DART is an open-source assimilation package that has been developed by NCAR since 2002 for data assimilation development, research, and education. DART can be coupled with full-complexity Earth system components due to the flexible interfaces provided. In addition, the DART package provides the modules to convert observation data from a variety of native formats, e.g., the Binary Universal Form for the Representation of meteorological data (BUFR) format and the Hierarchical Data Format (HDF), into the input format specified for the DART system (Anderson et al., 2009; Raeder et al., 2012). The recent version of DART (version 9.11.13) is capable of using the RTTOV, a fast RTM, for assimilating visible, infrared, and microwave satellite observations. Provided in the RTTOV, many satellite instruments on board the GEO and LEO satellites are also supported in the DART assimilation package, but the hyperspectral infrared sounders, e.g., the Cross-track Infrared Sounder (CrIS) and the Infrared Atmospheric Sounding Interferometer (IASI), are excluded (Hoar et al., 2020). The main data assimilation technique provided by DART is the ensemble Kalman filter (EnKF) in which the forecast error covariance is estimated using short-range ensemble forecasts. The derived forecast error covariance is fully multivariate and depends on the synoptic situation.

CESM version 2 (CESM v2.1.0) is used as the model component of the ensemble data assimilation system. CESM2 is the latest generation of a coupled climate–Earth modeling system developed by NCAR, consisting of the atmosphere, land surface, ocean, sea-ice, land-ice, river, and wave models. These component models can be coupled to exchange states and fluxes (Hurrell et al., 2013; Kay at al., 2015). In this study, atmosphere and land component models are actively coupled, but the ocean component (sea surface temperature) and sea-ice coverage are specified by data read from files. As the atmosphere model of CESM2, CAM version 6 (CAM6) is an atmospheric general circulation model (AGCM) with the Finite Volume (FV) dynamical core (Danabasoglu et al., 2020). CAM6 provides the short-term forecast (6 h forecast) of the atmospheric state, which is used as the background state in the DART assimilation scheme. The land model is the Community Land Model version 5 (CLM5). The atmospheric variables are directly updated by the information derived from the observations ingested in the DART assimilation process, while the land state is affected interactively by the updated atmosphere state because the two component models are coupled. The two active models (CAM6 and CLM5) are run with a nominal 1∘ (1.25∘ in longitude and 0.95∘ in latitude) horizontal resolution. CAM6 has 32 vertical levels from the surface level to the top at 3.6 hPa (about 40 km).

The baseline observation data are obtained from the National Centers for Environmental Prediction (NCEP) Automated Data Processing (ADP) global upper-air and surface weather observations that are available from the NCAR Research Data Archive (NCAR RDA) (https://rda.ucar.edu/datasets/ds337.0/, last access: 28 December 2021). These data are produced in the PrepBUFR format for assimilation in the diverse NCEP NWP systems and mainly consist of ground-based observations and satellite-based wind retrievals. The ground-based observations include land and marine surface reports, aircraft reports, and radiosonde and pilot balloon (pibal) measurements, which are transmitted via the Global Telecommunication System (GTS) coordinated by the World Meteorological Organization (WMO). The satellite-based retrievals are provided by the National Environmental Satellite Data and Information Service (NESDIS). They include oceanic wind derived from the Special Sensor Microwave Imager (SSMI) and upper wind from the LEO and GEO satellites. As the NCEP ADP dataset is provided in the BUFR format, it must be converted to the data format available in the DART assimilation system, using the modules provided in the DART data assimilation package.

AMSU-A is the microwave temperature sounder that is currently on board diverse sun-synchronous satellite platforms e.g., MetOp satellites (MetOp-A, MetOp-B, and MetOp-C), three satellites of the National Oceanic and Atmospheric Administration (NOAA), and the National Aeronautics and Space Administration (NASA) research satellite Aqua. These three LEO satellite constellations provide near-global coverage, even in data assimilation that has a sub-daily assimilation window; NOAA satellites circle in an early-morning orbit (around 06:00 local time), MetOp satellites have a mid-morning orbit (around 09:00 local time), and Aqua has an afternoon orbit (around 13:00 local time). As a cross-track scanning sounder, the AMSU-A instrument has a total of 15 channels that consist of 12 channels (AMSU-A channels 3–14) over the 50–58 GHz oxygen (O2) absorption band and three window channels (AMSU-A channels 1, 2, and 15) at 23.8, 31.4, and 89 GHz. The instrument measures 30 pixels in each swath with a spatial footprint size of 48 km in nadir. The channels over the O2 absorption band mainly provide information about the vertical structure of tropospheric and stratospheric temperature (Mo, 1999; Goldberg et al., 2001). In this study, observations of AMSU-A instruments on board four LEO satellites (i.e., NOAA-19, Aqua, MetOp-A, and MetOp-B) are assimilated within the DART data assimilation system.

As each AMSU-A channel has distinct spectral characteristics, it is necessary to carefully choose the channels to be assimilated in the DART data assimilation system. First, the three AMSU-A channels at 23.8, 31.4, and 89 GHz (i.e., channels 1, 2, and 15), distributed over the window region of the microwave spectrum, are not assimilated. These three window channels are mostly affected by the emitted radiances from the surface under clear-sky conditions, so there is almost no information about the atmosphere. However, AMSU-A channels 1 (23.8 GHz) and 2 (31.4 GHz) are highly sensitive to clouds, so they are used for the quality control in which clouds are detected. In addition, even though the AMSU-A channels 3 (50.3 GHz) and 4 (52.8 GHz) are located over the O2 absorption band used for the temperature sounding, they have a strong sensitivity to the surface, so they are not used in DART. Considering that the upper parts of the weighting function of AMSU-A channels 12 (57.29 ± 0.322 ± 0.022 GHz), 13 (57.29 ± 0.322 ± 0.010 GHz), and 14 (57.29 ± 0.3222 ± 0.0045 GHz) are above the top of the atmosphere (i.e., 3.6 hPa) in the CAM6, these three channels are also removed to prevent vertical interpolation errors that may occur in the forward modeling using the RTM. This leaves channels 5–11 (53.596 ± 0.115, 54.4, 54.94, 55.5, 57.29, 57.29 ± 0.217, and 57.29 ± 0.322 ± 0.048 GHz) as the ones which may be assimilated.

As this study aims to assimilate the AMSU-A observations under clear-sky conditions, the cloud-affected channels are filtered out in the quality control step. In other words, the tropospheric channels (channels 5–7) whose peak of the weighting function is below 200 hPa are rejected if the AMSU-A pixel is determined to be a cloud-affected pixel. To determine this, we calculate the cloud liquid water (CLW) derived from observations of AMSU-A channels 1 and 2 over the ocean, using the retrieval methodology suggested by Grody et al. (2001). The CLW is defined as follows: 1CLW=cos⁡θD0+0.754ln⁡285.0-BT23-2.265ln⁡285.0-BT312D0=8.2402.622-1.846cos⁡θcos⁡θ, where θ is the satellite viewing zenith angle. BT23 and BT31 are the brightness temperature of AMSU-A channels 1 and 2, respectively. If the retrieved CLW is larger than 0.2 mm, the AMSU-A pixel is judged to be cloud-contaminated, and then the three tropospheric channels (channels 5–7) are rejected.

In this study, seven candidate AMSU-A channels (i.e., channels 5–11) are assimilated differently, depending on the surface type. Channels 5, 6, and 7 are the main tropospheric channels. Their weighting functions peak below 200 hPa but also have a bit of sensitivity to the surface because of the broad vertical shape of the weighting functions. Thus, the quality of the analysis can be degraded by assimilating the three tropospheric channels over the land and sea-ice types whose surface information (e.g., surface temperature and surface spectral emissivity) is uncertain. For this reason, AMSU-A channels 5–7 are not assimilated over the land and sea ice. To identify sea-ice area, the sea-ice index (SII) is retrieved from observations of AMSU-A channels 1 and 3 over the high-latitude region (poleward of 50∘), using the retrieval algorithm suggested by Grody et al. (1999). The SII is derived as follows: 3SII=2.85+0.20BT23-0.028BT50, where BT50 is the brightness temperature of AMSU-A channel 3. Three tropospheric channels are turned off if the SII is larger than 0.1 in the latitudes beyond 50∘. However, as the surface information over the ocean is relatively reliable, seven candidate AMSU-A channels are assimilated under clear-sky conditions. The AMSU-A channel list for DART is summarized in Table 1.

As an example, Fig. 2a and b present the spatial distribution of the CLW and the SII retrieved from AMSU-A on board NOAA-19 on 12 August 2014. It is found that many regions over the ocean are covered by cloud-related systems (CLW > 0.2 mm) and also that sea ice (SII > 0.1) exists near the North and South Pole regions. Observations of AMSU-A channel 5 over the cloud region and sea-ice areas are rejected (Fig. 2c). The channel selection process is also applied to the other two AMSU-A channels (channels 6 and 7), which are likely affected by clouds and sea ice. In the pre-trial runs, it was found that the analysis quality is degraded if the AMSU-A observations are assimilated over Antarctica during the Southern Hemisphere winter season. This seems to be due to the complex topography of the Antarctic continent, extreme cold weather conditions, and large errors in the numerical model. Thus, AMSU-A observations are not used over the high-latitude region (> 60∘ S) during the Southern Hemisphere winter season, in order to prevent the degradation of the analysis quality.

In addition to the inter-channel error correlation (refer to Sect. 5), spatial error correlation between the observations at a close distance also exists due to different representativeness of the observed radiances and the model state, and the uncertain quality control process such as cloud detection (Ochotta et al., 2005; Bormann and Bauer, 2010). Thus, the analysis is likely to be sub-optimized if highly dense observations are assimilated without considering the spatial error correlations. A common treatment to counteract the spatial error correlation is spatial thinning, which is widely used in data assimilation systems operated by the NWP centers. To choose the optimal spatial thinning distance, we performed four extra assimilation runs in which different spatial thinning distance (i.e., 96, 192, 288, and 384 km) was applied. Except for the spatial thinning distance, these pre-trial runs were set up with the same assimilation factors, i.e., the estimated bias correction coefficients (refer to Sect. 4.3), the estimated observation errors (refer to Sect. 5), and the localization half-width of 0.075 (refer to Sect. 6). These distances are multiples of the AMSU-A field-of-view (FOV) footprint size (∼ 48 km in nadir). The thinning interval of 288 km resulted in the largest analysis impact, so that distance was used to thin the observations in this study.

The biases mainly come from systematic errors: instrument calibration errors, inaccuracies of the RTM, and uncertain preprocessing (e.g., cloud detection errors). The biases tend to depend on the time of day and on the season as well as the instrument scan angle and air mass. While random errors are considered by defining the observation errors used in the assimilation process, the biases should be removed before assimilating the satellite observations. In these experiments, the biases are estimated using the time-averaged departures between the observed radiances and the simulated radiances from the spatiotemporally collocated model field (background) because of the absence of reference data suitable to compare the satellite observations (Scheck et al., 2018). The use of the simulated radiances from the model background (i.e., 6 h forecast) may be questionable because the model background could be biased. However, it is effectively impossible to find sufficient reference observations for comparing with these satellite observations, so the biases are calculated using the departures between the observed radiances and the model-simulated radiances. To estimate the systematic biases coming from diverse error sources, in this study, two bias correction processes are performed separately: scan-bias correction and air-mass-bias correction, using the statistical bias correction methods suggested by Harris and Kelly (2002).

As a cross-track microwave sounder, AMSU-A scans 30 FOVs per scan line, which are distributed symmetrically about the nadir. The scan angles of 30 FOVs range between ±48.33∘. Thus, the observed radiance varies depending on the scan angle, even though the observation point is the same. The variation of AMSU-A radiance is due to the change in the optical path length between the Earth and the satellite instrument, called the limb effect. The variation of radiance along with the scan angle can be simulated in the RTTOV embedded in DART. However, the mean first-guess departures between the AMSU-A-observed radiances and forward-modeled radiances still increase with an increasing scan angle on the center of two near-nadir FOVs (15 and 16) (Fig. 3a), meaning that the residual scan-angle-dependent biases exist for each AMSU-A channel. Thus, the scan-bias correction is required to correct the residual scan bias for each AMSU-A channel. In this study, the scan-bias correction is performed using the pre-computed residual scan bias for each AMSU-A channel. There are two steps to estimate the residual scan bias for AMSU-A channels assimilated. First, the mean bias of the departure between the AMSU-A-observed radiances and forward-modeled radiances for each FOV is made with the data assimilation results derived from the pre-trial run. The pre-trial run was set up with the spatial thinning of 96 km (refer to Sect. 4.2) and the default localization half-width (0.15, refer to Sect. 6). The instrument noise errors were used as the observation errors within DART. Second, as the scan bias derived from the departures between the observed radiances and forward-modeled radiances likely includes the air-mass bias, the averaged residual scan bias is obtained by removing the mean bias of two near-nadir FOVs (15 and 16) from the bias for each FOV (1–30). In addition, as shown in Fig. 3b, it is also found that the residual scan biases have different patterns depending on the latitude band for AMSU-A channel 6 (not shown for other channels), suggesting that the use of globally averaged scan bias is likely to deteriorate the quality of AMSU-A data assimilation. Thus, the residual scan (bscan) bias for each AMSU-A channel is subdivided into 14 latitude bands as follows: 4biscanθ,ϕ=y-Hxbiθ,ϕ-y-Hxbiθ=0,ϕ, where the subscript i denotes the AMSU-A channel number (i=1,2,…,15), θ is the satellite scan angle, ϕ is the latitude band at an interval of 10∘ in the latitudes below 60 and 30∘ in the latitudes beyond 60∘, y is the AMSU-A radiance, xb is the background model state, and H is the observation operator. Prior to the air-mass-bias correction, the observed brightness temperatures of each AMSU-A channel are corrected using the estimated scan-bias coefficients.

The air-mass bias (bairmass) is predicted using the multivariate regression method. The biases are mainly due to uncertainties in the RTM, which tend to vary with the air-mass and surface characteristics. The predictors, used in the regression method, come from the model variables (i.e., 1000–300 hPa thickness, 200–50 hPa thickness, and surface temperature) that include information on air-mass and surface characteristics. The predictors regress to the first-guess departure between the satellite radiances and forward-modeled radiances as follows: 5biairmass=βi,0+∑j=1Nβi,jpi,j, where βi,0 indicates the constant component of bias bi, and βi,j denotes the bias correction coefficients of the predictor pi,j. The subscripts i and j denote the AMSU-A channel number and the predictor number (i.e., j=1,2…N), respectively.

For the tropospheric AMSU-A channels (channels 5–7), the air-mass bias is estimated with two model variables (i.e., 1000–300 hPa thickness and surface temperature) because the peak of the channel weighting function is positioned below the 200 hPa pressure level, and these channels have a bit of sensitivity to the surface. However, 200–50 hPa thickness is only employed for other upper-tropospheric and stratospheric AMSU-A channels (channels 8–11) whose peak of the weighting function is above 200 hPa. As the biases fluctuate with time, it is reasonable to update the regression coefficients and an intercept point periodically, rather than using the climatological-based coefficients that are estimated using the long-term model outputs. In this study, at each data assimilation cycle, the regression coefficients and an intercept point for each AMSU-A channel are computed using DART outputs for the last four cycles and then used to predict the air-mass biases. As shown in Fig. 4, the histograms of the first-guess departures of the MetOp-B channels 5–11 show a positive bias and a Gaussian distribution if the AMSU-A observations are not bias-corrected. In particular, channels 5 and 6 have a large positive bias of 1.0–1.5 K. However, the positive biases are almost removed through the bias correction process, meaning that the bias correction scheme works well (Table 2).

As the same conventional radiosonde measurements were assimilated in the two trial runs (i.e., CNTL and AMSU-A), the first-guess departure statistics between the radiosonde measurements and the spatiotemporally collocated background states (6 h forecast) can be used to assess the impact of the AMSU-A observations on the short-range forecast. Figure 6 shows the vertical structure of the standard deviation (SD) of the first-guess departure from the radiosonde temperature, zonal wind, and meridional wind as well as the number of the radiosonde measurements used.

For the temperature, the first-guess departure errors are significantly reduced below 300 hPa for the AMSU-A runs as compared with the errors for the CNTL run (Fig. 6a). Because the AMSU-A channels provide vertical information about the air temperature, the temperature error reduction is the direct impact derived by assimilating the AMSU-A observations in the AMSU-A run. In addition to the radiosonde temperature, the first-guess departure errors decrease for the two wind components (i.e., zonal and meridional winds) (Fig. 6b and c). In particular, the SDs of the two winds at the 200 hPa level are reduced by up to about 3.9 m s-1 in the AMSU-A run, compared to the error of about 4.1 m s-1 for the CNTL run. As the model background error covariance includes the multivariate correlation between different model parameters (e.g., temperature and winds), a change in one model parameter can change another model parameter in the assimilation process. In addition, model parameters are linked in the governing equations and the physical parameterizations, which are embedded in the CAM6. That is, the change in one parameter results in the adjustment of another parameter in the model time integration. Thus, the error reduction of the wind components is the indirect impact of the improved temperature field by assimilating the AMSU-A observations.

In addition to the first-guess departure analysis of radiosonde, the assimilation impact of the AMSU-A observations can be diagnosed by comparing the first-guess departures of the AMSU-A with the analysis departures between the AMSU-A observations and the model analysis state. In general, if the observations are successfully assimilated, the SD of the analysis departure is smaller than that of the first-guess departure because the background fields are improved by assimilating the observations. As shown in Fig. 7, the SDs of the analysis departure are significantly smaller than those of the first-guess departure for AMSU-A-assimilated channels (channels 5–11), regardless of the satellite platforms, meaning that the AMSU-A observations have a positive analysis impact. In particular, the gap between the SDs of two departures is large for the stratospheric AMSU-A channels (channels 9–11).

To assess the impact of the AMSU-A observations on the analysis derived from the DART data assimilation system, the analysis errors are computed between the DART analysis and the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis version 5 (ERA5) as the reference data. As the ERA5 is produced through the assimilation of all available observation data in the ECMWF data assimilation system and provides consistent maps without spatial gaps, the ERA5 is employed to assess the model-derived output. For four primary atmospheric parameters (i.e., 500 hPa geopotential height, temperature, zonal wind, and meridional wind), the departures between the DART ensemble-mean analysis and the ERA5 are computed. Then bias and standard deviation are derived from the long-term departures. In particular, the error of 500 hPa geopotential height is widely used to assess the overall performance of the model-derived output because large-scale atmospheric motion in the middle troposphere (500 hPa) is closely linked with lower-level atmospheric motion.

Figure 8 describes the mean bias and SD of 500 hPa geopotential height for the CNTL and AMSU-A run, depending on the latitudinal regions. Detailed error values are described in Table 3. For two trial runs, overall negative mean bias occurs. However, the bias difference varies depending on the latitudinal regions. Over the tropics and the Northern Hemisphere (30–90∘ N), two trial runs have a similar negative bias. However, over the Southern Hemisphere (30–90∘ S), the CNTL run has a larger negative bias than the bias for the AMSU-A run. Thus, similar global mean bias for two trial runs is caused by the offsetting between regionally different bias patterns.

Considering that the geopotential height is a primary function of the average air temperature between the surface and the pressure level, we assumed that the model temperature has a cold bias at least below the 500 hPa pressure level. As expected, it is found that a negative bias is presented in the temperature field for both two trial runs (not shown). In addition, as shown in Fig. 9, the first-guess departure of the radiosonde temperature for the two trial runs has large positive values, implying that a cold bias exists in the model temperature fields (6 h forecast). In Raeder et al. (2021), it was noted that the CAM6/DART-derived reanalysis has a cold bias in the troposphere. However, it is still unclear as to why the CAM6-based temperature fields have a cold bias. The bias issue in CAM6 will be an interesting study in future work.

Even though the AMSU-A observations, including the temperature information, are additionally assimilated in the AMSU-A run, the AMSU-A run has a negative temperature bias that occurs in the CNTL run. It is related to the bias correction applied to the AMSU-A observations in DART. As mentioned in Sect. 4.3, the AMSU-A radiances are corrected by eliminating the biases based on the departure between the observed radiances and the forward-simulated radiances from the model background field. In addition, in this study, the bias correction coefficients were even updated at each cycle, using the DART outputs from the last four cycles. Thus, the information on the model bias is included in the biases derived from the correction scheme, which gradually fits the observations to the model background over the sequent assimilation cycles. As a result, the model bias still exists in the AMSU-A run as well as the CNTL run.

However, the global-mean SD of 500 hPa geopotential height for the AMSU-A run is reduced to about 45 m as compared with the SD (about 57 m) for the CNTL run, meaning that the 500 hPa geopotential height predictions are improved by assimilating the AMSU-A observations (Table 3). In particular, the error is largely reduced over the Southern Hemisphere. That is, the analysis impact is more significant in the Southern Hemisphere. It is consistent with the consensus that the assimilation impact of satellite observations is larger in the Southern Hemisphere, where the conventional data are sparse (Terasaki and Miyoshi, 2017; Yamazaki et al., 2023). In contrast, over the tropics, the error reduction is relatively smaller than over the Northern Hemisphere. In the tropics, the analysis error (about 14 m) is quite small for the CNTL run, as compared with the large errors of about 42 and 83 m in the Northern Hemisphere and Southern Hemisphere, respectively. Following Judt (2020), it was demonstrated that the tropical atmosphere has longer predictability than the extratropical atmosphere. Thus, the AMSU-A observations are conservatively assimilated in the tropics due to the small forecast errors, leading to less analysis impact.

It is noted that the AMSU-A assimilation impact is neutral in the high-latitude region (> 60∘ S) over the Southern Hemisphere. In contrast, in the high-latitude region (> 60∘ N) over the Northern Hemisphere, the assimilation impact is significant. It is because the AMSU-A observations were not assimilated in the high-latitude region (> 60∘ S) over the Southern Hemisphere during the Southern Hemisphere winter season when the trial runs were conducted (mentioned in Sect. 4.1), resulting in the neutral analysis impact. It is still a challenging issue to assimilate the satellite radiances over the Antarctic continent because of the complex topography, extreme weather condition, and large errors in the numerical model. In particular, as the conventional observations are quite sparse in the high-latitude region, the forecast errors are relatively larger than the other latitudinal regions (i.e., the tropics and midlatitude region, shown in Fig. 10a). In addition, the trial period (11 August–30 September 2014) is the Southern Hemisphere winter season when the Antarctic continent was under extremely cold weather conditions. In fact, in the pre-trial run, we found that the analysis field was degraded near the Antarctic continent by assimilating the AMSU-A observations. Thus, to prevent the analysis degradation, the AMSU-A observations were rejected over the high-latitude region (> 60∘ S) in the Southern Hemisphere. The assimilation of the AMSU-A observation in the Antarctic region will be handled in future work.

Figure 11 shows the normalized difference of SD of temperature, zonal wind, and meridional wind between the AMSU-A run and CNTL run, depending on the latitudinal regions (i.e., global, Northern and Southern Hemisphere, and tropics). The SD difference is normalized by the SD for the CNTL run. A negative value means that assimilating the AMSU-A observations provide analysis benefit. In contrast, a positive value indicates that the analysis error increases for the AMSU-A run compared with the error for the CNTL run, implying a negative analysis impact of the AMSU-A observations.

For the temperature, the global-mean analysis errors are significantly reduced in the whole troposphere and lower stratosphere for the AMSU-A run, as compared with the CNTL run. Large error reduction occurs in the lower stratosphere (-29 % and -24 % in 100 and 200 hPa, respectively), which is consistent with the large gap between the SDs of the first-guess departure and the analysis departure for the stratospheric AMSU-A channels (channels 9–11) whose peak of the weighting function is above 200 hPa (shown in Fig. 7). Similar to the results of the 500 hPa geopotential height, a strong error reduction mainly occurs in the Northern and Southern Hemisphere, where the errors reduce up to about 25 % and 28 % in the 500 hPa pressure level (Fig. 11a). The error decrease trends are consistent with the trends of the first-guess departure errors of the radiosonde temperature measurements in which a significant error decrease occurs in the 500 hPa layer (Fig. 6a). However, in the lower stratosphere (100 hPa pressure level), the analysis error decreases up to about 51 % in the Southern Hemisphere.

For two wind components (i.e., zonal and meridional winds), similar to the results of the temperature, the global-mean analysis errors for the AMSU-A run overall decrease in the whole troposphere and lower stratosphere (Fig. 11b and c). It is noted that the magnitude of the error decrease tends to increase with height, reaching about -17 % in the 100 hPa level for the zonal and meridional wind. Moreover, most analysis impact is made in the Northern and Southern Hemisphere, except in the 100 hPa level, where the maximum error decrease occurs in the tropics. However, over the tropics, the analysis errors for the AMSU-A runs are larger than the errors for the CNTL run in the middle and lower troposphere. In the model humidity field, a positive analysis impact only occurs in the Southern Hemisphere (not shown) but is not as significant as the abovementioned parameters (i.e., 500 hPa geopotential height, temperature, and winds). As a further study, we plan to assimilate the Microwave Humidity Sounder (MHS), providing information on the vertical structure of humidity so that the initial condition of model humidity is improved.