Impact of Assimilating FY-3D MWTS-2 Upper Air Sounding Data on Forecasting Typhoon Lekima (2019)

In this study, the Fengyun-3D (FY-3D) clear-sky microwave temperature sounder-2 (MWTS-2) radiances were directly assimilated in the regional mesoscale Weather Research and Forecasting (WRF) model using the Gridpoint Statistical Interpolation (GSI) data assimilation system. The assimilation experiments were conducted to compare the track errors of typhoon Lekima from uses of the Advanced Microwave Sounding Unit-A (AMSU-A) radiances (EXP_AD) with those from FY-3D MWTS-2 upper-air sounding data at channels 5–7 (EXP_AMD). The clear-sky mean bias-corrected observation-minus-background (O-B) values of FY-3D MWTS-2 channels 5, 6, and 7 are 0.27, 0.10 and 0.57 K, respectively, which are smaller than those without bias corrections. Compared with the control experiment, which was the forecast of the WRF model without use of satellite data, the assimilation of satellite radiances can improve the forecast performance and reduce the mean track error by 8.7% (~18.4 km) and 30% (~58.6 km) beyond 36 h through the EXP_AD and EXP_AMD, respectively. The direction of simulated steering flow changed from southwest in the EXP_AD to southeast in the EXP_AMD, which can be pivotal to forecasting the landfall of typhoon Lekima (2019) three days in advance. Assimilation of MWTS-2 upper-troposphere channels 5–7 has great potential to improve the track forecasts for typhoon Lekima.


Introduction
The observations from satellite microwave temperature sounders play a key role in numerical weather prediction (NWP), typhoon monitoring, and climate change research [1,2]. Significant progress has been made in the NWP centers around the world regarding the direct assimilation of the satellite microwave and infrared radiances [3], and in the application of satellite observations for typhoon monitoring and forecasting. The AMSU-A clear-sky radiances are assimilated by a limited-area ensemble Kalman filter (EnKF) to improve forecast skills for tropical cyclone (TC) prediction [4]. The brightness temperatures from the Advanced Technology Microwave Sounder (ATMS) onboard the Suomi-NPP satellite were directly assimilated into the Hurricane Weather Research and Forecasting (HWRF) model, which improved the track and intensity forecasts of Hurricane Sandy (2012) [5]. Results from the assimilation of multiple satellite data also indicated that satellite radiances can improve the typhoon Chan-Hom (2015) forecast skills [6]. In conclusion, the incorporation of satellite data, especially microwave observations, has greatly improved the NWP forecast skills of the worldwide NWP centers [7][8][9].
Fengyun-3D (FY-3D), an operational satellite from the China Meteorological Administration (CMA), was launched in November 2017. There are 13 oxygen channels in the microwave temperature sounder-2 (MWTS-2) instrument onboard the FY-3D, which observe the atmospheric temperature from the surface to the upper stratosphere. However, the FY-3D MWTS-2 lacks the two window channels at 23.8 and 31.4 GHz (see Table 1), which are critical for retrieving the cloud liquid water path (CLW) [10] and for cloud detection in the Gridpoint Statistical Interpolation (GSI) data assimilation system [11]. In previous studies, the cloud detection methods for the MWTS-2 were developed mainly by setting the cloud products from other instruments [12,13]. Recently, a convenient cloud detection method was proposed by Yang and Weng (2019) [14] that uses the brightness temperatures of 23.8 and 31.4 GHz channels reconstructed from a machine learning method. The data from the FY-3C/D MWTS-2 and the microwave humidity sounder-2 (MWHS-2) were recently evaluated and assimilated into the Met Office operational global model. The results indicate that assimilating MWTS-2 and MWHS-2 can reduce the errors of the global forecasts [15]. Besides, FY-3D MWTS-2 channels have similar global O-B biases with those from NOAA-20 ATMS at the same frequencies, which suggests that MWTS-2 has a similar performance to ATMS [16]. Therefore, the FY-3D MWTS-2 can play a significant role in satellite data assimilation of the global/regional NWP model, tropical cyclone warm core retrievals [17], and detection of atmospheric gravity waves [18]. In order to further investigate the potential impact of assimilating FY-3D data into the regional NWP model, in this study observations from the MWTS-2 upper-troposphere channels were firstly assimilated into a regional model to quantify the impacts on typhoon forecast. Table 1. Channel number, frequency, noise equivalent differential temperature, bandwidth, and weighting function peak of the AMSU-A and FY-3D MWTS-2.

Channel
No. The remainder of this paper is organized as follows. Typhoon Lekima (2019) and model configurations are introduced in Section 2. The quality controls and data assimilation results are presented in Section 3. Finally, conclusions and discussion are given in Section 4.

Brief Description of Typhoon Lekima
Typhoon Lekima formed as a tropical depression (TD) at 0000 UTC on 2 August 2019 over the ocean to the east of the Philippines, and further strengthened into a typhoon (TY) around 1800 UTC on 6 August 2019. The best track of Lekima from the Regional Specialized Meteorological Center (RSMC), Tokyo, is shown in Figure 1. Lekima landed on the coast of Wenling City, Zhejiang Province, around 1800 UTC on 9 August 2019, and brought sufficient water vapor and abundant rainfall with an 85 kt (1 kt = 0.51 m/s) maximum sustained wind. Lekima caused severe province-wide rainstorms and disasters in Zhejiang, Jiangsu, and Shandong Provinces from 0000 UTC on 5 August to 1200 UTC on 17 August 2019 [19]. Therefore, it was essential to accurately forecast the landfall of Lekima two or three days in advance. phoon (TY) around 1800 UTC on 6 August 2019. The best track of Lekima from the Regional Specialized Meteorological Center (RSMC), Tokyo, is shown in Figure 1. Lekima landed on the coast of Wenling City, Zhejiang Province, around 1800 UTC on 9 August 2019, and brought sufficient water vapor and abundant rainfall with an 85 kt (1 kt = 0.51 m/s) maximum sustained wind. Lekima caused severe province-wide rainstorms and disasters in Zhejiang, Jiangsu, and Shandong Provinces from 0000 UTC on 5 August to 1200 UTC on 17 August 2019 [19]. Therefore, it was essential to accurately forecast the landfall of Lekima two or three days in advance.

Model and Experimental Setup
In this study, WRF version 3.8 [20][21][22] and GSI version 3.7 were used for regional forecast and data assimilation, respectively. For the data assimilation experiments, the three-dimensional variational (3D-Var) data assimilation (DA) method was adopted [23,24]. The forecast domain was centered at (25° N, 135° E) with a horizontal resolution WRF Domain and 1.5h DA window WRF Domain and 3h DA window

Model and Experimental Setup
In this study, WRF version 3.8 [20][21][22] and GSI version 3.7 were used for regional forecast and data assimilation, respectively. For the data assimilation experiments, the three-dimensional variational (3D-Var) data assimilation (DA) method was adopted [23,24]. The forecast domain was centered at (25 • N, 135 • E) with a horizontal resolution of 9 km. It is well known that WRF has a cumulus parameterization grey-zone between 1 and 10 km, which can lead to errors in small-scale convection systems [25,26], so the cumulus parameterization scheme was not used in this study. There were 51 Eta levels in the vertical direction for simulations from Eta = 1 to Eta = 0, and the WRF model maximum was set to 10 hPa. The WRF model was initialized by the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) analysis and forecasts with the spatial resolution of 0.5 • × 0.5 • [27]. The Duhia shortwave scheme [28], the rapid radiative transfer model (RRTM) scheme for longwave radiation [29], the Thompson microphysical scheme [30], the Noah land-surface model [31], and a three-dimensional (3D) scale-adaptive turbulent kinetic energy-based boundary layer scheme [32] were adopted in the WRF model (see Table 2). The advanced radiative transfer modeling system (ARMS) was Remote Sens. 2021, 13, 1841 4 of 12 coupled as an observation operator in the GSI to simulate the FY-3D MWTS-2 brightness temperatures [33]. The brightness temperatures of MWTS-2 channels 1-13 and the machinelearned 23.8 and 31.4 GHz were stored in the binary universal form for the representation of meteorological data as one data stream in the GSI system. Three experiments were conducted twice daily at 0600 and 1800 UTC, respectively, during the period from 5 August 2019 to 8 August 2019. Firstly, a model-free run without data assimilation (to be denoted as CTRL) was conducted. Then, two data assimilation experiments, EXP_AD and EXP_AMD, were carried out. The EXP_AD assimilated the conventional data and radiance data from NOAA-15, -18, -19, and the Meteorological Operations Platform A/B (MetOp-A/B) Advanced Microwave Sounding Unit-A (AMSU-A) channels 1-11 and 15 [34]. AMSU-A channels 12-14 were set by default not to be assimilated in the GSI because their weighting function peaks are above the WRF model maximum (10 hPa). The EXP_AMD is the same as EXP_AD except that it assimilated extra data from the FY-3D MWTS-2 upper-troposphere channels 5-7. In data assimilation experiments, a 3 h assimilation window was applied, instead of the 1.5 h window, because more MWTS-2 observations over the Northwest Pacific (Figure 1b) can be assimilated to improve the forecast skills for Lekima.

Quality Control
Latitudinal and scan-angle biases from instruments, calibrations, and models are sometimes larger than the atmospheric signal [35][36][37], which will cause a significant decrease in the quality of the analysis and forecast fields [38]. Therefore, a combined variational bias correction (VarBC) scheme was used to quantify and remove the latitudinal biases and scan-angle biases for the satellite radiances in the GSI. The quality control procedure of data thinning was also applied to the MWTS-2 data to reduce the correlated observation errors. A 120 km thinning box was selected for FY-3D MWTS-2 data thinning criteria, which is the same as for the Suomi-NPP ATMS [39]. A cloud detection method based on the retrieved CLW was carried out to remove the cloudy data points from the cloud-affected channels, MWTS-2 channel 4 (53.59 GHz), and channel 5 (54.4 GHz). Figure 2 shows the differences between the observations and the simulations (O-B) of the MWTS-2 channels 5, 6, and 7 at 0600 UTC on 7 August 2019 before and after bias corrections. Negative scan-angle biases of channel 5 were removed after the bias correction, and the cloudy data points around the typhoon Lekima were also rejected. The clear-sky mean O-B biases of FY-3D MWTS-2 channels 5-7 decreased from −1.8, −1.6 and −1.4 K to 0.27, 0.10 and 0.57 K, respectively, after bias corrections.
cloud-affected channels, MWTS-2 channel 4 (53.59 GHz), and channel 5 (54.4 GHz). Figure 2 shows the differences between the observations and the simulations (O-B) of the MWTS-2 channels 5, 6, and 7 at 0600 UTC on 7 August 2019 before and after bias corrections. Negative scan-angle biases of channel 5 were removed after the bias correction, and the cloudy data points around the typhoon Lekima were also rejected. The clear-sky mean O-B biases of FY-3D MWTS-2 channels 5-7 decreased from −1.8 , −1.6 and −1.4 K to 0.27, 0.10 and 0.57 K, respectively, after bias corrections.  Figure 3 shows the vertical root-mean-square errors (RMSEs) for the increments of temperature, water vapor mixing ratio, zonal wind (U), and meridional wind (V) at the initial time in the EXP_AD and EXP_AMD. The vertical RMSE value refers to the sensitivity of physical variables at different model levels after assimilating satellite radiance. The larger RMSE value implies the larger data impact on the model's initial condition.  Figure 3 shows the vertical root-mean-square errors (RMSEs) for the increments of temperature, water vapor mixing ratio, zonal wind (U), and meridional wind (V) at the initial time in the EXP_AD and EXP_AMD. The vertical RMSE value refers to the sensitivity of physical variables at different model levels after assimilating satellite radiance. The larger RMSE value implies the larger data impact on the model's initial condition. The RMSE values were slightly increased by EXP_AMD after adding FY-3D MWTS-2 channels 5-7 in data assimilation based on the EXP_AD. More changes were observed in the common physical variables by EXP_AMD, which resulted in slightly larger RMSE values. The RMSE values of increments for the water vapor mixing ratio in the EXP_AMD did not change much, compared with those in the EXP_AD, because there were only three MWTS-2 temperature sounding channels added to the EXP_AMD. Therefore, there was no significant improvement in forecast skills for typhoon intensity (not shown). However, the change of temperature led to the shift in the mass field, which resulted in the evolution of wind fields, especially zonal wind, from the 25th (~470 hPa) to the 34th (~210 hPa) model level. The shift in wind fields can directly affect the large-scale environmental steering flow of a typhoon, which will significantly influence the accuracy of the typhoon track forecasts. Therefore, there was no significant improvement in forecast skills for typhoon intensity (not shown). However, the change of temperature led to the shift in the mass field, which resulted in the evolution of wind fields, especially zonal wind, from the 25th (~470 hPa) to the 34th (~210 hPa) model level. The shift in wind fields can directly affect the large-scale environmental steering flow of a typhoon, which will significantly influence the accuracy of the typhoon track forecasts.

Track Forecasts
The large-scale environmental steering flow and the beta drift are the most critical factors for the motion of a tropical cyclone. To understand how the MWTS-2 upper-troposphere channels 5-7 affect the track forecast, the steering flows of EXP_AD and EXP_AMD were examined in this study. Firstly, the mean wind flow in deep layers is calculated by using the following formula [40,41]:

Track Forecasts
The large-scale environmental steering flow and the beta drift are the most critical factors for the motion of a tropical cyclone. To understand how the MWTS-2 uppertroposphere channels 5-7 affect the track forecast, the steering flows of EXP_AD and EXP_AMD were examined in this study. Firstly, the mean wind flow in deep layers is calculated by using the following formula [40,41]: u mean = 75u 300 +100u 400 +150u 500 +175u 600 +175u 700 +150u 850 825 v mean = 75v 300 +100v 400 +150v 500 +175v 600 +175v 700 +150v 850 825 (1) Then, the steering flow is defined as the average of the mean wind flows within 500 km of the typhoon center [41]. Figure 4a shows the three-day best track along with the forecast tracks of the CTRL, EXP_AD, and EXP_AMD from 0600 UTC on 7 August 2019. Only the EXP_AMD could forecast the landfall of typhoon Lekima, indicating its good performance in simulating the steering flow. Figure 4b,c shows the comparison of mean wind flows, wind speed (shaded) and calculated steering flows from the 48 h forecasts in the EXP_AD and EXP_AMD. The direction of the simulated steering flow changed from southwest in the EXP_AD to southeast in the EXP_AMD, which contributed to the northwest movement of typhoon Lekima. Figure 4a shows the three-day best track along with the forecast tracks of the CTRL, EXP_AD, and EXP_AMD from 0600 UTC on 7 August 2019. Only the EXP_AMD could forecast the landfall of typhoon Lekima, indicating its good performance in simulating the steering flow. Figure 4b,c shows the comparison of mean wind flows, wind speed (shaded) and calculated steering flows from the 48 h forecasts in the EXP_AD and EXP_AMD. The direction of the simulated steering flow changed from southwest in the EXP_AD to southeast in the EXP_AMD, which contributed to the northwest movement of typhoon Lekima.   cos (sin sin cos cos cos( )) DR where R represents the radius of the earth, and λ and φ are the longitude and latitude, respectively. In general, the EXP_AMD performed the best in the track forecast with the  Figure 5 shows the comparison of track errors of CTRL, EXP_AD, and EXP_AMD at eight different initial times. The track errors (D) are calculated by the distances between the model simulations (ϕ 2 , λ 2 ) and best tracks (ϕ 1 , λ 1 ) as follows: D = R · cos −1 (sin ϕ 1 sin ϕ 2 + cos ϕ 1 cos ϕ 2 cos(λ 2 − λ 1 )) (2) where R represents the radius of the earth, and λ and ϕ are the longitude and latitude, respectively. In general, the EXP_AMD performed the best in the track forecast with the leading time of 36 h to 72 h, followed by the EXP_AD, which indicated the good performance when the AMSU-A and MWTS-2 data were both assimilated into the model. The mean track errors and percentages of improvement, i.e., (EXP_AD-CTRL)/CTRL and (EXP_AMD-CTRL)/CTRL, are shown in Figure 6. There was no significant improvement from the initial time to the 30 h forecasts by the EXP_AD and EXP_AMD compared with that by the CTRL. However, the mean track errors, calculated by averaging track errors of eight different initial times from Figure 5

Conclusions and Discussion
In this study, the results show that the combined assimilation of the AMSU-A and MWTS-2 upper-troposphere channels 5-7 has great potential to improve the track forecast for typhoon Lekima. The main conclusions are as follows.
(i) Compared with the CTRL and EXP_AD, the EXP_AMD can significantly reduce the mean track errors by improving the zonal wind and meridional wind in the middle and upper layer for the analysis and forecasts. Assimilating FY-3D MWTS-2 upper-troposphere channels 5-7 shows added value for typhoon Lekima track forecasts.
(ii) The steering flow [42][43][44][45] plays an important role in the northwest movement of typhoon Lekima, which is conductive to forecasting typhoon Lekima's landfall three days in advance.
(iii) The VarBC scheme is effective to remove the latitudinal biases and scan-angle biases for the satellite radiances in the GSI. The clear-sky mean bias-corrected O-B values of FY-3D MWTS-2 channels 5, 6, and 7 are smaller than those without bias correction.
In this study, only the data from the MWTS-2 channels 5-7 were assimilated because there were considerable noise equivalent differential temperatures (see Table 1) in channels 8-13 (greater than 1.2 K). Moreover, the retrieved CLW can identify most cloudy data points (CLW > 0.05 mm), but some data points partially affected by the clouds were located near the edges of the cloud, which are hard to remove. Therefore, the data from the MWTS-2 channels 1-4 were not easily assimilated. In the future, we will assimilate the MWTS-2 channels 1-4 to improve the low tropospheric structures of the storm by removing the cloud-edge data points and optically thin clouds, based on the cloud detection method proposed by Niu and Zou (2020) [13].
Because only clear sky radiances from satellite temperature sounder channels were assimilated, there was no significant improvement in storm intensity forecasts in this study. In order to improve intensity forecasts, more effort should be made on assimilating observations within or surrounding the storm center, such as all-sky assimilation [46] and storm vortex initialization (VI) [47]. In the future, the data from the 183 GHz humidity channels of the microwave humidity sounder onboard the FY-3D will be assimilated into the GSI system to improve the performance in simulating the tropical cyclone threedimensional water vapor structures. Furthermore, other advanced assimilation methods, such as the ensemble Kalman filter or hybrid data assimilation methods, will also be used to further improve the tropical cyclone forecast skills.