EARLY ONLINE RELEASE

This is a PDF of a manuscript that has been peer-reviewed

and accepted for publication. As the article has not yet been

formatted, copy edited or proofread, the final published

version may be different from the early online release.

This pre-publication manuscript may be downloaded,

distributed and used under the provisions of the Creative

Commons Attribution 4.0 International (CC BY 4.0) license.

It may be cited using the DOI below.

The DOI for this manuscript is

DOI:10.2151/jmsj.2020-006

J-STAGE Advance published date: November 10th, 2019

The final manuscript after publication will replace the

preliminary version at the above DOI once it is available.

1

Typhoon Fanapi (2010) and Its Interaction with Taiwan 1

Terrain – Evaluation of the Uncertainty in Track, Intensity 2

and Rainfall Simulations 3

4

Yu-Feng LIN1, Chun-Chieh WU1*, Tzu-Hsiung YEN1, Yi-Hsuan HUANG1, 5

and Guo-Yuan LIEN2 6

7 1Department of Atmospheric Sciences, National Taiwan University, Taipei, 8

Taiwan 9

10 2Central Weather Bureau, Taipei, Taiwan 11

12

13

14

15

16

22 October, 2019 17

18

19

------------------------------------ 20

* Corresponding author Address: Chun-Chieh Wu, Department of Atmospheric Sciences, 21

National Taiwan University, No. 1, Sec. 4, Roosevelt Rd., Taipei 10617, Taiwan. 22

E-mail: cwu@typhoon.as.ntu.edu.tw 23

Current affiliation: Department of Atmospheric Sciences, National Taiwan University, No. 1, 24

Sec. 4, Roosevelt Rd., Taipei 10617, Taiwan. 25

2

Abstract 26

Using special data from the field program of “Impact of Typhoons on the Ocean in the 27

Pacific” (2010) and an ensemble Kalman filter–based vortex initialization method, this study 28

explores the impact of the Taiwan terrain on the uncertainty in forecasting track, intensity and 29

rainfall of Typhoon Fanapi (2010) based on ensemble simulations. The results show that the 30

presence of Taiwan topography leads to rapid growths of the simulation uncertainty in track and 31

intensity during the landfall period, in particular at the earlier landfall period. The fast moving 32

ensemble members show an earlier southward track deflection as well as the weakening of 33

intensity, resulting in a sudden increase of standard deviation in track and intensity. During the 34

period of offshore departure from Taiwan, our analysis suggests that the latitudinal location of 35

the long-lasting and elongated rainband to the south of tropical cyclone (TC) center has a strong 36

dependence on the latitude of the TC center. In addition, the rainfall uncertainty in southern 37

Taiwan is dominated by the uncertainty of simulated TC rainband, and the latitude of TC track 38

can be regarded as a good predictor of the rainband’s location at departure time. It is also found 39

that the rainband develop farther to the south as the topography is elevated. Considering the fact 40

that the rainband impinging the high mountains in the southern Central Mountain Range 41

generates the greatest accumulated rainfall, positions where the rainband associated circulation 42

and its interaction with topography appear to offer an explanation on the uncertainty of the 43

simulated rainfall. 44

45

Keywords Typhoon Fanapi; Taiwan topography; uncertainty; ensemble simulation; ensemble 46

Kalman filter.47

3

1. Introduction 48

The steep topography of Taiwan, especially the north-south oriented Central Mountain 49

Range (CMR), has a significant impact on the rainfall, track and intensity of landfalling 50

tropical cyclones (TCs). The complex topographic features of Taiwan and the interactions 51

between the environment, typhoon circulation and topography increase the difficulty in 52

predicting typhoon track, intensity change, and precipitation (e.g., Wu and Kuo 1999; Wu et 53

al., 2002; Jian and Wu 2008; Huang et al., 2011; Wu et al., 2013; Wu et al., 2015; Yu and Tsai 54

2017; Huang and Wu 2018). 55

The effect of topography on TC track has been extensively documented using both 56

real-case (e.g., Wu 2001; Jian and Wu 2008; Huang et al. 2011) and idealized numerical 57

simulations (e.g., Yeh and Elsberry, 1993a, 1993b; Lin et al. 1999; Lin et al. 2005; Huang et al. 58

2011; Wu et al. 2015; Huang and Wu 2018). One dynamical mechanism driving southward 59

track deflection is the acceleration of the tangential winds in the region between the 60

topography and storm center, referred to as the channeling effect (Jian and Wu 2008). Huang 61

et al. (2011) further investigated this effect using a simulation of Typhoon Krosa (2007) and a 62

set of numerical experiments with idealized designs of the vortex and environmental steering 63

flow. Performing also simulations with idealized designs, however, including a greater variety 64

of vortex and flow regimes, Wu et al. (2015) and Huang and Wu (2018) identified a second 65

mechanism for track deflection. They proposed that northerly asymmetric flow in the 66

midtroposphere is the primary feature causing the southward deflection of the simulated TC 67

tracks. 68

For the landfalling TCs, topography is considered to be a critical factor affecting the 69

structure and intensity change (e.g., Wu and Kuo 1999; Wu 2001; Chou et al. 2011; Yu and 70

Tsai 2017). Wu (2001) indicated that TC intensity is affected by the Taiwan terrain which cuts 71

4

off water supply to the boundary layer. In addition, Wu et al. (2009) demonstrated that the 72

frictional effect of terrain and the decrease of moisture and heat supply from the ocean are the 73

primary causes of the changes in the eyewall structure as well as the TC intensity. Based on 74

the microwave and radar images, Chou et al. (2011) also found all the 19 typhoons crossing 75

Taiwan from 2000-2010 weakened during the landfall period. 76

The steepness of the topography also plays a key role in substantially increasing the total 77

rainfall accumulation and in determining rainfall distribution over Taiwan (e.g., Wu and Kuo 78

1999; Lin et al. 2001; Wu 2001; Wu et al. 2002). Observational analyses of landfalling 79

typhoons in Taiwan have demonstrated that severe rainfall primarily occurs over mountains 80

due to orographic forcing (e.g., Yu and Cheng 2008). The distribution of rainfall is strongly 81

modulated by the topography of the CMR, with its southwest slope receiving especially large 82

amounts of rain (Chang et al. 1993), while the wind field associated with the TC also plays a 83

critical role (e.g., Wang 1989; Lee et al. 2006; Yu and Cheng 2013). 84

Previous studies have not only provided a better understanding of the impact of Taiwan 85

topography on landfalling typhoons but also improved the accuracy of track forecasts through 86

improvements on numerical weather models, advanced data assimilation schemes, and the 87

increased coverage of satellite observations (e.g., Wu et al. 2007; Chou et al. 2011; 88

Weissmann et al. 2011). Although there has been significant progress in TC simulations, 89

considerable uncertainties still exist in numerical TC simulations since the results of 90

simulations are affected by multiple factors, such as track variation, the complicated 91

interaction between TC circulation and topography, and cloud microphysics (Wu and Kuo 92

1999). Ensemble forecasts, which have been used extensively to study typhoon rainfall events 93

in the Taiwan region in recent years (e.g., Wu et al. 2010; Zhang et al. 2010; Fang et al. 2011; 94

Yen et al. 2011; Fang and Kuo 2013; Wu et al. 2013; Chen and Wu 2016; Lin et al. 2018), can 95

be used to quantify uncertainties in the track, intensity and rainfall of simulated TCs. Zhang et 96

5

al. (2010) demonstrated that simulations of extreme rainfall during Typhoon Morakot (2009) 97

are improved by the use of a convection permitting mesoscale ensemble data assimilation 98

system. In addition, through simulations of this typhoon using ensemble Kalman filter (EnKF; 99

Wu et al. 2010) with the Advanced Research WRF Model, Yen et al. (2011) demonstrated that 100

the storm translation speed is a key factor in determining the rainfall accumulation, and found 101

that a 55% increase in storm translation speed results in a 33% reduction in maximum rainfall 102

accumulations. Moreover, Fang et al. (2011) indicated that Taiwan topography plays an 103

important role in increasing rainfall variability in the case of Typhoon Morakot, but their 104

study did not explain how exactly the terrain affected the simulated variability. Wu et al. 105

(2013) investigated the impact of differences in TC track and rainfall distribution using 106

ensemble simulations of Typhoon Sinlaku (2008), demonstrating that uncertainties in rainfall 107

accumulation and distribution are strongly related to the variability of the ensemble tracks. 108

The interaction between a TC and the large scale flow is also a factor that influences 109

uncertainty in TC rainfall. For instance, Chen and Wu (2016) found that uncertainty in rainfall 110

distribution and accumulations increases in response to the mutual effects among the TC 111

circulation, topography and monsoon circulation. 112

Although the effect of Taiwan topography on the track, intensity and rainfall of the 113

landfalling TCs has been well investigated in many observational and numerical studies, yet 114

the uncertainty of such impact in the simulations has not been thoroughly explored in the 115

literature. To further understand the terrain’s effect on simulated uncertainties, ensemble 116

simulations on Typhoon Fanapi (2010) are conducted with a revised version of a WRF-based 117

EnKF data assimilation system (Wu et al. 2010, 2012), assimilating all available observation 118

data during the “Impact of Typhoons on the Ocean in the Pacific” (ITOP; D’Asaro et al. 2014) 119

field campaign in 2010. The purpose of this study is to evaluate the impact of CMR on the 120

uncertainty in the ensemble simulations of Fanapi and whether the uncertainties in track, 121

6

rainband positions and TC-associated rainfall have a mutual influence. A brief overview of 122

Fanapi is provided in Section 2. A description of the model configuration, experimental 123

design, and EnKF data assimilation scheme is given in Section 3. Section 4 presents the 124

simulation results, including comparisons of track, intensity, rainbands, and rainfall 125

distribution in the control and sensitivity experiments; in particular, the uncertainty in each 126

metric depicted by the ensemble simulations of each experiment group. Finally, a summary of 127

this study is shown in Section 5. 128

2. An overview of Typhoon Fanapi 129

Typhoon Fanapi formed as a tropical storm at 1200 UTC 15 September 2010 to the south 130

of the Ryukyu Islands (near 20.7°N, 127.6°E). It reached peak intensity with a minimum 131

central sea level pressure (MSLP) of 940 hPa at 2100 UTC 18 September as it moved 132

northwestward toward Taiwan, and made landfall at Hualien in eastern Taiwan at 0000 UTC 133

19 September. Prior to landfall, a significant southward deflection of the vortex track was 134

observed (Fig. 1a). The subtropical high was weak, and there was no obvious ambient steering 135

flow during this period (Fig. 1b). After landfall, Fanapi turned westward, leaving Taiwan at 136

1000 UTC 19 September and moving toward Fujian Province in China. 137

Figures 2a and 2b show the observed composite radar reflectivity at 0000 UTC and 0600 138

UTC 19 September 2010. It is evident that Fanapi was a compact TC with a clear eye and 139

organized inner eyewall, and remained symmetric upon landfall at Taiwan (Figs. 2a and 1a). 140

The radar reflectivity maximizes around the eyewall and in the outer rainband area 141

concentrated to the north side of Fanapi. This asymmetry is due to the interaction between the 142

typhoon circulation and the topography of northern Taiwan. Unsurprisingly, six hours after 143

Fanapi’s landfall, the eyewall nearly dissipated under the influence of the topography (Fig. 144

2b). During the same period, two primary bands of strong reflectivity were recognized. The 145

7

rainband active before landfall retains and become stronger and narrower. The other rainband 146

is located in the southwest quadrant of and closer to the storm’s center. This second banding 147

structure was roughly west–east elongating, extending from the Taiwan Strait toward the 148

mountains in the southern Taiwan. This rainband was possibly enhanced over the moist and 149

warm ocean and also through the interaction between the rainband-associated circulation and 150

its impinging on topography. The heaviest rainfall occurred in southwestern Taiwan, 151

especially on the windward side of the mountains due to orographic lifting. The 2-day 152

accumulated rainfall in Taiwan from 0000 UTC 18 September to 0000 UTC 20 September 153

2010 observed by using rain gauges showed two areas of maximum rainfall concentrated in 154

Yilan County (about 485 mm) and Kaohsiung-Pingdong Counties (about 951 mm on the 155

plains and 1127 mm in mountainous area) (Fig. 2c). 156

Changes in the track and structure of typhoon Fanapi as it passed across Taiwan included 157

southward track deflection before landfall, sudden changes of intensity and structure during 158

landfall, and changes in rainfall distribution. Some scientific issues related to the effect of 159

Taiwan topography on Typhoon Fanapi’s track, intensity and rainfall have been discussed in 160

the literature (e.g., Wang et al. 2013; Huang et al. 2016; Liou et al. 2016; Yang et al. 2018). 161

Wang et al. (2013) proposed that the asymmetric latent heating induced by Taiwan topography 162

can reduce Fanapi’s translation speed during its offshore departure and thus lead to heavy 163

rainfall over southwestern Taiwan. In addition, Huang et al. (2016) indicated that the double 164

intense rainfall peaks cannot be simulated as the CMR height is reduced or latent heating is 165

turned off in the model after Fanapi’s landfall. Although the above studies have demonstrated 166

that the topography of Taiwan has a significant impact on Typhoon Fanapi, they did not 167

explicitly address to which extent the CMR affects the simulation uncertainty as Fanapi 168

approached and passed across Taiwan. Therefore, in this study, we focus on the role of Taiwan 169

topography in the variability of track, intensity and rainfall simulations. 170

8

3. Methodology and experimental design 171

In this study, The Advanced Research WRF Model (version 3.2.1)–based ensemble 172

Kalman filter (EnKF; Evensen 1994, 2003) data assimilation system is used to simulate 173

Typhoon Fanapi. The system was initially developed by Meng and Zhang (2008a, b) and was 174

later considerably revised by Wu et al. (2010) to include additional parameters to improve the 175

vortex initialization and to add the vortex-following moving-nest capability for 176

high-resolution TC assimilation. The horizontal resolution in the experimental setup is 54 km 177

(121 × 91 grid points), 18 km (73 × 73 grid points), and 6 km (97 × 97 grid points) in the 178

first (D1), second (D2), and third (D3) domains respectively (Fig. 3). The D2 and D3 domains 179

follow the movement of the storm to resolve the structure of Fanapi. Thirty-five vertical levels 180

are used in the terrain-following sigma coordinate. The physical parameterizations used in the 181

simulation include the WRF single-moment 6-class microphysics scheme (WSM6; Hong and 182

Lim 2006), the Yonsei University (YSU) planetary boundary layer scheme (Hong et al. 2006), 183

the Rapid Radiative Transfer Model (RRTM) scheme (Mlawer et al. 1997) for longwave 184

radiation, and the Dudhia shortwave scheme (Dudhia 1989) for shortwave radiation. In the 185

two coarse domains (D1 and D2), cumulus convection is parameterized with the Grell–Freitas 186

ensemble scheme (Grell and Freitas 2014). 187

The simulations are initialized based on the National Centers for Environmental 188

Prediction (NCEP) global reanalysis at 1° by 1° at 1800 UTC 14 September 2010. The 28 189

ensemble members are produced by randomly perturbing the mean analysis on a transformed 190

streamfunction field as described in Zhang et al. (2006). The EnKF data assimilation system 191

update cycle is conducted every 1 h during the initialization from 1800 UTC 14 September to 192

0200 UTC 18 September 2010 using convectional data and additionally the aircraft 193

dropwindsonde observations collected during ITOP from the Dropwindsonde Observations 194

for Typhoon Surveillance near the Taiwan Region (DOTSTAR) Astra (Wu et al. 2005) and 195

9

three U.S. C-130 flights (Fig. 3). The TC center position, minimum central sea level pressure, 196

and axisymmetric wind profile structure are also assimilated as a special technique to improve 197

the initialization of the TC vortex, mainly following the method described in Wu et al. (2010, 198

2012), except that in this study the minimum central sea level pressure assimilation is added, 199

the storm motion vector assimilation is removed, and the axisymmetric wind profile is 200

constructed and assimilated at the C-130 flight level, 700-hPa, instead of at the surface. The 201

D2 domain is added from the beginning of the assimilation period (1800 UTC 14 September) 202

while the D3 is added later at 1200 UTC 15 September. 203

After the cycling EnKF data assimilation is performed, the control experiment (CTL) is 204

designed as a 3-day ensemble forecasts of 28 members conducted with the initial conditions 205

taken at 0200 UTC 18 September 2010. In order to improve our understanding of the physical 206

processes affecting the track, intensity and rainfall variability, a series of experiments with a 207

30% increase (H130) and a 50% reduction (H50) of the terrain height of Taiwan, and no 208

Taiwan terrain (H00) are conducted (Table 1). In addition to the 28-member ensemble 209

forecasts, the deterministic forecast initialized from the ensemble mean (the average of 28 210

members) analysis are also conducted for each experiment, denoted as CTL-M, H130-M, 211

H50-M, and H00-M, respectively. Furthermore, to examine the effect of model resolution and 212

to better resolve the terrain-induced precipitation, another high-resolution deterministic 213

forecast experiment, denoted Fix2km-M, is carried out using an independent domain setup 214

with 4 nested domains, all of which are fixed rather than vortex-following. Here the 215

horizontal resolutions and sizes of the second (D2), third (D3), and fourth (D4) domains are 216

18 km (124 × 91 grid points), 6 km (250 × 154 grid points), and 2 km (196 × 238 grid points), 217

respectively. The area of D4 is just to cover the Taiwan terrain as shown in Fig. 3 (while the 218

D2 and D3 in this experiment are not shown). The initial conditions of these fixed inner 219

domains are regridded from the moving nested analysis using the data from the finest 220

10

available domains. 221

4. Results 222

4.1 The effect of Taiwan topography on the simulation 223

Figure 4 shows the simulated ensemble tracks and ensemble mean of 28 members in the 224

experiments with terrain from 0200 UTC 18 September 2010. In CTL experiment, the 225

ensemble mean is generally consistent with the best track analyzed by the Central Weather 226

Bureau (CWB) of Taiwan although it is slightly south of the best track after Fanapi moves 227

away from Taiwan, while the ensemble members show a small spread throughout the 228

integration. The track forecast errors in CTL-M are relatively small during the entire 229

simulation period (the largest track error is about 120 km at 40 h), indicating that with EnKF 230

data assimilation performs reasonably well in analyzing the environmental flow and the TC 231

vortex. As the terrain height is increased by 1.3 times (H130), the southward track deflection 232

before landfall becomes more significant. This defection is caused by the northerly 233

asymmetric flow induced by the Taiwan terrain in the midtroposphere (figures not shown), 234

which is also found in idealized simulation in Wu et al. (2015) and Huang and Wu (2018). In 235

addition, as the terrain height is decreased by one half (H50), the ensemble tracks do not have 236

a significant direction shift as TCs approach and cross Taiwan. When the terrain of Taiwan is 237

removed (H00), the track of all members change markedly, especially during the landfall 238

period. The fact that the southward track deflection is hardly seen before landfall in the no 239

terrain experiment (H00) highlights the critical role of terrain in these simulations. 240

The intensity of CTL-M at the initial time is about 945 hPa, which is very close to CWB 241

observations (950 hPa) (Fig. 5; note the time goes from right to left). Prior to landfall, both 242

the observed and simulated intensities reach a minimum of about 940 hPa (4 h earlier in 243

CTL-M than from CWB), which is consistent with the statistical results in Brand and Blelloch 244

11

(1974) and the idealized simulation in Bender et al. (1987). After landfall, under the influence 245

of the Taiwan terrain, the TC intensity rapidly decreases with time both in observations and 246

simulations. The time evolution of simulated TC intensity of 28 ensemble members is 247

generally consistent with CTL-M, and the intensity averaged by 28 ensemble members is 248

nearly identical to the simulation in CTL-M at each time. This result indicates that the 249

simulation with initial condition integrated from 28 ensemble members (CTL-M) can also 250

represent the mean intensity tendency obtained from all ensemble members. In addition, as 251

the terrain height is increased, TC intensity decrease at a faster rate compared to those in 252

experiments with lower or no terrain. During the period from landfall in (at 1800 UTC 18 253

September) to departure from Taiwan (at 1200 UTC19 September), the increase of minimum 254

sea level pressure in H130-M, CTL-M, H50-M is 40.8 hPa, 33.7 hPa, 23.3 hPa, respectively. 255

However, in H00, the intensity of simulated TCs remains steady as the vortex makes landfall. 256

Figure 6 shows the 2-day accumulated rainfall from 0000 UTC 18 September to 0000 257

UTC 20 September 2010 for each experiment initialized from the ensemble mean. Although 258

the simulation in CTL-M captures the rainfall distribution, the extreme rainfall accumulations 259

over the plains and mountains in Kaohsiung and Pingdong Counties are underestimated, while 260

those over the mountains near Yilan County are overestimated (Fig. 6a). When the terrain 261

height is increased by 1.3 times (H130), the accumulated rainfall increases in all regions of 262

Taiwan, with the maximum rainfall in southern Taiwan increased by 47.5% in the mountains 263

and by 11% in the plain area and by 19.8% in the northeastern Taiwan as compared to CTL 264

(Figs.6a, d). In addition, if the terrain height is reduced by one half (H50), less rainfall is 265

produced in the northeast and southwest of Taiwan, but the rainfall pattern still can be 266

captured (Fig. 6c). When there is no terrain, without enough orographic lifting, the rainfall is 267

mainly contributed from by TC’s inner core and rainband and thus the rainfall distribution and 268

intensity show a marked difference (Fig. 6b). Results from cases with 2-km resolution show 269

12

that the maximum accumulated rainfall increases as the horizontal resolution increases, 270

especially in mountainous areas (Fig. 6e). While the maximum rainfall accumulation in 271

southwestern Taiwan increases from 804 mm in CTL-M to 1178 mm in the higher-resolution 272

simulation (Figs. 6a, e), which is very close to the observed value of 1127 mm (Fig. 2c), the 273

maximum rainfall accumulation in northeastern Taiwan also significantly increases to 1103 274

mm. It should be noted that even though the horizontal resolution is enhanced to 2 km, the 275

model is still unable to precisely capture the peak rainfall accumulation (mostly with an 276

overestimation in northeastern Taiwan and an underestimation for the plains in southwestern 277

Taiwan). This is consistent with the known limitations of models in simulation of heavy 278

rainfall, which has previously been attributed to the interaction of the storm and topography 279

and limitations in the physical parameterizations (e.g., Wu and Kuo 1999; Zhu and Zhang 280

2006; Tao et al. 2011). For a reliable assessment of the model performance, we also need to 281

take into account the insufficient rain gauge observations over mountainous areas in Taiwan, 282

where the largest errors in simulated rainfall occur. In all, the above results imply that the 283

terrain of Taiwan plays a vital role in affecting the variability in TC track and intensity as well 284

as the accumulated rainfall. 285

4.2 The impact of Taiwan topography on the uncertainty of track and intensity 286

simulation 287

Figure 7 shows the time evolution of ensemble mean track error (histograms) and 288

standard deviation of the storm center position (solid lines). The mean track errors show only 289

a gradual increase before landfall (at 0000 UTC 19 September). In H00, the track error rapidly 290

increases during the landfall period and can be explained by the absence of southward track 291

deflection as shown in Fig. 4, indicating that the topography has a significant impact on the 292

track error. The maximum track error is a deviation of around 80 km occurring at 3 h after 293

13

landfall (at 0300 UTC 19 September), after which the track error continues to decrease until 294

the TC moves into Fujian Province in China (at 1200 UTC 19 September). However, the track 295

errors in the CTL, H50 and H130 experiments have a similar but smaller variation as 296

compared to H00 after landfall. The standard deviation of the storm center position has a 297

sudden increase as compared to the original rising trend after landfall in CTL and H130 298

(especially for H130), while there is no noticeable increase in H50 and H00. The experiments 299

with higher terrain show larger variations in standard deviation. The reason for the sudden 300

increase in standard deviation during the earlier landfall period is that the higher terrain can 301

induce more significant southward track deflection, and thus some ensemble members with 302

faster translation speed are deflected southward earlier by the terrain compared to those 303

ensemble members with slower translation speed (Fig. 8a, d), resulting in a significant 304

variation in track position. In H50 and H00, the terrain does not have an obvious impact on 305

southward track deflection during landfall. Therefore, there is no noticeable change in 306

standard deviation (Fig. 8b, c). However, during the period after landfall from 0400 UTC to 307

1200 UTC 19 September, the standard deviation in CTL and H130 is significantly reduced, 308

but still maintaining an increasing trend in both H50 and H00. For the westward-moving TCs 309

approaching and crossing Taiwan, Wang (1980) and Yeh and Elsberry (1993b) suggested the 310

storm center would have a discontinuous track due to the formation of terrain-induced low 311

pressure trough on the lee side of CMR. This process is also found in our simulation as TCs 312

cross the terrain of Taiwan (figures not shown), resulting in the simulated TC center moving 313

toward the low pressure trough region. Therefore, a concentration of forecasted TC tracks 314

associated with smaller track spread is identified during this period (Fig. 8a, d). However, 315

when there is no terrain, this low pressure trough does not exist, and thus the standard 316

deviation increases continuously with time (Fig. 8b). In all, the presence of terrain not only 317

changes TC’s movement, but also increases the track uncertainty during the landfall period, 318

14

especially for the cases with higher terrain. 319

The time evolution of ensemble mean MSLP error and standard deviation of MSLP in 320

Fig. 9 reveals that the experiments show a small negative mean MSLP error at 4 h before 321

landfall. After this time, the mean MSLP error increases with time while the largest mean 322

MSLP error occurs at 2 h after landfall. In addition, the experiments with higher terrain show 323

a larger growth of mean MSLP error. In CTL and H130, the intensity is clearly weakened as 324

compared to observation after landfall (cf. Fig. 5), and thus there is a positive mean MSLP 325

error during the landfall period. In H00, the TCs maintain their intensity and show a negative 326

mean MSLP error after 0600 UTC 19 September with an average error of about -20 hPa. On 327

the other hand, the standard deviation in MSLP for CTL and H130 does not show significant 328

spread variation until 4 h before landfall, but the variation starts to increase after this time. 329

The sudden increase of standard deviation begins in H130 about 3-h earlier than in CTL, and 330

the increasing rate is also larger in H130 during the landfall period. The largest value of 331

standard deviation in CTL and H130 occurs at approximately 1~2 h after landfall, and then 332

reduces gradually with time. Meanwhile, the standard deviation is roughly reduced to the 333

same level as before landfall when the TCs leave Taiwan (at 1200 UTC 19 September). This 334

result can be explained by the difference in the TC centers and the timing when the TCs begin 335

to be influenced and weakened by terrain in each ensemble member (Fig. 8a, d). As all 336

ensemble members hit the Taiwan terrain, the degrees of their intensity reduction are quite 337

similar and therefore the standard deviation among all of them starts to reduce gradually. 338

Although TC intensity during the landfall period is also influenced by terrain in H50, the 339

weakening of the TC intensity is smaller and thus the terrain plays a rather minor role in 340

intensity variability. In H00, there is no change in intensity variability due to the lack of 341

influence from the terrain. In all, the existence of Taiwan terrain can result in the sudden 342

increase of intensity uncertainty, especially for the experiments with higher terrain. When the 343

15

terrain is removed, the forecast errors increase during the landfall period, while the intensity 344

uncertainty remains the same. 345

4.3 The uncertainties in rainfall simulation in the CTL 346

Figure 10 shows the 2-day accumulated rainfall associated with 28 ensemble members in 347

southern Taiwan from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 in CTL. Although the 348

ensemble members can capture the main rainfall characteristics as compared to observations, 349

the difference of maximum rainfall distribution as well as the rainfall variability among the 350

ensemble members can be clearly identified. For instance, CTL-M06 (the 6th ensemble 351

member in CTL) has less accumulated rainfall and its location of maximum rainfall shows a 352

northward deviation. In contrast, CTL-M24 (the 24th ensemble member in CTL) has more 353

accumulated rainfall and its location of maximum rainfall shows a southward deviation. In 354

CTL-M06, the track shows a northerly movement deviation as compared to the track in 355

CTL-M during the landfall period (Fig. 11a). To provide further insight into the processes of 356

the induced rainfall uncertainties, trajectory analyses of the hourly model output are 357

conducted using the Read/Interpolate/Plot (RIP) program. The twenty 6-h backward air parcel 358

trajectories released at 925 hPa in the rainfall region over the southwest of Taiwan show that 359

these air parcels originate from the outer TC circulation and move through Taiwan Strait 360

toward the southwest of Taiwan. The low-level jet-like flow defined as horizontal wind speed 361

beyond 35 m s-1 also extend from northern part of the Taiwan Strait to the storm’s southern 362

inner core region. The rainband develops in the south of TC’s center where the air flows slow 363

down and converge to the south of the TC, and this rainband associated with low-level jet-like 364

flow moves inland and impinges the topography. This suggests that heavy rainfall is induced 365

by the interaction between this rainband, TC circulation, and the Taiwan topography, 366

especially near Chiayi County, where the maximum rainfall is observed. However, the track in 367

16

CTL-M24 shows a southerly movement deviation (Fig. 11b) and the air parcels also induce 368

the confluence zone to the south of the TC associated with a developing convective rainband. 369

This rainband is located father to the south as compared to CTL-M06, and thus resulting in a 370

significant southward shift of the maximum rainfall in CTL-M24. 371

Figure 12 shows the simulated maximum radar reflectivity of 28 ensemble members as 372

the storm centers move away from the west coast of Taiwan (i.e., toward 120.1 E°) with the 373

axis of the TC rainband defined as the central point of the rainband’s width (Yu and Tsai 374

2017). It is found that the storm centers show spreads in latitude, while the TC rainbands and 375

their axis among the ensemble members also have similar spreads. This spread of TC 376

rainband is related to the rainfall uncertainty in southern Taiwan. To further examine the cause 377

of the rainfall uncertainty, the CTL ensemble members are classified into two groups 378

according to the distribution of these rainband’s axes: the north-biased rainband group (NG) 379

and the south-biased rainband group (SG) as compared to the rainband’s axis of CTL-M. Each 380

group contains 8 ensemble members, and the distribution of their rainband’s axis and the 381

corresponding of TC track and intensity for each member are shown in Fig. 13. It shows that 382

the higher (lower) latitude of rainband in NG (SG) are correlated to higher (lower) latitude of 383

storm center at the time as TCs leave Taiwan (Fig. 13a), and the correlation coefficient 384

between the latitudinal location of storm center and rainband is 0.8 among all the ensemble 385

members. Meanwhile, the higher latitude of rainband at the departure time corresponds with a 386

higher latitude of TC track and a rapid weakening of TC intensity during the landfall period 387

(Fig. 13b, c). The time-lag ensemble correlation (i.e., autocorrelation) of at least 0.6 exists 388

between the latitude of TC track and the latitude of storm center at the departure time (Fig. 389

14), suggesting that a higher the latitude of TC track from the initial time would expectedly 390

have a more northern location of the storm center at the departure time. However, there is no 391

significant correlation to the longitude of TC track. In addition, the correlation between MSLP 392

17

(maximum wind speed) and the latitude of storm center at the departure time shows that the 393

correlation rapidly increases (decreases) to 0.6 (-0.6) after landfall, implying that the stronger 394

(weaker) intensity after landfall would expectedly have a more southern (northern) location of 395

the storm center at the time as TCs leave Taiwan. 396

The simulated composite maximum radar reflectivity shows that a west-east elongated 397

rainband extends from southern Taiwan Strait toward the inland of southwestern Taiwan in 398

both groups, except that the latitude of rainband in SG is more southern than NG (Fig. 15). In 399

addition, the region of stronger composite radar reflectivity (> 35 dBZ) is more widespread in 400

SG than NG. Meanwhile, the low level westerly flow is also stronger and farther to the south 401

over the southwest of Taiwan in SG, and its impinging on the southern portion of the terrain 402

results in stronger orographic lifting as well as heavier rainfall in the windward side of the 403

mountains (Figs. 15b and 16b). In contrast, the region where the weaker low level flow 404

associated with narrower rainband interacts with the terrain is slightly farther to the north in 405

NG, which leads to less rainfall as well as the location of maximum rainfall shifts farther to 406

the north (Figs. 15a and 16a). Except for the difference in location of heavy rainfall, the 407

maximum rainfall also shows a difference over south of Taiwan where the maximum rainfall 408

difference can reach about 350 mm in the southern mountains (Fig. 16c). 409

The above results suggest that the uncertainty source of accumulated rainfall is primarily 410

determined by the variability of the TC’s rainband associated with the interaction between the 411

storm circulation and the terrain of Taiwan. In addition, the simulated track latitude from the 412

initial time can be regarded as a good predictor of the latitude of storm center as well as the 413

location of the TC rainband at the time when TC is leaving Taiwan. Although TC intensity 414

after landfall seems to be another predictor of the latitude of storm center at departure time, 415

the high correlation between intensity and the latitude of storm center is also related to the 416

latitude of TC track (i.e. the intensity for those northward landing members are weakened 417

18

more rapidly as they cross a higher and wider terrain to the north while the southward landing 418

members show a less intensity weakening as they cross a lower and narrower terrain to the 419

south). 420

4.4 The effect of terrain height on the rainfall uncertainty 421

The 2-day accumulated rainfall standard deviations indicate that the largest rainfall 422

variability is located in the east and south of Taiwan, especially in mountainous areas (Fig. 423

17). The accumulated rainfall variability increases as the terrain height increases and reaches 424

its maximum in H130 in the mountainous area in southern Taiwan. The area with the most 425

intense accumulation of rainfall shown in Fig. 6 coincides with the largest rainfall variance. 426

As compared to the distribution of each rainband’s axis in those experiments with terrain at 427

the departure time, it is found that the latitude of the storm centers from all ensemble mean 428

experiments (i.e., CTL-M, H50-M, and H130-M) and observation are in good agreement (Fig. 429

18). However, the distribution of each rainband’s axis has a large deviation in latitude among 430

the experiments with different terrain heights. The axis of the TC rainband in ensemble mean 431

shifts to the south as terrain height increases, and this result is also found in their ensemble 432

members. The coefficient of determination R2 of the regression among the 28 ensemble 433

members in CTL, H130, and H50 is 0.80, 0.79 and 0.91 respectively (Fig. 19), indicating that 434

there is a strong relationship between the latitude of storm center and the rainband’s latitude, 435

which can have significant impact on the rainfall distribution. In addition, the simulation 436

result also shows that the rainband’s axis in CTL-M biases farther to the south as compared to 437

observation. A further investigation of this bias involves the model’s parameterizations or the 438

existence of systematic error in the simulation, which is beyond the scope of this study. 439

To better understand how the Taiwan terrain affect the location of the TC rainband as TCs 440

leave Taiwan, the 925-hPa wind field and backward air parcel trajectories are shown in Fig. 441

19

20. The analyzed results show that as the terrain height is increased, the wind field becomes 442

more asymmetric with a long and narrow north-south flow over the Taiwan Strait. Meanwhile, 443

the low-level wind spirals into the inner core of the TC associated with high water vapor 444

mixing ratio. This low-level wind stream plays an important role in leading the convergence 445

line and the TC rainband to move southward away from the storm center. This can be clearly 446

seen in Fig. 21 where the convergence line induced by the asymmetric flow is located farther 447

to the south in the experiments with higher terrain (e.g., H130-M and CTL-M). In all, the 448

terrain of Taiwan not only changes the distribution and the total amount of the simulated 449

rainfall but also induces the asymmetric flow, which increases the variability in the location of 450

the TC rainband and further impacts rainfall uncertainty. 451

5. Summary 452

Typhoon Fanapi (2010) produced heavy rainfall over the mountainous areas of Taiwan, 453

with a maximum 2-day accumulated rainfall of 1127 mm over the southern ridge of CMR. 454

Extensive data collected during the ITOP program are used to provide a more feasible 455

delineation of the environment and typhoon vortex, and thus a unique opportunity to 456

investigate those key processes affecting TC evolution. In this study, a WRF-based EnKF data 457

assimilation system and all available ITOP data are integrated to evaluate the impact of CMR 458

on the uncertainty in the ensemble simulations of Fanapi as well as the sources of the 459

uncertainty. 460

The results show that the characteristics of the typhoon are generally well captured and 461

the track spread is small throughout the integration of the control experiment, indicating a 462

case with reasonably high predictability. In addition, the southward track deflection caused by 463

the northerly asymmetric flow in the midtroposphere before landfall is more significant in the 464

experiments with higher terrain, which is in agreement with previous studies in Wu et al. 465

20

(2015) and Huang and Wu (2018). Moreover, the total accumulated rainfall increases in all 466

regions of Taiwan as terrain height increases. The rainfall pattern is still captured in the 467

experiment with halved terrain while the total rainfall accumulations are decreased, and the 468

rainfall distribution is markedly changed when the topography is removed. With 2-km 469

horizontal grid resolution, the model is still unable to precisely identify the peak rainfall 470

amount especially in northeastern Taiwan and on the plains of southwestern Taiwan. 471

The ensemble members with faster translation speed are influenced earlier by terrain and 472

show an earlier southward track deflection and the weakening of intensity as compared to 473

those ensemble members with slower translation speed, resulting in a significant increase of 474

standard deviation in storm center position and MSLP, especially in the experiments with 475

higher terrain (e.g., H130 and CTL). In addition, the terrain-induced low pressure trough 476

formed on the lee side of CMR tends to cause the TC to move into the same region, and thus 477

resulting in smaller track spread as TCs cross the CMR in both CTL and H130. However, the 478

terrain does not have a clear impact on the southward track deflection and on the formation of 479

terrain-induced low pressure trough in H50 and H00, and thus there is no noticeable change of 480

standard deviation in storm center position and MSLP. These results indicate the important 481

impact of Taiwan topography on the simulated uncertainty in both track and intensity. 482

Further examination of the cause of the rainfall uncertainty in the ensemble, by 483

classifying two composite groups according to the distribution of the axis of the TC rainband 484

that the north-biased rainband group (NG) and the south-biased rainband group (SG), reveals 485

that the higher (lower) latitude of rainbands in NG (SG) group are correlated to higher (lower) 486

latitude of storm center at the time as TCs leave Taiwan. In addition, a significant correlation 487

is found between the latitude of storm center at departure time and earlier track latitude 488

starting at the initial time, suggesting that simulated track latitude at earlier time in the 489

forecast can be regarded as a good predictor of the storm center latitude as well as the location 490

21

of TC rainband when TC is leaving Taiwan. Moreover, both NG and SG groups show a 491

west-east elongated rainband extending from southern Taiwan Strait toward southern Taiwan, 492

except that the latitude of rainband associated with stronger low level flow and its interaction 493

with terrain in SG occurs more towards the south than in NG, and also leads to a stronger 494

orographic lifting as well as heavier rainfall in the southern CMR. These results suggest that 495

positions where the rainband associated circulation and its interaction with topography appear 496

to offer an explanation on the uncertainty of the simulated rainfall. 497

The 2-day accumulated rainfall standard deviations indicate that the largest rainfall 498

variability is found in the mountainous areas of eastern and southern Taiwan. The 499

accumulated rainfall variability increases as the terrain height is increased. In addition, the TC 500

vortex becomes more asymmetric in the cases with higher terrain (e.g., H130 and CTL), and 501

the low-level flow spirals from northern part of Taiwan Strait into the inner core of the TC 502

associated with high water vapor mixing ratio, which can contribute to the development of the 503

convective rainband in the south side of the simulated TC. Moreover, the convergence line as 504

well as TC rainband induced by the asymmetric flow has a more southward shift as the terrain 505

height increases. The presence of Taiwan topography does not only change the characteristics 506

of simulated rainfall but also induces the asymmetric flow, which increases the variability in 507

the location of the TC rainband and further impacts rainfall uncertainty. 508

In summary, forecasting typhoon-induced rainfall near Taiwan has long been a 509

challenging issue due to the complicated interactions among environmental flow, typhoon 510

circulation and the topography of Taiwan. This study highlights the impact of Taiwan 511

topography on the uncertainties in TC simulations. We note that the main source of 512

uncertainties in this study comes from the initial condition perturbations based on the EnKF, 513

while the uncertainties related to model errors are not considered under the current framework. 514

For example, the near-coastal rainfall peak on the plains of southwestern Taiwan is 515

22

underestimated generally in all members and experiments, which is likely attributed to the 516

model errors and thus limits the discussion in this study. Multi-model ensemble forecasts such 517

as applying different physical parameterization schemes or parameters in each member can be 518

useful in investigation of this issue. Therefore, considerable issues are yet to be studied in the 519

future to further understand the source of uncertainties associated with TC simulations as well 520

as the rainfall peaks near coastal and mountainous areas, including the initial errors and model 521

errors (e.g., the uncertainties of specific dynamics and physical processes), especially under 522

the influence of the topography of Taiwan. 523

23

Acknowledgements 524

This work is supported by the Ministry of Science and Technology of Taiwan under 525

Grants MOST 105-2628-M-002-001 and MOST 106-2111-M-002 -013 -MY3. Valuable 526

comments from Kosuke Ito, two anonymous reviewers that helped improve the quality of the 527

manuscript are highly appreciated. 528

24

References 529

Bender, M. A., R. E. Tuleya, and Y. Kurihara, 1987: A numerical study of the effect of island 530

terrain on tropical cyclones. Mon. Wea. Rev., 115, 130–155. 531

Brand, S., and J. W. Blelloch, 1974: Changes in the characteristics of typhoons crossing the 532

island of Taiwan. Mon. Wea. Rev., 102, 708–713. 533

Chang, C.-P., T.-C. Yeh, and J. M. Chen, 1993: Effects of terrain on the surface structure of 534

typhoons over Taiwan. Mon. Wea. Rev., 121, 734–752 535

Chen T.-C., and C.-C. Wu, 2016: The remote effect of Typhoon Megi (2010) on the heavy 536

rainfall over northeastern Taiwan. Mon. Wea. Rev., 144, 3109–3131. 537

Chou, K.-H., C.-C. Wu, P.-H. Lin, S. D. Aberson, M. Weissmann, F. Harnisch, and T. 538

Nakazawa, 2011: The impact of dropwindsonde observations on typhoon track forecasts 539

in DOTSTAR and T-PARC. Mon. Wea. Rev., 139, 1728–1743. 540

——, ——, Y. Wang, and C.-H. Chih, 2011: Eyewall evolution of typhoons crossing the 541

Philippines and Taiwan: An observational study. Terr. Atmos. Ocean. Sci., 22, 533-548. 542

D’Asaro, E. A., and Coauthors, 2014: Impact of typhoons on the ocean in the Pacific. Bull. 543

Amer. Meteor. Soc., 95, 1405–1418. 544

Dudhia, J., 1989: Numerical study of convection observed during the Winter Monsoon 545

Experiment using a mesoscale two dimensional model. J. Atmos. Sci., 46, 3077–3107. 546

Evensen, G., 1994: Sequential data assimilation with a nonlinear quasi-geostrophic model 547

using Monte Carlo methods to forecast error statistics. J. Geophys. Res., 99, 10 143–10 548

162. 549

——, 2003: The ensemble Kalman filter: Theoretical formulation and practical 550

implementation. Ocean Dyn., 53, 343–367. 551

Fang, X., and Y.-H. Kuo, 2013: Improving ensemble-based quantitative precipitation forecasts 552

for topography-enhanced typhoon heavy rainfall over Taiwan with a modified 553

25

probability-matching technique. Mon. Wea. Rev., 141, 3908–3932. 554

——, ——, and A. Wang, 2011: The impacts of Taiwan topography on the predictability of 555

Typhoon Morakot’s record breaking rainfall: A high-resolution ensemble simulation. Wea. 556

Forecasting, 26, 613–633. 557

Grell, G. A., and S. R. Freitas, 2014: A scale and aerosol aware stochastic convective 558

parameterization for weather and air quality modeling. Atmos. Chem. Phys., 14, 5233–559

5250. 560

Hong, S.-Y., and J.-O. J. Lim, 2006: The WRF Single-Moment 6-Class Microphysics Scheme 561

(WSM6). J. Korean Meteor. Soc., 42, 129–151. 562

——, Y. Noh, and J. Dudhia, 2006: A new vertical diffusion package with an explicit 563

treatment of entrainment processes. Mon. Wea. Rev., 134, 2318–2341. 564

Huang, C.-H., and C.-C. Wu, 2018: The impact of idealized terrain on upstream tropical 565

cyclone track. J. Atmos. Sci., 75, 3887–3909. 566

Huang, C.‐Y., I.‐H. Wu, and L. Feng, 2016: A numerical investigation of the convective 567

systems in the vicinity of southern Taiwan associated with Typhoon Fanapi (2010): 568

Formation mechanism of double rainfall peaks. J. Geophys. Res. Atmos., 121, 12647–569

12676. 570

Huang, Y.-H., C.-C. Wu, and Y. Wang, 2011: The influence of island topography on typhoon 571

track deflection. Mon. Wea. Rev., 139, 1708-1727. 572

Jian, G.‐J., and C.‐C. Wu, 2008：A numerical study of the track deflection of super‐Typhoon 573

Haitang (2005) prior to its landfall in Taiwan. Mon.Wea. Rev., 136, 598–615. 574

Lee, C.-S., L.-R. Huang, H.-S. Shen, and S.-T. Wang, 2006: A climatology model for 575

forecasting typhoon rainfall in Taiwan. Nat. Hazards, 37, 87–105. 576

Lin, Y.-L., J. Han, D. W. Hamilton, and C.-Y. Huang, 1999: Orographic influence on a drifting 577

cyclone. J. Atmos. Sci., 56, 534–562. 578

26

——, S. Chiao, T. A. Wang, M. L. Kaplan, and R. P. Weglarz, 2001: Some common 579

ingredients for heavy orographic rainfall. Wea. Forecasting, 16, 633-660. 580

——, S.-Y. Chen, C. M. Hill, and C.-Y. Huang, 2005: Control parameters for the influence of 581

a mesoscale mountain range on cyclone track continuity and deflection. J. Atmos. Sci., 582

62, 1849–1866. 583

Liou, Y., T. Chen Wang, and P. Huang, 2016: The inland eyewall reintensification of Typhoon 584

Fanapi (2010) documented from an observational perspective using multiple-doppler 585

radar and surface measurements. Mon. Wea. Rev., 144, 241–261. 586

Meng, Z., and F. Zhang, 2008a: Test of an ensemble Kalman filter for mesoscale and 587

regional-scale data assimilation. Part III: Comparison with 3DVAR in a real-data case 588

study. Mon. Wea. Rev., 136, 522–540. 589

——, and ——, 2008b: Test of an ensemble Kalman filter for mesoscale and regional-scale 590

data assimilation. Part IV: Comparison with 3DVAR in a month-long experiment. Mon. 591

Wea. Rev., 136, 3671–3682. 592

Mlawer, E. J., S. J. Taubman, P. D. Brown, M. J. Iacono, and S. A. Clough, 1997: Radiative 593

transfer for inhomogeneous atmospheres: RRTM, a validated correlated-k model for the 594

longwave. J. Geophys. Res., 102, 16663–16682. 595

Tao, W. K., and Coauthors, 2011: High-resolution numerical simulation of the extreme 596

rainfall associated with Typhoon Morakot. Part I: Comparing the impact of microphysics 597

and PBL parameterizations with observations. Terr. Atmos. Oceanic Sci., 22, 673–696. 598

Wang, C.-C., Y.-H. Chen, H.-C. Kuo, and S.-Y. Huang, 2013: Sensitivity of typhoon track to 599

asymmetric latent heating/rainfall induced by Taiwan topography: A numerical study of 600

Typhoon Fanapi (2010). J. Geophys. Res. Atmos., 118, 3292–3308. 601

Wang, S.-T., 1980: Prediction of the behavior and intensity of typhoons in Taiwan and its 602

vicinity (in Chinese). Chinese National Science Council Research Rep. 108, 100 pp. 603

27

——, 1989: Track, intensity, structure, wind and precipitation characteristics of typhoons 604

affecting Taiwan (in Chinese). Disaster Mitigation Research Rep. 80–73, NSC 80-04140- 605

P052–02B, National Science Council of Taiwan, Taipei, Taiwan, 285 pp. 606

Weissmann, M., and Coauthors, 2011: The influence of dropsondes on typhoon track and 607

midlatitude forecasts. Mon. Wea. Rev., 139, 908–920. 608

Wu, C.-C., 2001: Numerical simulation of Typhoon Gladys (1994) and its interaction with 609

Taiwan terrain using the GFDL hurricane model. Mon. Wea. Rev., 129, 1533–1549. 610

——, and Y.-H. Kuo, 1999: Typhoons affecting Taiwan: Current understanding and future 611

challenges. Bull. Amer. Meteor. Soc., 80, 67–80. 612

——, T.‐H. Yen, Y.‐H. Kuo, and W. Wang, 2002：Rainfall simulation associated with Typhoon 613

Herb (1996) near Taiwan. Part I: The topographic effect. Wea. Forecasting, 17, 1001–614

1015. 615

——, and Coauthors, 2005: Dropwindsonde Observations for Typhoon Surveillance near the 616

Taiwan Region (DOTSTAR): An overview. Bull. Amer. Meteor. Soc., 86, 787–790. 617

——, K.-H. Chou, P.-H. Lin, S. D. Aberson, M. S. Peng, and T. Nakazawa, 2007: The impact 618

of dropwindsonde data on typhoon track forecasts in DOTSTAR. Wea. Forecasting, 22, 619

1157–1176. 620

——, H.-J. Cheng, Y. Wang, and K.-H. Chou, 2009: A numerical investigation of the eyewall 621

evolution in a landfalling typhoon. Mon. Wea. Rev., 137, 21-40. 622

——, G.-Y. Lien, J.-H. Chen, and F. Zhang, 2010: Assimilation of tropical cyclone track and 623

structure based on the ensemble Kalman filter (EnKF). J. Atmos. Sci., 67, 3806–3822. 624

——, Y.-H. Huang, and G.-Y. Lien, 2012: Concentric eyewall formation in Typhoon Sinlaku 625

(2008) – Part I: Assimilation of T-PARC data based on the Ensemble Kalman Filter 626

(EnKF). Mon. Wea. Rev., 140, 506-527. 627

——, S.-G. Chen, S.-C. Lin, T.-H. Yen, and T.-C. Chen, 2013: Uncertainty and predictability 628

28

of tropical cyclone rainfall based on ensemble simulations of Typhoon Sinlaku (2008). 629

Mon. Wea. Rev., 141, 3517–3538. 630

——, T.-H. Li, and Y.-H. Huang, 2015: Influence of mesoscale topography on tropical 631

cyclone tracks: Further examination of the channeling effect. J. Atmos. Sci., 72, 3032–632

3050. 633

Yang, M.-J., Y.-C. Wu, and Y.-C. Liou, 2018: The study of inland eyewall reformation of 634

Typhoon Fanapi (2010) using numerical experiments and vorticity budget analysis. J. 635

Geophys. Res. Atmos., 123, 9604–9623. 636

Yeh, T.-C., and R. L. Elsberry, 1993a: Interaction of typhoons with the Taiwan orography. Part 637

I: Upstream track deflections. Mon. Wea. Rev., 121, 3193–3212. 638

——, and R. L. Elsberry, 1993b: Interaction of typhoons with the Taiwan orography. Part II: 639

Continuous and discontinuous tracks across the island. Mon. Wea. Rev., 121, 3213–3233. 640

Yen, T.-H., C.-C. Wu, and G.-Y. Lien, 2011: Rainfall simulations of Typhoon Morakot with 641

controlled translation speed based on EnKF data assimilation. Terr. Atmos. Oceanic Sci., 642

22, 647– 660. 643

Yu, C.-K., and L.-W. Cheng, 2008: Radar observations of intense orographic precipitation 644

associated with Typhoon Xangsane (2000). Mon. Wea. Rev., 136, 497–521. 645

——, and L.-W. Cheng, 2013: Distribution and mechanisms of orographic precipitation 646

associated with typhoon Morakot (2009). J. Atmos. Sci., 70, 2894–2915. 647

——, and C.-L. Tsai, 2017: Structural changes of an outer tropical cyclone rain band 648

encountering the topography of northern Taiwan. Quart. J. Roy. Meteor. Soc., 143, 1107–649

1122. 650

Zhang, F., Z. Meng, and A. Aksoy, 2006: Tests of an ensemble Kalman filter for mesoscale 651

and regional-scale data assimilation. Part I: Perfect model experiments. Mon. Wea. Rev., 652

134, 722–736. 653

29

——, Y. Weng, Y.-H. Kuo, J. S. Whitaker, and B. Xie, 2010: Predicting Typhoon Morakot’s 654

catastrophic rainfall with a convection permitting mesoscale ensemble system. Wea. 655

Forecasting, 25, 1816–1825. 656

Zhu, T., and D.-L. Zhang, 2006: Numerical simulation of Hurricane Bonnie (1998). Part II: 657

Sensitivity to varying cloud microphysical processes. J. Atmos. Sci., 63, 109–126. 658

30

List of Figures 659

1 (a) AMSR-E brightness temperatures (K) at 1736 UTC 18 Sep 2010 overlaid on 660

MTSAT-2 IR imagery at 1714 UTC 18 Sep 2010 (from the Naval Research Laboratory). 661

(b) Surface analysis map at 1800 UTC 18 Sep 2010 obtained from the Central Weather 662

Bureau (CWB) of Taiwan. The observed track of Fanapi issued by CWB is also shown 663

in (a). ………………………………………………………………….………… 34 664

2 Radar vertical maximum indicator reflectivity composites (dBZ; shaded) at (a) 0000 665

UTC and (b) 0600 UTC 19 Sep 2010. (c) Observed 2-day accumulated rainfall (mm; 666

shaded) from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 and terrain height (contour 667

intervals of 50 m and 500 m are shown in gray and of 1500 m, 2000 m and 3000 m are 668

shown in black). The numbers indicate the local maximum accumulated rainfall at the 669

locations with triangle signs. …………………………………………….……… 35 670

3 Geographical distribution of the data used for initialization, including radiosondes (red 671

circle), dropwindsonde data collected by DOTSTAR (blue square) and C-130 aircraft 672

(yellow square), surface station data (purple plus), cloud motion vector (green plus) and 673

aircraft reports (blue plus). For dropwindsonde observations, the plus, triangle, circle, 674

and multiple sign symbols indicate different flight missions. The track of Fanapi during 675

the assimilating period is plotted with red line. The outer domain (D1) and two moving 676

nested domains (D2 and D3) at 1200 UTC 15 Sep 2010 are shown by blue boxes; the 677

fourth domain (D4) in Fix2km-M experiment is shown by the black box. …..…. 36 678

4 Simulated tracks of Fanapi in experiments with the original (CTL), elevated (H130), 679

and reduced (H50) terrain height of Taiwan, and with Taiwan terrain being removed 680

(H00). Thin and thick lines present the ensembles and the ensemble means, respectively, 681

in each experiment, and the observed track from CWB is in black. Multiplication signs 682

indicate the initial TC center for the ensemble means. The small and large solid circles 683

31

are the center positions at 1200 and 0000 UTC, respectively. …………….……. 37 684

5 The simulations of ensemble MSLP (hPa) in CTL (blue) as well as their average value 685

of 28 ensemble members (dashed blue), H130 (purple), H50 (green), and H00 (orange). 686

Their ensemble means are shown by thick lines, and the observed MSLP from CWB is 687

shown by the black line. …………………………………………………………. 38 688

6 Simulated 2-day accumulated rainfall (mm; shaded) from 0000 UTC 18 Sep to 0000 689

UTC 20 Sep 2010 for ensemble means in (a) CTL-M, (b) H00-M, (c) HT50-M, (d) 690

H130-M, and (e) Fix2km-M. The numbers in each panel indicate the local maximum 691

accumulated rainfall at the locations with triangle signs. ………………..………. 39 692

7 The time evolution of ensemble mean track error (km, histograms) and standard 693

deviation of the storm center position (km, solid lines) in each experiment. ……. 40 694

8 The ensemble tracks of Fanapi in (a) CTL, (b) H00, (c) H50, and (d) H130. The TC 695

centers of each ensemble member (open circles) and ensemble mean (TC marks) at 696

0000 and 1200 UTC 19 Sep 2010 are plotted in each panel, respectively. Terrain 697

heights (m) of Taiwan used in each experiment are shaded by grey colors. ….…. 41 698

9 The time evolution of ensemble mean MSLP error (hPa, histograms) and standard 699

deviation of MSLP (hPa, solid lines) in each experiment. ………………………. 42 700

10 The 2-day accumulated rainfall (mm; shaded) of 28 ensemble members in southern 701

Taiwan from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 in CTL experiment. Their 702

sorting is based on the latitude of storm center (plotted by TC mark) at the time as TC 703

departs from the west coast of Taiwan. ………………………………………….. 43 704

11 Twenty 6-h backward air parcel trajectories beginning at 925 hPa are overlaid on the 705

simulated wind (m s-1; wind vector) and maximum radar reflectivity (dBZ; shaded) for 706

ensemble member of (a) CTL-M06 at 0700 UTC 19 Sep 2010 and (b) CTL-M24 at 707

1400 UTC 19 Sep 2010. The corresponding storm center at these times is also plotted 708

32

by TC mark. The blue shaded areas indicate the horizontal wind speed beyond 35 m s-1. 709

The track of two selected members and CTL-M are shown as black and blue line, 710

respectively. ……………………………………………………………………… 44 711

12 The simulated maximum radar reflectivity (>35 dBZ; shaded) of 28 ensemble members 712

as the storm center departs from the west coast of Taiwan and moves toward 120.1 E°. 713

Their sorting is based on the latitude of storm center (plotted by TC mark). The black 714

lines in each panel indicate the axis of the TC rainband which is defined as the central 715

point of the rainband’s width. …………………………………………………… 45 716

13 (a)The distribution of the axis of the TC rainband as the storm center moves to 120.1 E° 717

in NG (red) and SG (blue) and their corresponding (b) track and (c) MSLP of each 718

ensemble member. The CTL-M axis of the TC rainband is shown in black line and the 719

TC marks in (a) denote the corresponding storm centers at this time. The time in the 720

x-axis of (c) indicates the hours relative to the landfall time, and the red and blue thick 721

lines are the average MSLP of NG and SG members. …………………………… 46 722

14 Ensemble correlations between the latitude of storm center at 1000 UTC 19 Sep 2010 723

and storm center latitude (black), storm center longitude (red), MSLP (blue), and 724

maximum wind speed (green) over the 32-h period (0200 UTC 18 to 1000 UTC 19 Sep 725

2010). …………………………………………………………………………….. 47 726

15 The composite of maximum radar reflectivity (dBZ; shaded) and low level (1000 to 700 727

hPa) average wind speed (m s-1; wind vector > 25 m s-1) in (a) NG and (b) SG at the 728

time when the storm center moves to 120.1 E°. Contours of terrain height are plotted by 729

solid line with interval of every 300 m. ………………………………………….. 48 730

16 Composite of 2-day accumulated rainfall (mm; shaded) from 0000 UTC 18 Sep to 0000 731

UTC 20 Sep 2010 in (a) NG and (b) SG. The rainfall difference (mm; shaded) between 732

NG and SG are shown in (c). The numbers in (a) and (b) indicate the local maximum 733

33

accumulated rainfall at the locations with triangle signs. ………………………... 49 734

17 The spatial distribution of standard deviation of the 2-day accumulated rainfall (mm; 735

shaded) from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 in (a) CTL, (b) H00, (c) 736

H50 and (d) H130. Contours of standard deviation are plotted by solid lines with an 737

interval of 20 mm. ……………………………………………………………….. 50 738

18 The distribution of the axis of the TC rainband as the storm center moves to 120.1 E° in 739

CTL-M (blue), H130-M (purple), H50-M (green), and observed (black). The ensemble 740

members for each experiment are shown in the right subplots. TC marks denote the 741

corresponding storm centers of ensemble mean for each experiment and observation at 742

this time. ………………………………………………………………………….. 51 743

19 Latitude of the TC rainband versus latitude of the TC center for all ensemble members 744

as the storm center moves to 120.1 E° in CTL (blue), H130 (purple), and H50 (green). 745

Solid lines are least squares fits using all ensemble members from each 746

experiment. ………………………………………………………………………. 52 747

20 Twenty 6-h backward air parcel trajectories beginning at 925 hPa are overlaid on the 748

simulated wind (m s-1; wind vector) and water vapor mixing ratio (g kg-1; shaded) at 749

1000 UTC 19 Sep 2010 in (a) CTM-M, (b) H00-M, (c) H50-M, and (d) H130-M. The 750

red shaded areas indicate the horizontal wind speed beyond 35 m s-1. ………….. 53 751

21 The asymmetric flow vector (m s-1, arrow) and horizontal divergence (10-5 s-1, shaded) 752

at 850 hPa calculated from asymmetric flow at 1000 UTC 19 Sep 2010 in (a) CTM-M, 753

(b) H00-M, (c) H50-M, and (d) H130-M. TC marks denote the corresponding storm 754

centers for each experiment. ………………………………………………...…… 54 755

756

34

757

758

FIG. 1. (a) AMSR-E brightness temperatures (K) at 1736 UTC 18 Sep 2010 overlaid on 759

MTSAT-2 IR imagery at 1714 UTC 18 Sep 2010 (from the Naval Research Laboratory). (b) 760

Surface analysis map at 1800 UTC 18 Sep 2010 obtained from the Central Weather Bureau 761

(CWB) of Taiwan. The observed track of Fanapi issued by CWB is also shown in (a). 762

35

763

FIG. 2. Radar vertical maximum indicator reflectivity composites (dBZ; shaded) at (a) 0000 764

UTC and (b) 0600 UTC 19 Sep 2010. (c) Observed 2-day accumulated rainfall (mm; shaded) 765

from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 and terrain height (contour intervals of 50 766

m and 500 m are shown in gray and of 1500 m, 2000 m and 3000 m are shown in black). The 767

numbers indicate the local maximum accumulated rainfall at the locations with triangle signs. 768

36

769

FIG. 3. Geographical distribution of the data used for initialization, including radiosondes 770

(red circle), dropwindsonde data collected by DOTSTAR (blue square) and C-130 aircraft 771

(yellow square), surface station data (purple plus), cloud motion vector (green plus) and 772

aircraft reports (blue plus). For dropwindsonde observations, the plus, triangle, circle, and 773

multiple sign symbols indicate different flight missions. The track of Fanapi during the 774

assimilating period is plotted with red line. The outer domain (D1) and two moving nested 775

domains (D2 and D3) at 1200 UTC 15 Sep 2010 are shown by blue boxes; the fourth domain 776

(D4) in Fix2km-M experiment is shown by the black box. 777

37

778

FIG. 4. Simulated tracks of Fanapi in experiments with the original (CTL), elevated (H130), 779

and reduced (H50) terrain height of Taiwan, and with Taiwan terrain being removed (H00). 780

Thin and thick lines present the ensembles and the ensemble means, respectively, in each 781

experiment, and the observed track from CWB is in black. Multiplication signs indicate the 782

initial TC center for the ensemble means. The small and large solid circles are the center 783

positions at 1200 and 0000 UTC, respectively. 784

38

785

FIG. 5. The simulations of ensemble MSLP (hPa) in CTL (blue) as well as their average value 786

of 28 ensemble members (dashed blue), H130 (purple), H50 (green), and H00 (orange). Their 787

ensemble means are shown by thick lines, and the observed MSLP from CWB is shown by 788

the black line. 789

39

790

FIG. 6. Simulated 2-day accumulated rainfall (mm; shaded) from 0000 UTC 18 Sep to 0000 791

UTC 20 Sep 2010 for ensemble means in (a) CTL-M, (b) H00-M, (c) HT50-M, (d) H130-M, 792

and (e) Fix2km-M. The numbers in each panel indicate the local maximum accumulated 793

rainfall at the locations with triangle signs. 794

40

795

FIG. 7. The time evolution of ensemble mean track error (km, histograms) and standard 796

deviation of the storm center position (km, solid lines) in each experiment. 797

41

798

FIG. 8. The ensemble tracks of Fanapi in (a) CTL, (b) H00, (c) H50, and (d) H130. The TC 799

centers of each ensemble member (open circles) and ensemble mean (TC marks) at 0000 and 800

1200 UTC 19 Sep 2010 are plotted in each panel, respectively. Terrain heights (m) of Taiwan 801

used in each experiment are shaded by grey colors. 802

42

803 FIG. 9. The time evolution of ensemble mean MSLP error (hPa, histograms) and standard 804

deviation of MSLP (hPa, solid lines) in each experiment. 805

43

806

FIG. 10. The 2-day accumulated rainfall (mm; shaded) of 28 ensemble members in southern 807

Taiwan from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 in CTL experiment. Their sorting 808

is based on the latitude of storm center (plotted by TC mark) at the time as TC departs from 809

the west coast of Taiwan.810

44

811

FIG. 11. Twenty 6-h backward air parcel trajectories beginning at 925 hPa are overlaid on the 812

simulated wind (m s-1; wind vector) and maximum radar reflectivity (dBZ; shaded) for 813

ensemble member of (a) CTL-M06 at 0700 UTC 19 Sep 2010 and (b) CTL-M24 at 1400 UTC 814

19 Sep 2010. The corresponding storm center at these times is also plotted by TC mark. The 815

blue shaded areas indicate the horizontal wind speed beyond 35 m s-1. The track of two 816

selected members and CTL-M are shown as black and blue line, respectively. 817

45

818

FIG. 12. The simulated maximum radar reflectivity (>35 dBZ; shaded) of 28 ensemble 819

members as the storm center departs from the west coast of Taiwan and moves toward 120.1 820

E°. Their sorting is based on the latitude of storm center (plotted by TC mark). The black 821

lines in each panel indicate the axis of the TC rainband which is defined as the central point of 822

the rainband’s width. 823

46

824

FIG. 13. (a)The distribution of the axis of the TC rainband as the storm center moves to 120.1 825

E° in NG (red) and SG (blue) and their corresponding (b) track and (c) MSLP of each 826

ensemble member. The CTL-M axis of the TC rainband is shown in black line and the TC 827

marks in (a) denote the corresponding storm centers at this time. The time in the x-axis of (c) 828

indicates the hours relative to the landfall time, and the red and blue thick lines are the 829

average MSLP of NG and SG members. 830

47

831

FIG. 14. Ensemble correlations between the latitude of storm center at 1000 UTC 19 Sep 832

2010 and storm center latitude (black), storm center longitude (red), MSLP (blue), and 833

maximum wind speed (green) over the 32-h period (0200 UTC 18 to 1000 UTC 19 Sep 2010).834

48

835

FIG. 15. The composite of maximum radar reflectivity (dBZ; shaded) and low level (1000 to 836

700 hPa) average wind speed (m s-1; wind vector > 25 m s-1) in (a) NG and (b) SG at the time 837

when the storm center moves to 120.1 E°. Contours of terrain height are plotted by solid line 838

with interval of every 300 m. 839

49

840

FIG. 16. Composite of 2-day accumulated rainfall (mm; shaded) from 0000 UTC 18 Sep to 841

0000 UTC 20 Sep 2010 in (a) NG and (b) SG. The rainfall difference (mm; shaded) between 842

and SG are shown in (c). The numbers in (a) and (b) indicate the local maximum accumulated 843

rainfall at the locations with triangle signs. 844

845

50

846

FIG. 17. The spatial distribution of standard deviation of the 2-day accumulated rainfall (mm; 847

shaded) from 0000 UTC 18 Sep to 0000 UTC 20 Sep 2010 in (a) CTL, (b) H00, (c) H50 and 848

(d) H130. Contours of standard deviation are plotted by solid lines with an interval of 20 mm. 849

51

850

FIG. 18. The distribution of the axis of the TC rainband as the storm center moves to 120.1 E° 851

in CTL-M (blue), H130-M (purple), H50-M (green), and observed (black). The ensemble 852

members for each experiment are shown in the right subplots. TC marks denote the 853

corresponding storm centers of ensemble mean for each experiment and observation at this 854

time. 855

52

856

FIG. 19. Latitude of the TC rainband versus latitude of the TC center for all ensemble 857

members as the storm center moves to 120.1 E° in CTL (blue), H130 (purple), and H50 858

(green). Solid lines are least squares fits using all ensemble members from each experiment. 859

53

860

FIG. 20. Twenty 6-h backward air parcel trajectories beginning at 925 hPa are overlaid on the 861

simulated wind (m s-1; wind vector) and water vapor mixing ratio (g kg-1; shaded) at 1000 862

UTC 19 Sep 2010 in (a) CTM-M, (b) H00-M, (c) H50-M, and (d) H130-M. The red shaded 863

areas indicate the horizontal wind speed beyond 35 m s-1. 864

54

865

FIG. 21. The asymmetric flow vector (m s-1, arrow) and horizontal divergence (10-5 s-1, 866

shaded) at 850 hPa calculated from asymmetric flow at 1000 UTC 19 Sep 2010 in (a) 867

CTM-M, (b) H00-M, (c) H50-M, and (d) H130-M. TC marks denote the corresponding storm 868

centers for each experiment. 869

870

55

List of Tables 871

1 Summary of numerical experiments. “-M” denotes the deterministic forecast initialized 872

from the ensemble mean. ………..……………………………………………… 56873

56

Table 1. Summary of numerical experiments. “-M” denotes the deterministic forecast 874

initialized from the ensemble mean. 875

Experiment Topography in Taiwan Horizontal grid spacing

CTL, CTL-M Full topography 54/18/6 km

H00, H00-M No topography, ocean surface Same as in CTL

H50, H50-M 50% of the CTL Same as in CTL

H130, H130-M 130% of the CTL Same as in CTL

Fix2km-M Same as in CTL 2 km in D4

876

877