropsondes, but the position in the TCC analysis is CTL and TCC TCs, respectively. The black ex and asterisk (*) constrained by the use of a TC-centered framework. The denote the RMW and the radius of the 34-kt wind estimated major difference is that there is a track discontinuity by JTWC. during the first 12-h CTL forecast. This is possibly due to the inconsistency between the model dynamics and the the CTL analysis is unrealistically large, indicating a analysis field. Such a result also suggests that within the poor representation of the inner structure. conventional WRF-LETKF framework, an improperly When comparing the surface wind speed along the described TC structure in the background can introduce path of the reconnaissance flight carrying the SFMR an imbalanced correction, which is unfavorable to the (Fig. 13), it is evident that the TCC analysis is very model dynamics; therefore, significant adjustment to the successful in depicting the eyewall structure; the stron- dynamical balance occurs as the forecast is initialized. In ger wind speed on the eastern side of the TC (also visible comparison, there is no such spinup issue, but the TC in in Figs. 9d–f) supports the fact that the asymmetry the TCC forecast moves slightly faster to the west than is better established. With respect to the SFMR, the in CTL after the 1-day forecast. Hence, the forecast root-mean-square difference (RMSD) is reduced from track error in TCC is about 10 km larger than in CTL, 2 2 8.7 m s 1 with the CTL background to 5.3 m s 1 with the but after 60-h forecast time, the TC makes landfall at a 2 CTL analysis, and from 5.8 to 4.2 m s 1 for the TCC location closer to the observations (Fig. 14b) with a background and analysis, respectively. Overall, the position error in the landfall location of 44 km compared wind speed has been improved, except for the wind to 93 km with the CTL forecast. Given the similar en- speed corresponding to the SFMR time after 0300 vironmental conditions in CTL and TCC, the reason UTC 16 September, when both the CTL and TCC an- that the TC in the TCC forecast has a slight faster alyses show weaker wind speeds. However, we found westward movement is related to the TC’s deep vertical that the observed surface wind speed from OSCAT development, and thus the TC movement is influenced (Fig. 11) and the dropsondes (green exes in Fig. 13)is by the strong westward steering flow above 500 hPa generally weaker than the SFMR wind speeds during (figure not shown). this period. The difference between these observations Although the track prediction is less sensitive to the indicates the presence of uncertainty in the observa- assimilation strategy, the use of TC-centered assimilations. How to address these observation uncertainties tion has a major impact on the intensity forecast. As remains a challenging issue for verification and data shown in Fig. 15, the TC in the TCC forecast went assimilation. through a rapid intensification and has a lower MSLP and stronger maximum surface wind speed (MSWS) b. Forecast compared to the TC development in the CTL forecast. In this section, we investigate the impact of the TC- However, the TC in the TCC forecast is overintensified centered assimilation on TC predictions based on the during the first two days, despite the fact that it reaches APRIL 2018 L I N E T A L . 577

FIG. 13. (a) Surface wind speed of CTL (blue) and TCC (red) at 0000 UTC 16 Sep compared to SFMR observations (black line) from 2230 to 0400 UTC 15–16 Sep. The blue and red dashed lines are derived from the background ﬁeld, while the solid lines are from the analysis. The green ex marks are the dropsonde-observed surface wind speeds that were deployed from the C-130. (b) Radial distance to TC center.

an intensity level close to the JTWC values afterward observations, which gradually decrease during 17 and and sustains a longer mature stage similar to the ob- 18 September and are related to the shrinking eye of an servations. At 0000 UTC 17 September, the asymmetry intensifying TC. As pointed out by Jin et al. (2014), this with a stronger eastern sector remains in the SFMR may be related to the fact that the horizontal resolution observations (Fig. 16). The asymmetry of the TC is of 4 km may not be sufficiently high to maintain the size better represented in the TCC forecast compared to the of the eye; higher model resolution is needed to repre- CTL forecast, despite the wind speed around the eye- sent the inner-core structure. wall being much stronger. Comparing the forecast SWS Although the TC in the TCC forecast is over- to the SFMR observations (Fig. 16), the gradient of the intensified, error sources from the structure of the initial wind speed at the eyewall is generally better represented vortex and model uncertainty may add to the complexity in the TCC forecast. The sharp decay in MSLP and of the intensity forecast. For example, interactions be- MSWS 6 h earlier in the TCC forecast results from the tween the TC and ocean have not been considered in faster movement of the TC. However, the decaying this work, and a cold eddy is observed right beneath the tendency agrees better with the observations since the track of Fanapi on 17 September. (Mrvaljevic et al. landfall location is more accurately predicted. 2013), possibly weakening the TC rapid intensifica- In terms of the TC size defined by the radius of the tion within the coupled atmosphere–ocean modeling 34-kt wind speed (Fig. 15c), the TC in the TCC forecast framework. Nevertheless, results from Fig. 15 suggest exhibits a size variation similar to that in the observa- that the TC in the TCC forecast goes through a dy- tions, while the CTL forecast has larger fluctuations, namical development more similar to the observation especially during the first 12 h. This is also related to the than the CTL forecast. This may partially justify the model spinup issue, as shown in the CTL track pre- adjustment made by the TC-centered data assimilation diction. The TC in the TCC forecast generally has a framework, which can be considered to provide a better smaller RMW (Figs. 16 and 15d); however, both fore- dynamical structure to represent TC evolution. Further casts cannot capture the RMW evolutions shown in the investigation with models with more complete physics 578 WEATHER AND FORECASTING VOLUME 33

veriﬁed against independent surface wind measure- ments from the airborne SFMR and OSCAT satellite. Several key conclusions can be drawn from this study:

d Large TC position uncertainty has a major negative impact on the ensemble-based DA system, resulting in a less representative ensemble mean state through averaging and an ensemble-estimated background error covariance. d TC-centered (TCC) regional EnKF DA can help alleviate the negative impacts of position errors by limiting position uncertainty and improving the TC structure in the ensemble DA system. The performance of 6-hourly analysis–forecast cycles is also improved significantly d The improved DA analysis in TCC had a positive impact on Fanapi track forecasts during the first 12 h by mitigating the model spinup issues as well as improving the landfall position at 72 h. d The TCC DA impact on Fanapi intensity forecasts yielded mixed results. While the TC in the TCC forecast is overintensified during the first 2 days, it begins to resemble the JTWC best-track estima- tion afterward. The variations in TC structure from the TCC forecast generally agree better with the observations. In the idealized scenarios, the results indicate that the position uncertainty can degrade the analysis per-

FIG. 14. (a)The JTWC best-track data (black) at 0000 formance as the position uncertainty becomes large. UTC 16–20 Sep and the 3-day model track forecasts initialized at Assimilating an MSLP observation with large position 0000 UTC 16 Sep (solid line) and 17 Sep (dashed line) with the CTL spread in the background ensemble can create a large (blue) and TCC (red line) analyses. Open circles mark locations at innovation due to the weak background mean TC. 0000 and 1200 UTC. (b) Forecast track error against JTWC Also, this assimilation gives a false intensity adjustment best-track data. with a strong and broad horizontal-scale correction, because of the unrealistically large and broad-scale and dynamics will help us better understand the inten- ensemble spread over the inner core of the TC. Fur- sification process. thermore, if the background mean TC has a position error, although assimilating an MSLP observation is able to correct the TC position, an unrealistic asym- 7. Summary and discussion metric component will be introduced into the TC This study investigates the impact of TC position un- structure. The larger the position uncertainty, the certainty on ensemble-based TC data assimilation and larger the correction to the TC position; on the other prediction. We first demonstrate the concept of TC po- hand, this creates a stronger asymmetric component. sition error, uncertainty, and its effects on TC EDA This artificial inner-core structure also limits the ac- within an idealized framework, which allowed us to curacy of the TC analysis and affects the intensity identify the impact by varying the TC positions in terms prediction, as expected. In comparison, when the po- of the distance from a reference storm center and the TC sition spread and error are small, the intensity un- intensity. Based on a case study of Typhoon Fanapi certainty becomes noticeable, and observations can be (2010), we then examine the performance of a newly better used to correct the intensity error with storm- developed TC-centered regional EDA system coupled scale features. Therefore, these results suggest that the with the WRF Model. Dropsonde observations from the TC position uncertainty and error should be first taken ITOP field campaign were used in the DA system, while care of before assimilating observations for storm- the model analysis and forecasts of the TC structure are scale corrections. APRIL 2018 L I N E T A L . 579

FIG. 15. The JTWC best-track estimate from 0000 UTC 16 Sep to 0600 UTC 20 Sep and the 3-day model forecast initialized at 0000 UTC 16 Sep (solid line) and 17 Sep (dashed line): (a) MSLP, (b) MSWS, (c) radius of 34-kt wind speed, and (d) RMW. The black, blue, and red lines represent the JTWC, CTL, and TCC results, respectively. The black exes in (d) are the RMWs computed from the SFMR observations during three C-130 ﬂights during Typhoon Fanapi.

Using a case study of Typhoon Fanapi (2010), the Furthermore, the TC in the TCC analysis has a better CTL and TCC experiments were conducted to employ resemblance to the SFMR surface wind observations 2 the conventional and the TC-centered WRF-LETKF (RMSD of 4.2 m s 1) than the TC in the CTL analysis 2 systems, respectively. CTL can be considered to be the (RMSD of 5.3 m s 1). This reinforces the advantage of case with a large position uncertainty and error (P30I2) using the TC-centered DA. under the second scenario in the idealized experiment, The TC-centered approach has a larger impact on the while TCC is the case with a small position uncertainty TC intensity forecast than the track forecast, based on and error (P5I2) under the first scenario. The TCC the 3-day forecasts initialized from the CTL and TCC framework constrains the position error and minimizes analyses at 0000 UTC 16 September. Both the CTL and the position uncertainty, which leads to a stable anal- TCC forecasts demonstrated comparable and good ysis performance with an improved TC inner-core track predictions with track errors at 72 h smaller than structure. Also, MSLP observations are used more ef- 100 km. However, in the CTL experiment, the strong fectively to fine-tune the TC since the QC procedure corrections cause an imbalance between the model dy- rejects fewer observations. In contrast, the position namic and analysis corrections and introduce a signifi- error in the CTL analysis is reduced mainly with the cant adjustment in TC movement and development help of the dropsonde data collected during ITOP. during the first 12-h forecast, particularly degrading the With the large position uncertainty, the TC in the CTL track prediction and creating large variations in TC size. analysis at different times varies greatly from having In comparison, the TCC analysis provides conditions large innovation and strong corrections, depending on favorable for TC development. In terms of intensity, the the availability of the observations around the TC area. TC in the TCC forecast quickly intensifies and reaches 580 WEATHER AND FORECASTING VOLUME 33

FIG. 16. (a) Surface wind speed of the CTL (blue) and TCC (red) forecasts at 0000 UTC 17 Sep, compared to SFMR (black solid line) and dropsonde (green ex marks) observations along the C-130 flight path from 0000 to 0345 UTC 17 Sep. (b) As in (a), but from 2200 UTC 17 Sep to 0200 UTC 18 Sep. an intensity similar to the JTWC values. Although the also partially explain why an initial TC with accurate peak wind is overpredicted, the inner-core structure in intensity ends up being overestimated during the the TCC forecast is closer to the observations than the forecast. one derived from the CTL forecast. These results sug- We should point out that the results presented in this gest that the TC-centered EnKF framework can im- study are from the first real case study implementing the prove the intensity prediction by refining the inner-core TC-centered DA framework, which was proposed by structure. NH14, in an ensemble-based high-resolution TC as- Although TCC has shown overall improvement on TC similation and prediction system. The experiment and forecasts, there are some mixed results in intensity verification with independent observations of Typhoon forecasting; for example, the TC intensity in TCC has Fanapi (2010) have shown that the TC-centered been overestimated during the first two days of the framework provides a positive impact on the TC anal- forecast but becomes close to the observations after- ysis, especially for the inner-core structure. Although ward. These results indicate the difficulties in TC in- the effects of TCC on TC assimilation are clearly tensity prediction. As discussed in many previous demonstrated, a question remains unanswered in this studies, the evolution of TC intensity is a complicated study as to whether a TC with different characteristics, process involving various factors, and the errors associ- such as intensity, size, and verticality, will affect the ated with these factors make numerical simulation and feasibility of TCC. We acknowledge that a case study prediction very challenging. The accuracy of the initial may not be enough to prove that the TCC approach conditions of the TC structure is only one of the factors; would be beneficial for all different kinds of TCs. others such as the model physics (Tao et al. 2011) and However, as we demonstrated in our idealized study, dynamics (Jin et al. 2014) also play important roles. In large TC position uncertainty will dominate the results addition, Typhoon Fanapi was affected by a preexisting of the TC assimilation, indicating that no matter the oceanic cold eddy on 17 September (Mrvaljevic et al. characteristics of the TC, the position uncertainty is a 2013), which is the second day of our forecast, and the first-order problem that we should take care of before results presented in this study have not yet considered conducting high-resolution TC assimilations. Except the effects of air–sea coupling on TCs, as shown in Chen for the results shown in this study, the idealized et al. (2013) and Lee and Chen (2012, 2014). This may study of NH14 also showed that their storm-centered APRIL 2018 L I N E T A L . 581 assimilation DA framework has delivered better per- Holland, G. J., 1980: An analytic model of the wind and pres- formance compared to the conventional EnKF when the sure profiles in hurricanes. Mon. Wea. Rev., 108, 1212– , TC is tilted. To further explore the impacts of the TCC 1218, https://doi.org/10.1175/1520-0493(1980)108 1212: AAMOTW.2.0.CO;2. approach, our next step is to systematically investigate Hsiao, L., C. Liou, and T. Yeh, 2010: A vortex relocation scheme the capability of this newly developed regional EDA for tropical cyclone initialization in Advanced Research WRF. system with more real TC cases. Mon. Wea. Rev., 138, 3298–3315, https://doi.org/10.1175/ 2010MWR3275.1. Acknowledgments. We are very grateful for the tech- Hunt, H. R., E. J. Kostelich, and I. Szunyogh, 2007: Efficient data assimilation for spatiotemporal chaos: A local ensemble trans- nical support and discussions with Drs. Brandon Kerns form Kalman filter. Physica D, 230, 112–126, https://doi.org/ and Milan Curcic of University of Washington and 10.1016/j.physd.2006.11.008. Dr. Falko Judt of NCAR. We thank three anony- Jin, H., M. S. Peng, Y. Jin, and J. D. Doyle, 2014: An evaluation of mous reviewers for their constructive review comments, the impact of horizontal resolution on tropical cyclone pre- which helped to improve the manuscript. KJL and SCY dictions using COAMPS-TC. Wea. Forecasting, 29, 252–270, https://doi.org/10.1175/WAF-D-13-00054.1. are supported by Ministry of Science and Technology Judt, F., and S. S. Chen, 2016: Predictability and dynamics of Grants MOST-102-2111-M-008-202-MY3 and MOST- tropical cyclone rapid intensification deduced from high- 105-2111-M-008-023. KJL is also supported by Ministry resolution stochastic ensembles. Mon. Wea. Rev., 144, 4395– of Science and Technology Grant MOST-103-2917-I- 4420, https://doi.org/10.1175/MWR-D-15-0413.1. 008-003. SSC is partially supported by a research grant ——, ——, and J. Berner, 2016: Predictability of tropical cyclone intensity: Scale-dependent forecast error growth in high-resolution from NASA under OVWST NNX14AM78G. stochastic kinetic-energy backscatter ensembles. Quart. J. Roy. Meteor. Soc., 142, 43–57, https://doi.org/10.1002/qj.2626. REFERENCES Kawabata, T., M. Kunii, K. Bessho, T. Nakazawa, N. Kohno, Y. Honda, and K. Sawada, 2012: Reanalysis and reforecast of Chen,S.S.,J.F.Price,W.Zhao,M.A.Donelan,andE.J. Typhoon Vera (1959) using a mesoscale four-dimensional Walsh, 2007: The CBLAST-Hurricane Program and the variational assimilation system. J. Meteor. Soc. Japan, 90, 467– next-generation fully coupled atmosphere–wave–ocean 491, https://doi.org/10.2151/jmsj.2012-403. models for hurricane research and prediction. Bull. Kieu, C., and Z. Moon, 2016: Hurricane intensity predictability. Amer. Meteor. Soc., 88, 311–317, https://doi.org/10.1175/ Bull. Amer. Meteor. Soc., 97, 1847–1857, https://doi.org/ BAMS-88-3-311. 10.1175/BAMS-D-15-00168.1. ——, W. Zhao, M. A. Donelan, and H. L. Tolman, 2013: Di- Kleist, D. T., 2011: Assimilation of tropical cyclone advisory min- rectional wind–wave coupling in fully coupled atmosphere– imum sea level pressure in the NCEP Global Data Assimila- wave–ocean models: Results from CBLAST-Hurricane. tion System. Wea. Forecasting, 26, 1085–1091, https://doi.org/ J. Atmos. Sci., 70, 3198–3215, https://doi.org/10.1175/ 10.1175/WAF-D-11-00045.1. JAS-D-12-0157.1. Kunii, M., 2015: Assimilation of tropical cyclone track and wind Chen, Y., and C. Snyder, 2007: Assimilating vortex position with an radius data with an ensemble Kalman filter. Wea. Forecasting, ensemble Kalman filter. Mon. Wea. Rev., 135, 1828–1845, 30, 1050–1063, https://doi.org/10.1175/WAF-D-14-00088.1. https://doi.org/10.1175/MWR3351.1. Kurihara, Y., M. A. Bender, and R. J. Ross, 1993: An initialization Chou, K.-H., and C.-C. Wu, 2008: Typhoon initialization in a me- scheme of hurricane models by vortex specification. Mon. soscale model—Combination of the bogused vortex and the Wea. Rev., 121, 2030–2045, https://doi.org/10.1175/1520-0493 dropwindsonde data in DOTSTAR. Mon. Wea. Rev., 136, (1993)121,2030:AISOHM.2.0.CO;2. 865–879, https://doi.org/10.1175/2007MWR2141.1. ——, ——, R. E. Tuleya, and R. J. Ross, 1995: Improvements in the D’Asaro, E., and Coauthors, 2014: Impact of typhoons on the GFDL Hurricane Prediction System. Mon. Wea. Rev., 123, ocean in the Pacific. Bull. Amer. Meteor. Soc., 95, 1405–1418, 2791–2801, https://doi.org/10.1175/1520-0493(1995)123,2791: https://doi.org/10.1175/BAMS-D-12-00104.1. IITGHP.2.0.CO;2. Davidson, N. E., and Coauthors, 2014: ACCESS-TC: Vortex speci- ——, R. E. Tuleya, and M. A. Bender, 1998: The GFDL Hurricane fication, 4DVAR initialization, verification, and structure di- Prediction System and its performance in the 1995 hurricane agnostics. Mon. Wea. Rev., 142, 1265–1289, https://doi.org/ season. Mon. Wea. Rev., 126, 1306–1322, https://doi.org/ 10.1175/MWR-D-13-00062.1. 10.1175/1520-0493(1998)126,1306:TGHPSA.2.0.CO;2. Evensen, G., 1994: Sequential data assimilation with a nonlinear Lee, C.-Y., and S. S. Chen, 2012: Symmetric and asymmetric struc- quasi-geostrophic model using Monte Carlo methods to tures of hurricane boundary layer in coupled atmosphere– forecast error statistics. J. Geophys. Res., 99, 10 143–10 162, wave–ocean models and observations. J. Atmos. Sci., 69, https://doi.org/10.1029/94JC00572. 3576–3594, https://doi.org/10.1175/JAS-D-12-046.1. Hamill, T., J. S. Whitaker, M. Fiorino, and S. G. Benjamin, 2011a: ——, and ——, 2014: Stable boundary layer and its impact on Global ensemble predictions of 2009’s tropical cyclones ini- tropical cyclone structure in a coupled atmosphere–ocean tialized with an ensemble Kalman filter. Mon. Wea. Rev., 139, model. Mon. Wea. Rev., 142, 1927–1944, https://doi.org/ 668–688, https://doi.org/10.1175/2010MWR3456.1. 10.1175/MWR-D-13-00122.1. ——, ——, D. T. Kleist, M. Fiorino, and S. G. Benjamin, 2011b: Leslie, L. M., and G. J. Holland, 1995: On the bogussing of tropical Predictions of 2010’s tropical cyclones using the GFS and cyclones in numerical models: A comparison of vortex pro- ensemble-based data assimilation methods. Mon. Wea. Rev., files. Meteor. Atmos. Phys., 56, 101–110, https://doi.org/ 139, 3243–3247, https://doi.org/10.1175/MWR-D-11-00079.1. 10.1007/BF01022523. 582 WEATHER AND FORECASTING VOLUME 33

Liou, C.-S., and K. D. Sashegyi, 2012: On the initialization of Tenerelli, J. E., and S. S. Chen, 2001: High-resolution simula- tropical cyclones with a three-dimensional variational analy- tions of Hurricane Floyd using MM5 with vortex-following sis. Nat. Hazards, 63, 1375–1391, https://doi.org/10.1007/ mesh refinement. 18th Conf. on Weather Analysis and s11069-011-9838-0. Forecasting/14th Conf. on Numerical Weather Prediction/ Liu, Q., T. Marchok, H. Pan, M. Bender, and S. Lord, 2000: Im- 9th Conf. on Mesoscale Processes, Fort Lauderdale, FL, provements in hurricane initialization and forecasting at Amer. Meteor. Soc., JP1.11, https://ams.confex.com/ams/ NCEP with global and regional (GFDL) models. NOAA pdfpapers/23165.pdf. Tech. Procedures Bull. 472, 7 pp., http://www.nws.noaa.gov/ Torn, R. D., and G. J. Hakim, 2009: Ensemble data assimilation om/tpb/472.pdf. applied to RAINEX observations of Hurricane Katrina Lu, X., X. Wang, Y. Li, M. Tong, and X. Ma, 2017: GSI-based (2005). Mon. Wea. Rev., 137, 2817–2829, https://doi.org/ ensemble-variational hybrid data assimilation for HWRF for 10.1175/2009MWR2656.1. hurricane initialization and prediction: Impact of various error ——, ——, and C. Snyder, 2006: Boundary conditions for limited- covariances for airborne radar observation assimilation. area ensemble Kalman filters. Mon. Wea. Rev., 134, 2490– Quart. J. Roy. Meteor. Soc., 143, 223–239, https://doi.org/ 2502, https://doi.org/10.1175/MWR3187.1. 10.1002/qj.2914. Ueno, M., 1989: Operational bogussing and numerical prediction Mrvaljevic, R. K., and Coauthors, 2013: Observations of the cold of typhoon in JMA. JMA/Numerical Prediction Division wake of Typhoon Fanapi (2010). Geophys. Res. Lett., 40, 316– Tech. Rep. 28, 48 pp. 321, https://doi.org/10.1029/2012GL054282. Weng, Y., and F. Zhang, 2012: Assimilating airborne Doppler Navarro, E. L., and G. J. Hakim, 2014: Storm-centered ensemble radar observations with an ensemble Kalman filter for data assimilation for tropical cyclones. Mon. Wea. Rev., 142, convection-permitting hurricane initialization and prediction: 2309–2320, https://doi.org/10.1175/MWR-D-13-00099.1. Katrina (2005). Mon. Wea. Rev., 140, 841–859, https://doi.org/ Nehrkorn, T., B. K. Woods, R. N. Hoffman, and T. Aulign, 2015: 10.1175/2011MWR3602.1. Correcting for position errors in variational data assimilation. Wu, C.-C., and Coauthors, 2005: Dropwindsonde Observations for Mon. Wea. Rev., 143, 1368–1381, https://doi.org/10.1175/ Typhoon Surveillance near the Taiwan Region (DOTSTAR): MWR-D-14-00127.1. An overview. Bull. Amer. Meteor. Soc., 86, 787–790, https:// Nguyen, H. V., and Y.-L. Chen, 2011: High-resolution initialization doi.org/10.1175/BAMS-86-6-787. and simulations of Typhoon Morakot (2009). Mon. Wea. Rev., ——, G.-Y. Lien, J.-H. Chen, and F. Zhang, 2010: Assimilation of 139, 1463–1491, https://doi.org/10.1175/2011MWR3505.1. tropical cyclone track and structure based on the ensemble Poterjoy, J., and F. Zhang, 2011: Dynamics and structure of fore- Kalman filter (EnKF). J. Atmos. Sci., 67, 3806–3822, https:// cast error covariance in the core of a developing hurri- doi.org/10.1175/2010JAS3444.1. cane. J. Atmos. Sci., 68, 1586–1606, https://doi.org/10.1175/ Wu,T.-C.,H.Liu,S.J.Majumdar,C.S.Velden,andJ.L. 2011JAS3681.1. Anderson, 2014: Influence of assimilating satellite-derived ——, and ——, 2014: Intercomparison and coupling of ensemble atmospheric motion vector observations on numerical and variational data assimilation approaches for the analysis analyses and forecasts of tropical cyclone track and in- and forecasting of Hurricane Karl (2010). Mon. Wea. Rev., tensity. Mon. Wea. Rev., 142, 49–71, https://doi.org/ 142, 3347–3364, https://doi.org/10.1175/MWR-D-13-00394.1. 10.1175/MWR-D-13-00023.1. ——, ——, and T. Weng, 2014: The effects of sampling errors on Yang, S.-C., K.-J. Lin, T. Miyoshi, and E. Kalnay, 2013: Improving the EnKF assimilation of inner-core hurricane observations. the spin-up of regional EnKF for typhoon assimilation and Mon. Wea. Rev., 142, 1609–1630, https://doi.org/10.1175/ forecasting with Typhoon Sinlaku (2008). Tellus, 65A, 20804, MWR-D-13-00305.1. https://doi.org/10.3402/tellusa.v65i0.20804. Pu, Z.-X., and S. A. Braun, 2001: Evaluation of bogus vortex ——, S.-H. Chen, S.-Y. Chen, C.-Y. Huang, and C.-S. Chen, 2014: techniques using four-dimensional variational data assimila- Evaluating the impact of the COSMIC-RO bending angle tion. Mon. Wea. Rev., 129, 2023–2039, https://doi.org/10.1175/ data on predicting the heavy precipitation episode on 16 June 1520-0493(2001)129,2023:EOBVTW.2.0.CO;2. 2008 during SoWMEX-IOP8. Mon. Wea. Rev., 142, 4139– Rappaport, E. N., and Coauthors, 2009: Advances and challenges 4163, https://doi.org/10.1175/MWR-D-13-00275.1. at the National Hurricane Center. Wea. Forecasting, 24, 395– Zhang, F., Y. Weng, J. A. Sippel, Z. Meng, and C. H. Bishop, 2009: 419, https://doi.org/10.1175/2008WAF2222128.1. Cloud-resolving hurricane initialization and prediction through Skamarock, W. C., and Coauthors, 2008: A description of the Ad- assimilation of Doppler radar observations with an ensemble vanced Research WRF version 3. NCAR Tech. Note NCAR/ Kalman filter. Mon. Wea. Rev., 137, 2105–2125, https://doi.org/ TN-4751STR, 113 pp., http://dx.doi.org/10.5065/D68S4MVH. 10.1175/2009MWR2645.1. Tallapragada, V., and Coauthors, 2014: Hurricane Weather ——, ——, J. F. Gamache, and F. D. Marks, 2011: Performance of Research and Forecasting (HWRF) Model: 2014 scientific convection-permitting hurricane initialization and prediction documentation. NCAR Development Testbed Center during 2008–2010 with ensemble data assimilation of inner- Rep., 105 pp., https://dtcenter.org/HurrWRF/users/docs/ core airborne Doppler radar observations. Geophys. Res. scientific_documents/HWRFv3.6a_ScientificDoc.pdf. Lett., 38, L15810, https://doi.org/doi:10.1029/2011GL048469. Tao, W.-K., J. J. Shi, S. S. Chen, S. Lang, P.-L. Lin, S.-Y. Hong, Zou, X., and Q. Xiao, 2000: Studies on the initialization and simu- C. Peters-Lidard, and A. Hou, 2011: The impact of microphysical lation of a mature hurricane using a variational bogus data as- schemes on hurricane intensity and track. Asia-Pac. J. Atmos. similation scheme. J. Atmos. Sci., 57, 836–860, https://doi.org/ Sci., 47, 1–16, https://doi.org/10.1007/s13143-011-1001-z. 10.1175/1520-0469(2000)057,0836:SOTIAS.2.0.CO;2.