Multi‐Platform Observations of Severe Typhoon Koinu

Severe Typhoon Koinu passed south of Hong Kong on 8 and 9 October 2023, triggering the issuance of the Increasing Gale or Storm Signal No. 9, the second highest tropical cyclone (TC) warning signal in Hong Kong. Koinu was a difficult case for TC warning service due to its compact size and rather erratic movement over the coastal waters of Guangdong. To monitor Koinu's movement and wind structure, the Hong Kong Observatory utilized various observational platforms, including meteorological aircraft, ocean radar, and synthetic aperture radar on polar orbiting satellites. The paper presents major observations derived from these measurements. The aircraft probe and dropsonde data suggested boundary layer inflow, warm core structure, eyewall updraft, and high turbulence in the eyewall of the typhoon. The weather radar observations indicated the occurrence of a waterspout in the vicinity of the typhoon. Additionally, the study highlights the forecasting performance of the AI‐based Pangu‐Weather model, which could outperform the conventional global numerical weather prediction models in forecasting TC track in the region. The documentation of these observations aims to provide valuable references for weather forecasters and stimulate further research on forecasting this type of tropical cyclone.


Introduction
The South China coast is often affected by tropical cyclones originating from the South China Sea or the western North Pacific (Chen et al., 2018).Situated along the coast of southern China, Hong Kong is on average affected by six tropical cyclones moving within 500 km in a year.Tropical cyclones can bring high winds, heavy rainstorms and severe storm surges to Hong Kong (Choy et al., 2022a), causing significant damages and direct economic loss (Choy et al., 2020) and posing unique challenges to the forecasting and warning services (Choy et al., 2022b).Given the very sparse and limited observations over the ocean, and in support of tropical cyclone (TC) forecasting and warning services, Hong Kong Observatory (HKO) conducted various observational studies of TC in recent years (Chan et al., 2022;He, Li, et al., 2023;Tang et al., 2021).
September and October 2023 were eventful months for the occurrence of weather-related natural hazards over southern China.In early September, the region was affected by Super Typhoon Saola, which was a rather small TC yet the second most intense TC affecting the South China Sea since 1950.From the radar pictures of Hong Kong when Saola got close to the territory, it had a very compact eye and Saola's eyewall further shrank gradually in size when it passed south of Hong Kong, down to about 30 km in diameter.It is a rather special cyclone bringing hurricane force winds to Hong Kong and the case has been documented in a number of papers, including observational study (Chan, Choy, & Chiu, 2023), aircraft data (He, Chan, et al., 2023), and forecast (Chan, He, & Lui, 2023).About a week later, the remnant of TC Haikui brought torrential rain to the Pearl River Estuary and its neighboring areas.The HKO registered a record-breaking hourly rainfall of 158.1 mm ending midnight on 7 September 2023 since records began in 1884.It led to serious flooding at a number of places in Hong Kong.
On the evening of 8 October 2023, Severe Typhoon Koinu, surprisingly took on a northerly track toward Hong Kong for a few hours when it was located in the coastal waters south of the territory.The provisional best track of Koinu and its track near Hong Kong is shown in Figure 1.Koinu formed as a tropical depression over the western North Pacific to the east of the Philippines on 29 September 2023.It started to move northwestwards toward the seas east of Taiwan and intensified gradually in the next 4 days.Koinu intensified into a severe typhoon and reached its peak intensity on 2 October 2023 with a maximum sustained wind of 175 km/hr near its center.Koinu turned to move west-southwestwards on 4 October 2023 and weakened into a typhoon after skirting past the southern tip of Taiwan on 5 October 2023.While the satellite picture depicted the shrinking of Koinu's In view of public safety, the Increasing Gale or Storm Signal No. 9 in Hong Kong was issued for the second time of the year.Koinu was even more compact than Saola.From the radar pictures of Hong Kong, its eyewall had a diameter of less than around 30 km only.However, when being affected by the outer periphery of the eyewall of Koinu, a weather station at around 40 km south of Hong Kong, namely, Huangmao Zhou (HMZ), had registered the wind rising from around 75-90 km/hr to over 130 km/hr within an hour only.Winds at HMZ continued to increase afterward, reaching a peak of around 200 km/hr.Because of the relatively small size, Koinu was generally poorly captured by the global numerical weather prediction (NWP) models with a horizontal resolution of the order of 10 km only.
To monitor the movement and changes in the wind structure of Koinu, the HKO made use of a variety of measurement platforms, including meteorological aircraft, the newly installed ocean radar, and the newly available synthetic aperture radar (SAR) onboard polar orbiting satellites in real time.This paper summarizes the major observations from these measurements.A waterspout was suspected to occur in the vicinity of the territory in association with Koinu, and observations of which are also included in this paper.Finally there would be a short section on the forecasting aspect of Koinu.It turned out the AI-based Pangu-Weather (Bi et al., 2023) forecasts using deep-learning techniques with different sets of model initial field were capable of forecasting the proximity of Koinu to Hong Kong about 5 days ahead, and as such in medium term (3-5 days) the performance of Pangubased models was generally better than that of the conventional global NWP models, which has also been reported for the case of Saola (Chan, He, & Lui, 2023).
It is hoped that the documented materials would be useful for reference by weather forecasters in the future, and could stimulate further study on forecasting this type of rather special TC in autumn in the northern part of the South China Sea.

Data and Method
This study will analyze multi-platform observations from various sources.The first primary source is the in situ measurements from aircraft probe and Global Positioning System dropsondes.The second primary source is the remote sensing observations from weather radars, ocean radar, and SAR.

In Situ Observations
The aircraft probe data is probably the most significant data set for TC boundary layer observations because it directly measures the wind, turbulence, and other variables at high frequency (usually >4 Hz).However, one limitation is that the measurement can only be made along the flight path.Moreover, safety concerns prevent a routine probe measurement in the TC boundary layer (boundary layer measurements are usually available only under exceptional circumstances, e.g., when the aircraft is involved in a search and rescue exercise, which is the case for Koinu).To supplement the boundary layer information, dropsondes were deployed in Koinu.The dropsondes can provide the vertical profiles of wind, temperature, and other variables in and above the TC boundary layer.Similar to the probe measurement, the data collected by dropsonde only correspond to its descending path.
Three aircraft reconnaissance flights were conducted for Koinu, namely the morning and evening of 6 October, and the morning of 7 October.Air data probe collected useful data for these three flights.In particular, for the flight of 7 October, the aircraft was involved in a search and rescue exercise and low level flights had been conducted.For the two flights of 6 October, dropsondes were released near the center of Koinu.The probe data was collected by the Aircraft Integrated Meteorological Measuring System 20 Hz (AIMMS20), and the descriptions of the system can be found in Beswick et al. (2008) and Chan et al. (2011).The calculation of the turbulent kinetic energy (TKE) and eddy dissipation rate (EDR) (ε) using the aircraft data generally follow those in J. A. Zhang et al. (2011), Zhao et al. (2019), and He, Chan, et al. (2023).Dropsondes were released by the aircraft at a height of approximately 10 km to measure the vertical profiles of wind speed, wind direction, temperature, humidity, and pressure.The descriptions of the dropsondes are provided in Chan et al. (2018) and He et al. (2022).The ASPEN software was used to post-process the dropsonde data (J. A. Zhang et al., 2013).

Remote Sensing Observations
Although the in situ probe and dropsonde data are of high quality, they cannot provide a comprehensive spatial description of the TC structure.To complement such information, remote sensing observations including wind, turbulence (EDR), and rainfall/reflectivity observations from weather radars, wind and wave observations from ocean radar, and high-resolution nearsurface wind observations from SAR onboard satellites were also utilized and analyzed.The weather radar observations can reveal the 3D features of Koinu (e.g., the meteorological environments at the time of the waterspout) when it was close to Hong Kong, while the ocean radar and SAR can provide additional information on the wind and wave conditions in Koinu, for example, the SAR can provide a very-high-resolution (tens of meters) spatial distribution map of wind speed near the ocean surface (at a height of 10 m).
Composite radar reflectivity images from the two existing HKO's long-range weather radars installed at Tai Mo Shan and Tate's Cairn respectively were used for identifying the center of Koinu and assessing its eyewall structure.Low-level scans from the weather radar at Shenzhen were used for studying the characteristics of waterspout spawned from the rainbands of Koinu.Data from the first set of HF ocean radar operating at 5.275 MHz in Hong Kong paired up with another set in Shanwei operated by the South China Sea Bureau of the Ministry of Natural Resources of China were also used to study the wind pattern of Koinu when it came within 250 km of Hong Kong.The ocean surface wind patterns of Koinu were compared with those of weather radar imageries.Although extracting ocean wind and wave parameters from the second-order spectra of sea echoes from ocean radar measurements is very challenging to engineering applications (Barrick, 2008;Liu et al., 2007;Wu et al., 2003), recent advancements in the beam sampling method and retrieval algorithms improved the reliability of the derived wind and waves in TC situations (Li et al., 2012).The capability of ocean radar in capturing the TC center and estimating the radius of maximum wind (RMW) was validated against weather radar data.
In addition to EDR estimated from the in situ aircraft observation, EDR has also been estimated from the spectrum widths of the weather radars over southern China, based on the methods described in Chan et al. (2016).Only EDR data at 1-4 km above sea level are available.This study analyzes the results of 1 km height.
A SAR picture has been obtained when Koinu was located just to the south of Hong Kong.The SAR winds were directly obtained from the NOAA's STAR SOCD (Satellite Oceanography and Climatology Division) level-2 SAR data download site.The STAR SOCD TC products include the retrieved products from SAR images captured by different satellites, and the image analyzed in this paper is the scene captured by the Canadian Space Agency's RCM3 (RADARSAT Constellation Mission).The RCM was launched on 12 June 2019, with the three identical satellites working in concert to achieve daily access to 90% of the world's surface.

AI-Based Pangu-Weather Model
In addition to observations, the forecasting aspect of Koinu will also be briefly discussed with special attention paid to the recently developed AIbased Pangu-Weather model (Bi et al., 2023).The model was established by training deep neural networks using reanalysis weather data (i.e., ERA5 in 1979ERA5 in -2017) )

Wind Profile and Turbulence Intensity From Aircraft Observations
The plots of wind data from the aircraft data probe for the three flights are shown in Figure 2. In the first flight, winds were found to be lighter within the eyes and hurricane force winds were found just around the eyewall (Figure 2a).In the second flight, the flight pattern was more complex and there was not a clear pattern about the distribution of the winds (Figure 2b).Hurricane force winds were found at a limited number of locations only.At that moment, based on the aircraft data, Koinu might be relatively weak.
For the third flight, the flight level was lower and widespread hurricane force winds were recorded within the eyewall (Figure 2c).This is consistent with the weather radar observation and satellite imagery estimation that Koinu was a severe typhoon.Unfortunately, data were only available at the western half of Koinu so that a full picture of the distribution of winds could not be obtained.
The available dropsonde data are analyzed in Figure 3 for observations at around 03:00 UTC 6 October 2023, and Figure 4 for 12:00 UTC 6 October 2023.It could be seen from Figures 3 and 4 that there is a significant nearsurface inflow characterized by an inflow velocity from 10 to 20 m/s in the boundary layer of Koinu, suggesting that Koinu might continue to intensify over the northeastern part of the South China Sea, as indicated by He et al. (2022).On the other hand, many of the equivalent potential temperature profiles did not show much instability in the atmospheric boundary layer, that is, the equivalent potential temperatures were generally constant with height and sometimes even increasing with height.According to the results of He et al. (2022), this might suggest that the atmospheric instability on a large scale was rather limited to favor further strengthening of Koinu at that moment.The filled contour plots of equivalent potential temperature with respect to pressure levels are presented in Figure 5.The warm core structure of the TC similar to that shown in Emanuel (2018) is clearly depicted in the figure, which agrees with previous observations (He et al., 2022).
The vertical wind speed profiles from dropsondes on 6 October are shown in Figure 6.Similar to previous studies (e.g., He et al., 2022;He, Chan, et al., 2023), the semi-empirical model proposed by Vickery et al. (2009) to describe the wind profile in the TC boundary layer provides the best fit for the data points.The wind speed could be as high as around 33 m/s near the surface at a distance of about 24 km from the center of Koinu.The time series of the various meteorological parameters from the air data probe are shown in Figure 7 for the flight on the morning of 6 October, Figure 8 for the flight on the afternoon of 6 October, and Figure 9 for the flight on the morning of 7 October.In crossing/proximity of the eye of Koinu, the wind directions generally went through a circle and the wind speeds were rather low, which are expected structures of TC wind fields (Kepert, 2001).Near the eye of Koinu, the vertical velocity showed many fluctuations, and it could be both positive and negative.Nevertheless, the mean vertical velocity is generally positive and could reach around 3-5 m/s both in the boundary layer of Koiun below 500 m and well above the boundary layer at 10 km.This reflects that the eyewall updrafts can be found across a broad range of heights in TCs and agrees with Stern et al. (2016).
The turbulence intensity parameters, namely, TKE and EDR, were generally the largest in the eyewall of Koinu.This agrees well with previous observations and understandings of TCs (Ming et al., 2014;J. A. Zhang et al., 2009).The TKE per unit mass reached 6-10 m 2 /s 2 at wind speeds of 40-65 m/s and at heights of 400-500 m when the aircraft penetrated the western part of the eyewall of Koinu (Figure 9).This is slightly higher than those reported in J. A. Zhang et al. (2011), Zhao et al. (2019), and Sparks et al. (2019), in which the TKE was generally lower than 6 m 2 /s 2 in the TC eyewalls at similar wind speeds and altitudes.This, along with the EDR estimation based on the weather radar (to be shown in the next section), implies the special turbulence structure of Koinu, that is, there seems to be higher turbulence in the western and southern parts of its eyewall.Further research is needed to investigate whether this is a specific feature unique to this particular typhoon case and explore potential causes.In addition, it is observed that there is no significant wind speed dependence of TKE and EDR at this wind speed range (40-65 m/s; not shown).This has been well documented and discussed in previous studies (e.g., Black et al., 2007;French et al., 2007), which is primarily due to the saturation of sea surface roughness and drag coefficient at high wind speeds (>30 m/s; Donelan et al., 2004;Powell et al., 2003).
Local closure techniques are frequently employed for turbulence parameterization in NWP models, for example, the well-known Mellor-Yamada-Nakanishi-Niino (MYNN) and Mellor-Yamada-Janjić (MYJ) schemes for  (Janjić, 1994;Mellor & Yamada, 1982;Nakanishi & Niino, 2004).In the context of a local 1.5-order closure, the dissipation rate of TKE (ε) can be parameterized as a function of the 1.5 order of TKE (Stull, 1988).This implies a logarithmic relationship between ln(ε) and ln(TKE) with a slope of 1.5.This relationship aligns with the results shown in the present study.Specifically, the regression slope is in the order of 1.6 and the y-intercept is around 4 to 6.The values of regression slope closely agree with that predicted by the local 1.5-order closure parameterization.This provides an observational support for the use of the 1.5-order scheme in the parameterization of TCs in NWP models.More observations should be collected and analyzed to see if these parameters are universal or not.

Wind, Wave, and Turbulence Structure From Radar Observations
The weather radars in Hong Kong well captured the structure of Koinu.From the 3-km height imagery, the eyewall of Koinu showed up very well within the radarscope (Figures 10a and 11a).They serve as validation reference for testing of the capability of ocean radar in analyzing the wind structure of Koinu.The ocean surface wind speed and direction patterns of Koinu corresponding to the time of weather radar imageries are shown in Figures 10b and 11b.By comparing ocean radar derived wind field with the weather radar pictures, it could be seen that the ocean radar wind data captured the locations of Koinu's eye very well.The difference in the eye fix was just about 10-20 km.The wind speeds derived from the ocean radar were also generally reasonable, though underestimated as compared with the actual observations.On the morning of 7 October, maximum winds of around 20 m/s were found at a distance of about 24 km (Figure 10b) and 30 km from the center (Figure 11b).Both hinted the RMW of Koinu was of the order of 20-30 km.On the early morning of 8 October (Figure 11b), gale force winds generally prevailed over the measurement domain of the ocean radar.This wind pattern also appeared to be reasonable except that an annular ring of maximum wind speed near the center of Koinu could not be depicted in the derived field.It is probably due to Apart from the surface wind speed, ocean radar derived products also include significant wave height.On the morning of 7 October (Figure 12a), higher wave height of around 8 m was found to be associated with the strong south to southeasterly winds.Though actual observations are not available, such a pattern also appears to be reasonable.On the morning of 8 October (Figure 12b), the wave heights were generally about 6-8 m in the measurement domain.At both times, the wave heights were much lower (well below 2 m) near the center of Koinu because of weaker winds.This is also reasonable.
EDR map obtained from the spectral width data of the weather radars at a height of 1 km is shown in Figure 13.There was moderate to high turbulence with EDR from 0.3 to 0.7 m 2 /s 3 in the western and southern parts of Koinu's eyewall.This generally align with the aircraft observations.As for turbulence in the rainbands, the EDR was much lower on the eastern and southern parts of the rainband wrapping into the center of Koinu.

High-Resolution Near-Surface Wind Field From Satellite Observations
The SAR picture obtained when Koinu was located just to the south of Hong Kong is shown in Figure 14.From the figure, it clearly showed that Koinu was a very compact storm.The RMW of Koinu from the SAR image in the order of 20 km only, which is lower than most tropical cyclones with RMW >30 km.The comparison results between the SAR winds and the bouy winds could also be found in Figure 14.From the high R-value (0.94 for 6min time coincidences and 0.88 for 1-hr time coincidences), it could be seen that the SAR winds and the buoy winds are well correlated, though the buoy winds are generally lower, because they are 10-min mean winds    whereas the SAR winds are instantaneous.The availability of SAR wind data greatly help the monitoring and provision of warning service for TCs.

TC-Spawned Waterspout and Its Formation Environment
Associated with the outer rainbands of Koinu, an intense radar echo with a maximum reflectivity over 54 dBZ developed near the Dangan islands (about 20 km south of Hong Kong) at about 08:25 UTC on 9 October 2023.It moved west steadily and displayed a well-defined "hook" signatures based on 0.5°PPI scan of the Shenzhen Weather Radar (SZWR) at around 08:42 UTC on 9 October 2023 (Figure 15a).While a velocity couplet (curved arrow) was observed in Figure 15b in association with the hook echo, demonstrating a clear tornadic vortex signature (TVS) in the radial velocity products.The maximum gate-to-gate storm relative radial velocity (SRV) difference is about 26.5 m/s, suggesting the formation of a waterspout.Cross-section along the horizontal plane of velocity couplet in Figure 15c showing the maximum reflectivity core located at a height of around 2.5 km and the maximum velocity of the waterspout reaching above 21 m/s with height around 2 km in Figure 15d.The vortex tube related to the waterspout reached a height of about 2 km.Tracing the radar echo with reflectivity greater than 53 dBz as a proxy of tracking the waterspout, it seemed to have a rather long lifespan of over 3 hr, and the echo finally weakened after around 11:00 UTC that evening.In Hong Kong, waterspout occurred mostly in the summer months from June to August mostly caused by convective weather.The waterspout brewed from the rainbands of Koinu in the month of October was a relatively rare event.
The waterspout associated with Koinu can be regarded as a TC spawned tornado.Carroll-Smith et al. (2023) and Edwards (2012) pointed out several factors favorable to the formation of TC tornadoes which were found applicable in this case.These included (a) the waterspout was formed over the northeastern quadrant relative to Koinu's center due to climatologically enhanced instability and shear in that region, (b) SZWR's radar image (Figure 15c) suggested the presence of shallow-topped "miniature" supercell which was typically formed under high shear and helicity environment as well as concentration of buoyancy in the lowest few kilometers above ground, (c) the presence of baroclinic boundaries (in this case the Dangan islands or broadly speaking the coastal areas of Guangdong) providing stronger horizontal temperature gradient and higher lowlevel shear for spawning tornado.Further, some of the precursor requirements for tornado touchdowns as mentioned in Schneider and Sharp (2007) were observed in this case including a hook shape signature in the radar reflectivity image (Figure 15a) and the presence of a TVS (SRV) difference is about 26.5 m/s, which was a stronger indicator of tornadogenesis.While TC-spawned tornado was seldomly observed near the south China coast, the meteorological factors contributing to the formation of waterspout related to Koinu were found consistent with the results of past studies by Carroll-Smith et al. (2023), Schneider and Sharp (2007), and Edwards (2012).

Evaluation of Forecasting Performance of AI-Based Pangu-Weather Model
Koinu was very compact to be well resolved by the existing global NWP models.Despite some early runs of the global models indicated that Koinu might cross the northern part of the South China Sea as a relatively strong TC, most global models subsequently changed their stories on a consensus that Koinu would rapidly weaken over the northeastern part of the South China Sea and quickly dissipate due to intrusion of drier and cooler air of the northeast monsoon.This might have misled operational forecasting and warning services that Koinu might not pose significant threat to the Pearl River Estuary, even as close as a few days before its closest approach.Nevertheless, the Pangu-Weather model persistently forecast that Koinu would continue to track west along the coastal areas of Guangdong and reach the Pearl River Estuary.As a result, they have much better performance in terms of track error, especially in the forecast period of 2-5 days, as shown in the time series plots of Figure 16.This case study demonstrated the usefulness of Pangu-Weather initialized by different initial conditions.While Pangu-Weather was trained using ERA5 reanalysis, the results show that Pangu-Weather initialized by the various operational analyses is able to produce high-quality forecast for TC track, and potentially constitute a global ensemble system.The unique challenges in forecasting and warning services for Koinu, including a review to objective guidance from various models regarding the track, intensity, winds and rainfall, are documented in detailed in Chan and He (2023).
In addition to Saola, this is the second time that AI-based models outperformed conventional models in forecasting the movement of a midget TC.The mesoscale model TRAM (Y.Zhang et al., 2022) with a horizontal resolution of 9 km also performed much better than global models, in terms of both track forecast (Figure 16) and intensity forecast where multiple runs of TRAMS indicated an intense yet midget Koinu would manage to reach the Pearl River Estuary.This points to the need for improving the performance of global NWP models in forecasting the intensity and movement of TCs.

Conclusions
Koinu was a difficult case for TC warning service due to its compact size and rather erratic movement over the coastal waters of Guangdong.This paper documents the weather observations of Koinu from a number of platforms, including aircraft, radar and satellite.It is hoped that the paper could stimulate further study of this case, particularly the possibility of forecasting the compact structure and erratic track of Koinu.The added value of this work to the existing literature mainly lies in: 1. Due to various constraints such as limitations in observation techniques, there has been a limited amount of available data regarding turbulent processes occurring in the TC boundary layer.The primary source of significant existing data arises from research aircraft flights.However, due to safety constraints, only the United States has conducted direct aircraft observations in the boundary layer of two TCs worldwide before 2000 (J. A. Zhang et al., 2011).The aircraft observations made by the HKO during the eyewall penetration in the boundary layer of Koinu with wind speed up to 65 m/s would be a novel and unique contribution to the existing database of TC observations.The analysis of turbulence parameters such as TKE and EDR would also enhance our understanding of turbulence characteristics near the eyewall of an intense TC.Earth and Space Science 10.1029/2023EA003366 2. The use of ocean radar data in locating the center of a TC and estimating the radius of maximum winds has not been well demonstrated in the South China Sea region before.This case study of Koinu could serve as a validation reference for testing the capability of ocean radar in analyzing the wind structure of a TC. 3. TC-spawned tornado was seldomly observed near the south China coast.This study describes the occurrence of a waterspout associated with Koinu and the meteorological environment that triggers its formation.The documentation of this relatively rare event would be beneficial to the understanding and forecasting of TCspawned tornadoes in this region.
The main findings and corresponding discussions are given below.The dropsonde data indicated significant inflow in the boundary layer, and the Vickery model provided the best fit for wind speed profiles.The warm core structure of the TC was clearly depicted, while the equivalent potential temperature profiles showed limited atmospheric instability.The probe data revealed eyewall updraft up to about 5 m/s, along with high turbulence in the western and southern parts of the eyewall.A correlation between TKE and EDR was observed.These findings contribute to our understanding of wind and turbulence structure of tropical cyclones.Nevertheless, further investigation is necessary to ascertain whether these findings are specific to this particular TC or if they hold true for a broader range of TC events.Additionally, exploring the underlying causes behind these observed patterns will be a crucial aspect of future research.
The SAR picture obtained when Koinu was near Hong Kong revealed its compact nature.The SAR winds showed a good correlation with buoy winds, albeit slightly lower due to different temporal scales.The weather radar observations suggest the occurrence of a waterspout in the vicinity of Koinu.Typical characteristics of a TC-spawned tornado were displayed, with factors such as enhanced instability and shear, presence of a shallow-topped supercell, and the influence of baroclinic boundaries.The observations align with previous studies on tornado formation in TCs.Further research efforts are warranted to explore the fine-scale structure of TC spawned tornadoes and investigate the mechanism underlying their formation.This would be important for enhancing forecasting capabilities and facilitating early warning systems for these highly destructive phenomena.
Koinu's compact size posed challenges for global NWP models, leading to inconsistencies in track forecasts.However, the Pangu-Weather model and the mesoscale TRAM model demonstrated better performance in forecasting Koinu's track and intensity.This highlights the importance of enhancing global models for accurate TC intensity and movement predictions.
Based on the results so far in the TC season of 2023, AI-based Pangu-Weather models were found to have outperformed the conventional models in forecasting the track and intensity of TCs.Rather urgent need would be required to improve such models in order to provide effective support to the provision of forecast and warning services of TCs in this part of the world.

Figure 1 .
Figure 1.(top) Provisional best track of Koinu and its track near Hong Kong (in Hong Kong Time = UTC + 8 hr).The track of Koinu was slow and erratic during 6-8 October 2023.The center of Koinu passed very close to Huangmao Zhou on the night of 8 October 2023.(bottom) Infrared satellite image of Himawari-9 Satellite of the Japan Meteorological Agency at 00:00 UTC on 2 October (left) and 7 October (right) 2023.

Figure 2 .
Figure 2. Flight routes and selected horizontal wind data along the flights overlaid on images generated from Himawari-9 geostationary satellite of Japan Meteorological Agency.(a) Flight between around 02:00 and 03:30 UTC, 6 October 2023 on top of false color satellite image of 02:00 UTC; (b) flight between around 11:20 and 13:40 UTC, 6 October 2023 on top of infra-red satellite image of 12:00 UTC; and (c) flight between around 01:40 and 02:35 UTC, 7 October 2023, on top of false color satellite image of 02:00 UTC.

Figure 3 .
Figure 3. (a) Vertical profiles of tangential (red) and radial (blue) wind speeds in Typhoon Koinu at 03:00 UTC 6 October 2023.Tangential wind speed: anti-clockwise positive; radial wind speed: outflow positive.Red triangles represent dropsonde locations relative to the storm center.(b) Same as (a) but for potential temperature (red) and equivalent potential temperature (blue).
and is capable to provide medium-range global weather forecasting.Starting from mid-2023, the HKO has begun to trial run the opensourced Pangu-Weather model in real time.The model was initialized by five global NWP models' operational analyses, namely Pangu-ECMWF (European Center for Medium-Range Weather Forecasts), Pangu-NCEP (National Centers for Environmental Prediction), Pangu-DWD (Deutscher Wetterdienst in Germany), Pangu-MeteoFrance and Pangu-ECCC (Environment and Climate Change Canada).While traditionally, TC forecasting at HKO is mainly based on the consensus of four major global NWP models as an multi-deterministic-model ensemble approach, including ECMWF Integrated Forecast System, Japan Meteorological Agency, NCEP of the United States, and the Met Office in the United Kingdom (UKMO) Unified Model.

Figure 6 .
Figure 6.Fitting of vertical profiles of wind speeds in the lowest 1,000 m in Typhoon Koinu to the wind profile models, including the logarithmic law, Vickery et al. (2009) model, Gryning et al. (2007) model, Deaves and Harris (1978) model, and power law, (a-c) at 03:00 UTC 6 October 2023 and (d-f) at 12:00 UTC 6 October 2023.Panel (d) represents the distance to the storm center.

Figure 7 .
Figure 7. Time series of the (a) wind speed, (b) wind direction, (c) vertical wind speed, (d) turbulent kinetic energy (TKE), (e) cube root of the eddy dissipation rate (ε 1/3 ), (f) variation of ε with TKE, (g) distance to storm center, and (h) flight altitude based on the aircraft data in Typhoon Koinu between 02:00 and 03:30 UTC 6 October 2023.The black line in (f) represents the linear fit between ln(ε) and ln(TKE).

Figure 10 .
Figure 10.(a) Weather radar imagery at 3-km CAPPI showing Koinu's eye near 21.1°N, 115.4°E at around 02:36 UTC on 7 October 2023; (b) wind field derived from ocean radar around the same time also indicating an eye (purple star) at around 21.2°N, 115.4°E.

Figure 11 .
Figure 11.(a) Weather radar imagery at 3-km CAPPI showing Koinu's eye near 21.4°N, 115.0°E at around 18:12 UTC on 7 October 2023; and (b) wind field derived from ocean radar around the same time also indicating an eye (purple star) at around 21.3°N, 114.8°E.

Figure 14 .
Figure 14.Synthetic aperture radar winds captured by Canadian Space Agency's RADARSAT Constellation Mission (RCM3) on 8 October 2023 (left).Comparison of ocean winds between RCM3 winds and buoy observations (right).

Figure 15 .
Figure 15.(a) Reflectivity and (b) radial velocity based on 0.5°PPI scan of the SZWR at around 08:42 UTC on 9 October 2023.An intense radar echo displaying welldefined "hook" signatures in (a); while a velocity couplet (curved arrow) was observed in (b) in association with the hook echo in (a).(c) Cross-section along the horizontal plane of velocity couplet in (a) showing the maximum reflectivity core located at height of around 2.5 km; while the cross-section along the horizontal plane of velocity couplet showing the maximum velocity of the waterspout reaching above 21 m/s with height around 2 km in (d).

Figure 16 .
Figure 16.Root-mean-square error of model forecast positions of Koinu as a function of forecast hours.Forecasts are verified against Koinu's analysis positions based on HKO's operational warning track, and homogenized to have a common data set among different models.Dashed lines are verification results of forecasts from Pangu-Weather initialized respectively with the operational analyses of DWD, ECCC, ECMWF, MeteoFrance, and NCEP models.