Evaluation of Temperature and Humidity Proﬁles Retrieved from Fengyun-4B and Implications for Typhoon Assimilation and Forecasting

: Fengyun-4B (FY-4B) is the ﬁrst operational satellite from China’s latest generation of geostationary meteorological satellites. It is equipped with the Geostationary Interferometric Infrared Sounder (GIIRS), which is able to obtain highly accurate atmospheric temperature and humidity proﬁles through hyperspectral detection in long-and mid-wave infrared spectral bands. In this study, the accuracy of the FY-4B/GIIRS temperature and humidity proﬁle retrievals over two months is evaluated using radiosonde observations and ERA5 reanalysis data. We go a step further to investigate the impact of the satellite retrievals on assimilation and forecasts for Typhoons Chaba and Ma-on in 2022. Results reveal that the root-mean-square difference (RMSD) for the FY-4B/GIIRS temperature and humidity proﬁle retrievals were within 1 K and 1.5 g/kg, respectively, demonstrating high overall accuracy. Moreover, assimilating temperature and humidity proﬁles from FY-4B/GIIRS positively impacts model analysis and prediction, improving typhoon track and intensity forecasts. Additionally, improvements have been discovered in predicting precipitation, particularly with high-magnitude rainfall events.


Introduction
Satellite observation is known for its broad detection range and consistency, providing valuable data that complement conventional meteorological observation in regions including oceans, deserts, and plateaus.It is, therefore, a crucial component of the meteorological observation network [1,2].Satellite data are one of the central elements of the numerical weather prediction (NWP) system, and more than 90% of the data used in the European Centre for Medium-Range Weather Forecasts (ECMWF) operational system are satellite data [3].Studies demonstrate that the assimilation of satellite radiance data has significantly improved the forecasting skills of both global and regional NWPs [4][5][6].
Microwave (MW) and infrared (IR) sounders are two common spaceborne instruments used to obtain three-dimensional atmospheric state information.MW sounders can penetrate clouds and precipitation effectively but have a coarser spatial resolution than IR sounders.Hyperspectral IR sounders (HISs), having thousands of channels, are able to provide higher-resolution observations under clear-sky conditions, the characterization of the three-dimensional structure of the atmosphere in finer detail [7,8].The National Aeronautics and Space Administration (NASA) and the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) pioneered the development of a series of spaceborne HISs, including the Atmospheric IR Sounder (AIRS) on the Aqua satellite launched in 2002, the Cross-track IR Sounder (CrIS) on the Suomi National Polarorbiting Operational Environmental Satellite System Preparatory Project (SNPP) satellite launched in 2011, and the IR Atmospheric Sounding Interferometer (IASI) on the Meteorological Operational Satellite (MetOp) series launched in 2006, 2012, and 2018 [9][10][11].On board China's Fengyun-3D (launched in 2017) and Fengyun-3E (launched in 2021) satellites, a High-Spectral-Resolution IR Atmospheric Sounder (HIRAS) has also been put into operation [12].However, these HISs were operated on polar-orbiting satellites, which were not able provide a nearly continuous detection of weather systems over the same area.
In 2016 and 2021, China launched two second-generation geostationary meteorological satellites, respectively, Fengyun-4A (FY-4A) and Fengyun-4B (FY-4B), with the Geostationary Interferometric IR Sounder (GIIRS) on board.The GIIRS therefore became the first HIS onboard a geostationary orbiting meteorological satellite [13].It enables the high temporal and spatial resolution detection of weather systems over China and its surrounding areas and helps to obtain a more detailed atmospheric vertical structure, such as retrieving atmospheric temperature profiles for 1 km layers and moisture profiles for 2 km layers [14,15].Advancements such as the development of the fast radiative transfer model [16], the evaluation of bias characteristics [17], and the development of retrieval algorithms [18][19][20] have progressed the application of FY-4A/GIIRS data in NWP models [21][22][23].Atmospheric temperature and humidity profile products retrieved from spaceborne sounders make noteworthy contributions to weather monitoring and forecasting [24,25].Ma et al. [26] and Gao et al. [27] assessed temperature profile products from FY-4A/GIIRS using radiosonde observations.Feng et al. [28] investigated the influence of assimilating atmospheric temperature profiles obtained from FY-4A/GIIRS on typhoon analysis and forecast.They demonstrated that assimilated temperature retrievals generally improve the predictions of typhoon tracks, winds, and precipitation.Nevertheless, research pertaining to the temperature and humidity profiles of FY-4B/GIIRS remains relatively scarce, due to the limited research timeframe.
Compared to FY-4A, the GIIRS on FY-4B has a wider spectral range, higher spectral resolution in the long-wave IR band, and enhanced radiometric calibration accuracy and detection sensitivity [29].The temporal resolution of GIIRS has been improved from 2.5 h for FY-4A to 2 h for FY-4B, and the spatial resolution has been improved from 16 km to 12 km at nadir.To assess the precision of the FY-4B/GIIRS atmospheric temperature and humidity profile retrieval products and to analyze their impacts on data assimilation and NWP, this study evaluates the accuracy of the atmospheric temperature and humidity profiles retrieved from FY-4B/GIIRS and explores the impact of retrieval products on assimilation and forecasting with typhoon cases, using a regional numerical model.The paper is organized as follows.Section 2 presents the basic information about FY-4B temperature and humidity profile retrieval products and the evaluation of the retrieval accuracy.Section 3 describes the weather cases, the numerical experiment scheme, and the model configuration.The effects of assimilating GIIRS retrieval data for forecasting and the related mechanisms are elaborated upon in Section 4. Section 5 gives the summary and discussion.

Overview of Retrieval Products
The atmospheric temperature and humidity profiles retrieved from GIIRS used in this study are obtained using the neural network algorithm developed by the National Satellite Meteorological Center (NSMC) [30].This approach does not depend on the precision of prior information and provides more accurate products than the Dual-Regression method.The retrieval data share identical spatial and temporal resolution with the primary GIIRS data.A set of quality flags are attached to the retrieval data for the preliminary categorization of data quality.Flag "0" indicates invalid observations, which mainly includes observations contaminated with topography and precipitation.The rest of the observations are stratified by quality flags 1 to 4. Flags "1" and "2" indicate observations of clear skies and light cloudiness, respectively, while quality flags "3" and "4" show observations of medium cloudiness and heavy cloudiness, respectively.
Figure 1 shows the proportion of quality flags and the number of observations for the GIIRS temperature and humidity profile retrievals at various altitude levels, counted from 1 July to 31 August 2022.The retrievals are available at the 37 standard pressure levels, which range from 1000 hPa to 1 hPa.For the period under analysis, the quantity of observations increases with altitude, whereas the proportion of high-quality data (data with quality flags "1" or "2") decreases.The percentage of high-quality data above 550 hPa dropped to less than 90%, while the number of observations exceeded 10 million and continued to increase with altitude.

Overview of Retrieval Products
The atmospheric temperature and humidity profiles retrieved from GIIRS used in this study are obtained using the neural network algorithm developed by the National Satellite Meteorological Center (NSMC) [30].This approach does not depend on the precision of prior information and provides more accurate products than the Dual-Regression method.The retrieval data share identical spatial and temporal resolution with the primary GIIRS data.A set of quality flags are attached to the retrieval data for the preliminary categorization of data quality.Flag "0" indicates invalid observations, which mainly includes observations contaminated with topography and precipitation.The rest of the observations are stratified by quality flags 1 to 4. Flags "1" and "2" indicate observations of clear skies and light cloudiness, respectively, while quality flags "3" and "4" show observations of medium cloudiness and heavy cloudiness, respectively.
Figure 1 shows the proportion of quality flags and the number of observations for the GIIRS temperature and humidity profile retrievals at various altitude levels, counted from 1 July to 31 August 2022.The retrievals are available at the 37 standard pressure levels, which range from 1000 hPa to 1 hPa.For the period under analysis, the quantity of observations increases with altitude, whereas the proportion of high-quality data (data with quality flags "1" or "2") decreases.The percentage of high-quality data above 550 hPa dropped to less than 90%, while the number of observations exceeded 10 million and continued to increase with altitude.

Evaluation of Retrieval Accuracy
The accuracy of the temperature and humidity profiles retrieved from GIIRS during July and August 2022 was evaluated using radiosonde observations and the fifth-generation ECMWF reanalysis (ERA5) fields [31].FY-4B/GIIRS data are not assimilated by ERA5, so ERA5 and FY-4B/GIIR are independent of each other.The GIIRS retrievals obtained from NSMC were matched with the observations in the validation set.To ensure the accuracy of validation data, radiosonde observations that met specific quality standards (quality marker > 3) were filtered out before the assessment process.These observations were also deemed unsuitable for use in The National Centers for Environmental Prediction (NCEP) reanalysis processing.After filtration, we did not collect any humidity radiosonde observations above altitudes of 300 hPa.
Figure 2 shows the root-mean-square difference (RMSD) evaluation for GIIRS temperature and humidity retrieval from 1 July to 31 August 2022.The temperature profiles against ERA5 (Figure 2a) have greater accuracy at altitudes between 150 and 750 hPa, with RMSD substantially within 1 K for quality flags 1-2 and slightly higher for quality flags 3-4.This indicates that the retrievals have higher accuracy under clear skies or light cloudiness than medium or heavy cloudiness.The humidity profiles against ERA5 (Figure 2b) exhibit an increasing trend from top to bottom in RMSD, which substantially remained within 1.5 g/kg across all altitude levels, except for the data with quality flag 4.
The accuracy of the temperature and humidity profiles retrieved from GIIRS during July and August 2022 was evaluated using radiosonde observations and the fifth-generation ECMWF reanalysis (ERA5) fields [31].FY-4B/GIIRS data are not assimilated by ERA5, so ERA5 and FY-4B/GIIR are independent of each other.The GIIRS retrievals obtained from NSMC were matched with the observations in the validation set.To ensure the accuracy of validation data, radiosonde observations that met specific quality standards (quality marker >3) were filtered out before the assessment process.These observations were also deemed unsuitable for use in The National Centers for Environmental Prediction (NCEP) reanalysis processing.After filtration, we did not collect any humidity radiosonde observations above altitudes of 300 hPa.
Figure 2 shows the root-mean-square difference (RMSD) evaluation for GIIRS temperature and humidity retrieval from 1 July to 31 August 2022.The temperature profiles against ERA5 (Figure 2a) have greater accuracy at altitudes between 150 and 750 hPa, with RMSD substantially within 1 K for quality flags 1-2 and slightly higher for quality flags 3-4.This indicates that the retrievals have higher accuracy under clear skies or light cloudiness than medium or heavy cloudiness.The humidity profiles against ERA5 (Figure 2b) exhibit an increasing trend from top to bottom in RMSD, which substantially remained within 1.5 g/kg across all altitude levels, except for the data with quality flag 4.  The RMSD for the temperature and humidity profiles against radiosonde observations (Figure 2c,d) share fundamental qualities with the results obtained against ERA5, differing only in their respective values.
The evaluation of bias for GIIRS temperature and humidity retrieval is shown in Figure 3.The temperature profiles against ERA5 with quality flags 1-2 (Figure 3a) have a negative bias within 0.5 K below 800 hPa, and the humidity profiles (Figure 3b) show a negative bias within 0.3 g/kg below 900 hPa.Retrieval data with quality flags of 3-4 exhibit greater bias, especially below 500 hPa.
tions (Figure 2c,d) share fundamental qualities with the results obtained against ERA5, differing only in their respective values.
The evaluation of bias for GIIRS temperature and humidity retrieval is shown in Figure 3.The temperature profiles against ERA5 with quality flags 1-2 (Figure 3a) have a negative bias within 0.5 K below 800 hPa, and the humidity profiles (Figure 3b) show a negative bias within 0.3 g/kg below 900 hPa.Retrieval data with quality flags of 3-4 exhibit greater bias, especially below 500 hPa.The results validated against radiosonde observations (Figure 3c,d) show an analogous deviation distribution, albeit with differing values.Radiosonde observations present insurmountable difficulties when used to assess retrieval error characterization over the ocean due to their predominant distribution on land.Consequently, the results validated against ERA5 are selected here as the observation errors during assimilation.

Typhoon Cases
We evaluate two typhoon cases in this study (see Figure 4).Typhoon Chaba (2022) was generated at 0000 UTC 30 June 2022 in the central South China Sea.Chaba (2022) was characterized by structural asymmetry, leading to asymmetric distribution of wind and The results validated against radiosonde observations (Figure 3c,d) show an analogous deviation distribution, albeit with differing values.Radiosonde observations present insurmountable difficulties when used to assess retrieval error characterization over the ocean due to their predominant distribution on land.Consequently, the results validated against ERA5 are selected here as the observation errors during assimilation.

Typhoon Cases
We evaluate two typhoon cases in this study (see Figure 4).Typhoon Chaba (2022) was generated at 0000 UTC 30 June 2022 in the central South China Sea.Chaba (2022) was characterized by structural asymmetry, leading to asymmetric distribution of wind and precipitation under its influence.Meanwhile, it experienced a weak steering flow due to its distance from the Western Pacific Subtropical High, leading to variable tracks and challenging forecasts.Typhoon Ma-on (2022) evolved from a tropical depression in the maritime region east of the Philippines at 0600 UTC 21 August 2022, and subsequently moved into the South China Sea, where it underwent along with an intensification with a strong vertical wind shear.Ma-on (2022) caused serious disasters in the Philippines and brought precipitation and gales to southern China [32].
precipitation under its influence.Meanwhile, it experienced a weak steering flow due to its distance from the Western Pacific Subtropical High, leading to variable tracks and challenging forecasts.Typhoon Ma-on (2022) evolved from a tropical depression in the maritime region east of the Philippines at 0600 UTC 21 August 2022, and subsequently moved into the South China Sea, where it underwent along with an intensification with a strong vertical wind shear.Ma-on (2022)

Model Setups and Experimental Scheme
We used the advanced research version 4.2 of the Weather Research and Forecasting (WRF) Model and its data assimilation system (i.e., WRFDA) in this study.The initial and boundary conditions utilized are from the analysis and forecast fields of the NCEP Global Forecast System (GFS) [33].The horizontal resolution of the GFS dataset we used in this paper is 0.25° × 0.25°.The GFS system assimilated a significant number of meteorological observations from the Global Telecommunications System (GTS), including surface synoptic observations, METAR observations, ship observations, buoy observations, sound observations, aircraft reports, GPS refractivity, satellite observations, and satellite-retrieved winds.Triple-nested vortex-tracking grids were used in all experiments, as shown in Fig- ure 4. In this figure, the outermost and intermediate domains are shown, while the innermost domain dynamically trails the typhoon track.The horizontal grid resolutions for domains 1 to 3 are 27 km, 9 km, and 3 km, respectively.In addition, each of the three domains is equipped with 57 vertical levels, culminating in a model top of 10 hPa.
The physical parameterization adopted here follows Qin et al. [34]: the Tiedtke cumulus convection parameterization scheme [35], the Yonsei University boundary layer scheme [36], the Monin-Obukhov surface scheme, the Unified Noah Land Surface Model [37], the GCMs Rapid Radiative Transfer Model shortwave and longwave radiation scheme [38], and the WRF Single-Moment 6-class scheme [39].This set of physics options is suggested by the WRF for modeling tropical storms.
For each typhoon case, we conducted two experiments for comparative analysis.In the control (CTRL) experiment, no additional samples were assimilated due to the large number of observations already assimilated by the GFS analysis used as the initial field.

Model Setups and Experimental Scheme
We used the advanced research version 4.2 of the Weather Research and Forecasting (WRF) Model and its data assimilation system (i.e., WRFDA) in this study.The initial and boundary conditions utilized are from the analysis and forecast fields of the NCEP Global Forecast System (GFS) [33].The horizontal resolution of the GFS dataset we used in this paper is 0.25 • × 0.25 • .The GFS system assimilated a significant number of meteorological observations from the Global Telecommunications System (GTS), including surface synoptic observations, METAR observations, ship observations, buoy observations, sound observations, aircraft reports, GPS refractivity, satellite observations, and satellite-retrieved winds.Triple-nested vortex-tracking grids were used in all experiments, as shown in Figure 4.In this figure, the outermost and intermediate domains are shown, while the innermost domain dynamically trails the typhoon track.The horizontal grid resolutions for domains 1 to 3 are 27 km, 9 km, and 3 km, respectively.In addition, each of the three domains is equipped with 57 vertical levels, culminating in a model top of 10 hPa.
The physical parameterization adopted here follows Qin et al. [34]: the Tiedtke cumulus convection parameterization scheme [35], the Yonsei University boundary layer scheme [36], the Monin-Obukhov surface scheme, the Unified Noah Land Surface Model [37], the GCMs Rapid Radiative Transfer Model shortwave and longwave radiation scheme [38], and the WRF Single-Moment 6-class scheme [39].This set of physics options is suggested by the WRF for modeling tropical storms.
For each typhoon case, we conducted two experiments for comparative analysis.In the control (CTRL) experiment, no additional samples were assimilated due to the large number of observations already assimilated by the GFS analysis used as the initial field.The GIIRS temperature and QVAPOR (GIIRS_TQ) experiment assimilated the temperature and humidity profile data from FY-4B/GIIRS.Additionally, the following assimilation configuration was performed according to the error assessment in Section 2: observations with temperatures above 100 hPa and humidities above 300 hPa were discarded, and only data with quality flags 1 and 2 were retained.The observation errors are statistically given following the previous section, as shown in Figure 5.
The GIIRS temperature and QVAPOR (GIIRS_TQ) experiment assimilated the temperature and humidity profile data from FY-4B/GIIRS.Additionally, the following assimilation configuration was performed according to the error assessment in Section 2: observations with temperatures above 100 hPa and humidities above 300 hPa were discarded, and only data with quality flags 1 and 2 were retained.The observation errors are statistically given following the previous section, as shown in Figure 5.The assimilation experiment was performed to generate a 6 h forecast using the GFS analysis as the initial field, followed by three 6 h assimilation cycles.The time window for each analysis time was set to ±1 h, and a 48 h deterministic forecast was carried out for each analysis cycle.The assimilation period for Chaba began at 0000 UTC 1 July and ended at 1200 UTC 1 July, while that for Ma-on was from 1800 UTC 23 August to 0600 UTC 24 August.The event time configuration of the assimilation experiment is shown in Figure 6.The assimilation experiment was performed to generate a 6 h forecast using the GFS analysis as the initial field, followed by three 6 h assimilation cycles.The time window for each analysis time was set to ±1 h, and a 48 h deterministic forecast was carried out for each analysis cycle.The assimilation period for Chaba began at 0000 UTC 1 July and ended at 1200 UTC 1 July, while that for Ma-on was from 1800 UTC 23 August to 0600 UTC 24 August.The event time configuration of the assimilation experiment is shown in Figure 6.
ture and humidity profile data from FY-4B/GIIRS.Additionally, the following assimilation configuration was performed according to the error assessment in Section 2: observations with temperatures above 100 hPa and humidities above 300 hPa were discarded, and only data with quality flags 1 and 2 were retained.The observation errors are statistically given following the previous section, as shown in Figure 5.The assimilation experiment was performed to generate a 6 h forecast using the GFS analysis as the initial field, followed by three 6 h assimilation cycles.The time window for each analysis time was set to ±1 h, and a 48 h deterministic forecast was carried out for each analysis cycle.The assimilation period for Chaba began at 0000 UTC 1 July and ended at 1200 UTC 1 July, while that for Ma-on was from 1800 UTC 23 August to 0600 UTC 24 August.The event time configuration of the assimilation experiment is shown in Figure 6.7 displays the probability density functions (PDFs) of temperature and humidity OMB (i.e., the GIIRS observation minus the background fields) and OMA (i.e., the GIIRS observation minus the analysis fields) from the GIIRS_TQ experiments at various vertical levels in the Ma-on case.When the GIIRS retrievals are assimilated, PDFs of OMA exhibit a mean closer to zero and a minor standard deviation at all levels compared to OMB.Additionally, the probability density function curve appears smoother and more symmetric, resembling a better distribution, indicating that the 3DVAR analysis state becomes closer to the observations as the assimilation system effectively absorbs the GIIRS retrieval information.
GIIRS observation minus the analysis fields) from the GIIRS_TQ experiments at various vertical levels in the Ma-on case.When the GIIRS retrievals are assimilated, PDFs of OMA exhibit a mean closer to zero and a minor standard deviation at all levels compared to OMB.Additionally, the probability density function curve appears smoother and more symmetric, resembling a better distribution, indicating that the 3DVAR analysis state becomes closer to the observations as the assimilation system effectively absorbs the GIIRS retrieval information.It is evident from the spatial distribution of OMB and OMA at the 500 hPa level (Figure 8) that some difference exists between the model background state and the retrieved temperature and humidity, particularly over the continent, which decreases after assimilation.In addition, the temperature and humidity profiles retrieved from GIIRS are primarily distributed in the atmospheric regions surrounding the typhoon.The 3DVAR method can effectively adjust the temperature and humidity information in those areas.It is evident from the spatial distribution of OMB and OMA at the 500 hPa level (Figure 8) that some difference exists between the model background state and the retrieved temperature and humidity, particularly over the continent, which decreases after assimilation.In addition, the temperature and humidity profiles retrieved from GIIRS are primarily distributed in the atmospheric regions surrounding the typhoon.The 3DVAR method can effectively adjust the temperature and humidity information in those areas.These adjustments will impact the location and form of the weather system surrounding the typhoon and impact the typhoon's movement through factors such as steering flow.The relevant mechanisms are detailed in later sections.
These adjustments will impact the location and form of the weather system surrounding the typhoon and impact the typhoon's movement through factors such as steering flow.The relevant mechanisms are detailed in later sections.

Verification against ERA5
Figure 9 presents the three cycle-averaged RMSDs of the 48 h predicted relative humidity, temperature, and wind components of U and V for the Ma-on case in domain 1.The assimilated GIIRS temperature and humidity retrieval effectively suppresses error growth and exhibits lower RMSD.Additionally, this improvement is transferred to the U and V wind fields during the assimilation cycles and model integration (see Figure 9c,d).

Impacts of GIIRS Retrieval on Forecasts 4.2.1. Verification against ERA5
Figure 9 presents the three cycle-averaged RMSDs of the 48 h predicted relative humidity, temperature, and wind components of U and V for the Ma-on case in domain 1.The assimilated GIIRS temperature and humidity retrieval effectively suppresses error growth and exhibits lower RMSD.Additionally, this improvement is transferred to the U and V wind fields during the assimilation cycles and model integration (see Figure 9c,d).

Impact on Typhoon Track and Intensity Forecasts
Figure 10 shows the track forecasts of Typhoons Chaba and Ma-on, and the track forecast errors averaged three times for different experiments.The track error is defined as: where  and  are the latitude and longitude of the typhoon simulated by the model, which are automatically output by the WRF, and  and  are the latitude and longitude of the observed typhoon, which are obtained from the Best Track dataset provided by the China Meteorological Administration (CMA) [40,41].Track error in kilometers can be obtained using the internal functions of the NCAR Command Language (NCL).Both experiments successfully simulated Typhoon Chaba's change in motion from northwestward to northward.The GIIRS_TQ experiment partly corrected the westward shift seen in the CTRL experiment track, resulting in outcomes closer to those of the Best Track.The mean track forecast error (refer to Figure 10b) illustrated a consist-

Impact on Typhoon Track and Intensity Forecasts
Figure 10 shows the track forecasts of Typhoons Chaba and Ma-on, and the track forecast errors averaged three times for different experiments.The track error is defined as: where Lat model and Lon model are the latitude and longitude of the typhoon simulated by the model, which are automatically output by the WRF, and Lat OBS and Lon OBS are the latitude and longitude of the observed typhoon, which are obtained from the Best Track dataset provided by the China Meteorological Administration (CMA) [40,41].Track error in kilometers can be obtained using the internal functions of the NCAR Command Language (NCL).Both experiments successfully simulated Typhoon Chaba's change in motion from northwestward to northward.The GIIRS_TQ experiment partly corrected the westward shift seen in the CTRL experiment track, resulting in outcomes closer to those of the Best Track.The mean track forecast error (refer to Figure 10b) illustrated a consistently favorable influence from the assimilation of GIIRS temperature and humidity, except at the 48th hour.
In comparison to the CTRL experiment, the GIIRS_TQ experiment achieves an average decrease in track error of approximately 13.87%, with a maximum improvement of around 24.96% at the 18th hour.
Remote Sens. 2023, 15, x FOR PEER REVIEW 11 of 18 ently favorable influence from the assimilation of GIIRS temperature and humidity, except at the 48th hour.In comparison to the CTRL experiment, the GIIRS_TQ experiment achieves an average decrease in track error of approximately 13.87%, with a maximum improvement of around 24.96% at the 18th hour.The mean track forecast error in the Ma-on case (refer to Figure 10d) indicates that the GIIRS retrieval assimilation has a positive impact within an 18 h forecast timeframe, with a maximum forecast improvement of approximately 10.85%.Overall, in the two cases analyzed in this paper, the assimilation of GIIRS temperature and humidity retrieval products were shown to improve typhoon track forecasts and reduce the growth of forecast errors to some extent.
Environmental steering (advection) plays a crucial role in the motion of tropical cyclones.Figure 11 shows the 500 hPa steering flow for the Chaba case in the CTRL (blue vectors) and GIIRS_TQ (red vectors) experiments, along with their respective differences (green vectors) for various forecast periods.The calculation of the steering flow follows the method outlined by Wu et al. [42], which averages the wind vectors at the 500 hPa level within a 10° × 10° region centered on the tropical cyclone center.It is evident that the steering flows of both experiments at the initial moment indicate a northwestern direction.The mean track forecast error in the Ma-on case (refer to Figure 10d) indicates that the GIIRS retrieval assimilation has a positive impact within an 18 h forecast timeframe, with a maximum forecast improvement of approximately 10.85%.Overall, in the two cases analyzed in this paper, the assimilation of GIIRS temperature and humidity retrieval products were shown to improve typhoon track forecasts and reduce the growth of forecast errors to some extent.
Environmental steering (advection) plays a crucial role in the motion of tropical cyclones.Figure 11 shows the 500 hPa steering flow for the Chaba case in the CTRL (blue vectors) and GIIRS_TQ (red vectors) experiments, along with their respective differences (green vectors) for various forecast periods.The calculation of the steering flow follows the method outlined by Wu et al. [42], which averages the wind vectors at the 500 hPa level within a 10 • × 10 • region centered on the tropical cyclone center.It is evident that the steering flows of both experiments at the initial moment indicate a northwestern direction.Subsequently, the westward component decreases gradually, and the steering flow shifts towards the north, which corresponds to the track forecast results presented in Figure 10a.In addition, it is consistently observed that the steering flow for the GIIRS experiment includes an eastward component compared to the CTRL experiment, as indicated by the blue vectors, which corrects the westward deviation of the simulated track exhibited by the CTRL experiment in the fore-and mid-periods.
Remote Sens. 2023, 15, x FOR PEER REVIEW 12 of 18 Subsequently, the westward component decreases gradually, and the steering flow shifts towards the north, which corresponds to the track forecast results presented in Figure 10a.
In addition, it is consistently observed that the steering flow for the GIIRS experiment includes an eastward component compared to the CTRL experiment, as indicated by the blue vectors, which corrects the westward deviation of the simulated track exhibited by the CTRL experiment in the fore-and mid-periods.Figure 12 displays the central sea level pressure (CSLP) forecasts for Typhoons Chaba and Ma-on and the averaged CSLP forecast errors for different experiments.In the case of Chaba, both experiments overestimated the typhoon intensity.However, the GIIRS_TQ experiment forecasted the CSLP closer to the reference value during the typhoon's weakening phase (after 24 h; see Figure 12a).Furthermore, the GIIRS_TQ experiment had a minor average forecast error for CSLP during the forecast's mid-to late-periods (between 18 and 48 h; see Figure 12b).In the case of Ma-on, the GIIRS_TQ experiment simulated lower CSLP during the intensification phase of the typhoon that more closely matched the reference value, especially at its peak strength (between 18 and 48 h; see Figure 12c).In addition, similar to the Chaba case, the decrease in the mean forecast error of CSLP by the GIIRS_TQ experiment is primarily observed after 18 h, specifically during the weakening phase of the typhoon (see Figure 12d).Figure 12 displays the central sea level pressure (CSLP) forecasts for Typhoons Chaba and Ma-on and the averaged CSLP forecast errors for different experiments.In the case of Chaba, both experiments overestimated the typhoon intensity.However, the GIIRS_TQ experiment forecasted the CSLP closer to the reference value during the typhoon's weakening phase (after 24 h; see Figure 12a).Furthermore, the GIIRS_TQ experiment had a minor average forecast error for CSLP during the forecast's mid-to late-periods (between 18 and 48 h; see Figure 12b).In the case of Ma-on, the GIIRS_TQ experiment simulated lower CSLP during the intensification phase of the typhoon that more closely matched the reference value, especially at its peak strength (between 18 and 48 h; see Figure 12c).In addition, similar to the Chaba case, the decrease in the mean forecast error of CSLP by the GIIRS_TQ experiment is primarily observed after 18 h, specifically during the weakening phase of the typhoon (see Figure 12d).The intensity of the strengthening phase of a typhoon is significantly linked to it structure.Figure 13 displays the vertical cross-sections of the forecast fields across th center of Ma-on, including temperature and horizontal wind anomalies.Initially (see Fig ure 13a,d), the warm core structure of the typhoon is weak.There is no significant differ ence in temperature anomalies between the two experiments because GIIRS retrievals ar mainly located in the peripheral environmental field of the typhoon, which cannot directly impact the internal structure of the typhoon.The warm cores of both experiments becom stronger as the forecast time progresses, with differences between the two experiment becoming progressively more apparent.The CTRL experiment (see Figure 13b) and th GIIRS_TQ experiment (see Figure 13e) simulate two warm core structures at 0600 UTC 2 August.The warm core in the GIIRS_TQ experiment reaches a height of 13 km, which i higher than the 10 km in the CTRL experiment.
The warm core for both experiments further develops at 1800 UTC 24 August, when the CSLP of Ma-on reaches its lowest point in the Best Track data (see Figure 12c).Th GIIRS experiment shows a deeper and more compact warm core in contrast to the CTRL experiment (Figure 13f), with the upper part of the warm core extending up to 14 km and a straighter wind anomaly profile, indicating strong convection, which corresponds to th result in Figure 13c.Overall, although the GIIRS retrieval only provides temperature and humidity increments in the environmental field at the periphery of the typhoon, the effect can be gradually transferred to the inner regions of the typhoon as the model field is ad The intensity of the strengthening phase of a typhoon is significantly linked to its structure.Figure 13 displays the vertical cross-sections of the forecast fields across the center of Ma-on, including temperature and horizontal wind anomalies.Initially (see Figure 13a,d), the warm core structure of the typhoon is weak.There is no significant difference in temperature anomalies between the two experiments because GIIRS retrievals are mainly located in the peripheral environmental field of the typhoon, which cannot directly impact the internal structure of the typhoon.The warm cores of both experiments become stronger as the forecast time progresses, with differences between the two experiments becoming progressively more apparent.The CTRL experiment (see Figure 13b) and the GIIRS_TQ experiment (see Figure 13e) simulate two warm core structures at 0600 UTC 24 August.The warm core in the GIIRS_TQ experiment reaches a height of 13 km, which is higher than the 10 km in the CTRL experiment.

Impact on Precipitation Forecasts
Figure 14 illustrates the 24 h cumulative precipitation forecasts for the CT GIIRS_TQ experiments for Chaba and Ma-on against the reference values.Obser datasets (see Figure 14a,d) were obtained from the Global Precipitation Measurem gram precipitation data products [43,44].The precipitation in the Chaba case is p distributed in the northern region of the Beibu Gulf, Hainan Island, and the ocea face to the east of Hainan Island (see Figure 14a).The precipitation area of the GI experiment (see Figure 14c) is nearer to Hainan Island than that of the CTRL exp (see Figure 14b), which better agrees with the observations.Additionally, the GI experiment corrects the spurious convection in the lower left corner of the region o in the CTRL experiment and the overestimated precipitation to the east of 116°E.
The precipitation in the Ma-on case is primarily distributed over the south part of Guangdong Province, Hainan Island, and its northeastern sea surfa GIIRS_TQ experiment (Figure 14f) more accurately simulates the shape and loc the rainband in closer proximity to observations.Meanwhile, it effectively corr sporadic convection and false alarm range of 100 mm or more precipitation in the n portion of the South China Sea compared to the CTRL experiment.
Figure 15 shows the Threat Score (TS), the Equitable Threat Score (ETS), and t The warm core for both experiments further develops at 1800 UTC 24 August, when the CSLP of Ma-on reaches its lowest point in the Best Track data (see Figure 12c).The GIIRS experiment shows a deeper and more compact warm core in contrast to the CTRL experiment (Figure 13f), with the upper part of the warm core extending up to 14 km and a straighter wind anomaly profile, indicating strong convection, which corresponds to the result in Figure 13c.Overall, although the GIIRS retrieval only provides temperature and humidity increments in the environmental field at the periphery of the typhoon, the effects can be gradually transferred to the inner regions of the typhoon as the model field is adjusted during integration, ultimately impacting the typhoon's structure and intensity.

Impact on Precipitation Forecasts
Figure 14 illustrates the 24 h cumulative precipitation forecasts for the CTRL and GIIRS_TQ experiments for Chaba and Ma-on against the reference values.Observational datasets (see Figure 14a,d) were obtained from the Global Precipitation Measurement Program precipitation data products [43,44].The precipitation in the Chaba case is primarily distributed in the northern region of the Beibu Gulf, Hainan Island, and the oceanic surface to the east of Hainan Island (see Figure 14a).The precipitation area of the GIIRS_TQ experiment (see Figure 14c) is nearer to Hainan Island than that of the CTRL experiment (see Figure 14b), which better agrees with the observations.Additionally, the GIIRS_TQ experiment corrects the spurious convection in the lower left corner of the region observed in the CTRL experiment and the overestimated precipitation to the east of 116 • E.   The precipitation in the Ma-on case is primarily distributed over the southwestern part of Guangdong Province, Hainan Island, and its northeastern sea surface.The GIIRS_TQ experiment (Figure 14f) more accurately simulates the shape and location of the rainband in closer proximity to observations.Meanwhile, it effectively corrects the sporadic convection and false alarm range of 100 mm or more precipitation in the northern portion of the South China Sea compared to the CTRL experiment.
Figure 15 shows the Threat Score (TS), the Equitable Threat Score (ETS), and the Fractions Skill Score (FSS) for 24 h cumulative precipitation at different thresholds.The use of multiple scoring methods contributes to a more comprehensive assessment of precipitation forecast results.For the Chaba case (see Figure 15a-c), precipitation scores have improved for most thresholds from light rain (0.1-10 mm) to very heavy rainfall (>250 mm) due to GIIRS temperature and humidity retrieval assimilation.For the Ma-on case (see Figure 15d-f), GIIRS_TQ demonstrates a predominantly advantageous impact, particularly for precipitation thresholds under 10 mm and over 50 mm.

Summary and Discussion
Three-dimensional atmospheric temperature and humidity information of a high spatial and temporal resolution obtained from the GIIRS on board the Fengyun-4 satellite can help to improve the effectiveness of NWP.In this study, a cyclic assimilation and forecast experiment was conducted for two typhoon cases in 2022 based on the error assessment of FY-4B/GIIRS atmospheric temperature and humidity profile retrievals, and the impact of the GIIRS temperature and humidity retrievals on the forecast was investigated.
The assessment results for the atmospheric temperature and humidity profile retrievals against radiosonde observations and ERA5 demonstrate a low RMSD of approximately 1 K for altitude levels between 150 and 750 hPa for the temperature retrieval, and about 1.5 g/kg for the humidity retrieval for all levels.The analysis also demonstrates that observations impacted by medium and heavy cloudiness exhibit greater errors, primarily distributed above 300 hPa.
The results from the cyclic assimilation and forecasting experiments for the two typhoon cases demonstrate that, by assimilating temperature and humidity profiles retrieved from FY-4B/GIIRS, forecast errors in the temperature field, middle-and low-level humidity fields, and low-level wind field can effectively be reduced.Additionally, the assimilation improves the forecast for the typhoon track and intensity.Furthermore, the assimilation of FY-4B/GIIRS temperature and humidity retrievals effectively simulates the structure and location of rainbands, which are more closely aligned with the observations than the CTRL experiment.These yields improved scores in quantitative assessments of precipitation.
Although the GIIRS temperature and humidity retrieval data provide rich atmospheric three-dimensional information, it is worth mentioning that observations from cloudy regions were not assimilated due to their poor quality, which prevented detecting temperature and humidity structure inside the typhoon.Consequently, developing an all-sky retrieval algorithm for GIIRS is a promising avenue of research.Moreover, this paper only analyzes two typhoon cases, and more work needs to be carried out to confirm the effectiveness of FY-4B/GIIRS retrieval assimilation.In addition, FY-4B/GIIRS has a frequent detection rate, and its temporal resolution in China is 2 h.We plan to fully assimilate the GIIRS data at a higher frequency to maximize the product's potential.

Figure 1 .
Figure 1.The percentage and number of FY-4B/GIIRS retrievals with 4 types of QC flags at different atmospheric pressure layers.The statistics cover the period from 1 July to 31 August 2022.

Figure 1 .
Figure 1.The percentage and number of FY-4B/GIIRS retrievals with 4 types of QC flags at different atmospheric pressure layers.The statistics cover the period from 1 July to 31 August 2022.

Figure 3 .
Figure 3.The bias of (a) temperature and (b) humidity profile FY-4B/GIIRS retrievals against ERA5 and of (c) temperature and (d) humidity profile FY-4B/GIIRS retrievals against radiosondes.QF indicates the abbreviation for QC Flag.

Figure 3 .
Figure 3.The bias of (a) temperature and (b) humidity profile FY-4B/GIIRS retrievals against ERA5 and of (c) temperature and (d) humidity profile FY-4B/GIIRS retrievals against radiosondes.QF indicates the abbreviation for QC Flag.
caused serious disasters in the Philippines and brought precipitation and gales to southern China [32].

Figure 4 .
Figure 4.The experimental domains and the best tracks of Typhoons Chaba (ID: 2203) and Ma-on (ID: 2209) from the China Meteorological Administration (CMA).The shaded green areas plot the distribution of retrieval data for 00-02 UTC 1 July 2022.

Figure 4 .
Figure 4.The experimental domains and the best tracks of Typhoons Chaba (ID: 2203) and Ma-on (ID: 2209) from the China Meteorological Administration (CMA).The shaded green areas plot the distribution of retrieval data for 00-02 UTC 1 July 2022.

Figure 5 .
Figure 5. Vertical profiles of default observation errors (solid lines) and statistical observation errors (dashed lines) for (a) temperature and (b) humidity.

Figure 6 .
Figure 6.The event time configuration of the assimilation experiment.

Figure 5 .
Figure 5. Vertical profiles of default observation errors (solid lines) and statistical observation errors (dashed lines) for (a) temperature and (b) humidity.

Figure 5 .
Figure 5. Vertical profiles of default observation errors (solid lines) and statistical observation errors (dashed lines) for (a) temperature and (b) humidity.

Figure 6 .
Figure 6.The event time configuration of the assimilation experiment.Figure 6.The event time configuration of the assimilation experiment.

Figure 6 .
Figure 6.The event time configuration of the assimilation experiment.Figure 6.The event time configuration of the assimilation experiment.

4 .
Impacts of GIIRS-Retrieved Temperature and Humidity Profiles on Analysis and Forecasts 4.1.Analysis of OMB and OMA

Figure
Figure7displays the probability density functions (PDFs) of temperature and humidity OMB (i.e., the GIIRS observation minus the background fields) and OMA (i.e., the GIIRS observation minus the analysis fields) from the GIIRS_TQ experiments at various vertical levels in the Ma-on case.When the GIIRS retrievals are assimilated, PDFs of OMA exhibit a mean closer to zero and a minor standard deviation at all levels compared to OMB.Additionally, the probability density function curve appears smoother and more symmetric, resembling a better distribution, indicating that the 3DVAR analysis state becomes closer to the observations as the assimilation system effectively absorbs the GIIRS retrieval information.

Figure 7 .
Figure 7.The probability density functions (PDFs) of temperature OMB and OMA at (a) 850 hPa, (b) 500 hPa, and (c) 350 hPa and humidity OMB and OMA at (d) 850 hPa, (e) 500 hPa, and (f) 350 hPa from the GIIRS_TQ experiments at 0600 UTC 24 August 2022 for the Ma-on case (2022).N represents the assimilated retrieval counts.

Figure 7 .
Figure 7.The probability density functions (PDFs) of temperature OMB and OMA at (a) 850 hPa, (b) 500 hPa, and (c) 350 hPa and humidity OMB and OMA at (d) 850 hPa, (e) 500 hPa, and (f) 350 hPa from the GIIRS_TQ experiments at 0600 UTC 24 August 2022 for the Ma-on case (2022).N represents the assimilated retrieval counts.

Figure 8 .
Figure 8. Spatial distribution of (a) OMB and (b) OMA for temperature and of (c) OMB and (d) OMA for humidity at 500 hPa at 0600 UTC 24 August 2022 for the Ma-on case.Typhoon symbols show the location of the typhoon observation at the corresponding moment.

Figure 8 .
Figure 8. Spatial distribution of (a) OMB and (b) OMA for temperature and of (c) OMB and (d) OMA for humidity at 500 hPa at 0600 UTC 24 August 2022 for the Ma-on case.Typhoon symbols show the location of the typhoon observation at the corresponding moment.

Figure 9 .
Figure 9.The three cycle-averaged RMSDs of (a) temperature, (b) relative humidity, (c) zonal wind component, and (d) meridional wind component 48 h deterministic forecasts at different pressure levels (200, 500, and 850 hPa) for Typhoon Ma-on.The solid line and the dashed line indicate the CTRL experiment and GIIRS_TQ experiment, respectively.The RMSD is calculated against the ERA5 datasets.

Figure 9 .
Figure 9.The three cycle-averaged RMSDs of (a) temperature, (b) relative humidity, (c) zonal wind component, and (d) meridional wind component 48 h deterministic forecasts at different pressure levels (200, 500, and 850 hPa) for Typhoon Ma-on.The solid line and the dashed line indicate the CTRL experiment and GIIRS_TQ experiment, respectively.The RMSD is calculated against the ERA5 datasets.

Figure 10 .
Figure 10.Chaba's (a) 48 h track forecast started at 0000 UTC 1 July 2022 and (b) mean track errors, and Ma-on's (c) 48 h track forecast started at 1800 UTC 23 August 2022 and (d) mean track errors for CTRL experiment (blue), GIIRS_TQ experiment (red), and compared with the CMA Best Track data (black).

Figure 10 .
Figure 10.Chaba's (a) 48 h track forecast started at 0000 UTC 1 July 2022 and (b) mean track errors, and Ma-on's (c) 48 h track forecast started at 1800 UTC 23 August 2022 and (d) mean track errors for CTRL experiment (blue), GIIRS_TQ experiment (red), and compared with the CMA Best Track data (black).

Figure 12 .
Figure 12.Chaba's (a) central sea level pressure (CSLP) forecast started at 0000 UTC 1 July 2022 an (b) mean forecast errors, and Ma-on's (c) CSLP forecast started at 1800 UTC 23 August 2022 and (d mean forecast errors.The dotted lines indicate the Best Track data from CMA, the solid lines indicat the CTRL experiment and the dashed lines indicate the GIIRS_TQ experiment.

Figure 12 .
Figure 12.Chaba's (a) central sea level pressure (CSLP) forecast started at 0000 UTC 1 July 2022 and (b) mean forecast errors, and Ma-on's (c) CSLP forecast started at 1800 UTC 23 August 2022 and (d) mean forecast errors.The dotted lines indicate the Best Track data from CMA, the solid lines indicate the CTRL experiment and the dashed lines indicate the GIIRS_TQ experiment.

Figure
Figure 15d-f), GIIRS_TQ demonstrates a predominantly advantageous impact, particularly for precipitation thresholds under 10 mm and over 50 mm.

Figure 14 .
Figure 14.Twenty-four-hour cumulative precipitation from (a) observation, (b) CTRL, and (c) GIIRS_TQ for the case of Typhoon Chaba on 0000 UTC 2 to 0000 UTC 3 July 2022, and (d) observation, (e) CTRL, and (f) GIIRS_TQ for the case of Typhoon Ma-on on 1200 UTC 24 to 1200 UTC 25 July 2022.

Figure 14 .
Figure 14.Twenty-four-hour cumulative precipitation from (a) observation, (b) CTRL, and (c) GIIRS_TQ for the case of Typhoon Chaba on 0000 UTC 2 to 0000 UTC 3 July 2022, and (d) observation, (e) CTRL, and (f) GIIRS_TQ for the case of Typhoon Ma-on on 1200 UTC 24 to 1200 UTC 25 July 2022.

Figure 14 .
Figure 14.Twenty-four-hour cumulative precipitation from (a) observation, (b) CTRL, and (c) GIIRS_TQ for the case of Typhoon Chaba on 0000 UTC 2 to 0000 UTC 3 July 2022, and (d) observation, (e) CTRL, and (f) GIIRS_TQ for the case of Typhoon Ma-on on 1200 UTC 24 to 1200 UTC 25 July 2022.

Figure 15 .
Figure 15.(a) TS score, (b) ETS score, and (c) FSS score of 24 h cumulative precipitation prediction for Chaba case and (d) TS score, (e) ETS score, and (f) FSS score of 24 h cumulative precipitation prediction for Ma-on case.The black and gray bars represent the CTRL and GIIRS_TQ experiments, respectively.