Analysis and Forecast Using Dropsonde Data from the Inner-Core Region of Tropical Cyclone Lan ( 2017 ) Obtained during the First Aircraft Missions of T-PARCII

The inner core of Tropical Cyclone Lan was observed on 21− 22 October 2017 by GPS dropsondes during the first aircraft missions of the Tropical Cyclones-Pacific Asian Research Campaign for the Improvement of Intensity Estimations/Forecasts (T-PARCII). To evaluate the impact of dropsondes on forecast skill, 12 36-h forecasts were conducted using a Japan Meteorological Agency non-hydrostatic model (JMA-NHM) with a JMANHM-based mesoscale four-dimensional data assimilation (DA) system. Track forecast skill improved over all forecast times with the assimilation of the dropsonde data. The improvement rate was 8−16% for 27−36-h forecasts. Minimum sea level pressure (Pmin) forecasts were generally degenerated (improved) for relatively short-term (long-term) forecasts by adding the dropsonde data, and maximum wind speed (Vmax) forecasts were degenerated. Some of the changes in the track and Vmax forecasts were statistically significant at the 95% confidence level. It is notable that the dropsonde-derived estimate of Pmin was closer to the realtime analysis by the Regional Specialized Meteorological Center (RSMC) Tokyo than the RSMC Tokyo best track analysis. The degeneration in intensity forecast skill due to uncertainties in the best track data is discussed. (Citation: Ito, K., H. Yamada, M. Yamaguchi, T. Nakazawa, N. Nagahama, K. Shimizu, T. Ohigashi, T. Shinoda, and K. Tsuboki, 2018: Analysis and forecast using dropsonde data from the innercore region of Tropical Cyclone Lan (2017) obtained during the first aircraft missions of T-PARCII. SOLA, 14, 105−110, doi: 10.2151/sola.2018-018.)


Introduction
Analyses and forecasts of tropical cyclones (TCs) are important because TCs can result in devastating impacts such as violent wind, heavy rainfall, high waves, and storm surge.In the Western North Pacific (WNP), strong TCs occur more frequently than any other basin (D'Asaro et al. 2011).Previous researches and modeling efforts have improved TC track forecasts, and recent studies have reported improvements in TC intensity forecasts (DeMaria et al. 2014;Ito 2016;JMA 2018;Yamaguchi et al. 2017).However, track and intensity forecast errors in the WNP are still often too large for effective disaster prevention and mitigation.
To reduce forecast errors, in-situ observations of TCs are crucial, because they provide ground truth and impact the initialization of numerical weather prediction models through data assimilation (DA) systems.Although it is generally difficult to obtain TC observations over the ocean, several aircraft observation missions have been recently conducted in the WNP.The Dropwindsonde Observations for Typhoon Surveillance near the Taiwan Region (DOTSTAR) project (Wu et al. 2007) has measured environmental data around TCs operationally since 2003.Several studies have found that the DOTSTAR data contributed to improved TC track forecasts (Chou and Wu 2008;Chou et al. 2011).
In addition to the TC environment, the inner-core region plays a key role for both TC energetics (Wang and Wu 2004), and track (Yamada et al. 2016).As such, inner-core observations from aircrafts are critical in terms of TC intensity analyses.Forecast improvements in the WNP were demonstrated by The Observing System Research and Predictability Experiment (THORPEX) Pacific Asian Regional Campaign (T-PARC) in the autumn of 2008 (Chou et al. 2011;Weissmann et al. 2011).Recently, Wong et al. (2014) documented the potential forecast impact of flightlevel inner-core observations by the Hong Kong Observatory.However, the importance of aircraft-based inner-core observations for forecast quality is still an open question because such observations are limited (Aberson 2008;Aberson et al. 2017).
On 21−22 October 2017, 26 GPS dropsondes were dropped from 13.8 km in the inner-core region of TC Lan (2017), during the time of the lowest central sea level pressure (Pmin) of any WNP TC in 2017.These were the first Japanese aircraft missions as part of the Tropical cyclones-Pacific Asian Research Campaign for Improvement of Intensity estimations/forecasts (T-PARCII).During these missions, the eyewall of the intense TC was penetrated with a Gulfstream II jet.
In this paper, we briefly introduce the observed pressure and wind speed, and then evaluate the impacts of inner-core observations on forecast skill using the Japan Meteorological Agency (JMA) non-hydrostatic model (NHM) (Saito 2012;Saito et al. 2007).For DA, we employed the JMA-NHM-based variational DA (JNoVA) system, which is a sophisticated mesoscale 4D-Var system (Honda and Sawada 2009;Honda et al. 2005).The benefits of inner-core observations described in this paper are encouraging, yet at the same time they remind us of the overall challenge of TC intensity analysis and forecast.

Aircraft missions for TC Lan
TC Lan (2017) formed north of Palau on 16 October and moved northward on 17 October after some meandering.It made landfall on Shizuoka Prefecture, Japan, at 1800 UTC 22 October with a record-breaking gale wind radius of > 800 km and subsequently transitioned into an extratropical cyclone at 0000 UTC 23 October.
During the first aircraft missions of T-PARCII, dropsonde

Analysis and Forecast Using Dropsonde Data from the Inner-Core Region of Tropical Cyclone Lan (2017) Obtained during the First Aircraft Missions of T-PARCII
Nakanishi-Niino level-3 closure model (Nakanishi and Niino 2004).The physical processes retained (Abarca and Corbosiero 2011) in the adjoint and simplified nonlinear models are similar to those retained in the forward model, although they include several simplifications such as the replacement of the convective scheme with large-scale condensation and moist convective adjustment.
The length of the assimilation window was set to 3 h, as in the operational regional DA system, and observations were divided into four hourly slots.

Experimental setup and validation
To elucidate the impact of inner-core dropsonde data on forecast skill, we conducted two DA experiments with JNoVA and subsequent 36-h forecast experiments with JMA-NHM.We refer to the DA experiment with the inner-core observations and subsequent forecast experiment as TPARCII and the ones without the inner-core observations as CTRL.The DA experiments were performed over 12 3-h cycles from 0300−0600 UTC 21 October through 1200−1500 UTC 22 October, and the subsequent 12 forecast experiments used the objective analysis at the end of each DA cycle as initial conditions.The 10, 13, and 5 dropsonde1 observations deployed in T-PARCII were assimilated into the first, second, and eighth DA cycles.For simplicity, the forecast time of x hour is henceforth described as FTx.A tracking algorithm for the simulated TC is the same as in Ito et al. (2015), except that the TC center is defined by the method of Braun (2002).Although we focused on the TC position and intensity, the analysis increment and the precipitation forecast skill are described in the supplemental materials.
The data used in both experiments included conventional observations recorded by surface stations, radiosondes, ships, aircraft (including dropsondes from DOTSTAR), and vertically integrated precipitable water measurements derived from groundbased GPSs, wind profilers, Doppler radar radial velocity data, and TC bogus data.As for dropsonde observations deployed in T-PARCII, observed zonal wind, meridional wind, temperature, and relative humidity were assimilated at the designated pressure levels of 200, 250, 300, 350, 400, 500, 600, 700, 800, 850, 900, and 925 hPa as well as sea-level pressure.Sea-level pressure was estimated from the lowest recorded vertical level of the dropsonde corrected to sea-level.We considered the GPS-derived horizontal displacement as the dropsonde drops because the dropsonde may orbit around the TC center significantly (Aberson 2008;Aberson et al. 2017).If the difference between the first-guess and observed value was greater than 10 times that of the standard observation error, the observation was not assimilated.
A paired-sample two-tailed t-test was employed to check if a difference was statistically significant, except for FT36, for which only two samples were available.The sample number was replaced by the effective sample number to account for the persistence of sequential errors by using a lag-1 positive autocorrelation coefficient, similar to the method of Kuhl et al. (2013) and Ito et al. (2016).An individual absolute error was used as a sample element for applying the statistical test to intensity forecast skill.Because TC Lan became an extratropical cyclone at 0000 UTC 23 October, the number of samples used for validation decreased with increasing forecast time.

Inner-core observations and intensity analysis
Figures 1b and 1c show the locations of the assimilated dropsonde data below 1 km on 21 and 22 October, respectively.The blackbody temperature from HIMAWARI-8 infrared satellite imagery is overlaid at 06:52:30 UTC 21 October and 01:15:00 UTC 22 October when the lowest sea-level pressure on each day was observed by the dropsondes at almost the same time (rectan-observations were deployed at 0510−0715 UTC 21 October and 0101−0136 UTC 22 October (Fig. 1a).This period covers the lowest Pmin of the year according to the Regional Specialized Meteorological Center (RSMC) Tokyo best track analysis (hereafter, BTA).During the two flight missions, 4, 5, and 17 dropsondes were deployed in the eye, eyewall, and surrounding region, respectively.In the meantime, the surveillance mission of DOTSTAR was conducted during the evening of 21 October.
The dropsondes used were newly developed by Meisei Electric Co., LTD., and Nagoya University.The dataset of T-PARCII has already been subjected to basic quality control measures.A trial experiment showed that aircraft-based dropsonde data generally agrees well with ground-based sonde data.However, there are inevitable uncertainties about the quality of dropsonde data, particularly for data obtained under extreme weather conditions.

JMA-NHM and JNoVA
The frameworks of JMA-NHM and JNoVA are outlined in this subsection, but more detailed explanations can be found in Honda and Sawada (2009) and JMA (2013).The domain is discretized into 817 × 661 grid points with a 5-km grid spacing in the forecast model and the high-resolution model of the JNoVA system, while the same domain is discretized into 273 × 221 grid points with a 15-km grid spacing in the simplified model of the JNoVA system.The calculation domain was the same as that used in the JMA operational regional model as of 2017.The high-resolution forward model employs a horizontally explicit and vertically implicit scheme as a dynamical core with six-category bulk microphysics (Ikawa and Saito 1991), the modified Kain-Fritsch convective scheme (Kain and Fritsch 1990), a clear-sky radiation scheme (Yabu et al. 2005), and a cloud radiation scheme (Kitagawa 2000).Boundary layer turbulence is determined by the Mellor-Yamada- gles in Figs.1b and 1c, respectively).It is evident that the dropsonde observations were located very close to the center of TC Lan, which enables us to evaluate Pmin accuracy.After applying an altitude adjustment from the lowest observation level to sealevel through the hydrostatic relationship, observations indicate sea-level pressures of 925 hPa and 930 hPa, respectively (Figs. 2a  and 2b).Surface pressure variation among two dropsondes within the eye on 21 October (dropsonde identifi cations (IDs): R and P) was small, such that 925 hPa is a reasonable estimate of Pmin (Fig. 2a).
Figure 2c shows that the largest wind speed observed was 80 m s −1 at z = 1.3 km in the inward fl ank of the eyewall, south of the TC center on 21 October.For low-level wind speed, we observed 57.6 m s −1 at z = 14.2 m, which corresponds to 55.3 m s −1 at 10 m, assuming a surface roughness length of 0.002 m (Powell et al. 2003) and neutral stability.This is merely the maximum wind speed within our observations and is not necessarily equivalent to the offi cial Vmax in BTA, defi ned as the 10-minute sustained wind speed at 10 m.The weak wind speed of 4 m s −1 was observed in the eye on 21 October (ID: R).In contrast, the observed wind near the center was around 40 m s −1 on 22 October (ID: V) (Fig. 2d).
Compared with real-time analysis (hereafter, RTA) and BTA by RSMC Tokyo, the dropsonde-derived Pmin are in better agreement with RTA (Fig. 3a).The quality control of dropsonde data took time, such that the level-2 dropsonde data was not available at the time of the publication of BTA on 8 December 20172 .Pmin and Vmax in RTA were corrected by as much as 30 hPa and 15 kt, respectively (Figs. 3a and 3b).While forecast "error" is evaluated against BTA in Section 3, the impacts of this discrepancy between RTA and BTA on forecast skill will be discussed in Section 4.

Track forecast
Figure 4 shows the mean forecast error of TC track and the root mean squared differences (RMSDs) of Pmin and Vmax averaged over 12 forecasts.Figure 4a demonstrates that dropsonde data improved track forecasts across all FTs in TPARCII.The maximum improvement rate of 16% was achieved at FT27 and FT30.A statistical test reveals that the difference in track forecast error at FT30 was statistically signifi cant at the 95% confi dence level.These results suggest the potential utility of inner-core dropsonde observations to improve TC track forecast skill.
For the forecast initialized at the end of the fi rst DA cycle, improved track forecasts are exemplifi ed by the difference in TC center positions between CTRL and TPARCII (57 km displacement to the north in TPARCII) at landfall time (Fig. 5a).The simulated TC track in CTRL was generally displaced to the southwest of the corresponding TC center in BTA.Although a westward displacement bias was still observed, landfall time was correctly predicted in TPARCII.
Figures 5b and 5c show the zonal and meridional component of the steering wind, defi ned as the mass-weighted mean wind between 300 and 700 hPa within a 300-km radius of the TC center.The predicted steering vector in TPARCII exhibits better agreement with the steering vector in both analysis fi elds.The time-mean difference of the zonal and meridional component between CTRL and TPARCII was 0.016 m s −1 and 0.481 m s −1 , respectively.One may think that this is too small to account for   the northward shift of 57 km.However, multiplying the difference of the meridional component by 36 h, we obtain a displacement of 62 km due to advection, which roughly corresponds to the 57-km northward shift.Because the difference between CTRL and TPARCII stems solely from the assimilation of dropsondes in the inner core region, the inner-core observations serve to yield more accurate wind forecasts.

Intensity forecast
For TC Lan intensity forecasts, RMSDs of Pmin are reduced in TPARCII for FT27-36, while they are larger for FT3-24 (Fig. 4b).RMSDs of Vmax are generally larger (up to 2 m s −1 ) in TPARCII.These results indicate that the simulated intensities in TPARCII are not in better agreement with BTA than the CTRL, except for the Pmin forecast for FT27-36.Statistically, the degeneration of Vmax at FT30-33 reached the 95% confidence level.
Pmin and Vmax in the analysis field generated by JNoVA are shown in addition to the forecasts in Fig. 6.The intensities in TPARCII were generally weaker during cycles 1−4 than in the CTRL because the observed wind was generally weaker than the first-guess (see supplemental material A), except for Vmax in the analysis of the first cycle that incorporated the strong wind observations (IDs: F and G, Fig. 2c).While BTA showed intensification from 21 October to 22 October, decay or neutral change was indicated by the dropsonde-derived Pmin, RTA, and simulated values.One notable difference between CTRL and TPARCII is the magnitude of decay of Pmin at 0300−0600 UTC 22 October.The large deviation in TPARCII from BTA yielded the large forecast errors, although the simulated values are in better agreement with RTA.

Discussion
The differences in TC intensity between RTA and BTA dataset were considerable, and the dropsonde-derived Pmin more closely agrees with RTA.Thus, forecast skill is re-evaluated with respect to RTA to quantify the impact of uncertainties in the reference data (Fig. 7).
When RTA is employed as the reference, track forecast errors decrease by roughly 10% in both experiments, which can be explained by the TC bogus used in JNoVA constructed from RTA.Nevertheless, the improvement in TPARCII is robust and significant at the 95% confidence level for FT30.This result is presumably because the track forecast error is sufficiently larger than uncertainties in the position analysis.In contrast, the Pmin forecast skill in TPARCII becomes comparable for FT0-18 and better for FT21-36 with respect to CTRL.Recall that completely different results are obtained with BTA as the baseline (Figs.4b  and 7b).Smaller differences in Vmax forecasts are also seen in Figs.4c and 7c.However, none of the Pmin and Vmax forecast differences between CTRL and TPARCII were statistically significant at the 95% confidence level.
The better agreement in Pmin between RTA and the dropsonde-based analysis implies that what we think of as "intensity forecast errors" against BTA is just an uncertain estimate of forecast errors.In other words, TC intensity forecast "improvement" may not necessarily mean that the forecast is more accurate with respect to true TC intensity.

Concluding remarks
Using dropsonde data from T-PARCII, we conducted DA and forecast experiments using JMA-NHM and JNoVA.Our results are encouraging in that statistically significant improvement in the track forecast was obtained using inner-core dropsonde observations with high-resolution systems.Although model TC intensity forecast skill was generally degraded, this may be an artifact of the quality of BTA.Dropsonde-derived Pmin was closer to RTA, and intensity forecast skill becomes as good or better with dropsondes when RTA is used as the verification baseline.In other words, potential uncertainties in BTA may constrain the reliability of evaluating TC intensity forecast skill.There are inevitable uncertainties in the intensity analysis because newly developed dropsondes were deployed in the strong TC for the first time without supportive ground-based observations.Additionally, we have not considered the interference between dropsondes and TC bogus.These issues may become more evident as further forecast improvements are requested and should be investigated by future TC observation missions.

Fig. 1 .
Fig. 1.(a) Locations of dropsondes (colored dots) assimilated below 1 km around TC Lan (2017) with the TC center position from BTA identified by the hash marks on the solid black line.The dashed line indicates the center position after extratropical transition.(b) Close-up view of dropsonde locations in T-PARCII plotted over an infrared satellite image from 06:52:30 UTC 21 October.The closest surface dropsonde observation time to the satellite image time is indicated by a rectangle.(c) Same as (b) but at 01:15:00 UTC 22 October.

Fig. 4 .
Fig. 4. (a) Track forecast errors relative to BTA data averaged over the forecasts.The results are shown for CTRL (in blue) and TPARCII (in red).(b), (c) Same as (a) but for RMSD of Pmin and Vmax forecasts, respectively.(d) Number of samples for each forecast time.

Fig. 5 .
Fig. 5. (a) TC center positions of forecast experiments initialized at the end of the first DA cycle, as well as BTA and RTA.Mass-weighted steering flow of TC Lan in (b) zonal and (c) meridional directions for the forecast initialized from the analysis field at the end of the first DA cycle and the analysis.