Dropsonde Data Impact on Rain Forecasts in Taiwan Under Southwesterly Flow Conditions with Observing System Simulation Experiments

Abstract

This paper conducts an observing system simulation experiment (OSSE) to assess the impact of assimilating traditional sounding and surface data, along with dropsonde observations over the northern South China Sea (SCS) on heavy rain forecasts in Taiwan. Utilizing the hybrid ensemble transform Kalman filter (ETKF) and the three-dimensional variational (3DVAR) data assimilation (DA) system, this study focuses on an extreme precipitation event near Taiwan on 22 May 2020. The event was mainly influenced by strong southwesterly flow associated with an eastward-moving southwest vortex (SWV) from South China to the north of Taiwan. A nature run (NR) serves as the basis, generating virtual observations for radiosonde, surface, and dropsonde data. Three experiments—NODA (no DA), CTL (traditional observation DA), and T5D24 (additional dropsonde DA)—are configured for comparative analyses. The NODA experiment shows premature and weaker precipitation events across all regions compared with NR. The CTL experiment improved upon NODA’s forecasting capabilities, albeit with delayed onset but prolonged precipitation duration, particularly noticeable in southern Taiwan. The inclusion of dropsonde DA in the T5D24 experiment further enhanced precipitation forecasting, aligning more closely with NR, particularly in southern Taiwan. Investigations of DA impact reveal that assimilating traditional observations significantly enhances the SWV structure and wind fields, as well as the location of frontal systems, with improvements persisting for 40 to 65 h. However, low-level moisture field enhancements are moderate, leading to insufficient precipitation forecasts in southern Taiwan. Additional dropsonde DA over the northern SCS further refines low-level moisture and wind fields over the northern SCS, as well as the occurrence of frontal systems, extending positive impacts beyond 35 h and thus improving the rain forecast.

1. Introduction

Taiwan is a subtropical island located on the southeastern side of the Eurasian continent with a unique climate regulated by the East Asian monsoon, including the cold-season northeast monsoon and the warm-season southwest monsoon. These monsoons significantly influence seasonal rainfall in Taiwan [1,2,3]. The mei-yu season in Taiwan usually occurs between mid-May and mid-June during the transition between the cold and warm seasons [4,5,6,7]. The predominant weather system during this period is mei-yu fronts which are usually characterized by significant moisture gradients on both sides and strong southwesterly winds on the southern side of the front [7,8]. The mei-yu frontal system not only contributes to precipitation; it also often interacts with other weather systems such as the low-level jet (LLJ), the southwesterly flow, and the marine boundary layer jet (MBLJ), leading to heavy precipitation in Taiwan [9,10,11,12]. A crucial role that the LLJ, southwesterly flow, and MBLJ commonly play in precipitation is in transporting warm, moist air from tropical oceans to the frontal regions, significantly influencing precipitation in Taiwan [8,11,13,14,15,16,17,18].

Studies investigating the relationship between southwesterly flow and precipitation in Taiwan suggest that precipitation intensity is correlated with the position of the southwesterly flow axis, with stronger precipitation occurring when the axis passes through southern Taiwan [11,16]. Furthermore, the low-level wind speed and humidity of the southwesterly flow upstream of Taiwan are the two most essential factors in determining intense precipitation events in southern Taiwan during a mei-yu season [19], and the absence of either factor makes such events less likely to occur [20,21]. Therefore, it can be inferred that sufficient wind and moisture data from observations around Taiwan are important for accurate precipitation forecasts during southwesterly flow events. However, due to Taiwan’s island location, observational data from the surrounding seas, particularly upstream of the southwesterly flow, the northern South China Sea (SCS), are not easy to obtain. When weather systems occur over the upstream ocean, the scarcity of observational data often leads to large forecast errors [22]. A possible solution to address this issue is the deployment of dropsondes, which can provide direct measurements of atmospheric pressure, temperature, dew point temperature, wind speed, and wind direction over this area. Many studies have shown that despite the relatively high cost, the high-quality data of dropsonde observations make them valuable for various purposes in many field projects [23,24,25]. Since the Dropwindsonde Observations for Typhoon Surveillance near the Taiwan Region (DOTSTAR) program, dropsonde data have been incorporated into the simulations of typhoon, resulting in improved typhoon track and precipitation forecasts in Taiwan [26,27]. This approach has also proven beneficial for other weather systems, with targeted observations enhancing tropical cyclone (TC) forecasts (10–40% improvement) more than forecasts for other weather systems (0–20% improvement) [28,29,30].

In addition, dropsondes have previously been used in data assimilation (DA) experiments for precipitation events associated with a southwesterly flow event in Taiwan in Chien and Chiu [31], which utilized observing system simulation experiments (OSSEs) to assess the impact of dropsonde DA on the forecast of a heavy rainfall event over Taiwan. This rain event was triggered by a decaying tropical cyclone that made landfall in southeastern China, which led to the formation of strong southwesterly flow over the northern SCS. The results indicated that dropsonde data can help the model simulations by creating better initial conditions through DA, and the minimum requirement for improving rainfall forecasts was to deploy 12 dropsonde observations at 225 km separation distance with a 12 h time interval over the northern SCS. While the OSSEs assessed the impact of dropsonde DA on heavy precipitation associated with the southwesterly flow event, the spatial distribution of the synthetic synoptic-scale surface and sounding observations over land was idealized and evenly spaced in that study (see Figure 3 of that paper). This idealized configuration of observations, which deviates from the actual locations of meteorological stations, may raise questions about the validity of the results. It thus requires further verification. Moreover, the OSSEs used a conventional ensemble transform Kalman filter (ETKF) DA system [32,33], which was shown to be inferior to the hybrid ETKF–three-dimensional variational (3DVAR) DA system in terms of analysis performance [34,35]. Therefore, based on the experimental design by Chien and Chiu [31], but using the new hybrid DA system along with the synthetic data that better align with the actual observational system, this paper aims to assess the DA impact on rainfall forecasts in Taiwan. Furthermore, we selected a heavy rainfall event that has been extensively studied regarding the factors contributing to heavy rainfall [21,36], which is advantageous for this study. The event occurred in late May 2020, mainly caused by a southwest vortex (SWV) [37,38,39] moving from South China to Taiwan. This current paper aims to understand how the dropsonde DA affects the performance of the hybrid ETKF-3DVAR DA system and the subsequent rainfall forecasts in Taiwan by answering the following questions:

• Can a precipitation case dominated by the southwesterly flow associated with an SWV achieve better rainfall forecasts through DA of traditional observations and extra dropsonde observations?
• What are the key factors contributing to the improvement in forecasting accuracy?
• How long and how far can the DA affect the forecasts?

In addition, this paper intends to extend the results based on the two southwesterly flow cases, one associated with a TC [31] and the other associated with the SWV of the current paper. The case description and experimental design are detailed in Section 2. Section 3 presents the precipitation forecast results. The DA impacts are discussed in Section 4. Section 5 offers a summary and conclusions.

2. Case Description and Experimental Design

In 2020, Taiwan experienced the highest accumulated precipitation during the first half of mei-yu season (15–31 May) in 42 years (1979 to 2020), marked by an extreme heavy rainfall event around 22 May that surpassed a 36 h accumulated rainfall record [36]. During this period, a stationary mei-yu frontal system was situated in the vicinity of Taiwan (Figure 1a), leading to atmospheric instability. Additionally, the front was associated with a long and wide cloud band extending from the southwest of Taiwan to the Indochina Peninsula, indicating abundant moisture and convection in the region (Figure 1b). The 850 hPa fields indicate the extension of an SWV from South China towards the northern side of Taiwan at 0000 UTC 22 May 2020 (Figure 1c), leading to the development of widespread and intense southwesterly flow on its southern side. As the SWV approached Taiwan, precipitation gradually increased from 21 May, reaching its peak around 1200 UTC 22 May before dissipating (Figure 2a). The precipitation occurred primarily on the windward side of the southwesterly flow, particularly in the southwestern region, with some mountainous areas experiencing rainfall intensities exceeding 300 mm in 12 h.

To examine whether a better representation of the upstream southwesterly flow through dropsonde DA can play a role in rain forecasts for Taiwan, an OSSE study was conducted for this heavy rain event. The first step in the OSSE was to obtain a nature run (NR; Figure 3a) that closely resembled the actual atmospheric conditions. To produce candidates for selection, a 48-member ensemble was first created at 1200 UTC 18 May 2020 based on the ETKF system. Then, the Weather Research and Forecasting (WRF) model version 3.8.1 [40] was run for 5.5 days (green line in Figure 3b) for each member. The initial and boundary conditions of these simulations were obtained from the ERA5 (the fifth generation of atmospheric reanalysis) data [41] of ECMWF (European Centre for Medium-Range Weather Forecasts) with a 0.25° × 0.25° resolution. The model included two nested domains with 15 and 3 km horizontal resolutions (green boxes in Figure 4) and 45 vertical levels, with two-way interactions between adjacent domains. Domain 2 started 2 days later than domain 1. The Yonsei University planetary boundary layer (PBL) scheme [42] was applied during the simulations. The 48 WRF ensemble member runs were repeated four times to produce a 192-member ensemble. These members had different combinations of microphysics schemes (MP) between the WRF double-moment 5-class and 6-class [43] and cumulus parameterization schemes (CPS) between Tiedtke [44] and new Tiedtke [45]. The CPS was not used in domain 2. The NR was chosen from the best member run among the 192 ensemble member simulations, as judged objectively by the ETS (equitable threat score) [46] and bias of rain simulations in Taiwan in domain 2 and, subjectively, by the synoptic weather pattern in domain 1.

The second step involved generating synthetic synoptic-scale sounding (red dots) and surface (purple dots in Figure 4) observations using data from domain 1 of the NR. Gaussian-distributed observational errors were added to the synthetic observations using the default observation error statistics from the obserr.txt file in the WRF-Var system, which provides specific error values for various observation types and variables. This approach, following the methodology outlined by Wang et al. [47], ensures that the synthetic observations more closely mimic the characteristics of real-world meteorological measurements, including their inherent uncertainties and imperfections. The locations of the above stations were based on the actual station positions, resembling more realistic conditions than the idealized settings in Chien and Chiu [31]. In addition, synthetic observations from 24 dropsondes over the northern SCS (green crosses in Figure 4), with a horizontal separation of 135 km, were generated via the same method. These dropsonde data include geopotential height, wind speed, wind direction, temperature, and dew point temperature at 45 pressure levels.

The third step included the DA process and the forecast of the three experimental runs, namely, NODA, CTL, and T5D24, each utilizing different numbers and intervals of OSSE data as shown in

Table 1. The design of the experimental runs, including their names and the data (sounding, surface, and dropsonde observations) that were assimilated during the ETKF DA process.

Experiment Names	Sounding Data/ Interval/Number	Surface Data/ Interval/Number	Dropsonde Data/ Interval/Number
NODA	no	no	no
CTRL	yes/12 h/197	yes/6 h/1408	no
T5D24	yes/12 h/197	yes/6 h/1408	yes/12 h/24

. During the spin-up (1200 UTC 18 May to 0000 UTC 19 May 2020) and the DA periods (0000 UTC 19 May to 0000 UTC 21 May 2020), the model was performed in domain 1 only, while in the forecast period (0000 UTC 21 May to 0000UTC 24 May 2020), all three domains were involved (Figure 3b). During the DA period, the Hybrid ETKF-3DVAR DA system [34,35] was employed nine times (eight cycles), with static/flow-dependent background error covariances weights of 25/75%, and horizontal ensemble covariance localization of 1125 km. The static background error covariance matrix was generated using the WRFDA gen_be tool via the NMC method with the CV5 option, and the localization scale was chosen based on prior research [34,35]. No data were assimilated in NODA, while 197 sounding and 1408 surface observation data were assimilated with 12 h and 6 h intervals, respectively, in CTL. Comparisons between the CTL and NODA experiments can examine the impact of traditional observations. In T5D24, dropsonde data over the aforementioned 24 locations, in addition to the sounding and surface observations of the CTL, were assimilated 5 times at a 12 h interval. Chien and Chiu [31] performed several other dropsonde experiments and suggest that 12 dropsondes with a 225 km separation distance over the northern SCS (T5D12) set a minimum requirement for enhancing the model regarding rainfall forecasts. However, in this study, we had to increase the number of dropsonde observations in the same area. The rationale behind this adjustment lies in the differences of the activity and position of the low-pressure systems between the previous case and the current one. In the previous case, the low-pressure system (tropical cyclone) was located to the north of Taiwan and moved from east to west during the DA period. The influence of dropsonde DA over the northern SCS can be transmitted downstream toward the low-pressure system by the southwesterly flow. Therefore, even with a limited number of dropsonde observations, structures of the low-pressure system can be significantly improved in the initial conditions, leading to improved precipitation forecasts. However, in the present case, the low-pressure system was located over South China and moved from west to east during the DA period. Improvements to the initial structure and the path of the low-pressure system can already be obtained in the experiment using traditional observation data (CTL). In order to achieve superior precipitation forecasts compared to the CTL, more dropsonde observations were required to further enhance the initial conditions over the northern SCS as suggested by a comparison study between the current T5D24 and T5D12. Therefore, this study opted for the T5D24 assimilation method, as Chien and Chiu [31] had also indicated its optimal effectiveness in adjusting initial conditions over the northern SCS. In addition, the locations of dropsonde deployment are consistent with the sensitive areas found in Chien and Chiu [36].

The National Centers for Environmental Prediction (NCEP) final reanalysis data (FNL) served as the initial and lateral boundary conditions for the forecast experiments. The temporal resolution of the FNL data was 6 h, with a horizontal resolution of 0.5° × 0.5°. The WRF model, version 4.1.2 [48], was used for the forecast experiments, with a three nested domain configuration (black boxes in Figure 4). The domain horizontal resolutions were 45, 15, and 3 km, with 45 eta levels in the vertical. A two-way interaction method between adjacent domains was implemented. The Goddard microphysics scheme [49] was chosen for microphysics process and the Yonsei University scheme [42] for the PBL process in all domains. The Tiedtke scheme [44] was utilized for CPS in domains 1 and 2. This physics combination, different from those of the NR, was selected to ensure continuity with previous research [31]. The reasons that different WRF versions with different model physics and grids were used for the NR and the experimental runs in this study are based on the general methodology of conducting an OSSE [50].

3. Precipitation Forecasts

This section presents comparisons of rainfall forecasts among the three experiments using NR as the truth. It is therefore necessary to make comparisons first between the simulation of NR and the observations.

3.1. Comparisons of NR with Observations

The observed precipitation from 21 to 23 May 2020 (Figure 2a) indicated heavy rainfall, with the strongest precipitation in the south and secondary intensity in the northwest and central mountainous areas of Taiwan. Although the NR (Figure 2b) exhibited slightly stronger precipitation than the observations in the southern mountainous regions and northern coastal areas during the three 12 h periods with stronger precipitation intensity from 0000 UTC 21 to 1200 UTC 22 May, the spatial correlation coefficients (SCC) of precipitation between the NR and the observations reached 0.69, 0.71, and 0.57, respectively (red numbers in Figure 2b). During other periods, although the SCC was lower, the root mean square error (RMSE) was also smaller. These findings suggest that the NR captured well the spatial and temporal distribution of precipitation and simulated rainfall that was comparable to the observations.

The observed sounding at Pingtung Airport (location denoted by the red dot in Figure 2a) revealed moist air from the surface to about 300 hPa at 0000 UTC 21 May 2020 (Figure 5a). Wind directions changed from southeasterly to southwesterly, gradually shifting to northwesterly with increasing wind speed. NR (Figure 5b) in general showed similar temperature and moisture patterns to the observed ones. Although winds were slightly stronger, the overall variations of wind speed and direction with height were similar to the observations. The time series of Pingtung Airport soundings displayed peak low-level southwesterly winds at about 0000 UTC 22 May (Figure 5c), with relative humidity in the middle and lower levels remaining close to saturation throughout. NR (Figure 5d) exhibited peak low-level wind speed around 1200 UTC 22 May, slightly later than observed, with a distribution of relative humidity similar to the observation. In general, the overall trends in wind speed and relative humidity were comparable with the observations. Since NR reproduced reasonably well the actual atmospheric conditions, especially the two most important factors affecting precipitation in Taiwan, i.e., the low-level moisture and southwesterly winds, this set of OSSE data was used for subsequent experiments.

3.2. Precipitation Verification

Time series of precipitation intensity of NR averaged over the entire domain (the TW region) in the leftmost column of Figure 2a show that the 72 h simulation can be divided into three time periods (Figure 6a). The first period (P1) from 0000 UTC to 1500 UTC 21 May 2020 represents the time before significant precipitation. The second period (P2) from 1500 UTC 21 May to 1500 UTC 22 May 2020 exhibits the strongest average precipitation intensity in TW, reaching a maximum of 10 mm h⁻¹. The third period (P3) from 1500 UTC 22 May to 0000 UTC 24 May 2020 shows a rapid decline in precipitation intensity, falling below 1 mm h⁻¹. For easy comparisons, the 12 h accumulated precipitation of Figure 2b is thus replotted for the three periods and displayed in Figure 7a. In P1, precipitation primarily occurs in the southern mountainous regions, with relatively weaker intensity in the central and northern parts. In P2, intense precipitation extends from the southern regions to the northwestern coastal areas, experiencing precipitation of up to 300 mm in 24 h. Over western Taiwan, the 24 h accumulated precipitation exceeds 90 mm. In P3, precipitation intensity rapidly diminishes across Taiwan, with only sporadic strong precipitation in isolated mountainous areas. If we delineate two regions with stronger precipitation intensity in NR (the southern/northern rain area, sRA and nRA, highlighted in green and red boxes in Figure 7a, respectively), it is evident that both sRA (Figure 6b) and nRA (Figure 6c) experience heavy precipitation during P2, with the most intense precipitation intensity occurring around 0000 UTC to 1200 UTC 22 May. The maximum rain intensity reaches up to 25 mm h⁻¹ in sRA and 15 mm h⁻¹ in nRA.

For NODA, precipitation intensity gradually increases in all three regions from the initial time until approximately 0600 UTC 22 May, after which it decreases and concludes around 1200 UTC on 22 May (Figure 6d–f). This suggests that the major precipitation event in NODA occurs and ends earlier than in NR. Additionally, NODA exhibits weaker precipitation intensity, with maximum intensities in each region approximately half of that in NR. Figure 7b shows that NODA concentrates precipitation in the central and southwestern mountainous areas extending toward the coastal regions during P1–P2. This rainfall pattern is similar to that of NR, but the precipitation intensity is insufficient. Furthermore, there is almost no precipitation in P3, indicating the poor precipitation forecasting capability of NODA.

The precipitation time series of CTL (Figure 6g–i) illustrates an improvement in the issue of premature precipitation forecasting seen in NODA. While the onset of major precipitation in CTL lags NR by approximately 6 h across all three regions, CTL exhibits a slightly larger precipitation intensity during P1 and a prolonged precipitation duration in P2 than NODA. These improvements are more pronounced in nRA (Figure 6i) than in sRA (Figure 6h), resulting in improved accumulated precipitation during P1 and P2 (Figure 7c), compared with those of NODA (Figure 7b). In P3, CTL produces a better precipitation pattern, but the rain amount is not enough in the north and too much in the south.

T5D24 (Figure 6j–l) shows a better onset of precipitation than CTL. The onsets of precipitation in all three regions are similar to those of NR. In sRA, there is a significant increase in precipitation intensity with prolonged duration (Figure 6k), closely resembling that of NR and suggesting that T5D24 further enhances precipitation in southern Taiwan. Figure 7d illustrates that T5D24 produces more precipitation than CTL during P1–P2. Although T5D24 slightly overestimates precipitation in several northern areas during P1, especially on the northeast coast, its precipitation is closer to NR than that of CTL in most areas. However, in P3, T5D24 also struggles to accurately predict precipitation. It is thus evident that dropsonde DA can effectively help the model in improving precipitation, but the effect can only last approximately 39 h (P1–P2). Overall, the observed precipitation intensity weakened during P3, and the forecasting abilities of the model in the three experiments show minimal differences. This paper therefore focuses mainly on the first two periods in the following discussions.

The threat scores (TS) for the ensemble mean precipitation verified against NR show that regardless of regions or periods, CTL yields TS greater than or close to those of NODA at all thresholds (Figure 8a–f). This result indicates that assimilating traditional sounding and surface observation data effectively enhances precipitation forecasting capabilities in the Taiwan region, with more substantial improvements observed in the 25–100 mm threshold range. The TS of T5D24, although exhibiting lower values than those of CTL for nRA during P1 (Figure 8e) and for TW at the 25–60 mm thresholds during P2 (Figure 8b), are overall greater than or equal to those of CTL in other periods and thresholds. Especially for thresholds exceeding 75 mm during P2 (Figure 8b,d,f), T5D24’s TS are significantly higher than those of CTL. In sRA (Figure 8c,d), T5D24 outperforms CTL at almost all thresholds. From these results, it can be inferred that assimilating sounding and surface observation data in the model can improve precipitation forecasts across Taiwan, and further inclusion of dropsonde data can enhance rain predictions, especially in sRA and for the heavy rainfall scenario.

To further evaluate the rain forecasts from all the 32 members of each experiment against NR, the performance diagrams [51] were plotted. Figure 9a reveals that during P1 at the 5 mm threshold, there is minimal difference in scores among experiments, indicating comparable forecasting capabilities at this threshold. At the 20 mm threshold (Figure 9b), NODA exhibits small bias scores (BS) across all regions, signifying underestimation of precipitation. The assimilation of sounding and surface observation data in CTL increases BS across all regions, reducing underestimation of precipitation. The increase in success ratio (SR) in CTL contributes to an increase in TS, representing a reduction in false alarms. Furthermore, the inclusion of dropsonde DA (T5D24) results in increased BS and probability of detection (POD) across all regions, with all TS approaching 1, suggesting a reduction in most instances of precipitation underestimation, making T5D24’s precipitation forecasts optimal. At the 50 mm threshold (Figure 9c), the scores in sRA and TW are similar to those of the 20 mm threshold. However, in nRA, while CTL continues to mitigate underestimation issues of NODA, T5D24 exhibits increased BS and decreased SR. These results indicate that T5D24 is slightly inferior to CTL in forecasting heavy rain in nRA owing to the overestimation problem. It is likely related to the fact that nRA is farther from the dropsonde locations than sRA, so the dropsonde DA may not provide much help in forecasting heavy rain there in early times (P1).

During P2, all regions exhibit a similar trend at the 90 mm threshold (Figure 9d); PODs become higher as more data are assimilated, indicating improved detection and forecasting capabilities (as shown in TS). At the 180 mm (Figure 9e) and 300 mm (Figure 9f) thresholds, the trends in all three regions are also similar; with more assimilated data, BS and TS increase, indicating that additional assimilation data reduces the underestimation of precipitation. The difference among the regions is that sRA shows the least decrease in SR and the most increase in POD, suggesting that the effect of increased precipitation on increased forecasting capability is most pronounced in sRA, followed by that in TW, and the least in nRA.

In summary, during P1, NODA exhibits precipitation underestimation at thresholds above 20 mm, which is mitigated by the inclusion of sounding and surface DA. The addition of dropsonde DA further improves this situation, with only minor overestimation in nRA at the 50 mm threshold. During P2, all three experiments show precipitation underestimation, with T5D24 demonstrating the best forecasting capability for heavy precipitation.

4. The Impact of DA

To understand the impact of DA and the factors contributing to the differences in precipitation forecasts, we compare the initial and forecasted fields among the three experiments in this section.

4.1. Initial Fields

The 850 hPa fields of NR (Figure 10a) show that at the initial time the center of the SWV is located approximately at 29.5° N, 119° E in South China, with the low-pressure area extending southwestward from the center. The region of abundant low-level moisture lies on the south side of the vortex, extending southwestward and southeastward towards Taiwan. The enhanced low-pressure zone associated with the SWV results in increased southwesterly winds extending from southwestern Taiwan to the northern SCS. In NODA (Figure 10b), compared to that of NR, the SWV’s center is shifted southward, and the low-pressure area is smaller, as is the region of abundant moisture. These results indicate suboptimal vortex structure in NODA’s initial fields, leading to weaker southwesterly winds over the SCS in NODA, adversely affecting precipitation in Taiwan. With the assimilation of sounding and surface observation data (Figure 10c), CTL produces a better vortex position and structure, with moisture and wind fields closer to NR in South China. However, there is still slightly less moisture than in NR over the northern SCS, especially in the waters surrounding Hainan Island, leading to the forecasts of smaller precipitation in southern Taiwan. Further inclusion of dropsonde data in T5D24 (Figure 10d) produces better moisture and wind fields over the northern SCS, improving the suboptimal initial fields of CTL. As a result, because the environmental field upstream of Taiwan in T5D24 is closest to NR, T5D24 yields the best precipitation forecasts.

At the 500 hPa level, NR (Figure S1a) shows a pressure trough extending southward to the west of Taiwan and an area of strong upward motion near the low-level SWV center. The trough is associated with prevailing westerly winds in South China, causing the low-level SWV to move eastward. In NODA, due to the southward shift of the low-level SWV center (see Figure 10b), the mid-level strong upward motion area is also southward (Figure S1b) compared to NR. Furthermore, the pressure trough near Taiwan is overestimated in NODA, causing northwesterly winds in the South China region. This condition leads to a path for the low-level SWV to move more southward compared to NR, resulting in poor precipitation forecasts in Taiwan. With the assimilation of sounding and surface observation data (Figure S1c), CTL noticeably fixes the issue of the pressure trough near Taiwan, with mid-level winds similar to those of NR, leading to a better forecast of the SWV movement. With further inclusion of dropsonde data, T5D24 (Figure S1d) shows minimal differences in mid-level initial fields compared to CTL, indicating that dropsonde data have not much effect on the SWV which is located in South China.

The RMSE of each member’s initial fields against NR is averaged over the South China (SC) and the northern SCS (nSCS) regions (red and green boxes in Figure 10a, respectively). Results show that NODA has larger RMSEs in the initial field, with the RMSEs in the SC region (Figure 11a–c) being greater than those in the nSCS region (Figure 11d–f), indicating that the environmental fields in the SC region contain more uncertainty owing to the presence of severe weather systems there, making the analysis of the initial field in this area more challenging. With the inclusion of sounding and surface observation data, however, the RMSEs in the SC region are effectively reduced across almost all variables and heights (Figure 11a–c), including moisture and wind fields. In the nSCS region, except for the low-level moisture (Figure 11d), the RMSEs of CTL are also slightly smaller than those of NODA (Figure 11d–f). This result suggests that the ability of improving the analysis of low-level moisture over the northern SCS is limited when assimilating only conventional observational data. However, when the extra dropsonde data over the nearby ocean are assimilated, T5D24 exhibits significantly smaller RMSEs in both the moisture and wind fields in the nSCS region than CTL (Figure 11d–f). In contrast, in the SC region (Figure 11a–c), the low-level moisture and wind fields for T5D24 do not exhibit reduced RMSEs compared to those of CTL. This discrepancy can be attributed to the inferior initial field structure in the southwestern part of the SWV in T5D24 compared to CTL (Figure 10c,d). However, the environmental fields in this area are likely to have a limited influence on precipitation over Taiwan. It is thus clear that the assimilation of traditional observation data effectively helps improve the environmental fields of the SWV in South China, leading to the improvement of the wind fields in the northern SCS. However, to further improve the moisture and wind fields there, the use of local dropsonde data is required. Additionally, dropsonde data can extend their influence toward the environmental fields in South China within some distance. Furthermore, Figure 11 clearly shows that the boxplots of T5D24 are generally narrower than those of NODA and CTL, especially over the nSCS region. This finding indicates that the variability among members in the T5D24 is significantly lower than in the NODA and CTL. The reduction in member variability suggests that the results of the T5D24 are more reliable, whereas NODA and CTL exhibit greater uncertainty among their members.

4.2. Forecast Fields

The low-level environmental fields of NR at the beginning of the major precipitation period (1500 UTC 21 May 2020; Figure 12a) show that the SWV is passing through the north side of Taiwan, with the low-pressure area exhibiting a northeast–southwest orientation. Under this condition, strong and moist southwesterly flows extend from the northern SCS to western Taiwan, causing subsequent precipitation in Taiwan during P2. In the ensemble mean of NODA (Figure 12b), due to the adverse effects of the suboptimal SWV structure at the initial time, the vortex has a north–south orientation at this time, resulting in significant differences in the wind field over the northern SCS compared to NR, leading to substantial differences in precipitation between NODA and NR. In the ensemble mean of CTL (Figure 12c), benefiting from the improved initial SWV structure over South China, the eastward movement of the SWV is better simulated, results in a much better structure of the vortex and wind fields at this time. However, due to the insufficient moisture in the northern SCS at the initial time, the forecasted moisture is also lower at the current time, causing the precipitation in CTL to remain below NR during P2. In the ensemble mean of T5D24 (Figure 12d), the improved forecasts of the moisture field over the northern SCS help improve the forecast of heavy rainfall intensity in P2.

The time series of the zonally averaged (119.8–120.2° E) 850 hPa geopotential height and horizontal wind (Figure 13) shows that in NR (Figure 13a), a pressure trough is moving southward over time near the center longitude of 120° E, with three lows approximately at 0080 UTC and 2000 UTC on the 21st and 0800 UTC on the 22nd, likely associated with semidiurnal pressure oscillation. Influenced by the significant differences in vortex structure between NODA and NR, as aforementioned, the ensemble mean of NODA (Figure 13b,e) exhibits large differences in latitude and time for the three lows compared to NR, with the second low stronger and deviating southward and the third low deviating northward. These improper pressure forecasts directly affect the wind forecasts, further resulting in inadequate precipitation forecasts of NODA. The ensemble means of both CTL (Figure 13c) and T5D24 (Figure 13d) produce better forecasts of the latitude and timing of the lows. Particularly, the southwesterly winds in the south during P1 and the geopotential height in the north during P2 show smaller differences compared to those of NR (Figure 13f,g). Their geopotential height and wind forecasts agree well with NR, indicating that correcting the forecast of the SWV over South China through assimilation of observation data can improve the pressure and wind forecasts upstream of Taiwan.

Further analyzing the time series of the 850 hPa mixing ratio and moisture flux shows that NR (Figure 14a) has a large mixing ratio at early times of P1, and after a short period (~6 h) of decreasing, moisture starts to increase significantly with abundant moisture flux on the south side of the SWV at the beginning of P2. This result is also suggested by the strong and moist southwesterly flow on the south side of the vortex (Figure 12a). As the pressure trough gradually pushes southward during P2, the location of the maximum moisture flux also shifts southward, reaching approximately 400 m s⁻¹ g kg⁻¹, providing favorable conditions for heavy precipitation upstream of Taiwan. Influenced by the deficient water vapor distribution and the misplaced SWV at the model initial time, the ensemble mean of NODA (Figure 14b,e) starts with less water vapor and moisture flux than NR, and it consequently produces inadequate moisture flux in the forecast, resulting in poor precipitation forecasts. The ensemble mean of CTL, also affected by the initial water vapor deficiency over the northern SCS (Figure 10c), has less water vapor to the south of 24° N at early times (Figure 14c,f) and less precipitation during P1. In P2, the water vapor and moisture flux are mainly affected by the eastward movement of the SWV. Since CTL produces better forecasts of the SWV, its forecasts of moisture flux during P2 are better than those of NODA. However, because CTL’s moisture flux is still insufficient, the precipitation intensity is slightly underestimated. The ensemble mean of T5D24 (Figure 14d,g) produces better water vapor and moisture flux at the initial time, resulting in improving precipitation forecasts in southwestern Taiwan during P1. During P2, the moisture flux of T5D24 is closest to NR such that T5D24 can correct the underestimation problem of CTL and produce the best rain forecasts among the experiments.

Next, we calculated the member mean RMSEs of moisture fluxes against NR at each grid point, using 32 ensemble members of each experiment. These moisture fluxes were averaged both vertically (from the surface to 700 hPa) and temporally (over a ±7 h window centered at each analysis time). The reason for this time window is because moisture flux is located upstream of the southwesterly flow. It takes a certain amount of time for the flux in this region to influence rainfall downstream in Taiwan. In other words, the forecast performance of the moisture flux over a period of time is crucial for accurately predicting rainfall in Taiwan during that period. The time window of ±7 h is thus selected, corresponding to the shortest time period (the P1) discussed in our study regarding rainfall. The skill scores (SS), defined as follows, were then computed to examine how much improvement CTL achieved with respect to NODA:

$${SS}_C=1-{\displaystyle \frac{{RMSE}_{CTL}}{{RMSE}_{NODA}}},$$

(1)

where ${RMSE}_{CTL}$ and ${RMSE}_{NODA}$ were the RMSE of CTL and NODA, respectively. A similar ${SS}_T$ was also computed for the improvement of T5D24 relative to CTL:

$${SS}_T=1-{\displaystyle \frac{{RMSE}_{T5D24}}{{RMSE}_{CTL}}},$$

(2)

where ${RMSE}_{T5D24}$ was the RMSE of T5D24. Figure 15a shows that at 1500 UTC 21 May 2020, the x-component of moisture flux (qu) exhibits predominantly large positive ${SS}_C$, with many areas exceeding the 95% significance level, over the coastal areas of South China extending to the ocean north and east of Taiwan and around Luzon Island as well. The y-component (qv; Figure 15b) of moisture flux has large positive ${SS}_C$ over South China near the SWV location, the northern SCS, and the ocean southeast to Taiwan. The ${SS}_T$ of qu show positive values over almost the entire northern SCS region (Figure 15c), but only a few of them exceed the 95% significance level. The most important positive ${SS}_T$ region of this plot is the southwestern part of the northern SCS, which is the upstream of Taiwan’s rainfall. Similarly, the ${SS}_T$ of qv (Figure 15d) have a similar pattern over the upstream. These results suggest that the DA of traditional observations significantly helps the simulation of low-level moisture fluxes at the beginning of major precipitation, especially over the upstream area of Taiwan. With additional dropsonde DA, the forecast of low-level moisture flux can be further improved, but the improvement is moderate.

Lastly, we calculated the areal mean RMSEs of vertically and temporally averaged moisture fluxes for each ensemble member in the three experiments against NR over the SC and nSCS regions at each full hour. ${SS}_C$ was then computed for each member of CTL with respect to each member of NODA, resulting in 1024 combinations. Furthermore, because the computation was conducted at each full hour, P1 overall encompassed 8192 samples (8 h), P2 had 24576 samples (24 h), and P3 contained 27648 samples (27 h). Since the moisture fluxes over the Taiwan Strait are more critical to rainfall in Taiwan, the RMSEs and SS over this area (the TWS region) were also included in the analysis. The above processes were also adopted for ${SS}_T$. Over the SC region, the ${SS}_C$ are high in P1, with a median of 0.38 for qu (Figure S2a) and 0.24 for qv (Figure S2b). The overall ${SS}_C$ decrease over time and become nearly evenly distributed between positive and negative values in P3. The ${SS}_T$ median values remain close to zero for all three periods. This result suggests that the assimilation of sounding and surface observations helps the forecasts of low-level moisture flux over the SC region by a factor of 0.1–0.3 on average, with the improvement on qu being more significant than on qv. However, the positive effect in general decreases over time and can only last about 40 h. When the additional dropsondes are assimilated in T5D24, there is minimal difference in SS.

The scenario is slightly different over the nSCS region. The ${SS}_C$ for qv (Figure S2d) are, in general, larger than those of qu (Figure S2c), and the positive effect of DA can last longer. The ${SS}_C$ of qv in P2 are even higher than in the other two periods. This result leads to the improvement of rainfall forecast in P2 for CTL and T5D24 because the improvement in qv can help produce better moisture transport by the southwesterly flow upstream of Taiwan. The ${SS}_T$ of qu and qv gradually decrease from P1 to P3. Over an area even closer to Taiwan, the TWS region exhibits quite high ${SS}_C$ in qu among the three regions. The improvement of RMSEs for CTL with respect to NODA can, on average, reach a factor of 0.37 in P1 (Figure S2e). The ${SS}_C$ of qv are also relatively high for P1 through P3 (Figure S2f). The above results indicate that the assimilation of extra dropsonde observations generally can provide more assistance for qu than for qv over the northern SCS region. It is therefore concluded that the dropsonde observations over the northern SCS can have a great impact on the forecast of moisture transport upstream of Taiwan and consequently help the rain forecast in Taiwan.

The duration of positive impact obtained from DA can be more easily evaluated via Figure 16, which presents the median of SS at each full hour of the model simulation. To obtain a fair evaluation, the duration is defined as the lasting hours during which the median SS of both qu and qv maintain a positive impact continuously from the initial time. Since the SS may fluctuate between consecutive hours, the above judgement is made based on an 11 h running mean. Furthermore, the first 12 h are excluded because fluctuations may also result from the spin-up problem.

The duration of positive impact from the DA of traditional observations (${SS}_C$) is approximately 40 h over the SC region (Figure 16a,b) and approximately 43 h over the TWS region (Figure 16e,f). In contrast, ${SS}_T$ in these two regions remains close to zero and does not pass the significance level throughout the forecast period, indicating that dropsonde data provide virtually no improvement to moisture transport in the SC and TWS regions (Figure 16a,b,e,f). However, over the nSCS region (Figure 16c,d), the durations of positive ${SS}_C$ become longer (65 h), and those of ${SS}_T$ can last approximately 35 h. The longer durations of ${SS}_C$ may be related to the fact that this area is farther away from the SWV center, leading to less uncertainty in the analysis.

4.3. Forecasts of the Front

Previous studies [21,36] have documented that precipitation of this case is not only affected by the transport of moisture but also the frontal activity. That is, precipitation is more intense when the front is closer to southern Taiwan (e.g., sRA). It is therefore necessary to analyze frontal activity in NR and the three experiments. The 1000 hPa fields show that NR exhibits a linear narrow convergence zone (Figure S3a) with large virtual potential temperature gradient (Figure S3b) over the northern Taiwan Strait (around 25° N) at 0000 UTC 22 May 2020, indicating the location of the front. Subsequently, the front moves southward into the Taiwan Strait. By 0700 UTC 22 May (Figure S3c,d), the front has moved to the vicinity of sRA, causing an increase in precipitation intensity in sRA (Figure 6b). At 1500 UTC 22 May (Figure S3e,f), the front moved further south, away from sRA, leading to a rapid weakening of precipitation intensity. The averaged latitude of the front at each hour from 0000 UTC to 1500 UTC 22 May (thick black line in Figure 17) further shows that the front in NR moves southward at a roughly consistent speed while passing through the Taiwan Strait, with only a few instances of slightly retreat. These results confirm the close relationship between the frontal activity and precipitation intensity in southern Taiwan.

To understand the impact of DA on frontal activity, we investigated and compared the frontal locations among the experiments. The frontal detection method by Chien and Chiu [21,36] was employed to locate the frontal latitude (Lat_F) of each member in the ensemble forecasts. This method is briefly described in the following. More detailed information can be found in Chien and Chiu [21,36]. The method first searches for regions of strong horizontal wind convergence ($\nabla ·V$ < −7.1 × 10⁻⁴ s⁻¹) and large virtual potential temperature gradient ($\left|{\nabla \theta }_v\right|$ > 0.14 K km⁻¹) at 1000 hPa. If a grid point simultaneously satisfies these two conditions, it is identified as a frontal point. These frontal points are then verified for continuity with sufficient length to resemble a realistic front, and the averaged latitude, Lat_F, is noted. If not, the sample is registered as no front. The detection is conducted over the sea area between 118–120° E and 21–25° N (the red box in Figure 10d) from 0000 UTC to 1500 UTC 22 May 2020. Each experiment comprises 512 samples (16 h × 32 members).

The time–latitude plot of Lat_F reveals that the frontal speeds in NODA are mostly too fast after 0300 UTC 22 May (Figure 17, gray boxplots). Consequently, at 0700 UTC 22 May, the area favorable for frontal activity is mainly located south to the Taiwan Strait (Figure S4a,b). As a result, precipitation is too weak and terminates too early in NODA (Figure 6d–f). Fronts in CTL are better simulated than in NODA (Figure 17, red boxplots). However, their southward-moving speeds are, in general, too slow. At 0700 UTC 22 May (Figure S4c,d), the area favorable for frontal formation, particularly the large virtual potential temperature gradient region, remains over the northern Taiwan Strait (Figure S4d), leading to delayed precipitation events in CTL (Figure 6h). The mean frontal positions in T5D24 (Figure 17, blue boxplots) are mostly closer to that of NR. The average frontal speed is similar to that of NR. At 0700 UTC 22 May, the area favorable for frontal activity is near sRA, approximately around 23.5° N (Figure S4e,f), supporting sustained precipitation and stronger precipitation intensity in T5D24 (Figure 6k). The RMSEs of the hourly ensemble mean Lat_F were further calculated against that of NR during the 16 h. They are 1.41, 0.61, and 0.58 degrees for NODA, CTL, and T5D24, respectively. Moreover, the total amounts of samples in Figure 17 are 138, 135, and 215, out of the possible 512 samples, for NODA, CTL, and T5D24, respectively. With more samples identified as the case of having a front, T5D24 has the best forecasts of frontal activity in the simulations. These results indicate that the assimilation of traditional observational data can improve the forecasts of frontal locations and precipitation. Further improvement in the occurrence of frontal systems and precipitation can be obtained when additional dropsonde data are assimilated.

5. Summary and Conclusions

This paper presents an observing system simulation experiment (OSSE) aimed at evaluating the impact of assimilating traditional sounding and surface data and extra dropsonde observations in the northern SCS on Taiwan’s rainfall forecasts, using the hybrid ensemble transform Kalman filter (ETKF) and the three-dimensional variational (3DVAR) data assimilation (DA) system. The selected case is an extreme precipitation event that occurred in Taiwan around 22 May 2020, with the influence of a southwest vortex (SWV) that has been well examined in previous studies. Virtual observations for conventional radiosonde, surface, and dropsonde data were generated based on a nature run (NR) that closely resembled real atmospheric conditions. Subsequently, a 32-member WRF ensemble was created, employing the Hybrid ETKF-3DVAR DA system that assimilated the NR data for 48 h and ran for 72 h. To determine the impact of these assimilated data, three experiments were configured: NODA did not assimilate observation data; CTL assimilated traditional radiosonde observations at 12 h intervals and conventional surface observations at 6 h intervals; and T5D24 assimilated 24 extra dropsondes as well as the traditional observations at 12 h intervals. Comparisons between CTL and NODA reveal the impact of traditional sounding and surface observations, and those between T5D24 and CTL identify the impact of dropsonde observations.

The heavy rainfall event indicated three significant periods: a pre-precipitation phase (P1), a peak precipitation phase (P2), and a declining precipitation phase (P3). Intense rainfall characterized P2, extending from southern regions to northwestern coastal areas, with peak intensities exceeding 300 mm in 24 h. However, precipitation intensity rapidly diminished in P3, with sporadic rainfall in isolated mountainous areas. The NODA experiment shows premature and weaker precipitation events across all regions compared with NR. The CTL experiment improved upon NODA’s forecasting capabilities, albeit with delayed onset but prolonged precipitation duration, particularly noticeable in southern Taiwan. The inclusion of dropsonde DA in the T5D24 experiment further enhanced precipitation forecasting, aligning more closely with NR, particularly in southern Taiwan. Threat score analyses underscored the effectiveness of assimilating traditional sounding and surface observation data, with CTL exhibiting notable improvements over NODA, especially in moderate to heavy precipitation thresholds. T5D24 outperformed CTL in most cases, particularly in southern Taiwan and during heavy rainfall scenarios, albeit with occasional overestimations. Performance diagrams revealed that T5D24 demonstrated optimal forecasting capabilities, mitigating underestimation issues and reducing false alarms, particularly evident during heavy rainfall events.

The comparisons of environmental fields between CTL and NODA demonstrate that assimilating traditional radiosonde and surface observations significantly improves the SWV structure over South China at the initial time, leading to notable improvements in the wind fields in South China and some improvements in the northern SCS. However, the moisture field shows no improvement in the northern SCS. The improved SWV structure and wind fields at the initial time result in better predictions of the SWV locations and the associated southwesterly winds. The improvement can last for about 40 h in South China and over 65 h in the northern SCS. Additionally, the location of the front over the Taiwan Strait is better forecasted during the latter part of P2. Such improvement plays a significant role in precipitation forecasts, especially for heavy rainfall events. Despite the improvements, the forecasted precipitation intensity of CTL remains insufficient in southern Taiwan due to the initial shortage of moisture in the northern SCS. Further assimilating data from 24 dropsondes over the northern SCS with a temporal interval of 12 h leads to additional improvements in the low-level moisture and wind fields in the initial field over the northern SCS. Consequently, the forecast of low-level moisture transport in the southwesterly flow shows improvements, with the extra positive effects lasting beyond 35 h. Furthermore, the occurrence of the front over the Taiwan Strait are also better forecasted. As a result, the forecasting capability for heavy precipitation in southern Taiwan and other regions is also improved.

In conclusion, the distributions of low-level wind and moisture fields in the upstream region significantly influence precipitation in Taiwan during southwesterly flow events. A forecasting system must address both aspects for better precipitation forecasts. The results of this current case indicate that when Taiwan is affected by a dominant weather system over the South China region, assimilating traditional observational data alone can effectively help the wind forecasts near Taiwan. However, to improve precipitation forecasts, local observation data from dropsondes are essential to adjust the initial moisture field over the upstream ocean. These conclusions are consistent with those of the previous study [31], although the main causes of the southwesterly flow of that event were different to the current one. In addition, the current paper applies superior model configuration with observations at actual station locations and presents more advanced statistical analyses to better support the conclusions. Furthermore, the duration analysis reveals that dropsonde DA extends the positive impact window, particularly over regions critical to precipitation, offering a valuable contribution to the advancement of forecasting capabilities. These findings advocate for the continued exploration and integration of advanced DA methods, emphasizing the utility of dropsonde observations for improved precipitation predictions in areas with similar meteorological conditions.

This study has an important limitation worth noting, namely, the exclusion of satellite radiance DA. Satellite radiance offers observations over large areas, particularly in regions such as oceans, where traditional observation networks have sparse coverage. The absence of radiance DA may have led to an overestimation of the dropsonde impact in this study because if radiance data were incorporated into the DA system, they could partially offset the contribution from dropsonde data. Given these considerations, future studies should incorporate satellite radiance DA into their experimental designs. This would not only allow for a more accurate assessment of the relative contribution of dropsonde observations but also facilitate a comprehensive evaluation of the combined effects of different types of observations in improving rainfall forecasts for Taiwan.