Estimation of Tropical Cyclone's Intensity Using TRMM/TMI Brightness ...

Journal of the Meteorological Society of Japan, Vol. 85, No. 4, pp. 437--454, 2007

437

Estimation of Tropical Cyclone’s Intensity Using TRMM/TMI Brightness Temperature Data

Shunsuke HOSHINO and Tetsuo NAKAZAWA Typhoon Research Department, Meteorological Research Institute, Japan Meteorological Agency, Tsukuba, Japan (Manuscript received 31 March 2006, in final form 25 March 2007)

Abstract A new method for the estimation of tropical cyclone (TC) intensity utilizing 10, 19, 21, 37 and 85 GHz channel TRMM Microwave Imager (TMI) data from 1999 to 2003 is developed. As a first step, we investigated the relationship between the TRMM/TMI brightness temperature (TB) parameters, which are computed in concentric circles, or annuli of different radius in different TMI frequencies, and the TC maximum wind speed from the TC best track data, and/or observed by microwave scatterometers (QuikSCAT and SeaWinds). In contrast to the previous studies, we found that the parameters with lower frequency channels of 10 or 19 GHz give higher correlation. This would be because that TBs of lower frequencies, that have less sensitivity to rain than those of higher frequencies, reflect the speed of sea surface wind more directly in the TC case. The highest correlation coefficient obtained is 0.7, and the root mean square error (RMSE) of the regression between a parameter of highest correlation case is found to be 6 ms1 . We developed a TC intensity estimation method, based on the multiple regression equations using a few parameters. After choosing 3 parameters out of all possible combinations, we computed the regression coefficients and chose 10 regression equations, sorted by the lower RMSEs. Finally, we evaluated our estimation method using independent verification data during 2004. The RMSEs are found to be about 8 ms1 in the entire basin for the best track data, and about 6 ms1 for the best track data in the northwestern Pacific. Whereas, for the microwave scatterometer data in all basins RMSE is found to be about 7 ms1 . We also found that the temporal TC intensity change in our method shows good agreement with the TC best track data.

1.

Introduction

The tropical cyclone (TC) is the weather system that causes severe disaster to our society by bringing strong wind and heavy rainfall. Thus, it is important to develop an estimation technique of tropical cyclone intensity for prediction of the TC track and intensity, and disaster prevention. Most operational centers estiCorresponding author: Shunsuke Hoshino, Typhoon Research Department, Meteorological Research Institute, Japan Meteorological Agency, 1-1 Nagamine, Tsukuba, Ibaraki 305-0052, Japan. E-mail: [email protected] ( 2007, Meteorological Society of Japan

mate TC intensity by the ‘‘Dvorak technique’’ (Dvorak 1975). This technique is based on a ‘‘pattern recognition’’ of satellite visible and infrared imagery, and usually shows a good estimate. However, it is subject to analysis because the relationship between cloud patterns and physical parameters that characterize tropical cyclones is not obvious. The difficulties to distinguish eyewall clouds and upper-level dense clouds that extend from the eyewall using infrared image, especially during the nighttime, sometimes cause misclassification of the TC pattern, which make a misestimate of TC intensity. There are several studies to improve the Dvorak Technique. For example, Velden

438

Journal of the Meteorological Society of Japan

et al. (1998) developed an ‘‘Objective Dvorak Technique’’ using automated estimation, by identifying cloud pattern and the temperature difference between the eye and the surrounding eyewall. Recently, satellite-borne microwave sensors such as microwave radiometers (Special Sensor Microwave Imager (SSM/I) on Defense Meteorological Satellite Program (DMSP), Tropical Rainfall Measurement Mission (TRMM) Microwave Imager (TMI), Advanced Microwave Scanning Radiometer-E (AMSR-E) on Aqua), microwave scatterometers (Seawinds on QuikSCAT and Advanced Earth Observing Satellite (ADEOS)-II, hereafter we call simply as ‘QuikSCAT’ and ‘SeaWinds’ respectively), and microwave sounders (Advanced Microwave Sounding Unit (AMSU) on the National Oceanic and Atmospheric Administration (NOAA) polar satellite series) became operational. These sensors provide valuable information for TC intensity in terms of rainfall, water vapor, sea surface wind, sea surface temperature, vertical profile of air temperature and moisture, and so on. In addition, these data are very useful for TCs over the ocean where only few observations are available. There are several studies to estimate intensity of TCs using satellite-borne microwave sensors. Velden and Smith (1983), Velden (1989) and Velden et al. (1991) developed a method to estimate minimum sea level pressure (MSLP), and maximum wind speed by assessing the intensity of warm core with Microwave Sounding Unit (MSU) on NOAA satellites. Using microwave radiometer data, Cecil and Zipser (1999) investigated the relationship between polarization corrected temperature (PCT) of the 85 GHz channel, and present and future maximum wind speed and tendency of TC intensity. PCT is the parameter, proposed by Spencer et al. (1989), which represents the radiation eliminating the radiation from ocean using polarization diversity, and is calculated as: PCT ¼ 1:818TBv 0:818TBh where TBv is the brightness temperature (TB) for the vertically polarized channel, and TBh is that of the horizontally polarized channel at the 85 GHz of the SSM/I. The study by Cecil and Zipser (1999) revealed that mean PCT values are highly correlated

Vol. 85, No. 4

with maximum wind speed at 24 hours later (correlation coefficients ¼ 0:65). In addition, they showed that mean PCT values and ratio of the area below 250 K are highly correlated with simultaneous TC intensity (correlation coefficients are 0.62 and 0.54, respectively). They use PCT as a parameter characterizing the TC intensity. Although they do not use lower frequency channels, we think that lower frequency channels also can be useful, so it is worthy to investigate the relationship between the TC intensity and the parameters using low frequency channels, too. The main objective of this study is to develop a method for the TC intensity estimation using TRMM/TMI data, not only based on highfrequency data sets but also using lowfrequency channels. In this study we first compute the correlation between TBs of TRMM/TMI, and maximum wind speed of TCs, from the best track data in Section 3. We also used the ocean surface data from QuikSCAT and SeaWinds on ADEOS-II for the maximum wind speed information. Then we propose a method to estimate TC intensity in Section 4. 2.

Data

In the present study, we investigated global TCs, which developed to tropical storm stage (when maximum wind speed is over 35 kt), over the all basins since July 1999 (thus, the period that QuikSCAT data have became available), until the TC season in 2003 (2002–2003 TC season for the Southern Hemisphere). Over the northwestern Pacific, TCs during 2004 are also analyzed, and the results are presented in Section 4. We utilized TMI 1B11 TB data, which are available at GSFC DAAC site,1 and JAXA / EORC Tropical Cyclone Database site2. For the analysis of maximum wind speed in TC, we used two different data sources. One is from the TC best track data, and another is the microwave scatterometer data. TC best track data sets are available from the following operational centers; Japan Meteorological 1 2

http://lake.nascom.nasa.gov/data/dataset/TRMM/ index.html http://sharaku.eorc.jaxa.jp/TYP_DB/index_e .shtml

August 2007

S. HOSHINO and T. NAKAZAWA

Agency (JMA) for the northwestern Pacific Ocean, National Hurricane Center for the northeastern Pacific Ocean and the northern Atlantic Ocean, Central Pacific Hurricane Center for the north-central Pacific Ocean and Joint Typhoon Warning Center, for the northern Indian Ocean and the southern hemisphere. QuikSCAT and SeaWinds provide sea surface wind data at 10 m sea level, which are retrieved from microwave scatterometer on-board QuikBird (launched in 1999) and ADEOS-II (launched in 2002 and become unavailable in October 2003) satellites, respectively. These satellites are polar-orbit, so each satellite can observe a tropical cyclone up to twice a day (and observation interval is about 12 hours). Standard data products of QuikSCAT are available from NASA Jet Propulsion Laboratory (JPL), with 25 km resolution, and the accuracy of wind speed is 2 ms1 below 20 ms1 and is about 10% when wind speed is above 20 ms1 to 30 ms1 . In this study, QuikSCAT Version 3 products, and Seawinds Version 3 products, by Remote Sensing Systems3 are used. These Version 3 products also have 25 km resolution and the wind speed is up to 70 ms1 (Wentz et al. 2001). Using these data, the maximum wind speed within 3.0 latitude from the center of TC is selected and will be referred hereafter as ‘the maximum wind speed’. Although this size (3.0 latitude from the center of TC) is determined arbitrarily, Fujii (1998) shows that radii of maximum wind speed is smaller than 300 km for 51 typhoons that landed on Japan (i.e., mature or extratropical transition stage), so this size would be reasonable. The position (latitude and longitude) of TC center at satellite observation time is calculated by interpolating from position data from the best track data sets. 3.

Relationship between brightness temperature (TB) parameters and maximum wind speed

3.1 Methodology Following Cecil and Zipser (1999), we computed the TB parameters in concentric circles, and annuli from the TC center and compared it with the maximum wind speed using the TC best track data, and microwave scatterometer data. TB parameters are computed for each TB 3 http://www.ssmi.com/

439

of 10, 19, 21 and 37 GHz and PCT of 85 GHz (PCT85), over ocean. The details of TB parameter computations are described in the next sub section. The time collocation with TMI observation is within 3 hours for the TC best track data, and 6 hours for the microwave scatterometer data. This time collocation is based on a half of the interval of data (i.e., every 6 hours for the best track data, and almost every 12 hours for the microwave scatterometer data). For scatterometer data, this time collocation could appear insufficient to observe the rapid change of TC intensity, but we considered it reasonable to get the statistical relationship. Hereafter, for convenience, the following abbreviations for data are used. First, the source of maximum wind speed is introduced. Namely, ‘BT’ means the maximum wind speed from the TC best track data, and ‘SCAT’ from the microwave scatterometers. Then the basin name follows, such as: ‘ALL’ for all basins, ‘WP’ for the northwestern Pacific, ‘EP’ for the northeastern Pacific, ‘AT’ for the northern Atlantic and ‘SH’ for the southern hemisphere (TCs over northern Indian Ocean and north Central Pacific Ocean are not analyzed, due to the absence of significant number of cases). Thus, ‘BT_WP’ is the data set with a maximum wind speed by the TC best track over the northwestern Pacific, and ‘SCAT_ALL’ with a maximum wind speed by microwave scatterometers over all basins. Number of cases in this study is presented in Table 1. It is noted that maximum wind speed data from microwave scatterometers are used over all basins (SCAT_ALL). a.

TB parameters TB parameters are computed by selecting a frequency, coverage and type of computation. We used two choices for the selection of cover-

Table 1. The number of cases used in a comparison TCs

Observations

BT_ALL BT_WP BT_EP BT_AT BT_SH

110 41 19 22 28

911 339 148 223 201

SCAT_ALL

109

579

440


Fig. 1. Example of calculation regions of parameters. Image is PCT85 for Typhoon Jelawat (2000) at 17:51 UTC 5 August 2000. Circles are concentric circles whose radii are every 0.5 lat to 2.0 lat. Cross is the center of Jelawat.

age. One is a concentric circle, with different radius every 0.5 latitude to 2.0 latitude. Another is an annuli enclosed by circles every 0.5 latitude (e.g., annulus with 0.5 latitude inner radius and 1.0 latitude outer radius) (Fig. 1). The mean (MEAN), minimum (MIN), maximum (MAX) TB, and ratio of pixels over the threshold (AREA) TB, are computed in the computation coverage area. For AREA, a parameter is computed every 10 K threshold, shown in Table 2. Hereafter, each parameter is named as follows: Firstly, frequency (e.g., TB10H for horizontal polarized TB of 10 GHz), or PCT85 is presented. Secondly, a type of parameter (i.e.,

Table 2. The minimum and the maximum threshold temperatures in computing AREA for each frequency channels, and PCT85 [K]. Frequency 10 GHz 19 GHz 21 GHz 37 GHz PCT85

Minimum

Maximum

110 190 190 210 180

200 260 270 270 270

Vol. 85, No. 4

Fig. 2. Example of naming of calculation regions of parameters. Circles are concentric circles whose radii are every 0.5 lat to 2.0 lat. Cross is the center of the TC.

MEAN, MIN, MAX or AREA) is presented. For AREA, a threshold number follows. Finally, a type of computation coverage (C or A) follows. Computation region is written as ‘C’, followed by radius for circles, and ‘A’, followed by inner radius and outer radius for annuli. For example, ‘C10’ represents a circle whose radius is 1.0 latitude and ‘A1015’ represents annulus whose inner radius is 1.0 latitude and outer radius is 1.5 latitude (Fig. 2). Thus, ‘TB10H_MEAN_C10’ denotes mean TB of 10 GHz-H in circle of 1.0 latitude radius. 3.2 Results The highest 10 correlated parameters with the maximum wind speed for BT_ALL, BT_WP and SCAT_ALL are shown in Table 3. Correlation coefficients and root mean square errors (RMSEs) are also given. The highest correlated parameters are TB10H_MIN_C05 for BT_ALL, TB10H_AREA140_C10 for BT_WP, and TB10H_MEAN_C15 for SCAT_ALL. Scatter plots of these three cases are shown in Fig. 3. Horizontal axis in each figure is the highest correlated parameter (i.e., TB10H

August 2007


Table 3. The high correlated parameters in each data set [(a) BT_ALL, (b) BT_WP, (c) SCAT_ALL]. The correlation coefficients, and the RMSEs (when linear relationship is hypothesized, unit: ms1 ) are also shown. Parameter

Corr.Coef.

RMSE

0.744 0.722 0.717 0.715 0.711 0.707 0.704 0.700 0.699 0.698

8.66 8.97 9.04 9.06 9.11 9.17 9.20 9.25 9.26 9.28

Corr.Coef.

RMSE

0.704 0.702 0.700 0.699 0.696 0.694 0.689 0.689 0.687 0.687

6.08 6.09 6.10 6.12 6.14 6.16 6.20 6.20 6.21 6.22

Corr.Coef.

RMSE

0.707 0.700 0.695 0.688 0.686 0.686 0.683 0.682 0.681 0.681

7.78 7.85 7.91 7.98 8.00 8.01 8.04 8.05 8.05 8.05

TB10H_MIN_C05 TB10V_MIN_C05 TB10H_MEAN_C10 TB10H_MEAN_C05 TB10H_AREA140_C10 TB10V_MEAN_C05 TB10H_AREA130_C10 TB10V_MEAN_C10 TB10V_AREA200_C10 TB10H_AREA150_C10

441

The following noticeable characteristics are inferred from these results: 1. There is a tendency that parameters of 10 and 19 GHz channels give higher correlation with the TC maximum wind speed. 2. The most correlated parameter for each frequency is MEAN and AREA (and a few MINs), not MAX (MIN for PCT85 case). 3. The parameters in the surrounding region (1.0 or 1.5 latitude away from the TC center) tend to give higher correlation with the TC intensity. 4. There are differences in character among basins. We will consider each of these results in the following subsections.

(a) BT_ALL Parameter TB10H_AREA140_C10 TB10H_MIN_C10 TB10H_MIN_C05 TB19H_MEAN_C10 TB19H_MIN_C10 TB10H_MEAN_C10 TB10H_AREA130_C10 TB21V_AREA270_C10 TB10H_AREA150_C10 TB10V_AREA200_C10 (b) BT_WP Parameter TB10H_MEAN_C15 TB10H_MEAN_C20 TB10H_MEAN_A0515 TB10V_MEAN_C15 TB10H_AREA120_C15 TB10H_AREA110_A0515 TB10H_AREA120_A0515 TB10H_AREA110_C20 TB10H_AREA110_C15 TB10H_MEAN_A0510

(c) SCAT_ALL

_MIN_C05 for BT_ALL, TB10H_AREA140 _C10 for BT_WP, and TB10H_MEAN_C15 for SCAT_ALL, respectively) and vertical axis is the maximum wind speed from the TC best track data, or microwave scatterometer data. The highest correlation coefficients are about 0.7, and the RMSEs are 6–8 ms1 .

a.

The tendency that parameters of 10 and 19 GHz channels give higher correlation with the TC maximum wind speed The noticeable feature in Table 3 is that most of the parameters with higher correlations comes from the lowest frequency, 10 GHz. For both BT_ALL, and SCAT_ALL, all 20 parameters are from 10 GHz. For BT_WP, we found seven 10 GHz cases, two 19 GHz cases and one 21 GHz case. In most of the previous studies, such as Cecil and Zipser (1999), only PCT85 is treated as a noticeable parameter in estimating TC intensity, because it is widely recognized that PCT85 is related with the activity of convective clouds of TC. On the other hand, our study shows that parameters from lower frequencies are also highly correlated with TC intensity. PCT85 is considered to have good sensitivity to ice particle and to be related with activity of convection in tropical storms. Meanwhile, low frequency channels are sensitive to sea foam and water content in the atmosphere. Sea foam is produced as the result of wave breaking, which is related with energy transfer between sea and atmosphere, so coverage of sea foam is related with sea surface wind speed, and the difference of temperature between sea surface and atmosphere (Ross and Cardone 1974). According to Rose et al. (2002), the difference of emissivities between 36.5 GHz and 10.8 GHz to foam-covered sea are small (emissivities are about 0.85 for horizontal polarization and 0.94 for vertical polarization in both cases for TMI).

442


Vol. 85, No. 4

Fig. 4. The sensitivities of TB to physical values observed by microwave (NASA 1987). This shows that the sensitivities to the wind speed (via sea foam) are little between 10 GHz and 37 GHz, but the sensitivity of 37 GHz to the liquid clouds is larger than that of 10 GHz.

So the difference of the sensitivities, to the wind speed, would be small in the ‘clear-sky’ case. This is shown in Fig. 4 (NASA 1987). Meanwhile, as shown in Fig. 4, the sensitivity to liquid cloud water gets greater as frequency gets higher. This means that in the TC situtation (usually accompanied by rain and clouds), TB of lower frequency reflects the wind speed more directly, because it is less affected by rain or clouds than the TB of higher frequency. This could be because the trend that lower frequencies (especially 10 GHz) have higher correlation with the maximum wind speed. b.

Fig. 3. Scatter plots of the highest correlated parameter (horizontal) and maximum wind speed from the TC best track data or microwave scatterometer data (vertical) for each data set. The dashed line is the regression line. (a) BT_ALL (horizontal axis is TB10H _MIN_C05), (b) BT_WP (horizontal

The higher correlation with MEAN and AREA, not MAX A relationship between cloud water content and TB (Fig. 5) shows that TBs at 19 GHz and 37 GHz get higher, and PCT85 gets lower as cloud water content increases, indicating the same trend as an organization of the TC proceeds. This suggests that MAX (MIN for PCT85) should be highly correlated with maximum wind speed. However, other parameters show higher correlation. Cecil and Zipser (1999)

axis is TB10H_AREA140_C10, and (c) SCAT_ALL (horizontal axis is TB10H _MEAN_C15).

August 2007


Fig. 5. Relationship between brightness temperature and PCT for each SSM/I frequencies and oceanic cloud water content (Spencer et al. 1989).

also got the result that MEAN and AREA of PCT85 are more highly correlated than MIN, so our result is consistent. The possible reason is that MAX (MIN for PCT85) is likely to be high (low) even when there is a small active convection, or a few pixels with heavy rainfall in the region, but it does not represent the TC’s overall intensity. In addition, the effect of attenuation and scattering with cloud water content or rain drops affect brightness temperatures to saturate in the TC, with heavy rain case over the ocean (Figs. 5 and 6). MEAN is computed in the whole corresponding region, and MIN (MAX for PCT85) tells that all other points are larger (smaller) than this number, thus indicating the lowest activity. AREA indicates the ratio of an active area, hence it may be a more appropriate indicator to represent the TC activity. c.

The higher correlation with the parameters in the surrounding region In spite of high correlation with the maximum wind speed in C05 (especially for

443

Fig. 6. Brightness temperature—rain rate relationships for each SSM/I frequencies (Spencer et al. 1989).

BT_ALL), for BT_ALL and BT_WP, we find that the majority is in C10. But for SCAT_ALL, the dominant parameters area is in A0515 or C15. Here, we speculate that the overall TC intensity is not only related with the core region, but also the surrounding region, just outside of the eyewall. Of course, this may be related with the average size of eyes or the TCs, so we might have to consider the size of eyes or the TCs in the future work. d.

Different character among basins We noticed that the RMSEs over the northwestern Pacific (BT_WP) are lower than those over all basins (BT_ALL) (6.1 ms1 vs. 8.7 ms1 ) (See Tables 3a and b). Cecil and Zipser (1999), compared the correlations in the northwestern Pacific, northeastern Pacific and northern Atlantic, and showed the different TC characteristics in each basin. Their result is based on the best track data, hence it is possible that the result is affected by the analysis in each centers, but it is also possible that charac-

444


Vol. 85, No. 4

Fig. 7. Scatter plots of TB10H_MEAN_C10 (horizontal) and maximum wind speed by best track (vertical) for each basin. The dashed line is the regression line. (a) WP, (b) AT, (c) EP, and (d) SH.

teristics of the TC are different in each basin. It should also be noted that, the maximum wind speed in the TC best track for the northwestern Pacific, issued by the JMA, is a 10-minute averaged one, but the other centers are using 1 minute averaged one. The JMA uses an independent table from Dvorak one, based on the JMA analysis (Koba et al. 1990), to convert a CI number in the Dvorak technique to the mean sea level pressure (MSLP) and maximum wind speed. Dvorak (1984) uses different tables for the northwestern Pacific and the northern Atlantic for the MSLP. Velden et al. (1991) compared the relationships between warm core intensity, MSLP and wind speed of TCs in the northwestern Pacific and the northern Atlantic basins, and implies the difference of characteristics between basins, but also pointed out the

possibility that this difference is caused by the difference of distributions of intensities. Scatter plots of TB10H_MEAN_C10, and maximum wind speed by the TC best track data for each basin (Fig. 7), may imply the differences in the TC characteristics in each basin. Lonfat et al. (2004) investigated the rain distribution of the TCs and showed different patterns for each basin. The differences in patterns may be partly responsible for the differences in rela-

Table 4. The number of cases used as verification data

BT SCAT

TCs

Observations

15 13

157 96

August 2007


Table 5. The parameters and coefficients Vmax ¼ aPi þ bPj þ cPk þ d for each data sets a

Pi

of

the

‘candidates

445 of

regression

equation’:

b

Pj

c

Pk

d

PCT85_MIN_A0510 PCT85_MAX_A1020 TB19V_MIN_A1020 PCT85_MEAN_C10 TB10H_MAX_C10 PCT85_MIN_A0510 PCT85_MAX_A1020 PCT85_MAX_A1020 TB19V_AREA250 _A1520 PCT85_MIN_A1020

44.29 283.18 5.19 116.84 49.78 52.37 168.90 266.50 62.51

c

Pk

d 176.95 190.89 213.38 239.44 159.75 25.57

1 2 3 4 5 6 7 8 9

0.39 0.37 0.43 0.51 0.38 0.42 0.40 0.43 0.40

TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05 TB10H_MIN_C05

0.10 0.066 0.57 0.10 0.27 0.30 0.23 0.032 0.49

TB10H_MAX_C10 TB10H_MAX_C10 TB10H_MIN_A1020 TB10H_MAX_C10 TB10H_MIN_A1020 TB10H_MIN_A1020 TB10H_MIN_A1020 PCT85_MIN_A0510 TB10H_MIN_A1020

0.060 1.1 0.36 0.26 0.051 0.030 0.71 1.0 0.087

10

0.41

TB10H_MIN_C05

0.30

TB10H_MIN_A1020

0.031

51.50

(a) BT_ALL a

Pi

b

1 2 3 4 5 6

0.14 0.21 0.21 0.28 0.13 0.23

TB10H_AREA140_C10 TB19H_MIN_A0510 TB37H_MEAN_A0515 TB19V_MEAN_A0515 TB10H_AREA140_C10 TB10H_AREA140_C10

0.21 0.25 0.28 0.29 0.21 0.19

TB19H_MIN_C05 TB19H_MIN_C05 TB19H_MIN_C05 TB19H_MIN_C05 TB19H_MIN_C05 TB19H_MIN_C05

0.56 0.47 0.50 0.51 0.53 0.14

7 8 9 10

0.20 0.14 0.20 0.24

TB10H_AREA140_C10 TB10H_AREA140_C10 TB10H_AREA140_C10 TB19H_MIN_C05

0.14 0.15 0.14 0.22

TB19H_MIN_A1020 TB19H_MIN_C05 TB19H_MIN_A1020 TB19H_MIN_A0510

0.36 0.19 0.36 0.53

PCT85_MAX_C05 PCT85_MAX_C05 PCT85_MAX_C05 PCT85_MAX_C05 TB37V_MAX_C15 PCT85_AREA260 _C10 PCT85_MAX_C05 TB19H_MIN_A1020 PCT85_MAX_C05 TB37V_MAX_C15

Pj

c

Pk

d

TB19V_AREA250 _C15 TB19V_AREA250 _C15 TB37V_MIN_A0515 TB21V_MEAN _A1020 TB19H_AREA250 _A1020 TB19H_MEAN_C05 TB10H_MAX_C15 TB19H_MIN_A1020 TB10V_MIN_A0515 PCT85_MIN_A1020

239.75

Pj

104.55 37.49 104.55 201.90

(b) BT_WP a

Pi

b

1

0.56

PCT85_MEAN_C15

1.0

TB10H_MEAN_C15

0.35

2

0.51

1.1

TB10H_MEAN_C15

0.48

3 4

0.72 0.67

PCT85_AREA230 _A0515 PCT85_MEAN_C15 PCT85_MEAN_C15

0.78 0.78

TB10H_MEAN_C15 TB10H_MEAN_C15

0.26 0.35

5

0.72

PCT85_MEAN_C15

0.78

TB10H_MEAN_C15

0.19

6 7 8 9 10

0.73 0.67 0.67 0.67 0.53

PCT85_MEAN_C15 PCT85_MEAN_C15 PCT85_MEAN_C15 PCT85_MEAN_C15 PCT85_MEAN_C15

0.63 0.60 0.75 0.78 0.62

TB10H_MEAN_C15 TB10H_MEAN_C15 TB10H_MEAN_C15 TB10H_MEAN_C15 TB10H_MEAN_C15

0.087 0.084 0.17 0.40 0.033

137.71 201.15 158.21 259.63 261.28 238.43 215.58 175.24 193.17

(c) SCAT_ALL

tionships between the parameters and the maximum wind speed in each basin. These observational results suggest that the difference between basins comes from the difference of environmental condition, such as the sea surface

temperature, the global circulation, the effect of landmass or islands, and so on. This suggest that the choice of adequate parameter sets for each basin contributes for accurate TC intensity estimation.

446


4.

Vol. 85, No. 4

Estimation of maximum wind speed using multiple regression candidates

4.1 Methodology As described in the pervious section, it is shown that we can estimate maximum wind speed using a single parameter, with a certain accuracy (with 6–8 ms1 RMSEs). We may estimate TC intensity more precisely with multi parameters using available TB data. To examine the performance of our method, we introduce the following procedure to develop a reliable method of estimation. First, multiple estimating equations with a few parameters are prepared. Then multiple ‘candidates of maximum wind speed’ are computed from these equations for each TMI observation. The number of parameters calculated in Section 3 is too big to check all combination of parameters, and it is possible that there are little meaningful combinations when the colinearity between each parameter is large. So, at first, we made the surrogate set of parameters containing the 30 (this number is decided arbitrary) parameters that can explain the total parameter set, with eliminating the combinations of parameters which have high correlation. These 30 parameters (cf., P1 ; P2 ; . . . ; P30 ) are selected from all parameters computed in Section 3, with the genetic variable selection algorithm using the package ‘‘subselect’’ of R statistical software (R Development Core Team 2005; Cerdeira et al. 2005). Most highly correlated parameters in each data set (i.e., TB10H_MIN _C05 for BT_ALL, TB10H_AREA140_C10 for BT_WP and TB10H_MEAN_C15 for SCAT _ALL) are included in the selected 30 parameters. In the next step, three parameters (Pi , Pj and Pk ) are selected from these 30 parameters to compute a regression equation, Vmax ¼ aPi þ bPj þ cPk þ d:

Fig. 8. Scatter plots of estimated maximum wind speed by the TMI using the ‘regression candidates’ for each data set (horizontal) and the maximum wind speed by best track or the microwave scatterometer (vertical). The solid line is the line for perfect coincidence in both estimated maximum wind speed

We then compute the regression coefficients, a, b, c and d, by using available maximum wind speed data (i.e., the TCs from July 1999 to Deand those by the TC best track / scatterometer data. The dashed lines show the difference in the maximum wind speed, which are within 5 ms1 . (a) BT_ALL, (b) BT_WP, and (c) SCAT _ALL.

August 2007


447

cember 2003), and repeat this process for all combination of three parameters. All regression equations are sorted based on the RMSEs, and the highest 10 equations are selected as ‘candidates of regression equation’.

Fig. 9. Scatter plots of estimated maximum wind speed by the TMI using a ‘regression candidates’ for each data set (horizontal), and the maximum wind speed by best track or microwave scatterometer (vertical) for the TCs over the northwestern Pacific Ocean in 2004. The solid line and dashed lines are the

4.2 Verification By using these 10 ‘candidates of regression equation’ for each TMI observation, respective ten ‘candidates of estimated maximum wind speed’ are computed from the regression coefficients. The mean value of these candidates is adopted as the final estimated maximum wind speed by TMI. The data over the northwestern Pacific in 2004 are used to verify the performance of our estimation method. Number of verification data is shown in Table 4. The selected parameters, and their coefficients of ‘candidates of regression equation’ for each data set are shown in Table 5. The results of cross validation for each data set is shown in Fig. 8 [(a) BT_ALL, (b) BT_WP and (c) SCAT_ALL]. In these figures, horizontal axis is the maximum wind speed, estimated by TMI, and vertical axis is the one for the TC best track data or microwave scatterometer data. The solid line is the line for perfect coincidence in both estimated maximum wind speed, and those by the TC best track data. Dashed line shows the difference within 5 ms1 . The RMSEs by cross-validation are 8.14 ms1 for BT_ALL, 5.01 ms1 for BT_WP and 6.71 ms1 for SCAT_ALL. Figure 9a shows the verification of the final estimated maximum wind speed by TMI for BT_ALL, using the TC data over the northwestern Pacific Ocean in 2004. The horizontal axis is the maximum wind speed estimated by TMI, and the vertical axis is by the TC best track data. Figure 9b is the same as Fig. 9a, but for BT_WP, and Fig. 9c is for SCAT_ALL (here, vertical axis is the maximum wind speed by the scatterometer data). The RMSEs are 8.13 ms1 for BT_ALL, 6.34 ms1 for BT_WP, and 6.86 ms1 for SCAT_ALL, respectively. The time series plots of estimated maximum wind speed and observed maximum wind speed for Super Typhoons Dianmu, Chaba and Songda

same as Fig. 8. (a) BT_ALL, (b) BT_WP, and (c) SCAT_ALL.

448


Fig. 10. Time series plots of maximum wind speed estimated by the TMI, best track and microwave scatterometer of (a) Super Typhoon Dianmu (2004), (b) Super Typhoon Chaba (2004), (c) Super Typhoon Songda (2004). The solid line is a maximum wind speed by the TC best track data, the dashed line with mark ‘X’ is for the microwave scatterometer data, the ‘Estimation 1’ (open square) denotes an estimated maxi-

Vol. 85, No. 4

(TCs over northwestern Pacific Ocean in 2004) are shown in Fig. 10 [(a) Dianmu, (b) Chaba and (c) Songda]. In these figures, the solid line is the maximum wind speed by the best track, the dashed line with X is by the microwave scatterometer, ‘Estimation 1’ (open square) indicates the estimated maximum wind speed by the TMI with regression candidates for BT_ALL, ‘Estimation 2’ (closed square) shows the estimated with regression candidates for BT_WP, and ‘Estimation 3’ (solid bullet) denotes the estimated with regression candidates for SCAT_ALL. These figures show that the variations of the maximum wind speed estimated by the TMI are sometimes found to be larger, but implies that tendency of intensity change corresponds reasonably well with the maximum wind speed by the TC best track data, or microwave scatterometer data. For SCAT_ALL, there seems to be a trend of underestimating maximum wind speed, and the errors become larger especially when estimated winds are higher than 40 ms1 (Fig. 9c). In cross-validation, the RMSE is found to be 6.21 ms1 and the corresponding bias (estimated wind speed minus best track one) is 0.26 ms1 when estimated values are below 40 ms1 . On the other hand, the RMSE is 7.72 ms1 and the bias is 0.59 ms1 , when the estimated winds are higher than 40 ms1 . This contrast in bias shows the clear evidence of underestimation, with a large error under the strong wind conditions. It may be due to the insufficient number of cases with the maximum wind speed over 40 ms1 in the data sets for making regressions (33% of SCAT_ALL data). Also, we would like to mention that in the weak TC stages, our method tends to overestimate the maximum wind speeds when the best track data show below 20 ms1 (Figs. 8 and 9). There would be two possible reasons for this: 1)

mum wind speed by the TMI with regression candidates for BT_ALL, the ‘Estimation 2’ (closed square) shows an estimate with regression candidates for BT_WP and the ‘Estimation 3’ (solid bullet) denotes estimate with a regression candidates for SCAT_ALL.

August 2007


Fig. 11. Histogram of maximum wind speed for BT_ALL data set.

The lack of data in weak (early developing or dissipating) stage. Figure 11 is the histogram of maximum wind speed of the BT_ALL data set, and this shows that few data in weak stage are available. 2) The difficulty in positioning of the center due to an unorganized cloud pattern. For example, Fig. 12 shows an image of PCT85 distribution of Typhoon Francisco at 12:27 20 September 2001 (the early stage). The maximum wind speed estimated by the TMI is 39.3 ms1 , but the maximum wind speed in the TC best track data was 18.0 ms1 , and 19.0 ms1 by QuikSCAT. This may be related with the accuracy in positioning of the center and the TB patterns (discussed in details in a later part), i.e., organization of clouds with the TC in the earlier or dissipating stage, implying the limitation of our estimation technique. For 23 cases in which the estimated maximum wind speed was greater than that by microwave scatterometer by more than 10 ms1 , we analyzed and examined the horizontal pattern and distribution of PCT85. With subjective classification, that will be mentioned in the next paragraph, we found that there are eight cases with strong asymmetry (e.g., shown in Fig. 13a), one case without organization of clouds (shown in Fig. 13b), and six cases with active rain bands (e.g., shown in Fig. 13c). On

449

Fig. 12. Distribution of PCT85 of Typhoon Francisco (2001) at 12:27 UTC 20 September 2001, during the early stage of development. The cross mark is the center of Francisco. Estimated maximum wind speed is 39.3 ms1 , but is 18.0 ms1 at 12:00 by best track, and 19.0 ms1 at 08:03 by QuikSCAT.

the other hand, pattern recognitions are also done for 24 cases having estimated maximum wind speed is over 10 ms1 weaker than maximum wind speed by microwave scatterometer, (except for the cases that estimated values are over 40 ms1 ). There are 10 cases with strong asymmetry (e.g., shown in Fig. 13d), 11 cases with active rain bands (e.g., shown in Fig. 13e), four cases with a small size (major axis is smaller than 2.0 latitude), and seven cases without active convectional clouds nor organized eyewall were observed. In the last seven cases, the patterns in the extratropical transition stage are included (e.g., shown in Fig. 13f ). In this study, we subjectively classified the ‘asymmetry’ pattern as characterized by the low PCT (i.e., active convention), organized in a quadrant or semicircle region, but without circumscribing the center. On the other hand, we classified the ‘rain band’ pattern with a spiral nature of active convection, that extends widely, around the center. Thus there is a slightly different structure between the ‘asymmetry’ pattern, and the ‘rain band’ pattern, but sometimes the classification is difficult, so the ambiguities with the classification remain.

450


Vol. 85, No. 4

(a) Overestimating—Asymmetry: Fung-Wong (2002) at 04:32 UTC 21 July 2002: Vest ¼ 27:7 ms1 , VQS ¼ 17:6 ms1

(b) Overestimating—Unorganized: Jose (1999) at 06:06 UTC 18 October 1999: Vest ¼ 27:4 ms1 , VQS ¼ 17:4 ms1

(c) Overestimating—Rain band: Wutip (2001) at 09:58 UTC 27 August 2001: Vest ¼ 36:5 ms1 , VQS ¼ 25:7 ms1

(d) Underestimating—Asymmetry: Olga (2001) at 21:48 UTC 1 December 2001: Vest ¼ 21:2 ms1 , VQS ¼ 32:2 ms1

(e) Underestimating—Rain band: Dujuan (2003) at 14:48 UTC 30 August 2003: Vest ¼ 36:0 ms1 , VQS ¼ 46:5 ms1

(f ) Underestimating—Extratropical Transition: Bavi (2002) at 17:09 UTC 11 October 2002: Vest ¼ 31:8 ms1 , VQS ¼ 44:4 ms1

Fig. 13. Characteristic patterns of PCT85 in overestimating or underestimating case. The pattern name, the TC’s name, the observation time of TMI, estimated maximum wind speed ðVest Þ, and maximum wind speed by scatterometer ðVQS Þ are shown in each caption.

August 2007


451

Fig. 14. Scatter plots of maximum wind speed estimated by the TMI with regression candidates for each basin (horizontal), and maximum wind speed by best track (vertical). (a) WP, (b) AT, (c) EP, and (d) SH. Lines are the same as Fig. 8.

Similar investigation as described above revealed that all of seven cases (included three cases with concentric eyes) with overestimating are the cases of small TCs, whose length of major axis is smaller than 2.0 latitude. On the other hand, six of 11 cases with underestimating are the cases of strong asymmetric type, and 4 of the 11 cases have active rain bands. These results are based on a subjective classification without having much quantitative manner. But it suggests the possibility of improvement by considering the pattern of the TC structure. Another noticeable factor is the time collocation, which is set to 3 hours for the TC best track data, and 6 hours for microwave scatterometer data in this study. If a TC develops rapidly, the time collocation may affect the regression equations.

In this study, the positions of the TCs are decided by interpolation of the TC best track data, which is the smoothed TC track data. So the interpolated position may not be accurate when the TC moves unsmooothly, or movement speed changes rapidly. This disagreement of the position may affect the parameter value. In the operational use, it is desired to use the TMI image, or the latest visible (or IR) imagery for the TC center information. The maximum wind speed, estimated with regression candidates for BT_ALL, and for BT_WP, sometimes does not agree well (e.g., as is shown in Fig. 10). In such a case, there is a tendency that the estimated maximum wind speed by BT_ALL is stronger than that by BT_WP. We examined the change of the RMSEs, by constructing the individual regres-

452

Journal of the Meteorological Society of Japan Table 6. The RMSEs of estimations by specific regression candidates for each basin [ms1 ]

Data sets

Table 7. The candidates of the maximum wind speed in the case of Olga (2001) at 21:48 UTC 1 December 2001 (Fig. 13d). Each number in the first column corresponds to the regression number in Table 5c. [ms1 ]

RMSE

BT_WP BT_AT BT_EP BT_SH

5.01 6.49 7.52 9.01

BT_ALL (all basins) BT_ALL (each basin)

8.14 6.83

Fig. 15. The scatter plots of maximum wind speed, estimated by the TMI with regression candidates for each basin (horizontal), and a maximum wind speed by best track (vertical). Lines are the same as Fig. 8.

sion candidates in each basin. Scatter plots of cross-validation of BT_AT, BT_EP and BT_SH are shown in Fig. 14, and the RMSEs are shown in Table 6. We can see the improvement of the RMSEs not only for WP, but also for AT. A scatter plot of the maximum wind speed, estimated with the ‘basin-dependent’ regression candidates, and those in the TC best track data is shown in Fig. 15. As seen from Table 6, the errors become smaller for BT_ALL (i.e., basin-independent). The RMSE reduced from 8.14 ms1 to 6.83 ms1 . Presently, the validations with verification data sets have not been done, but it is expected

Vol. 85, No. 4

Candidate 1 2 3 4 5 6 7 8 9 10

21.6 20.3 37.7 22.5 19.3 16.8 22.5 21.8 20.5 18.7

that the accuracy will be higher than the case, using BT_ALL regression candidates. It is possible that there is a candidate with smaller error than our ‘estimated maximum wind speed’ from these 10 candidates. For example, each value of wind speed candidates in Fig. 13d case are shown in Table 7. The estimated maximum wind speed (thus, mean value of candidates) is 21.2 ms1 , which is about 10 ms1 weaker than the maximum wind speed by microwave scatterometer (32.2 ms1 ), but candidate 3 is 37.7 ms1 , whose error is about 5 ms1 , which is smaller than the difference between the maxium wind speed by microwave scatterometer and ‘estimated maximum wind speed’. In addition, it is characterized in that the value of candidates spread over 20 ms1 in this case. So, it may be helpful to use the median, and spread of ten estimated maximum wind speeds, as an indicator for the robustness and reliability of the estimation for operational use. Figure 16a is a scatter plot whose horizontal axis is the percentage of the standard deviation of 10 ‘candidates of the maximum wind speed’ divided by the maximum wind speed in the best track, and the vertical axis is the error (‘estimated maximum wind speed’ minus the maximum wind speed in the best track) for BT_WP data sets. The symbols show the intensity classification, i.e., dots, plus, ‘x’ marks, diamonds mean that the maximum wind speed in the best track is under 20 ms1 , 20–30 ms1 , 30–

August 2007


453

40 ms1 and over 40 ms1 , respectively. Figure 16b is similar to 16a, using the maximum wind speed observed by scatterometers for SCAT _ALL data sets, instead of by the best track. These figures show that error tends to be larger as the ratio of standard deviation to the maximum wind speed get larger, especially when the maximum wind speed is weak. 5.

Fig. 16. The scatter plots of the standard deviation of the ‘candidates of the maximum wind speed’, divided by the observed maximum wind speed (horizontal), and the error of the ‘estimated maximum wind speed’ (thus, the ‘estimated maximum wind speed’, minus the observed wind speed) (vertical). Symbols show the intensity classification, i.e., dots, plus, ‘x’ marks, and diamonds mean that the maximum wind speed in the best track is under 20 ms1 , 20–30 ms1 , 30–40 ms1 , and over 40 ms1 , respectively. (a) BT_WP and, (b) SCAT_ALL.

Summary

We developed a method to estimate tropical cyclone (TC) intensity using TRMM Microwave Imager (TMI) brightness temperature (TB) data at 10, 19, 21, 37 and 85 GHz channels during 1999 to 2003. First, we investigated the relationship between TRMM/TMI TB and maximum wind speed of the TC. We identified the TB parameters of higher correlations with the maximum wind speed in the TC best track data or by microwave scatterometers (QuikSCAT and SeaWinds), by calculating in concentric circles or annuli of different TMI frequencies. We found that the parameters with lower frequency channels of 10 or 19 GHz give higher correlation with TB MEAN, MIN and ratio of AREA for threshold TB within 1.0 latitude circle from the TC center. The highest correlation coefficient is 0.7 and the root mean square error (RMSE) of the regression with highest correlation case is 6 ms1 . The proposed TC intensity estimation method is based on the multiple regression equations using selected parameters. We evaluated our estimation method using independent verification data during 2004. The RMSEs are about 8 ms1 for the best track data in all basins, about 6 ms1 for the best track data in the northwestern Pacific, and about 7 ms1 for the microwave scatterometer data in all basins. The error in the microwave scatterometer data grows when an estimated maximum wind speed becomes higher than 40 ms1 . We found that the temporal TC intensity change in our method shows good agreement with the TC best track data. For operational use, the proposed methodology may need further improvement. We found that big errors are related with a strong asymmetry, spiral band pattern and little inner core, and so on. So the parameterization or classifica-

454


tion of such characteristics may contribute to get more accurate estimation. The current method has a deficiency to estimate the TC intensity in the asymmetric cloud pattern or the less organization of active convective cloud situation, such as the earlier formation, dissipating or extratropical transition stage. However, there is a limitation for the TMI observation. A TMI can only observe within 40 degree N/S latitude so that it cannot observe most of the extratropical transition stage in higher latitude. To overcome this issue, the method to estimate the TC intensity with other satellite-borne microwave sensors like AMSR-E and SSM/I is required. In addition, we are interested in estimating method of MSLP of the TC, because MSLP is also widely used as an indicator of the TC intensity. Dvorak (1984) has pointed out that CI number in the Dvorak technique is more related with MSLP than maximum wind speed. So it is meaningful to investigate the relationship between microwave radiometer data with the method of MSLP. Warm core temperature information, derived from AMSU may be useful to estimate MSLP of the TC. Acknowledgements We thank Dr. A. Shibata and Dr. C. S. Velden for helpful comments to our research, and Dr. Rajendran for improving our English expression in writing this paper. References Cecil, D.J. and E.J. Zipser, 1999: Relationships between tropical cyclone intensity and satellitebased indicators of inner core convection: 85GHz ice-scattering signature and lightning. Mon. Wea. Rev., 127, 103–123. Cerdeira, J.O., P.D. Silva, J. Cadima, and M. Minhoto, 2005: subselect: Selecting variable subsets. R package version 0.9-1. URL http:// cran.r-project.org/doc /packages/subselect.pdf Dvorak, V.F., 1975: Tropical cyclone intensity analysis and forecasting from satellite imagery. Mon. Wea. Rev., 103, 420–430. Dvorak, V.F., 1984: Tropical cyclone intensity analysis using satellite data. NOAA Tech. Rep. NESDIS 11, NOAA. Fujii, T., 1998: Statistical analysis of the characteristics of severe typhoons hitting the Japanese main islands. Mon. Wea. Rev., 126, 1091–1097.

Vol. 85, No. 4

Koba, H., T. Hagiwara, S. Osano, and S. Akashi, 1990: Relationships between CI number from Dvorak’s technique and minimum sea level pressure or maximum wind speed of tropical cyclone (in Japanese). J. Meteor. Res., 42, 59– 67. Lonfat, M., F.D. Marks, Jr., and S.S. Chen, 2004: Precipitation distribution in tropical cylclones using the Tropical Rainfall Measuring Mission (TRMM) microwave imager: A global perspective. Mon. Wea. Rev., 132, 1645–1660. NASA, 1987: High-Resolution Multifrequency Microwave Radiometer, Earth observing system volume IIe, Instrument Panel Report. NASA, Washington, D.C., 59 pp. R Development Core Team, 2005: R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-900051-07-0. URL http://www .r-project.org/ Rose, L.A., W.E. Asher, S.C. Reising, P.W. Gaiser, S. Germain, K.M., D.J. Dowgiallo, K.A. Horgan, G. Farquharson, and E.J. Knapp, 2002: Radiometric measurements of th microwave emissivity of foam. IEEE Trans. Geosci. Remote Sens., 40, 2619–2625. Ross, D.B. and V. Cardone, 1974: Observations of oceanic whitecaps and their relation to remote mesurements of surface wind speed. J. Geophys. Res., 79, 444–452. Spencer, R.W., H.M. Goodman, and R.E. Hood, 1989: Precipitation retrieval over land and ocean with the SSM/I: Identification and characteristics of the scattering signal. J. Atmos. Oceanic Technol., 6, 254–273. Velden, C.S., 1989: Observational analyses of north Atlantic tropical cyclones from NOAA polarOrbiting satellite microwave data. J. Appl. Meteor., 28, 59–70. Velden, C.S., B.M. Goodman, and R.T. Merrill, 1991: Western north Pacific tropical cyclone intensity estimation from NOAA polar-orbiting satellite microwave data. Mon. Wea. Rev., 119, 159– 168. Velden, C.S., T.L. Olander, and R.M. Zehr, 1998: Development of an objective scheme to estimate tropical cyclone intensity from digital geostationary satellite infrared imagery. Wea. Forecasting, 13, 172–186. Velden, C.S. and W.L. Smith, 1983: Monitoring tropical cyclone evolution with NOAA satellite microwave obsevations. J. Clim. Appl. Meteor., 22, 714–724. Wentz, F.J., D.K. Smith, C.A. Mears, and C.L. Gentemann, 2001: Advanced algorithms for QuikScat and SeaWinds/AMSR. Proc. of IGARSS 2001, IEEE, ed., IEEE, 1079–1081.