Performance of windshear/microburst detection algorithms using numerical weather prediction model data for selected tropical cyclone cases

Three windshear/microburst detection algorithms are used operationally at the Hong Kong International Airport. Their performance is studied in the present article by assuming the availability of the complete set of meteorological data from numerical weather prediction model output. Only selected tropical cyclone cases are considered. The performance is evaluated using pilot reports of windshear. It turns out that the glide path scan windshear detection algorithm has the best overall performance by considering the area under curve using the relative operating characteristic curves. This may be related to the better geometry in covering the glide paths of the airport for this particular algorithm.


| INTRODUCTION
Low-level windshear and turbulence are hazardous weather conditions to the arriving and departing aircraft at an airport. At the Hong Kong International Airport (HKIA), the major windshear and turbulence reports from the pilots were related to terrain-disrupted airflow disturbances. Such terrain-induced windshear and turbulence happen mostly in the spring and summer. In the spring, the prevailing easterly airstream under the stable boundary layer is disrupted by the complex terrain of Lantau Island, a mountainous island to the southeast of HKIA, leading to various microscale airflow features like mountain waves and vortices over the airport. In summer, the terrain disruption of the airflow is very much related to the southwest monsoon or east/southeasterly airstream associated with tropical cyclones. In this article, we would focus on terrain-induced windshear and turbulence at HKIA relating to tropical cyclones. The alerting of low-level windshear/turbulence at HKIA is mainly based on remote-sensing (including Doppler LIght Detection and Ranging (LIDAR) systems, Terminal Doppler Weather Radars (TDWR) and wind profilers) and surface (including automatic weather stations and buoys) observations. The short-term forecasting of such aviation hazards has been studied for selected spring-time cases, as reported in Chan et al. (2021Chan et al. ( , 2022Chan et al. ( , 2023. In the present study, the forecasting of windshear/turbulence for tropical cyclone situations would be studied. Moreover, the previous studies focus on the two-runway system (2RS) of HKIA, with the north and south runways only. The third runway of HKIA started to be constructed in 2016, and has been put into operation since 8 July 2022. With limited number of pilot reports at the third runway, the windshear/turbulence for aircraft arriving at the new north runway of HKIA from the west, namely, 07LA(3RS), would also be studied for the first time. Furthermore, the existing windshear detection algorithms based at HKIA would be applied to the numerical weather prediction (NWP) model forecasting data. In the previous studies of spring-time cases, only the LIDAR-based windshear detection algorithm (called GLYGA, namely, Glide-path scan windshear detection algorithm) is considered. In the present study, the performance of microburst detection algorithms developed based on the concept of the 1st and 2nd Terminal Doppler Weather Radars (TDWR) of the Hong Kong Observatory (HKO) respectively would also be applied to model data. These three features are the novel in the windshear/ turbulence study for HKIA and have not been reported in the literature before.

| TROPICAL CYCLONES UNDER CONSIDERATION
Terrain-disrupted airflow at HKIA mostly occurs when tropical cyclones are located to the southwest of Hong Kong, bringing strong to gale force east to southeasterly winds to Hong Kong. High winds of such wind directions would be disrupted by the terrain of Lantau Island, bringing about high and low speed wind streaks at the airport area. This is the typical meteorological condition for terrain-induced windshear/turbulence at HKIA.
Three tropical cyclones are studied in the present article, namely, Lionrock (2021), Chaba (2022) and Mulan (2022). All of them are located to the southwest of Hong Kong for some periods of time and brought strong to gale force winds the region near to HKIA. Only the recent cases were selected when the third runway of HKIA was under construction/in use. Their tracks and intensity changes could be found in Figure 1.

| RUNWAY NOMENCLATURE AND METEOROLOGICAL EQUIPMENT
HKIA originally had two runways. The locations of the two runway system corridors, namely, 07LA, 07RA and 25RA for arrival, and 07RD and 25LD for departure, are F I G U R E 1 Three tropical cyclone cases under study in the present article.
F I G U R E 2 Meteorological equipment for HKIA and the runway nomenclature.
F I G U R E 3 Topography near to HKIA.
shown in Figure 2. The third runway was built further north of the two-runway system, forming a three-runway system (3RS) in HKIA. The new 07LA is labelled as 07LA (3RS) in this article and its location is shown in Figure 2. The topography near to HKIA is shown in Figure 3.
The windshear detection algorithms considered in this article are adopted from those used in the windshear detection meteorological instruments. GLYGA assumes that the equipment providing the data (namely, the LIght Detection And Ranging [LIDAR] in reality) is located near the centre of the respectively runway. There are three long-range LIDARs at HKIA, and their locations are given in Figure 2. In reality, the LIDAR uses infra-red laser beam to measure the winds along the glide paths. In this article, the NWP modelled data are outputted in the same format as the LIDAR data and then processed by the LIDAR's windshear detection algorithm, assuming that there is a meteorological instrument located near the centre of each runway to provide the complete set of headwind data along the glide path (which in reality is not the case due to measurement range limitation of laser beam especially in humid and/or rainy situations).
TDWR is used for microburst detection at HKIA. There are two TDWRs near HKIA, namely, the 1st one at Tai Lam Chung, and the 2nd one at Brothers' Point. The mission of TDWR is to detect the low-level windshear along the runway (i.e., the radar Doppler winds parallel to the runway) for protecting the taking-off and landing of aircrafts. Therefore, both TDWR stations have to be sited along the direction of the runway with an unobstructed view as shown in Figure 2. In this article, the NWP model data are used to represent the radial wind measurements from the 1st TDWR, that is, assuming that there is a meteorological instrument at the location of the 1st TDWR providing radial wind data over the airport area completely in all weather conditions (which in reality would not be the case due to measurement range limitation of microwave radar, especially in dry conditions). The two TDWRs use different microburst detection algorithms. They are labelled as microburst detection algorithm for 1st TDWR and microburst detection algorithm for 2nd TDWR. Both algorithms are applied to the meteorological instrument at the location of the 1st TDWR only, that is, the measurement geometry of the 2nd TDWR is not considered.

| NUMERICAL WEATHER PREDICTION MODEL
The NWP model used in the study is the same as that adopted in Chan et al. (2021Chan et al. ( , 2022Chan et al. ( , 2023, namely, Regional Atmospheric Modelling System (RAMS) version 6.3. It is nested from the NCEP global model at a spatial resolution of about 10 km. Dynamic downscaling is employed so that in the innermost domain, the spatial resolution is reduced to 40 m. The first model level is about 30 m above sea surface. The stretching coefficient for the grid space in the vertical direction is 1:1.08 (i.e., the ratio of the model level heights between the adjacent model levels). The simulation is essentially a large eddy simulation (LES) of terraindisrupted airflow at HKIA.
It is noted that the innermost domain is able to cover the whole alerting region of the two runway systems, namely, up to three nautical miles from the runway ends. However, the new 07LA(3RS) is only partially covered, up to about 2.5 nautical miles only. When more pilot reports of 07LA(3RS) are collected in the future, the coverage of this runway corridor would be extended in the future simulation studies.

| WINDSHEAR AND MICROBURST DETECTION ALGORITHMS
GLYGA assumes that the meteorological equipment near the centre of the runway is able to measure the headwind profiles along the runway corridors up to 3 nautical miles away from the runway end. The headwind profile is processed by an algorithm to detect automatically the significant headwind changes. A change of headwind of 10 knots or more is flagged and would be considered for the issuance of windshear alert based on the windshear severity factor. Multiple windshear may occur over a particular runway corridor. In particular, both headwind gains and losses could occur at the same time. Headwind loss is considered to have a priority higher than headwind gain. Details of the algorithm could be found in Shun and Chan (2008).
Microburst detection algorithms for the 1st and the 2nd TDWR are designed for TDWRs situated outside of the airport as radar cannot detect signals too close to its boresight in where radar main and side lobes are still forming. If the TDWR is placed in the airport area, special design on the algorithms is required to generate continuous data in range near the radar boresight.

| MICROBURST DETECTION ALGORITHM OF THE 1ST TDWR
The microburst detection algorithm of the 1st TDWR is radial-shear based. A microburst is a pool of intense cold air which bursts downwards and outspreads when hitting the ground, causing a significant outflow of airstreams with abrupt changes in wind direction and speed at the surface level. The basic principle of the 1st TDWR is to identify microburst by searching such abrupt wind changes associated with the surface outflows at the lowest PPI scan. There are four major steps in its microburst detection algorithm which are (1) identifying divergence region, (2) identifying surface outflow, (3) identifying microburst precursors and (4) generating microburst shapes and runway alerts (Tse et al., 2019).
i. Identifying divergence region: the radar searches windshear segments along its radial direction beamby-beam at the lowest PPI scan with a search window of several radar range gates. When a monotonic increasing of velocity is observed in consecutive range gates in the outwards direction, a windshear segment will be identified. The identified windshear segments will then be grouped and combined into different divergence regions according to their spatial distribution, for example, significant overlapping in range between two segments. The magnitude of the divergence region is determined by the maximum velocity difference among the combined segments. ii. Identifying surface outflow: the next is to check temporal correlation of the divergence regions to ensure persistence and reduce the false alarm rate. For those pass the temporal check will be identified as the surface outflow. iii. Identifying microburst precursors: Based on the conceptual model of microburst (mainly for thunderstorm-induced type), the radar detects features, such as descending reflectivity core (e.g., ≥57 dBZ) or descending velocity structures, like convergence/divergence aloft, as the microburst precursors. For those divergence regions do not pass the temporal check in Step 2, a second check will be made and those divergence regions could still be identified as the surface outflow if a spatial correlation with the microburst precursors is observed. iv. Generating microburst shapes and runway alerts: The final step is quality check of the surface outflows, for example, distance from a storm cell with reflectivity of at least 30 dBZ, and so forth. Microburst (headwind loss ≥30 knots) or windshear (headwind loss ≥15 and <30 knots) would be generated and runway alerts would be calculated on a pro-rata basis according to the overlapping of microburst/ windshear shape with the runway/arena.

| MICROBURST DETECTION ALGORITHM OF THE 2ND TDWR
The microburst detection algorithm for the 2nd TDWR is basically the same as step (1) identifying divergence region of the 1st TDWR. In order to identify the divergence region based on radial velocity data, there are four steps, (1) detect segment, (2) check validity of segments, (3) create potential features, and (4) check correlation.
Steps (1) and (2) are performed beam-by-beam at the lowest PPI scan.
i. Detect segment: This step searches for divergence segments with monotonic increasing of velocity along a radial beam by using a sliding window to determine start and end points of each divergence segment. ii. Check validity of segments: Each divergence segment identified in Step (1) is checked to ensure that its length is long enough, say 500 m; the moving average of velocity value is monotonic increasing along the segment; velocity value of the start/end points of the segment falls within a defined range of the median of velocity values of data points near the start/end points. iii. Create potential features: Segments in 1-dimension are then grouped into 2-dimension cluster based on their spatial relationship, that is, proximity in range and azimuth. Each cluster is then checked against other site-dependent parameters such as number of segments it contains, size, and the maximum value of velocity difference of the segments. Clusters which can pass the checking is considered as potential feature. However, those which cannot pass the checking are discarded. iv. Check correlation: The potential features are checked with other features (potential features, windshear or microburst feature) in previous two scans. If any part of the potential feature can overlap with any features in the previous two scans spatially, the potential feature is considered as a valid windshear or microburst feature which is categorised by the maximum value of velocity difference of the segments in the feature.
Each steps of the algorithm contains parameters which are configurable. For example, the 4th step was disabled in 2016. Runway alert is determined based on the microburst feature which the highest value of velocity difference over the runway.

| SIMULATION EXAMPLE AND COMPARISON WITH ACTUAL OBSERVATION
When east to southeasterly winds of significant wind strength in association with tropical cyclones prevail over the region of HKIA, high and low wind streaks may occur downstream of gaps and mountains respectively of Lantau Island. When the aircraft flies through these wind streaks, low-level windshear may be encountered. If the windshear has a headwind loss of 30 knots occurring with precipitation, microburst alert may be issued. Such microburst may be different from the conventional definition of microburst associated with thunderstorms.
With reasonable representation of mountainous terrain of Lantau Island, the NWP model has good skills in reproducing these wind streaks. An example of the wind streaks in the case of Lionrock (2021) is shown in Figure 4a. This is based on the 0.6-degree elevation plan position indicator (PPI) scan of the TDWR. The corresponding NWP-based radial velocity pattern is given in Figure 4b. It is readily shown that the wind streaks are simulated rather well, and the simulated radial velocity would form the basis for input into the microburst detection algorithms in the further studies.
The simulated radial velocity pattern in Figure 4b is considered to compare favourably with the actual observations in Figure 4a in the aspect of the successful simulation of the high and low wind speed streaks downstream of the valleys and hills respectively of Lantau Island. However, in general the simulated radial velocity is slightly smaller than that in the actual observation. This may be related to (i) difference in the wind direction in the simulation and the actual observation; slight difference of the direction could contribute to a change of the magnitude of the radial velocity with the respect to the position of the radar; and (ii) still rather insufficient representation of the wind speed profile of the atmospheric boundary/surface layer of the tropical cyclone in the numerical model, particularly about the height and magnitude of the low level jet, because of the still rather limited understanding of the boundary layer of the cyclone. However, given these limitations/uncertainties, the general pattern of the simulated radial velocity appears to be rather satisfactory, and we may try out the use of such data as input to the microburst/windshear detection algorithms, particularly when the change of the radial velocity is more important than the absolute magnitude of the velocity.
With the 3-components of the wind available from the numerical weather prediction model output, the simulated radial velocity of the LIDAR/TDWR is calculated by resolving the 3-D wind along the line of sight of the equipment, and this forms the basis, namely, the radial velocity, for inputting into the windshear/microburst detection algorithms for the generation of windshear/ microburst alerts. The numerical weather prediction model is run at a spatial resolution of 40 m. It has sufficiently fine horizontal resolution to resolve the complex terrain near the HKIA, which is the main factor for generating the mechanical turbulence. From previous studies, the vortices/waves associated with terrain-disrupted airflow was simulated very well, though the exact timing and location of the vortex/wave might not be reproduced in comparison with LIDAR/TDWR observations. The whole process is just dynamic downscaling of the global model to microscale without addition of more observational data. The impact of observational data on the microscale simulation is being studied and would be reported in the future papers.

| WINDSHEAR/TURBULENCE REPORTS AND ALGORITHM PERFORMANCE
When the aircraft encounters any windshear or turbulence, the pilot is encouraged to report the details of the event to HKO, which is called the 'Pilot Report'. Those reports of windshear/turbulence form the basis for the assessment of the various algorithms and is important for enhancing HKO's windshear and turbulence alerting service. Their details could be found in Table 1 for the six runway corridors under consideration. The reports are used to calculate the relative operating characteristics (ROC) curves of the three algorithms. ROC diagram is a presentation of the performance of the algorithm with the probability of detection against false alarm rate or the percentage of time on alert. The algorithm would have the best performance when the curve is closer to the upper left corner of the figure. The area under curve (AUC) is a quantitative indication of the performance of the algorithm. When AUC is closer to 1, its performance is better. The algorithm is found to have skills when the ROC curve is above the diagonal of the figure, with an AUC above 0.5. Both the arrival and the departure runway corridors are considered. The results are shown in Figure 5. As 07LA(3RS) just commenced operation in 2022, fewer pilot reports are available.
To assess the performance of the various algorithms, the area under curve (AUC) of each ROC curve is considered. All the ROC curves start from the origin at the lower left corner of the diagram. However, the algorithm may not be able to reach 100% probability of detection even by lowering the alerting threshold to very low value.
In the case, a horizontal line is drawn from the highest possible probability of detection to the far right vertical axis of the diagram, and the area under the resulting curve is calculated.
A summary of the AUC for the three algorithms at the six runway corridors is given in Table 2. In conclusion, if meteorological observations are 100% available under all weather conditions, GLYGA appears to be best algorithm based on the present results, with the area reaching 0.8 or above in all cases. For the two microburst detection algorithms, it is more difficult to draw conclusion as to which one is better. Given the present results, subject to site availability in the airport, maybe it is preferable to locate the meteorological equipment near the centre of the runway to measure the headwind profile, and to capture the windshear by directly measuring the headwind changes, in order to archive better result in windshear detection.

| EXAMPLES OF ALERTING
Two examples of windshear alerting are shown in more details. The first one occurs at around 17:33 UTC, 8 October 2021 over 07LA during Lionrock (2021). The GLYGA algorithm (Figure 6a) readily captures the windshear gain and loss. As in the real-life situations, the microburst detection algorithms give quite a number of microburst 'bandages' over the airport area (Figure 6b,c). But some 'bandages' manage to cover the 07LA runway corridor. As a result, in the present case study, both microburst detection algorithms are able to provide alerts. Another example is again Lionrock (2021), namely, around 03:03 UTC of 8 October 2021, over 07LA. The GLYGA algorithm is able to capture some headwind changes (Figure 7a). However, the microburst 'bandages' are limited to locations closer to the terrain of Lantau Island. As a result, both microburst detection algorithms are not able to provide alerts over 07LA. This is a typical case showing the better performance of GLYGA. Following the development of windshear alerting service at HKIA, microburst detection of the 1st TDWR was introduced first. This was followed by GLYGA, and then the microburst detection of the 2nd TDWR. This study aims at answering the question on which type of windshear detection algorithm and detection geometry is better, assuming that it is possible to provide 100% data coverage for the respective algorithm. This question cannot be answered with real observations, but an NWP-based simulation study may shed some lights. Based on three real cases of terraininduced windshear at HKIA in tropical cyclone situations, it appears that the detection capability and geometry from GLYGA is better. This may be related to the direct consideration of headwind along the runway corridor. In the microburst detection algorithms, the wind measuring equipment is further away from HKIA and the 0.6-degree PPI scan is not directly aligned with the runway corridor geometry. This may limit the performance of windshear detection to some extent. However, there are some practical limitations of setting up a microwave radar near the centre of the runway in order to perform the glide path scans as in the case of a LIDAR. Physical space constraint is one issue. The electromagnetic interference with the navigation signals is another important topic. This is a preliminary study with three tropical cyclone cases only. Also, in the period of COVID-19, the number of flights and thus pilot reports are very much limited. Further study with a larger dataset would be conducted in the future to compare the performance of the various windshear detection algorithms.