Impact of All-Sky Assimilation of Multichannel Observations from Fengyun-3F MWHS-II on Typhoon Forecasting

Abstract

All-sky radiance assimilation can increase the utilization of satellite observations in cloudy regions and improve typhoon forecasts. This study focuses on the newly launched FengYun-3F satellite equipped with the Microwave Humidity Sounder II (MWHS-II) and develops an all-sky assimilation capability for its radiance data. A series of assimilation experiments were conducted to evaluate their impacts on the forecast of Typhoon Yagi (2024), demonstrating that all-sky assimilation leads to reductions in track error (23.14%) and improvements in precipitation forecasts (Equitable Threat Score increase of 16.92%) compared to clear-sky assimilation. Furthermore, a detailed comparison of assimilation experiments shows that using only the 183 GHz humidity channels yields limited improvement in tropospheric humidity, whereas assimilating the 118 GHz temperature channels significantly enhances temperature and wind forecasts. Combined assimilation of both frequency bands synergistically maintains accurate track and intensity predictions while further improving precipitation prediction. These findings demonstrate the value of multichannel all-sky assimilation and inform future satellite data assimilation strategies.

1. Introduction

The Microwave Humidity Sounder is highly sensitive to cloud ice and water vapor and plays a vital role in acquiring information on hydrometeors, precipitation, and wind [1]. Meteorological satellites equipped with microwave instruments provide effective observational supplements over oceanic regions where conventional observations are sparse. However, fewer than 30% of the available satellite observations are effectively assimilated into numerical weather prediction systems [2]. These discarded observations are typically contaminated by cloud and precipitation signals. Clouds and precipitation are discontinuous in time and space and are formed by complex, nonlinear physical processes [3]. As a result, many studies adopt clear-sky assimilation strategies, which discard observations affected by clouds and assimilate only those from cloud-free regions to maintain the reliability of the analysis [4,5,6,7,8]. Although this approach has demonstrated notable improvements in numerical weather prediction accuracy, it inevitably results in the exclusion of substantial cloud-related observational information. To retain these observations that are closely related to weather structures such as typhoons and heavy rainfall, all-sky assimilation has been progressively developed [8,9], which incorporates satellite observations under all sky conditions (including those affected by clouds and precipitation) into the numerical weather prediction system. Compared to clear-sky assimilation, the all-sky assimilation simulates a more complete radiative transfer process and retains more cloud-related information. This advantage enables the model to better simulate the formation of clouds and precipitation. Consequently, all-sky assimilation improves the analysis of temperature, humidity, and cloud distributions, thereby enhancing the forecasts of typhoon intensity and precipitation [10,11,12,13,14].

A primary challenge in conducting all-sky assimilation is reducing the impact of various error sources on the assimilation process. Due to the limited accuracy of the numerical model and radiative transfer calculations in cloud-covered areas [15], there is often a low consistency between the background fields and satellite observations in identifying cloudy regions. These representativeness errors are treated as part of the observation error, which affects the Gaussianity of the observation error distribution [16]. If the observation errors are not properly specified, the quality of the analysis fields can be degraded. To address this, Geer and Bauer [10] proposed a symmetric observation error model that dynamically allocates observation errors based on cloud amount. This model offers a robust threshold for quality control and enables the estimation of observation error magnitudes in all-sky assimilation, thereby reducing the non-Gaussian characteristics of observation-minus-background (O−B) departures. Its performance can be further enhanced by refining cloud indicators and adjusting the size of the statistical sample [17,18].

Another critical issue in all-sky assimilation is the construction of the observation operator. The radiative transfer models are typically employed as part of the observation operator to provide the required forward, tangent-linear, and adjoint models [19]. Initially, some radiative transfer models were limited to clear-sky, non-scattering conditions. However, as assimilation techniques have evolved towards all-sky conditions, these models have gradually been extended to incorporate fast scattering solvers [20,21]. At present, the most widely adopted radiative transfer models include RTTOV (Radiative Transfer for TIROS Operational Vertical Sounder), developed by EUMETSAT, and CRTM (Community Radiative Transfer Model), developed by the Joint Center for Satellite Data Assimilation in the United States. RTTOV has been progressively enhanced to support microwave radiance simulations under cloudy and rainy conditions through the RTTOV-SCATT module. This module employs the delta-Eddington approximation to account for the scattering effects of hydrometeors on microwave radiation and has proven to be an effective tool for simulating brightness temperatures to reveal typhoon structures [22]. With the aid of the advanced radiative transfer model, the European Centre for Medium-Range Weather Forecasts (ECMWF) has begun to incorporate radiance observations from multiple satellites over cloudy regions into its assimilation system. Correspondingly, the number of satellites included in operational all-sky data assimilation has increased more than threefold from 2007 to 2020 [23].

In addition, the channel sensitivity of microwave instruments is another key factor influencing the performance of the data assimilation. Some channels display strong sensitivity to lower tropospheric humidity and are, therefore, designed to detect information on tropospheric water vapor. However, this sensitivity also makes them vulnerable to surface emissivity interference. As a result, some operational centers have to discard land-based observations or improve surface emissivity models in order to reduce contamination from surface radiation [24]. Recognizing the limitations of conventional humidity-sensitive channels, satellite microwave instruments have been equipped with more diverse channels to compensate for these deficiencies and to enhance both the coverage and the diversity of atmospheric information acquired. The Microwave Humidity Sounder (MWHS) onboard the FengYun (FY) satellites, for instance, innovatively includes 118 GHz oxygen absorption channels. Assimilating the observations at this frequency has been shown to improve numerical weather prediction and to serve as a complement to traditional humidity channels [25,26,27,28]. However, the sensitivity characteristics of these channels during assimilation and their performance in simulating large-scale weather structures under all-sky conditions remain unclear.

This study focuses on the channel sensitivity in all-sky microwave assimilation and aims to better understand how the sensitivity influences numerical forecasts, particularly for typhoons. The findings are expected to provide helpful information for optimizing multichannel joint assimilation strategies in typhoon forecasting and to enhance the utilization of satellite observations under cloudy conditions for tropical cyclone warnings.

The structure of this paper is as follows. Section 2 introduces the observational data and assimilation methods used in the study. Section 3 provides an overview of Typhoon Yagi (2024) and describes the experimental design for data assimilation and numerical forecasting. Section 4 analyzes the impact of all-sky assimilation of FY-3F MWHS-II radiance data on typhoon forecasting and discusses the differences caused by various channel selection strategies. Finally, Section 5 and Section 6 present the overall discussion and conclusions of the study.

2. Materials and Methods

2.1. MWHS-II Radiance Data

As the successor to FY-3C, the MWHS-II instrument onboard FY-3F features improved detection accuracy and sensitivity [29] and provides both descending and ascending track observations. The instrument consists of 15 channels, with their basic parameters listed in

Table 1. Basic parameters of MWHS-II channels. Central Frequency: The dominant or nominal frequency of a spectral band or signal. Field of View (FOV): The spatial extent observable by a sensor during a single measurement. For MWHS-II, each cross-track scan acquires 98 discrete views. Resolution: The minimum distinguishable feature size detectable by the instrument. Bandwidth: The spectral range covered by a channel.

Channel Number	Central Frequency (GHz)	Field of View (FOV)	Resolution (km)	Bandwidth (MHz)
1	89.0 (V)	98	30	1500
2	118.75 ± 0.08 (H)	98	30	20
3	118.75 ± 0.2 (H)	98	30	100
4	118.75 ± 0.3 (H)	98	30	165
5	118.75 ± 0.8 (H)	98	30	200
6	118.75 ± 1.1 (H)	98	30	200
7	118.75 ± 2.5 (H)	98	30	200
8	118.75 ± 3.0 (H)	98	30	1000
9	118.75 ± 5.0 (H)	98	30	2000
10	166.0 (V)	98	15	1500
11	183.31 ± 1 (H)	98	15	500
12	183.31 ± 1.8 (H)	98	15	700
13	183.31 ± 3 (H)	98	15	1000
14	183.31 ± 3 (H)	98	15	2000
15	183.31 ± 7 (H)	98	15	2000

. The MWHS-II includes five humidity-sounding channels centered around the 183.31 GHz water vapor absorption line, designed to provide detailed vertical profiles of atmospheric moisture. In addition, the instrument is equipped with eight oxygen absorption channels near 118.75 GHz, primarily used for atmospheric temperature sounding. These channels are unique to MWHS-II and are not available on other similar microwave humidity sounders such as MHS or ATMS. Although primarily intended for temperature retrieval, these channels are also affected by clouds and precipitation due to liquid water absorption/emission and scattering by ice particles [30]. The instrument includes window channels at 89 GHz and 116 GHz, which are generally used for surface emissivity sensing and precipitation detection.

Figure 1 shows the distribution of weighting functions for MWHS-II channels, which were simulated using the RTTOV forward model based on transmittance calculations from the U.S. Standard Atmosphere profile [31]. As illustrated in Figure 1, channels 1, 10, and 7–9 have weighting function peaks close to the surface, while channels 2–4 peak in the upper troposphere or even the stratosphere. The remaining channels have weighting function peaks mainly within the troposphere. In this study, Level-1 MWHS-II data are used, and the assimilated channels include channels 4–7 and 11–15. Channels 8 and 9 are excluded from assimilation due to their peak sensitivity being too close to the surface, which could introduce greater uncertainty into the analysis.

2.2. All-Sky Data Assimilation Method

All experiments in this study are conducted by the Weather Research and Forecasting (WRF V4.6.0) model and its data assimilation system (WRFDA), both developed by the National Center for Atmospheric Research (NCAR). WRFDA provides various modern data assimilation techniques, including three-dimensional variational (3D-Var), four-dimensional variational (4D-Var), and ensemble Kalman filter (EnKF) methods. In this study, the 3D-Var method is primarily used, which obtains the optimal analysis field by minimizing a cost function as follows:

$$J\left(x,x_b\right)={\displaystyle \frac{1}{2}}{\left(x-x_b\right)}^T{B}^{-1}\left(x-x_b\right)+{\displaystyle \frac{1}{2}}{\left(y-H\left(x\right)\right)}^T{R}^{-1}\left(y-H\left(x\right)\right)$$

(1)

In the 3D-Var framework, $x$ is the atmospheric state vector, and $x_b$ is the background field, which typically provides either the output of WPS/REAL (cold start) or the forecast results from WRF (warm start). $H$ is the nonlinear observation operator that delivers model variables to observation space. In satellite radiance assimilation, this operator is typically implemented via a radiative transfer model. $y$ is the observation vector, $B$ and $R$ represent the background error covariance matrix and the observation error covariance matrix, respectively. The default control variable option (CV5) is applied in all experiments. CV5 utilizes stream function (Ψ), unbalanced velocity potential (Χ_u), unbalanced temperature (T_u), pseudo relative humidity (RH_s), and unbalanced surface pressure (P_s,u). The pseudo relative humidity is defined as Q/Q_b,s, where Q_b,s is the saturated specific humidity from the background field. The flow-dependent background error covariance matrix is computed using the National Meteorological Center (NMC) method [32] based on the differences between 24 h and 12 h forecasts over the entire month of September 2024.

RTTOV-SCATT (the scattering module of Radiative Transfer for TOVS) version 12.1 serves as the radiative transfer model to simulate MWHS-2 brightness temperature observations. RTTOV-SCATT is a radiative transfer module developed by ECMWF for simulating microwave brightness temperatures in cloudy and precipitating conditions. The simulated brightness temperatures from RTTOV-SCATT are a combination of observations from both clear-sky and cloudy/rainy conditions [33]:

$$L_B^{Total}=\left(1-C\right)L_B^{Clear}+C{L}_B^{Rainy}$$

(2)

where $C$ is the effective cloud reflection, and $L$ is the simulated bright temperature. RTTOV-SCATT internally calls RTTOV’s clear sky mode to obtain the bright temperature ($L_B^{Clear}$) and the vertical distribution of the clear sky transmittance, then finds the Mie scattering lookup table to compute the bright temperature ($L_B^{Rainy}$) for the cloud or rainy conditions based on the clear sky transmittance. The final result ($L_B^{Total}$) is obtained by linearly combining the simulated bright temperatures for both clear sky and cloud cover. At present, WRFDA cannot support RTTOV-SCATT as a radiative transfer model directly. Therefore, the WRFDA system is extended by developing a new coupling module between the radiative transfer mode and the assimilation system. Additional hydrometeor profiles required for RTTOV-SCATT are also computed, enabling the assimilation system to use both the forward, tangent linear, and adjoint operators of RTTOV-SCATT.

2.3. Symmetric Observation Error Model

A major challenge in all-sky assimilation is the inaccuracy of both forecasts and brightness temperature simulations in cloud-covered regions. These issues lead to inconsistencies between observations and the background field when identifying cloud-covered areas. Geer and Bauer proposed a symmetric observation error model that dynamically distributes the observation error through a function of symmetric cloud amount [10]. The model adopts a symmetric cloud factor $C_{sym}$ as a criterion to assess the consistency between the simulated and observed brightness temperatures in cloud-covered areas:

$$C_{sym}={\displaystyle \frac{{SI}_{fg}+{SI}_{ob}}{2}}$$

(3)

The scattering indices ${SI}_{fg}$ and ${SI}_{ob}$ are, respectively, calculated from the simulated brightness temperature of the background field and the observed brightness temperature. For MWHS-II, the scattering index is derived by taking the difference between two window channels. A larger value of the symmetric cloud factor ($C_{sym}$) indicates better agreement between the model and the observations in cloud-covered areas. When both the observation and the model are under clear-sky conditions, $C_{sym}$ is equal to zero. Finally, observation errors over ocean and land can be determined as follows:

$$g\left(C_{sym}\right)=\left\{\begin{array}{c}{g}_{clr} \in C_{sym}<C_{clr}\hfill \\ g_{clr}+\left(g_{cld}-g_{clr}\right)\hfill \\ g_{cld} \in C_{sym}>C_{cld}\hfill \end{array}\right.\left({\displaystyle \frac{C_{sym}-C_{clr}}{C_{cld}-C_{clr}}}\right)\in C_{clr}{ \le C}_{sym}\le C_{cld}$$

(4)

where $g_{clr}$ represents the observation error assigned to clear sky conditions, $g_{cld}$ is the observation error assigned to cloudy conditions and $C_{clr}$, $C_{cld}$ are the thresholds of the symmetric cloud factors under the two weather conditions. A total of 148,952 observations (77,828 over land and 71,124 over ocean) collected over a two-week period were used to calculate the standard deviation of O−B (observation minus background) as a function of symmetric cloud amount. The statistical domain covered southern China, adjacent regions, and the South China Sea/western Pacific to align with Typhoon Yagi’s activity range. Given FY-3F daily overpasses at 03:00 and 15:00, we included observations within ±3 h of these times. Land and ocean observations were separated using the surface type code embedded in the satellite data so that statistics for land and ocean regions were computed separately. The simulated brightness temperatures were generated by RTTOV-SCATT using 15 h forecasts based on NCEP FNL (Final) Operational Global Analysis data as the first guess. Figure 2 shows the all-sky standard deviations over ocean and land for each frequency channel (118 GHz and 183 GHz) as dashed lines, along with the corresponding observation error functions as solid lines.

As shown in Figure 2, both the standard deviations and normalized observation errors over land are generally larger than those over the ocean. Channels 4–6 show smaller standard deviations compared to other channels, likely because they observe higher atmospheric layers, with channel 4 peaking even above the tropopause. This pattern is also evident in the weighting functions shown in Figure 1, where higher peaking channels are less affected by clouds and precipitation, resulting in smaller standard deviations.

3. Experimental Designs

3.1. Overview of Typhoon Case

Typhoon Yagi (2024) was selected as the simulation case for the assimilation experiments using FY-3F data. The best track of the typhoon is illustrated in Figure 3. Typhoon Yagi formed as a tropical storm east of the Philippines at 1200 UTC on 1 September 2024 and intensified into a typhoon by approximately 2300 UTC on 3 September. It continued to strengthen and reached its peak intensity at 0000 UTC on 5 September, with a maximum wind speed of 65.85 m/s and a minimum central pressure of 922 hPa. At 1600 UTC on 6 September, the typhoon made landfall in Wenchang, Hainan Province, with a maximum wind speed of 62 m/s (region 4 in Figure 3). After a brief weakening, the typhoon rapidly intensified again at around 1800 UTC on the same day, increasing from 52 m/s to 60 m/s in a short period. This was the fastest of three intensification phases during its lifecycle and was identified as an “explosive” intensification [34]. After the intensification, Typhoon Yagi made its second landfall in Xuwen, Guangdong Province (region 1 in Figure 3), at 2200 UTC on 6 September, with a maximum wind speed of 58 m/s. It still maintained a wind speed of 58 m/s when it made the final landfall in northern Vietnam at around 1500 UTC on 7 September. Typhoon Yagi caused severe precipitation in Guangdong, Guangxi, and Hainan provinces and was the strongest autumn storm at landfall in China since records began in 1949. The typhoon completed two intensity upgrades (Very strong typhoon, Violent typhoon) within 15 h, and the extremely fast rate of intensification caused difficulties for numerical weather forecasting.

3.2. Experimental Settings

The Advanced Research WRF (ARW) version 4.6.0 was employed to conduct the experiments in this study. The model configuration included 241 × 301 grids, with a horizontal resolution of 9 km, 38 eta levels in the vertical direction, and a model top level at 50 hPa. The model employed the following parameterization: the Purdue Lin microphysics scheme [37], the Kain–Fritsch cumulus parameterization scheme [38], the Dudhia shortwave radiation scheme [39], the RRTM longwave radiation scheme [40], the Yonsei University (YSU) planetary boundary layer scheme [41], and the Unified Noah land surface scheme. The study set up a total of seven experiments to analyze the impact of all-sky assimilation of FY-3F MWHS-II data on typhoon forecasting. The configuration of each experiment is listed in

Table 2. Assimilation experiment configurations.

Experiment Name	Experiment Configuration
CON	No data assimilation
118_183_CLR (CLR)	MWHS-II 118 GHz + 183 GHz clear-sky data
118_CLR	MWHS-II 118 GHz clear-sky data
183_CLR	MWHS-II 183 GHz clear-sky data
118_183_AS (AS)	MWHS-II 118 GHz + 183 GHz all-sky data
118_AS	MWHS-II 118 GHz all-sky data
183_AS	MWHS-II 183 GHz all-sky data

:

The experiments 118_183_CLR and 118_183_AS are hereafter referred to as the CLR and AS experiments in Section 4.1 to evaluate the impact of all-sky assimilation based on the FY-3F MWHS-II data. In Section 4.2 a detailed comparison was conducted among all-sky and clear-sky assimilation experiments (118_CLR, 183_CLR, 118_183_CLR, 118_AS, 183_AS, and 118_183_AS) to investigate the channel sensitivity, especially how different channels under all-sky conditions influenced the analysis fields and forecasts. The model integration time was from 0000 UTC on 5 September to 0000 UTC on 8 September 2024. The control experiment involved no data assimilation, while all the assimilation experiments used a background field generated by a 15 h spin-up simulation starting from 0000 UTC on 5 September, initialized with FNL analysis data. The assimilation analysis time was set to 1500 UTC on 5 September, with a 6 h assimilation window. The atmospheric variables listed in

Table 3. All atmospheric variables involved in the assimilation process.

Atmospheric Variable Name	Unit
Temperature	K
Water vapor mixing ratio	g/kg
Pressure	hPa
Zonal wind component (u) and Meridional wind component (v)	m/s
2 m temperature	K
2 m water vapor mixing ratio	g/kg
10 m Zonal wind component (u) and Meridional wind component (v)	m/s
Surface pressure	hPa

are converted into brightness temperatures through the radiative transfer model and involved in the assimilation process.

After minimizing the three-dimensional variational cost function, the analysis increments are propagated back to the background field via the adjoint model of the assimilation system, thereby updating the background state. The resulting analysis fields from each assimilation experiment were then used to update the lateral boundary conditions and to serve as the first guess for a 57 h forecast, ending at 0000 UTC on 8 September 2024.

3.3. Quality Control

For both clear-sky and all-sky assimilations, the same general radiance quality control procedures are applied to the MWHS-2 data: (1) radiance data located over mixed surface types is rejected, (2) radiance data within five scan positions of each swath edge is removed [30], (3) radiance with O−B departures exceeding 3σ is rejected, where σ represents the specified standard deviation of the observation errors, (4) all radiance data are thinned to a minimum spacing of 45 km, (5) data exceeding the channel-specific altitude thresholds is rejected (Channel 11: 1500 m; Channels 12–13: 1000 m; Channels 14–15: 800 m). In addition, for clear-sky assimilation experiments, a cloud detection procedure is used to eliminate cloud-contaminated observations: (1) radiance data with an absolute brightness temperature difference between Channels 1 and 10 greater than 10 K is removed, (2) radiance with a liquid cloud water path exceeding 0.2 g·m⁻² is removed. For all-sky assimilation, a symmetric observation error model is applied to observations influenced by clouds and precipitation. Bias correction is implemented using the variational bias correction (VarBC) method, which estimates observation system biases dynamically within a Bayesian framework [42]. This approach allows the O−B departures to better conform to the assumption of Gaussian error distributions, thereby improving assimilation performance.

4. Results from FY-3F MWHS-2 Experiments

Figure 4 displays the distribution of assimilated observations from two representative channels (CH 7 and CH 13) under different sky conditions. Quality-controlled observations increased from 1291 (CH 7) and 954 (CH 13) in clear-sky assimilation to 2142 and 1584, respectively, in all-sky assimilation. All-sky assimilation increased the utilization rate of observations, with the newly assimilated data distributed not only around the typhoon eye but also within the spiral rainbands and other regions covered by thick clouds. However, the number of additional observations within the typhoon core remains relatively limited compared to other cloud-covered areas. This may be owing to larger discrepancies between the simulated typhoon structure in the background field and the observations, leading to higher observation-minus-background (O−B) errors and, consequently, more rejections during quality control. Nevertheless, the additional observations in surrounding regions still have the potential to influence the regions near the typhoon, contributing to improvements in the forecast accuracy.

Figure 5 illustrates the spatial distribution of observation-minus-background (O−B) and observation-minus-analysis (O−A) differences at the analysis time (1500 UTC on 5 September 2024). In both the clear-sky and all-sky assimilation experiments, the O−A values are noticeably smaller than the corresponding O−B values, indicating that the analysis fields are closer to the observations than the background fields. Although the AS experiment shows slightly larger mean O−B values compared to the CLR experiment, it achieves smaller mean O−A values, suggesting better overall performance in reducing analysis errors. For Channel 14, the standard deviation of O−A in the AS experiment is slightly higher than that in the CLR experiment (Figure 5f,h). However, this discrepancy can be attributed to the larger sample size used in the all-sky assimilation, which results in greater sample-based variability. Given this context, the difference in standard deviation is considered acceptable.

4.1. Forecast Verification

4.1.1. Typhoon Track and Intensity Forecasts

The International Best Track Archive for Climate Stewardship (IBTrACS) is used as the best track dataset to verify the typhoon forecast performance of the control, clear-sky, and all-sky experiments. Figure 6 presents the differences between the forecasts and the best track in terms of multiple evaluation metrics. The typhoon track in the CON experiment exhibits a northward deviation over time (Figure 6a). Although the CLR experiment shows smaller track errors than the CON experiment within approximately 18 h after assimilation (Figure 6b), the average improvement remains within 10 km. After 0900 UTC on 6 September 2024, its track error even exceeds that of the CON experiment. In contrast, the AS assimilation effectively suppresses this deviation and maintains track errors below 120 km throughout the forecast period. On average, the AS experiment reduces track error by 23.2 km compared to the CON and by 31.8 km compared to the CLR experiment. For the near-surface maximum horizontal wind speed (Figure 6d), there is little difference among the three experiments during the first 27 h. However, a notable negative bias emerges after 1800 UTC on 6 September 2024, coinciding with a rapid intensification phase of Typhoon Yagi [34]. Neither the CON nor the CLR experiments capture this intensification trend, resulting in underestimated wind speeds and large negative biases. The AS experiment, however, captures the intensification process, leading to much smaller wind speed errors (Figure 6d). The maximum wind speed error in the AS experiment is reduced by an average of 1.33 m/s compared to the other two experiments, with the maximum improvement reaching 6.0 m/s. In terms of the minimum sea level pressure (Figure 6c), the AS experiment also maintains relatively lower errors during the intensification period of Typhoon Yagi, and this improvement is more consistent over time. Compared to the CON and CLR experiments, the maximum reduction in MSLP error reaches 3.30 hPa and 3.29 hPa, respectively. Based on these evaluations, it is concluded that the all-sky assimilation exhibits a clear advantage in capturing the rapid intensification of the typhoon, and the improvement it brings is both substantial and persistent.

4.1.2. Thermodynamic and Dynamic Structure

Figure 7 shows the vertical root mean square error (RMSE) profiles of vector winds, temperature, and specific humidity over the entire simulation domain at 24 h (1500 UTC on 6 September 2024) and 48 h (1500 UTC on 7 September 2024) forecast lead times, using ECMWF reanalysis as the reference to assess the impact of assimilation. The CLR assimilation experiment provides only limited improvements, with reductions mainly in temperature and humidity within the troposphere (Figure 7c,d,g,h). In comparison, the AS experiment displays more evident benefits, especially for vector winds (Figure 7a,b,e,f), where improvements extend from 200 hPa to the surface, the RMSEs of both zonal wind (U) and meridional wind (V) are reduced by an average of 0.59 m/s and 0.37 m/s, respectively, compared to the CLR experiment. The enhanced wind performance may contribute to better typhoon forecasts. Additionally, the AS experiment produces lower RMSE in temperature than the CLR case, both above 400 hPa and below 700 hPa, with mean reductions of 0.027 K and 0.046 K, respectively (Figure 7c,g). The advantage in the upper levels becomes more distinct at longer forecast ranges, reaching a maximum reduction of 0.069 K. For specific humidity (Figure 7d,h), the two experiments show similar performance in the early forecast period, while the AS experiment gradually demonstrates clearer advantages later on, with a vertical pattern similar to that of temperature. Although the RMSE is averaged across the full simulation domain, the overall enhancements in wind and temperature with all-sky assimilation remain noticeable and may help explain its superior performance in typhoon forecasts.

4.1.3. Precipitation Forecast

From 0000 UTC on 5 September to 0000 UTC on 8 September 2024, Typhoon Yagi brought accumulated precipitation amounts of 300 to 500 mm to parts of Hainan, Guangdong, and Guangxi provinces in China (region 1, 2, 4 in Figure 8). Figure 8 shows the spatial distribution of 24 h accumulated precipitation from 0000 UTC on 6 September to 0000 UTC on 8 September 2024. During this period, the model demonstrates visible overestimations in precipitation forecasts. In the 24 h accumulation from 0000 UTC on 6 September to 0000 UTC on 7 September, the CLR experiment shows a precipitation distribution similar to the CON experiment (Figure 8b,c). Compared to the CLR experiment, the AS assimilation experiment (Figure 8d), due to a smaller northward displacement of the predicted typhoon center, results in a narrower coverage of the spiral rainbands over Guangdong and Guangxi provinces, with rainfall locations more consistent with observations. In the subsequent 24 h period from 0000 UTC on 7 September to 0000 UTC on 8 September, the CON experiment still exhibits clear overestimations (Figure 8f). Although the CLR experiment (Figure 8g) reduces overestimation in Guangdong and Guangxi, it intensifies the overestimation in Yunnan Province (region 3 in Figure 8). Meanwhile, the AS experiment (Figure 8h) further reduces excessive precipitation in Guangdong and Guangxi and does not exhibit overestimation in Yunnan Province, indicating that the all-sky assimilation produces more accurate forecasts in both rainfall magnitude and spatial distribution.

To quantify this improvement, the Equitable Threat Score (ETS) is calculated for all three experiments. ETS is a widely used metric for evaluating binary precipitation forecasts by measuring improvement over random forecasts. Figure 9 shows that the performance advantage of the AS experiment is evident. Compared to the CLR experiment, the AS experiment consistently yields higher ETSs for thresholds greater than 10 mm. Moreover, as the accumulated precipitation threshold increases, the ETS of the AS experiment exceeds that of the CLR experiment by approximately 10% to 20%, indicating that all-sky assimilation offers a more pronounced advantage in forecasting extreme precipitation. Between 0000 UTC 6 September and 0000 UTC 7 September (Figure 9a), the ETS of the AS experiment exceeds that of the CLR experiment by at least 10.29% (0.036), and this advantage further increases with higher thresholds, reaching a maximum of 19.47% (0.029). On the following day (0000 UTC 7 September to 0000 UTC 8 September 2024, Figure 9b), the AS experiment still maintains an improvement of at least 11.53% (0.028) over the CLR experiment, with the maximum reaching 21.33% (0.020). This improvement may result from better adjustments to temperature and moisture fields, which correspond to lower RMSE in the forecasts of temperature and specific humidity in the AS experiment (Figure 7c,d,g,h).

The study also found that during the period from 0000 UTC on 7 September to 0000 UTC on 8 September 2024, the ETS of the CLR experiment was lower than that of the CON experiment. To avoid the potential bias introduced by the inherent limitations of ETS (high sensitivity to spatial displacement), the Fraction Skill Score (FSS) was used as an objective alternative for evaluation. The FSS is defined as [43]:

$$FSS=1-{\displaystyle \frac{\frac{1}{N}\sum _{i=1}^N{(P_{fi}-P_{oi})}^2}{\frac{1}{n}\left[\sum _{i=1}^N{P}_{fi}^2+\sum _{i=1}^N{P}_{oi}^2\right]}}$$

(5)

$P_{fi}$ and $P_{oi}$ represent the forecasted and observed probabilities at the $i$-th grid point exceeding a given threshold, and $N$ is the total number of grid points in the verification region. The $FSS$ ranges from 0 (no overlap) to 1 (perfect overlap) between forecasted and observed precipitation areas. Figure 10 shows the FSSs for all experiments. Under this more balanced verification metric, the CLR experiment demonstrated better forecast performance than the CON experiment across most precipitation thresholds. The superior performance of the AS experiment is still evident.

4.2. Channel Sensitivity

4.2.1. Comparison of Analysis Increments

The 118 GHz channels of FY-3F MWHS-II, designed for oxygen absorption, are highly sensitive to atmospheric temperature, whereas the 183 GHz channels, targeting water vapor absorption, are more responsive to clouds and precipitation. To compare the impact of all-sky assimilation from these two types of channels on typhoon forecasts, additional sensitivity experiments were conducted. Figure 11 illustrates the geopotential height increments from six assimilation experiments—three under clear-sky conditions (118_CLR, 183_CLR, 118_183_CLR) and three with all-sky assimilation (118_AS, 183_AS, 118_183_AS). Under clear-sky assimilation, the 118_CLR experiment (Figure 11b) produces a positive increment northeast of the typhoon center, which is not present in the other two clear-sky experiments. This positive increment becomes more pronounced in the all-sky assimilation experiments using 118 GHz (118_AS) and 118_183 GHz (118_183_AS) channels (Figure 11e,f). The enhancement of positive increments in the northeastern part of Typhoon Yagi contributes to the southward shift in the subtropical high, effectively limiting the northward deviation of the typhoon and correcting its track forecast. From the perspective of geostrophic adjustment theory, the positive impact of assimilating the observations from the 118 GHz temperature-sensitive channels can be further explained. In synoptic-scale motion, the wind field tends to adjust to the pressure field, and based on the hydrostatic balance and the ideal gas law, the distribution of pressure is closely related to the temperature field. Therefore, assimilating temperature channel observations can directly optimize the configuration of the pressure field, enhancing the simulation of large-scale atmospheric motions. Previous studies have also shown that temperature features have a more significant influence on synoptic-scale systems compared to moisture [44,45].

Temperature and moisture play critical roles by shaping the typhoon’s warm-core structure through latent heat release, thereby affecting intensity forecasts. The 118 GHz and 183 GHz channels are specifically designed to detect temperature and moisture, and their distinct sensitivities to atmospheric state variables and vertical levels may lead to different assimilation effects. To compare their impact, additional sensitivity experiments are conducted. Figure 12 and Figure 13 present the temperature and moisture analysis increments at various pressure levels from two all-sky assimilation experiments (118_AS and 183_AS), following the numerical ordering of the figure labels. The 118 GHz channel, mainly designed for temperature sensing, produces temperature analysis increments at all three examined levels, with the largest increment appearing at 300 hPa (Figure 12a), matching the peak levels of the weighting functions for channels 5 and 6. These results reflect the vertical sensitivity of the channel to temperature and are consistent with the weighting function profiles of the MWHS-II instrument. The temperature increments in the 118_AS experiment help bring the mid-to-upper tropospheric temperatures closer to observations, contributing to a better representation of the typhoon’s warm-core structure.

The 183 GHz channel, as a traditional water vapor absorption band, exhibits high sensitivity to humidity, which is evident in the moisture analysis increments across different levels in the 183_AS experiment (Figure 13). The increment is most pronounced in the mid-troposphere around 500 hPa (Figure 13e), corresponding to the peak of the weighting function for the 183 GHz frequency, as shown earlier in Figure 1. Although primarily designed to detect humidity, the 183 GHz channel also shows some sensitivity to temperature, especially under hydrometeor scattering conditions, resulting in slight temperature increments in the lower to mid-troposphere (Figure 12f). However, these increments are of small magnitude and may have a limited impact on forecasts. Meanwhile, mid-tropospheric water vapor serves as a key energy source for typhoon development. Through condensation into hydrometeors, especially cloud water, latent heat is released, thereby promoting storm intensification [46,47]. This sequence of physical processes highlights how moisture-sensitive channels, such as the 183 GHz band, enhance microphysical processes and provide favorable conditions for intensifying typhoons.

4.2.2. Forecasts Comparison

Figure 14a–d present the typhoon forecast results from three all-sky assimilation experiments: 118_AS, 183_AS, and 118_183_AS. As shown in Figure 14a,b, the 183_AS experiment produces the largest track errors, generally exceeding 120 km. The 118_AS experiment yields the smallest errors, and the 118_183_AS experiment falls in between. For both the 118_AS and 183_AS experiments, the track errors mostly remain below 120 km, with the 118_AS experiment achieving an additional average reduction of approximately 5 km compared to the 183_AS experiment. This comparison further supports the conclusion that the observations from the 118 GHz channel play a primary role in improving typhoon track forecasts by adjusting the dynamical structure in the initial field and reducing the northward deviation of the storm. In terms of intensity, the 118_183_AS experiment provides the best overall improvement in minimum sea level pressure and maximum wind speed forecasts (Figure 14c,d). This enhancement becomes more apparent during the rapid intensification of Typhoon Yagi around 1800 UTC on 6 September 2024. The mean SLP error in the 118_183_AS experiment is reduced by 0.13 hPa compared to 183_AS and by 0.67 hPa compared to 118_AS. For maximum wind speed, the mean error is reduced by 1.28 m/s compared to 183_AS and by 0.44 m/s relative to 118_AS. The combined assimilation observations from both channels increase the volume of assimilated observations and enhance the representation of water vapor and temperature structures in the initial field. This leads to improved simulation of thermodynamic and dynamic processes associated with the typhoon. Such improvement is particularly notable during the rapid intensification stage, where the joint assimilation further supports the formation of the warm-core structure and surrounding circulation, contributing to more accurate intensity forecasts.

Figure 15a,b presents the ETSs of 24 h accumulated precipitation forecasts from all channel assimilation experiments for the period from 0000 UTC 6 September to 0000 UTC 8 September 2024. The results show that the 183_AS experiment provides limited improvement over the 183_CLR experiment on 6–7 September and only begins to show some advantage on 7–8 September. Based on the analysis increments of specific humidity and the precipitation forecast results, the study concludes that although the 183 GHz all-sky assimilation improves the humidity field and contributes to better precipitation intensity prediction, this improvement cannot fully correct the displacement in rainfall caused by inaccurate typhoon track forecasts. As a result, the ETSs remain low. In contrast, the 118_AS experiment effectively improves the typhoon track forecast, leading to a more accurate simulation of rainfall distribution around the storm, which helps reduce the spatial error in precipitation and results in ETSs that are 8.7% (0.016) higher than those of the 183_AS experiment for precipitation thresholds above 10 mm. In the 118_183_AS experiment, the observations from 118 GHz channels ensure relatively accurate positioning of rainfall, while the added contribution of the 183 GHz channel helps refine the moisture field, enabling the model to improve the accuracy of precipitation intensity on top of the spatial accuracy. Under the combined influence of these two factors, the 118_183_AS experiment ultimately achieves the highest ETSs among all assimilation experiments, with higher ETS of 13.12% (0.024) over the 183_AS experiment and 4.10% (0.008) over the 118_AS experiment for precipitation thresholds above 10 mm. These findings demonstrate that the joint assimilation of the observations from both channels successfully leverages their respective strengths. The complementary, vertical sensitivities of the two channels enhance the representation of key atmospheric variables such as moisture and temperature in the initial conditions, which not only improves typhoon forecast accuracy but also strengthens the model’s ability to predict precipitation.

To further examine whether the temperature and humidity increments in the analysis field influenced the subsequent forecasts, the study also presents the RMSE profiles of wind, temperature, and specific humidity over the simulation domain at 24 h (Figure 16a–d) and 48 h (Figure 16e–h) forecast lead times after all-sky assimilation. As shown in Figure 16c,g, the forecast temperature RMSE of the 118_AS experiment is, respectively, 0.042 K and 0.079 K lower than that of the 183_AS experiment above 400 hPa, which is consistent with the earlier comparison of their respective analysis increments. Channels 5–6 of the 118 GHz band are sensitive to the middle and upper troposphere (Figure 1), and this is reflected in the strong positive temperature increments at high altitudes in the analysis field (Figure 12a), allowing the forecasts based on the 118_AS experiment analysis to remain closer to real atmospheric conditions. In addition, the 118_AS experiment yields zonal and meridional wind RMSE values that are 0.54 m/s and 0.46 m/s smaller than those of the 183_AS experiment at 24 h forecast and 0.52 m/s and 0.19 m/s at 48 h forecast (Figure 16a,b,e,f). These results indicate that the improved wind forecasts in all-sky assimilation are primarily driven by the assimilation of the observations from the 118 GHz channel. This channel not only improves the wind field but also effectively corrects the mid-to-upper level temperature field, thereby enhancing the representation of the typhoon’s thermal and dynamical structure and enabling the forecast to better capture the rapid intensification process. On the other hand, as a water vapor sounding channel, the 183_AS experiment shows superior performance in specific humidity forecasts, with RMSE values approximately 0.14 g/kg (24 h) and 0.08 g/kg (48 h) lower than those of the 118_AS case (Figure 16d,h), particularly in the lower and middle troposphere. This pattern is consistent with the moisture increments observed in the analysis fields (Figure 13e,f). The difference in forecast performance between the two channels is largely attributed to their distinct vertical sensitivities and the specific atmospheric variables they detect. These characteristics have a persistent impact on forecast outcomes. The observations from the 118 GHz channel excel in improving mid-to-upper pressure level temperature and wind fields, thus enhancing the simulation of the typhoon’s thermodynamic and dynamic structure, while the observations from the 183 GHz channel provide stronger constraints on low-level moisture, improving thermal conditions and contributing to typhoon intensification and rainfall prediction. The joint assimilation of the observations from both channels takes full advantage of their respective strengths: the observations from the 118 GHz channel enhance track prediction and warm-core structure through temperature and wind optimization in the upper layers, while the observations from the 183 GHz channel enhance moisture transport and latent heat release through improved humidity fields in the lower atmosphere. As a result, the combined assimilation strategy leads to coordinated improvements in both typhoon intensity changes and rainfall forecasts, effectively enhancing the overall forecast skill.

5. Discussion

As an exploratory study evaluating the all-sky assimilation of MWHS-II radiance from the FengYun-3F satellite for typhoon forecasting, we successfully demonstrated its potential capability in improving three track and intensity forecasts of the typhoon and revealed the influence of channel sensitivity on assimilation performance. However, some limitations remain in this study.

Firstly, the early forecasts after all-sky assimilation showed suboptimal performance in certain individual metrics. For instance, the early-period prediction of the minimum sea level pressure during typhoon development was not satisfactory. This may be explained from two aspects. On the one hand, surface emissivity has been demonstrated to interfere with the simulation of brightness temperatures, resulting in negative impacts on the assimilation of surface-related variables [48,49]. Given that all-sky assimilation incorporates a larger volume of observations, the impact of surface emissivity becomes more pronounced. On the other hand, imbalances introduced by the assimilation process may cause the model to take longer to stabilize, during which the forecasts can be highly unstable. This issue is becoming more evident in rapid-cycling assimilation systems [50].

Secondly, the all-sky symmetric observation error model for all-sky assimilation requires further improvement. The application of the all-sky symmetric observation error model helps to reduce the impact of representativeness errors and ensures the quality of data. However, the current all-sky symmetric observation error model still has room for further investigation and improvement. Factors such as the terrain characteristics, weather conditions, and the choice of the symmetric cloud factor may all influence the construction of the model. As a result, further developments of the all-sky symmetric observation error model are needed to meet the high-quality data requirements of all-sky assimilation.

Additionally, future work will focus on applying advanced assimilation algorithms, such as four-dimensional variational (4D-Var) assimilation, to the all-sky framework. By incorporating a time window, 4D-Var dynamically aligns observations with the background field, providing a better fit over time. Its ability to capture error growth through model dynamics makes it well-suited for evolving weather systems. The use of an advanced assimilation algorithm is expected to further enhance the effectiveness of all-sky assimilation.

6. Conclusions

This study developed the all-sky assimilation capacity for FengYun-3F MWHS-II radiance based on the WRFDA system. The impact of FY-3F MWHS-II all-sky assimilation on typhoon forecasts was systematically evaluated, with particular attention to the response characteristics of different channels to vertical layers and atmospheric state variables, as well as the forecasting implications of channel selection. The main findings are summarized as follows.

• Model forecasts based on the all-sky assimilation analysis of FY-3F MWHS-II radiance provided a more realistic simulation of Typhoon Yagi (2024) in closer agreement with observations. The all-sky assimilation experiment demonstrated a more robust performance compared to the clear-sky assimilation. The average errors in track, minimum sea level pressure, and max wind speed were substantially lower (31.8 km, 0.18 hPa, and 1.33 m/s), and the ETSs for the typhoon precipitation were 10–20% higher at thresholds above 10 mm. These results indicate that assimilating satellite radiance from cloud-covered regions improves forecast stability and accuracy.
• All-sky assimilation experiments using observations from different channels show varying impacts on forecasting the typhoon’s dynamic and thermal structures. Observations from the temperature-sensitive 118 GHz channel primarily improve the dynamical structure by adjusting the geopotential height field and suppressing the typhoon’s northward displacement. In contrast, the 183 GHz channel regulates moisture in the initial field, enhancing latent heat release and thermodynamic energy, which supports stronger intensity forecasts and increased rainfall.
• Joint all-sky assimilation of 118 GHz and 183 GHz channel observations leverages their complementary strengths, providing stronger constraints on atmospheric dynamics and moisture and improving initial analysis and forecast accuracy. Compared to the assimilation of individual channels, the joint all-sky assimilation reduced the minimum sea level pressure error by up to 0.67 hPa and the max wind speed error by up to 1.28 m/s. For precipitation thresholds above 10 mm, ETSs increased by 13.12% relative to the 183 GHz experiment and 4.10% relative to the 118 GHz experiment.