Impact of YunYao GNSS-RO Refractivity Data Assimilation on Typhoon Forecasts: A Case Study of Typhoon BEBINCA (2024)

Abstract

The accuracy of numerical weather prediction largely depends on the quality of the initial conditions. Global Navigation Satellite System radio occultation (GNSS-RO) observations, with their high vertical resolution, play an important role in reducing initial condition errors. In this study, multiple simulations with different initialization times were conducted during the development of Typhoon BEBINCA using the WRF-GSI assimilation system to evaluate the impact of YunYao GNSS-RO observations on improving extreme weather simulation performance and to investigate the sensitivity of refractivity assimilation to different cloud microphysics parameterization schemes. The results show that assimilating YunYao GNSS-RO data significantly improves the consistency between the model initial fields and observations and enhances the analysis quality in the middle and upper troposphere. Compared with ERA5 reanalysis data, the assimilation experiments better reproduce the spatial and temporal evolution of key atmospheric variables, and the improvements persist from 36 h to 120 h forecast lead time. Statistical results from multiple initializations show that the maximum RMSE reductions exceed 0.2 K for temperature, 0.1 m s⁻¹ for wind speed, and geopotential height shows consistent improvements throughout the entire atmosphere. In addition, the assimilation experiments improve the simulation of Typhoon BEBINCA’s track and intensity. Statistical results from multiple initializations indicate that the 84 h track error is reduced by approximately 30 km on average, and the minimum central pressure bias is also reduced. Sensitivity experiments further show that the WSM6 microphysics scheme performs better in track forecasting, while the Thompson scheme is more suitable for intensity forecasting. Overall, YunYao GNSS-RO assimilation effectively improves typhoon forecast accuracy and demonstrates strong potential for operational applications.

1. Introduction

Since the COSMIC mission, Global Navigation Satellite System Radio Occultation (GNSS-RO) observations, with their unique advantages of high accuracy, high vertical resolution, all-weather capability, and negligible systematic bias, are capable of providing globally distributed thermodynamic profile observations [1,2,3] and have become an important reference for numerical weather prediction (NWP) and climate models [1,4]. When occultation observations are assimilated into NWP systems, bending angle or refractivity is typically used as the observed variable [5,6], which requires the development of two types of observation operators to map model variables into observation space. The refractivity observation operator converts model variables—temperature, pressure, and specific humidity—into atmospheric refractivity based on the Smith–Weintraub formula and then interpolates the model fields to the geometric heights of the ray tangent points. The bending angle observation operator further calculates the corresponding bending angles using the Abel integral under the assumption of spherical symmetry [7].

A large number of studies have demonstrated that the assimilation of GNSS-RO observations can significantly improve atmospheric initial analyses and subsequent forecast accuracy across multiple temporal and spatial scales [5,8,9,10,11,12]. Major operational NWP centers, including the National Centers for Environmental Prediction (NCEP), the European Centre for Medium-Range Weather Forecasts (ECMWF), the UK Met Office, and the Japan Meteorological Agency (JMA), have incorporated refractivity or bending angle observations as routine data sources, and long-term operational applications have confirmed their stable positive impact on analysis quality and forecast performance [8]. Such assimilation can systematically improve the representation of key thermodynamic variables, including temperature, geopotential height, and humidity, with particularly pronounced impacts in the upper troposphere and lower stratosphere [9,13,14], and can reduce 48 h temperature forecast errors by approximately 10–20% in the 300–50 hPa layer over the Southern Hemisphere [9]. In addition, multiple studies have shown that GNSS-RO assimilation also plays an important role in improving the prediction of regional severe convective weather and extreme weather events. For example, assimilating FY-3C GNOS refractivity observations into the China Meteorological Administration’s global/regional assimilation and prediction system (GRAPES) improves the 24 h precipitation threat score (TS) of heavy rainfall events over the Yangtze-Huai River basin by approximately 4–7% [15]. Recent evaluations based on the Hurricane Weather Research and Forecasting (HWRF) model indicate that assimilating COSMIC-2 observations can reduce the mean absolute error (MAE) of minimum sea-level pressure by approximately 8–12% at key forecast lead times, while improving tropical cyclone intensity forecasts by about 10% [11].

The YunYao GNSS-RO constellation currently represents China’s largest commercial meteorological satellite program and is also one of the world’s leading radio occultation systems in terms of data production capacity. The constellation is planned to comprise 90 satellites, including 72 in Sun-synchronous orbit and 18 in low-inclination orbit, of which 46 had been successfully launched by March 2025. Observing System Simulation Experiments (OSSEs) indicate that assimilating a daily total of 1.2 × 10⁵ radio occultation profiles can reduce geopotential height errors at 100 hPa by up to 50%, and as the number of GNSS-RO observations increases, NWP performance is expected to continue improving [16,17]. By 2024, twelve YunYao satellites had been integrated into the China Meteorological Administration’s comprehensive observation system, providing core data support for numerical weather prediction, space weather monitoring, and marine environment applications.

In this study, the Weather Research and Forecasting (WRF) model, coupled with the Gridpoint Statistical Interpolation (GSI) variational data assimilation system, is employed to investigate the effects of YunYao GNSS-RO refractivity assimilation. A multi-faceted assessment is performed, targeting both the refinement of meteorological element accuracy and the enhancement of typhoon prediction. Moreover, the study is extended through sensitivity experiments on physical parameterizations to identify the optimal model configurations for effective data assimilation.

2. Materials and Methods

2.1. Model Description

The WRF model, developed jointly by the NCEP, the National Center for Atmospheric Research (NCAR) and other institutions, is a mesoscale numerical weather prediction and non-hydrostatic atmospheric simulation system [18]. Its fully compressible governing equations and advanced numerical schemes enable efficient simulation of complex atmospheric processes from kilometer-scale to regional-scale. The model can be coupled with various assimilation systems, allowing observational data to be incorporated to optimize the initial conditions and thereby reduce forecast errors caused by initial condition uncertainties [19]. Accordingly, in this study, experiments were carried out by assimilating YunYao GNSS-RO refractivity observations, using the initial and boundary conditions generated by the WRF model coupled with an assimilation system, to analyze the impact of refractivity observation assimilation on typhoon forecasts.

The GSI data assimilation system, developed under the leadership of the NCEP, is a third-generation variational assimilation system that seeks the optimal analysis by minimizing a cost function constrained by both the background field and observations [20]. Unlike traditional spectral methods, GSI performs the analysis directly at model grid points, avoiding computational errors associated with spectral transformations, which makes it particularly suitable for mesoscale numerical weather prediction applications [21]. In this study, the GSI assimilation system was employed to assimilate YunYao GNSS-RO observations. The system establishes the physical relationship between refractivity and the model state variables—temperature, pressure, and specific humidity—through advanced observation operators, ultimately producing an optimized initial analysis field that provides improved initial conditions for subsequent numerical forecasts.

2.2. Experimental Design

Typhoon BEBINCA, which made landfall in Shanghai, China, at 23:30 UTC on 15 September 2024, is selected as the target case in this study. Numerical simulations are conducted using the WRF model 4.3. The model is configured with a single domain at a horizontal resolution of 27 km. The spatial extent of this domain corresponds to the area shown in Figure 1a, which encompasses most of the life cycle of Typhoon BEBINCA from genesis to landfall. The model top is set at 1 hPa with 56 vertical levels. Physical parameterization schemes employed in the simulations include the WRF Single-Moment 6-class (WSM6) scheme for cloud microphysics [22], the Kain–Fritsch (KF) scheme for cumulus convection [23], the Rapid Radiative Transfer Model (RRTM) for longwave radiation [24], the Dudhia scheme for shortwave radiation [25], the Yonsei University (YSU) scheme for the planetary boundary layer [26], the MM5 Monin-Obukhov scheme for the surface layer [27], and the Noah Land Surface Model (Noah LSM) for land surface processes [28] (

Table 1. Configuration of physical process parameterization schemes in the WRF model.

Physical Process	Parameterization Scheme
cloud microphysics	WSM6
Cumulus Convection	KF
Longwave Radiation	RRTM
Shortwave Radiation	Dudhia
Planetary Boundary Layer	YSU
Surface Layer	MM5 M-O
Land Surface Processes	Noah LSM

).

Regarding the assimilation experiment design, four synoptic times—00:00, 06:00, 12:00, and 18:00 UTC—on both 13 and 14 September 2024 are selected as independent initialization times. A total of eight sets of assimilation (DA) and no-assimilation (NODA) experiments are conducted. To statistically evaluate the assimilation impacts, each experiment is integrated for a 120 h forecast from its initialization time. The NODA experiments are driven by the Global Forecast System (GFS) analysis fields as the initial fields and the DA experiments utilize the GSI system with the three-dimensional variational method to update the initial fields at the initialization time. The assimilation time window is set to ±3 h centered on the analysis time. The background error covariance in the assimilation system is derived using the NMC method [29], which computes the statistical differences between forecasts valid at the same time but initialized at different times, whereas the observation errors are prescribed using the default error model of the GSI system [21]. It should be emphasized that the NCEP GFS analysis field (f000) used as the background in this study is an operational product that has already assimilated a comprehensive suite of conventional observations (e.g., radiosondes, surface stations) and satellite measurements (e.g., radiances, operational GNSS-RO). Therefore, the NODA experiments represent the forecast initialized from a robust, multi-platform observing system. Consequently, the differences between the DA and NODA experiments explicitly demonstrate the incremental added value of the YunYao GNSS-RO data assimilation on top of a realistic operational environment, particularly its impact on meteorological variables and typhoon track predictions.

Considering the limitations of the model top height and vertical resolution in regional models, it is difficult to meet the high-altitude accuracy requirements for bending angle retrieval. And the derivation of bending angles relies on the Abel integral, which tends to introduce substantial errors in the upper levels. Moreover, its path-integral characteristics complicate error propagation, hindering stable assimilation within regional models. Consequently, existing studies on assimilating GNSS-RO observations into regional numerical models [15,30,31,32] mostly adopt refractivity as the assimilation variable. Therefore, to enhance the stability of the assimilation process and its applicability within regional numerical models, refractivity from the YunYao GNSS-RO observations is utilized as the assimilation variable in this study.

Additionally, to further investigate the impacts of physical parameterizations on assimilation performance, sensitivity experiments focusing on cloud microphysics schemes are conducted using one specific set of experiments (initialized at 00:00 UTC 14 September 2024) as a representative case. Two widely used cloud microphysics schemes were selected for comparison: The WSM6 scheme [22] is a single-moment scheme that can simultaneously simulate six types of hydrometeors, including water vapor, cloud water, rainwater, ice crystals, snow, and graupel. It has a clear structure, is widely applied, and is suitable for simulating large-scale precipitation processes. The Thompson scheme [33], on the other hand, features more complex ice-phase processes and particle size distribution prediction, describes microphysical processes in greater detail, and is also widely used in typhoon and convective weather simulations. These two schemes have their own characteristics in terms of theoretical assumptions and parameterization processing. Incorporating both into the experiment will help evaluate the comprehensive impact of cloud microphysical processes on assimilation effects and typhoon forecasts.

Therefore, as shown in

Table 2. Sensitivity Experiment Design for GNSS-RO Refractivity Assimilation and Cloud Microphysics Schemes.

Experiment Setup	WSM6_NODA	WSM6_DA	THOM_NODA	THOM_DA
Assimilation of YunYao GNSS-RO Refractivity	NO	YES	NO	YES
Cloud Microphysics Scheme	WSM6	WSM6	Thompson	Thompson

, four sensitivity experiments were designed as follows:

(1)

WSM6_NODA: The WSM6 cloud microphysics scheme is adopted, utilizing background field data provided by the GFS with a horizontal resolution of 0.25° and a time interval of 6 h, used as initial and boundary conditions. These data are interpolated onto the simulation domain by the Weather Research and Forecasting Preprocessing System (WPS) and WRF, with no data assimilation involved.

(2)

WSM6_DA: Based on the WSM6_NODA configuration, YunYao GNSS-RO refractivity observations were assimilated at the forecast initial time using the GSI three-dimensional variational assimilation system.

(3)

THOM_NODA: The Thompson cloud microphysics scheme is adopted, with all other settings completely consistent with the WSM6_NODA experiment. By comparing with WSM6_NODA, the differences in forecasts under different cloud microphysics schemes without assimilation can be evaluated.

(4)

THOM_DA: Based on the THOM_NODA configuration, YunYao GNSS-RO refractivity observations are assimilated using the GSI system, with the assimilation settings consistent with those in WSM6_DA. By comparing with WSM6_DA, it is possible to further evaluate the differences in forecasting effects of cloud microphysical schemes after assimilating YunYao GNSS-RO refractivity data.

2.3. Observational and Reanalysis Data

The observational data used in this study were obtained from the YunYao GNSS-RO constellation. Through networked observation, the constellation achieves full coverage observation of the global atmosphere and ionosphere, generating more than 300,000 radio occultation profile data daily, which provides sufficient and high-quality occultation observation data support for the experiment.

The experiment selected the fifth-generation reanalysis data (ERA5) [34] from the ECMWF as the objective reference field to evaluate the simulation performance of meteorological elements in the sensitivity experiments. ERA5 data has high temporal and spatial resolution (1 h, 0.25° × 0.25°), covers the global range, and provides multi-level atmospheric state variables, including key meteorological elements such as temperature, specific humidity, geopotential height, wind speed, and surface pressure. It is widely used in numerical model verification and climate diagnostic analysis. Based on ERA5, the root-mean-square error (RMSE) for all the DA and NODA experiments are calculated to quantitatively characterize the model simulation errors. Furthermore, the paired t-test [35] is applied to evaluate the significance of the differences in these metrics between the two sets of experiments, thereby determining the statistical reliability of the assimilation improvements.

In addition, the Best Track Dataset for Tropical Cyclones from the China Meteorological Administration (CMA) [36] was used to verify the model’s ability to simulate typhoon tracks and intensities. This dataset is a widely recognized and highly authoritative source of data for tropical cyclone research in the Northwest Pacific (including the South China Sea, the sea areas north of the equator and west of 180° E). It systematically contains records of the position and intensity of all tropical cyclones in this region every 6 h since 1949, mainly including key parameters such as central longitude and latitude, maximum sustained wind speed, and minimum sea level pressure.

3. Results

3.1. Initial Field Assimilation Results

To ensure the reliability of subsequent analyses on the role of YunYao GNSS-RO refractivity observations assimilation in typhoon track forecasting, this section first examines the GSI diagnostic outputs of the assimilation experiments to verify whether the assimilation algorithm operates as expected.

Figure 1a shows, as an example for the forecast initialized at 00:00 UTC on 14 September 2024, the observation points (black dots) and the assimilated observation points (red dots) when the GSI assimilation system entered the third outer loop while assimilating the YunYao GNSS-RO refractivity data. From the perspective of the spatial distribution of observations, the YunYao GNSS-RO refractivity data that pass quality control and are adopted by the assimilation system have good spatial coverage in the simulation area and are highly representative. In particular, there are a certain number of valid assimilated observations near the landing location of Typhoon BEBINCA (approximately 31.2° N, 121.5° E), which is helpful for improving the accuracy of the initial field in this key area. A further analysis was conducted on the variation in the difference between the initial field (before and after assimilation) and the observations with height, as shown in Figure 1b. Among them, the blue solid line and blue dashed line represent the profile of the average relative error (Mean) and the profile of its standard deviation (STD) of the atmospheric refractivity background value compared with the observation value in the simulation area, respectively, while the red solid line and red dashed line represent the corresponding profiles of the analysis value, respectively. The statistical results are the mean values within each 1 km height layer. The observation value refers to the refractivity of an occultation profile at a specific height, and the background value and analysis value are obtained from the GSI diagnostic files.

The results show that the overall distribution of the Mean between the background values and the observed values is within the range of ±0.5%, reaching a maximum at an altitude of approximately 2 km, which is about 0.5%. The STD of each altitude layer peaks around 5 km, exceeding 2.0%. For this specific initialization time, a total of 5892 YunYao GNSS-RO observations were successfully assimilated. After assimilating the refractivity data, both the average relative error and its standard deviation between the analyzed values and the observed values are significantly reduced. The distribution of the average relative error is generally narrowed to −0.2% to 0.2%, with the maximum value appearing near the surface, where the average relative error reaches about −0.3%. The STD above 10 km drops to below 0.5%, and that below 10 km is about 0.5% to 1.0%. The STD of each altitude layer is generally reduced by about 0.25% compared with the background values, and the reduction near 5 km exceeds 1%, indicating that the assimilation of YunYao GNSS-RO refractivity data has effectively enhanced the consistency between the background field and the observations in the entire vertical layer. In general, the YunYao GNSS-RO refractivity data show a high assimilation application rate, and assimilating them in the GSI system can effectively reduce the uncertainty of the background field and improve the consistency between the analysis field and the observations.

Figure 2 shows the distribution of analysis increments in the model space for the experiment initialized at 00:00 UTC on 14 September 2024, displaying the variation characteristics of (a) temperature, (b) zonal (U) wind, (c) meridional (V) wind at the 500 hPa level and (d) specific humidity at the 700 hPa level of the initial field before and after assimilation. Warm colors indicate that the variable increased after assimilation, while cold colors indicate a decrease. Overall, the spatial structure of the analysis increments is continuous and reasonably distributed, reflecting the ability of the GSI assimilation system to coordinately correct multiple variables such as temperature, wind field, and humidity. In the 500 hPa temperature increment field (Figure 2a), most of the simulation area shows a negative increment. The largest negative increment is found in the top-left corner, reaching a magnitude of around 2 K. The corresponding increments of 500 hPa U wind (Figure 2b) and V wind (Figure 2c) respectively exhibit increments distributed in the north–south and east–west directions with respect to the temperature increment, with the maximum increment reaching approximately 3 m s⁻¹. The spatial phase relationship shows that the wind field analysis increments are largely perpendicular to the temperature gradient. This pattern is consistent with the thermal wind response under geostrophic balance and reflects the physical consistency of the assimilation increments. Furthermore, the 700 hPa specific humidity increment (Figure 2d) shows positive and negative values consistent with the assimilated observation sites, with a maximum magnitude of 0.002 kg kg⁻¹. The response of low-level humidity indicates that observational information is effectively transmitted to thermal and dynamic fields, resulting in assimilation results that maintain physical coherence in the vertical direction. This demonstrates that the assimilation process can reasonably incorporate observational information and significantly improve the quality of the model initial field, while preserving geostrophic, thermal wind, and humidity balance constraints.

3.2. Forecast Results of Meteorological Elements

To quantitatively evaluate the impact of assimilation on the overall forecasting capability of the model, we first compare the forecast results before and after assimilation with the ERA5 reanalysis data. As a high-quality global reanalysis dataset, ERA5 has high spatial and temporal resolution and consistency across multiple elements, and can be regarded as a “background truth” close to the real atmospheric state. By comparing the spatial distribution and temporal evolution of key meteorological variables, including geopotential height, temperature, humidity, and wind fields, we systematically analyze the improvement effect of assimilation on the model’s ability to depict large-scale circulation and major weather systems. The results are aggregated from multiple forecast initializations, in order to reduce the uncertainty associated with individual forecasts and improve the robustness and representativeness of the evaluation results.

Figure 3 shows the distribution of the difference in RMSE for temperature (T), wind components (U and V), specific humidity (Q), relative humidity (RH), and geopotential height (Z) between the data assimilation (DA) and no-data assimilation (NODA) experiments in the simulation domain, as a function of forecast lead time and pressure level (DA-NODA). In the figure, cold colors indicate negative differences, meaning that the RMSE of the DA experiment, which assimilates GNSS-RO refractivity data, is smaller than that of NODA, reflecting an improvement in forecast performance; warm colors indicate positive differences, implying the opposite. Black dots denote grid points where the differences pass the paired t-test at the p < 0.15 level, indicating statistical significance and that the differences are unlikely to result from random sampling variability. Overall, the results show that data assimilation generally reduces the model errors for most meteorological variables across pressure levels and forecast lead times, thereby enhancing the overall forecast accuracy.

In the temperature field (Figure 3a), the temperature RMSE exhibits overall improvement at nearly all vertical levels and forecast lead times. The improvements are most evident in the lower troposphere (1000–850 hPa) after 60 h, the mid-troposphere (600–500 hPa) after 24 h, and the upper troposphere (300–200 hPa) during most of the forecast period. This indicates that data assimilation effectively improves the evolution of the thermal structure in the simulation domain, particularly leading to a more realistic depiction of the temperature gradient and stratification near the tropopause.

Data assimilation also improves the simulation of the dynamic field. The RMSE of wind components (Figure 3b,c) shows a significant and persistent reduction after 36 h. The improvement in the U wind is mainly concentrated in the middle–lower troposphere (approximately 1000–500 hPa) and at upper levels (350–200 hPa), with the most pronounced reduction occurring during 60–72 h, reaching a maximum decrease of more than 0.1 m s⁻¹. The improvement in the V wind becomes larger than that of the U wind after 80 h, with the greatest improvement occurring between 1000–600 hPa. These results indicate that assimilation of GNSS-RO refractivity data better captures the large-scale circulation pattern and continuously improves the evolution of the mid–upper tropospheric dynamic structure, thereby enhancing medium-range forecast skill.

For the moisture fields, the RMSE of specific humidity (Figure 3d) shows the most significant improvement in the 800–600 hPa layer, with persistent reductions from 36 h to 120 h forecast lead time, and a maximum improvement exceeding 0.04 g kg⁻¹. The RMSE of relative humidity (Figure 3e) exhibits a similar pattern to that of specific humidity, with the maximum reduction occurring around 500 hPa, reaching 0.3%. This indicates that assimilation of refractivity data strongly constrains the water vapor field, improving not only the accuracy of absolute moisture content but also the temperature–moisture consistency, leading to a more realistic vertical humidity structure.

In the geopotential height field (Figure 3f), the RMSE shows stable and consistent improvement throughout the entire atmosphere after 6 h, with the maximum improvement exceeding 0.4 m. This indicates that the assimilation process adjusts the atmospheric mass field and height structure. The improvement in geopotential height also reflects a more accurate representation of the large-scale circulation in the model.

Overall, although the RMSE at the initial time is slightly increased, as the forecast lead time increases, the assimilation of YunYao GNSS-RO refractivity data exerts positive impacts on the thermal, moisture, and dynamic fields. These improvements become more pronounced during the later forecast period, and as indicated by the dots in the figures, the vast majority of these improvements pass the significance test. Aggregated results indicate that data assimilation effectively refines the initial state. This improvement persists throughout the model integration, leading to consistently higher skill in medium-range numerical weather predictions.

3.3. Typhoon Forecast Results

In addition meteorological variables, the specific impact of YunYao GNSS-RO refractivity data assimilation on extreme weather processes was further evaluated by comparing typhoon track and intensity simulations between the DA and NODA experiments for forecasts initialized at different times. By matching the model-predicted typhoon centers with the CMA Best Track Dataset, the effects of assimilation on track deviations and intensity errors (with minimum sea-level pressure as the intensity indicator) were quantitatively assessed. Furthermore, comparing the evolution of track deviations over different forecast lead times provides insight into the ability of assimilation to regulate the stability of medium- and long-term forecasts and to influence the error growth rate.

As shown in Figure 4a–h, the CMA Best Track Dataset indicates that Typhoon BEBINCA moved along a southeast–northwest track and eventually made landfall near Shanghai. Simulated landfall locations across most initialization times exhibited a southward bias relative to observations, primarily clustering along the Zhejiang coast. This trend was reversed only in the 00:00 UTC on 13 September initialization, which favored a more northerly landfall in both DA and NODA experiments.

Comparison of the simulated tracks from the DA and NODA experiments across different initialization times shows that, for all cases other than the forecast initialized at 18:00 UTC on 13 September 2024, the DA tracks are consistently closer to the Best Track dataset and yield more accurate landfall positions than those in the NODA experiment. This demonstrates that assimilation of YunYao GNSS-RO refractivity data positively influences typhoon track simulation. Among all cases, the forecast initialized at 12:00 UTC on 13 September 2024 (Figure 4c) exhibits the largest difference between the DA and NODA experiments. In this case, the DA-simulated track closely follows the observed track, whereas the NODA experiment shows a pronounced southward bias, resulting in a substantially more southerly landfall. This indicates that the assimilation process exerts its strongest guiding effect on the typhoon track for this particular initialization.

From the aggregated statistics of multiple forecast initializations, the track errors in both the DA and NODA experiments increase gradually with forecast lead time, showing a typical error growth behavior. In the initial forecast stage, the track errors in both experiments are less than 50 km, but the DA experiments shows slightly smaller errors than the NODA experiments by about 2 km, indicating that data assimilation improves the initial position and early track forecasts of the typhoon. Starting from about the 24 h forecast lead time, the advantage of the DA experiments becomes more evident, and the improvement in track error gradually increases with forecast lead time. At around the 84 h forecast lead time, the track error in the DA experiments is reduced by approximately 30 km compared to the NODA experiments, and this improvement remains relatively consistent in the later forecast period. These results indicate that the assimilation of GNSS-RO data not only improves the initial conditions but also suppresses the growth of track forecast errors to some extent, thereby improving the stability of medium- and long-range typhoon track forecasts.

In addition, the assimilation also shows a certain improvement in the simulation of typhoon intensity. The mean intensity evolution in both experiments shows a similar trend, with the typhoon first intensifying and then weakening. During most forecast lead times, the DA experiments shows a slight improvement compared with the NODA experiments, with the mean central pressure error reduced by up to approximately 0.5 hPa. Although at some individual forecast times the DA experiments perform similarly to or slightly worse than the NODA experiments, the overall statistical results indicate that the assimilation of YunYao GNSS-RO refractivity data still has a weak but consistently positive impact on typhoon intensity simulation.

Overall, the results from multiple forecast initializations indicate that the assimilation of GNSS-RO refractivity data has a more significant impact on typhoon track forecasts than on intensity forecasts. The main improvements are reflected in reducing the southward track bias, improving landfall location forecasts, and slowing the growth rate of track errors with forecast lead time, which is beneficial for improving medium- and long-range typhoon track prediction.

It should be noted that although the current study provides a robust evaluation of Typhoon BEBINCA through eight independent forecast experiments and statistical significance testing, the scope remains focused on a single tropical cyclone event. Future work is planned to incorporate a larger ensemble of typhoon cases and diverse extreme meteorological processes to further verify the consistency and generalizability of the YunYao GNSS-RO data’s impact. These efforts, including long-term seasonal assessments and multi-source data assimilation experiments, will provide a more comprehensive understanding of the data’s value in enhancing global and regional numerical weather forecasts.

3.4. Sensitivity Tests on Parameterization Schemes

To further investigate the sensitivity of occultation assimilation effects to model parameterization schemes (as detailed in Section 2.3), two additional experiments were initiated on 14 September 2024 using the Thompson microphysics scheme. These include one with YunYao GNSS-RO assimilation (THOM_DA) and one without data assimilation (THOM_NODA). These results are compared with the typhoon track simulation experiments in Section 3.1, Section 3.2 and Section 3.3 that used the WSM6 scheme (DA and NODA experiment). In this section, they are denoted as WSM6_DA and WSM6_NODA, respectively. In this way, the impact of different cloud microphysics parameterization schemes on the assimilation effects and typhoon forecast performance could be evaluated.

As shown in Figure 5a, the simulated tracks of the two non-assimilated experiments, THOM_NODA and WSM6_NODA, showed little difference before landfall; starting from landfall, the THOM_NODA track was more northerly and closer to the best track. In contrast, the THOM_DA and WSM6_DA experiments diverged significantly from their non-assimilated counterparts as early as the second time step. While THOM_DA exhibited a southward shift, WSM6_DA trended further north, aligning most closely with the best track—a consistency that became even more pronounced at the time of landfall. Figure 5b shows the variation in typhoon track error and central pressure error over time. The solid lines indicate the track errors, while the colored bars represent central pressure errors. The track errors of the four groups of experiments were the smallest in the initial stage of simulation, approximately 25 km, and gradually increased as the typhoon developed. From 06:00 UTC on 15 September to before landfall, the WSM6_DA experiment consistently had the best track forecasting performance, with the averaged track error reduction of more than 30 km compared to the WSM6_NODA, THOM_NODA, and THOM_DA experiments. At the time of typhoon landfall at 00:00 UTC on September 16, the track deviation was reduced by nearly 50 km compared to the non-assimilated WSM6_NODA.

In terms of typhoon intensity, before 12:00 UTC on September 14, the simulated central pressures from the four experiments showed only small differences, with errors all below 3 hPa compared to the Best Track dataset. As the typhoon developed rapidly, the errors in the WSM6_NODA, WSM6_DA, and THOM_NODA experiments gradually became larger. In contrast, the THOM_DA experiment, which assimilated YunYao GNSS-RO refractivity data, significantly improved the simulation of the typhoon’s central pressure: at 00:00 UTC on 15 September, its central pressure error was approximately 3.5 hPa, about 5 hPa lower than that of the other experimental groups. In the subsequent period shown in the figure, the central pressure error in THOM_DA remained consistently lower than in the other three experiments, with reductions of over 10 hPa compared to the experiment with the largest error. At the time of typhoon landfall, the central pressure error in this experiment was also reduced by roughly 2–4 hPa relative to the other experiments.

In summary, although the THOM_DA experiment performed the best in simulating the central pressure, its track error was higher than that of the THOM_NODA experiment without assimilation. The above results indicate that the improvement effect of assimilation on typhoon forecasting is closely related to the setting of microphysical schemes. The WSM6 scheme combined with refractivity assimilation improves both track and intensity forecasts. In contrast, although the Thompson scheme significantly improves central pressure after assimilation, its impact on track forecasts is limited and may even be worse than that of the non-assimilated experiment. Therefore, in numerical forecasting for typhoons, it is necessary to comprehensively consider the coordination between assimilation strategies and physical process schemes to synergistically improve the forecasting capabilities of typhoon track and intensity.

4. Conclusions

This study focuses on evaluating the impact of YunYao GNSS-RO refractivity assimilation on typhoon simulation performance. Using the WRF model coupled with the GSI data assimilation system, multiple forecast experiments with and without assimilation were conducted for Typhoon BEBINCA during 13 and 14 September 2024. The effects of refractivity assimilation on key meteorological variables, typhoon track, and intensity forecasts were systematically analyzed. On this basis, sensitivity experiments using two cloud microphysics schemes (WSM6 and Thompson) were further performed to evaluate the influence of different parameterization schemes on the assimilation effects. The results are as follows:

• Data assimilation improves the consistency between the initial atmospheric fields and observations in the middle and upper troposphere. Both the mean relative error and the standard deviation of the background field relative to observations are significantly reduced, providing a more accurate initial atmospheric state for typhoon forecasting.
• Comparisons with ERA5 reanalysis data show that assimilating YunYao GNSS-RO data effectively improves the forecast accuracy of key meteorological variables. The assimilation experiment shows higher simulation accuracy for geopotential height, humidity, temperature, and wind speed. Statistical results from multiple initializations show that the improvements persist from 36 h to 120 h forecast lead time, with maximum RMSE reductions exceeding 0.2 K for temperature and 0.1 m s⁻¹ for wind speed, while geopotential height shows consistent improvement throughout the entire atmosphere.
• Assimilation of YunYao GNSS-RO data significantly improves the simulation accuracy of Typhoon BEBINCA’s track and intensity. Compared with observations, the tracks in seven assimilation experiments are closer to the best track. Statistical results from multiple forecast initializations show that the 84 h track error is reduced by approximately 30 km on average, and the minimum central pressure bias is also reduced.
• Cloud microphysics schemes have a significant influence on the assimilation effects. The WSM6 scheme is more beneficial for improving track forecast accuracy, whereas the Thompson scheme performs better for intensity forecasting. In operational forecasting, the assimilation strategy should be optimized in combination with the characteristics of physical parameterization schemes to achieve overall improvements in typhoon track and intensity forecast performance.

In summary, the assimilation of YunYao GNSS-RO refractivity data can effectively improve the performance of numerical forecasting, and its application is of great value for enhancing the accuracy of operational forecasting.