Impact of AMSU-A and MHS radiances assimilation on Typhoon Megi (2016) forecasting

2.1 WRF model and WRFDA system

The WRF [29] Model version 3.8.1 is used in this prediction. It is a non-hydrostatic, limited-area primitive equation model with a variety of physical parameterization schemes. The physical options in the WRF used in this analysis are as follows: the WRF Single-Moment 6-class microphysics scheme [30], the New Kain-Fritsch cumulus convection parameterization scheme [31], MM5 similarity for surface layer physics [32], Noah land surface model for land surface processes [33], and the Yonsei University planetary boundary layer scheme [34]. The Rapid Radiative Transfer Model [35] and Dudhia scheme [36] were used for longwave and shortwave radiation, respectively. The model domain is centered 25°N and 130°E with a grid spacing of 30 km and 200 × 125 grid points (Figure 1). There were 35 vertical levels in the model, with the top of the atmosphere at 50 hPa.

Figure 1

Computational domain for the experiments.

The WRF three-dimensional variational (3D-Var) data assimilation (WRFDA) [37] system was used in this study. It is capable of assimilating data from a variety of observational platforms, including satellite-based radiance measurements. All DA experiments are carried out using the 3DVar method. The WRF 3DVar approach [38] is based on the minimization of a cost function defined as follows:

where x represents analysis field, x _b represents background field, y represents observation field, and B and R are background error (BE) covariance and observation error covariance, respectively. The observation field and the background field must be compared using observation operator H, which projects model variables into observation space. In the case of radiance assimilation, the community radiative transfer model (CRTM) [39], developed by the Joint Center for Satellite Data Assimilation (JCSDA) was used as the forward operator to simulate radiances. The CV3 default option in WRFDA 3DVAR used the National Meteorological Center method [40] to obtain BE covariance statistics for five control variables including the stream function ψ, the unbalanced velocity potential x _u, the unbalanced temperature T _u, the pseudo-relative humidity q, and the unbalanced surface pressure P _su. The averaged 24 and 12 h forecast differences that vary at the same time are used to approximate the BE covariances. Additionally, the default observation errors of each satellite channel provided by WRFDA were used in this study.

2.2 ATOVS data

ATOVS instruments are sensors on the NOAA series of polar-orbiting satellites. The AMSU-A is a multi-channel radiometer with 15 channels operating at frequencies ranging from 23.8 to 89.0 GHz that is primarily used to provide temperature profiles. The swath width of each channel is 2,343 km, with 30 pixels measured in each swath and a footprint of around 24 km radius at nadir [17]. The MHS is a five-channel cross-track scanning microwave radiometer of the ATOVS instrument suite that is designed to retrieve profiles of atmospheric water vapor. It operates in the 89–190 GHz frequency range. Its swath distance is approximately 1,920 km, and each scan line has 90 pixels across the earth view, corresponding to a circle of diameter approximately 16 km at nadir [41,42].

AMSU-A and MHS are currently onboard the NOAA (NOAA 15–19) and the MetOp (MetOp A-C) satellite series providing the measurements of the atmospheric temperature and humidity. The radiances measured by AMSU-A onboard NOAA-15 and -18, as well as MHS onboard NOAA-18, are used in this article.

4.1 QC procedures

The satellite data was subjected to a series of QC procedures, including channel selection, cloud, precipitation detection, data thinning, and bias correction. All of these QC techniques were aimed at eliminating radiances that were significantly different from CRTM simulations.

Channels 1–4 and 15 from AMSU-A were rejected because their weighting function peaks were near the Earth’s surface. Since the WRF model top is approximately 50 hPa, AMSU-A channels 10–14 in assimilation were omitted because they have a higher contribution at the model’s top levels. Channels 1 and 2 of MHS are two window channels that are most impacted by radiation from the Earth’s surface in clear sky situation, but also by the water vapor emission and ice scattering if clouds are present [42]. In summary, the AMSU-A channels 5–9 for NOAA-15 and 5–8 for NOAA-18, as well as MHS channels 3–5, were chosen in radiances assimilation experiments.

Due to the limb impact of the data, observations at limb scan locations (Field of views of 1–3 and 28–30, 1–8 and 83–90, for AMSU-A and MHS, respectively) are not utilized. Furthermore, precipitation-affected AMSU-A and MHS data were excluded by eliminating data where the absolute value of the window channel innovation exceeded a 3K threshold (for channels 1 and 15 of the AMSU-A and channels 1 and 2 of MHS, respectively). AMSU-A and MHS radiances affected by liquid and/or ice clouds were eliminated for liquid water path >0.2 g/m². To prevent potential correlations between horizontally adjacent observations, the raw AMSU-A and MHS radiance data are thinned using a 60 km thinning mesh. Table 2 lists the selected channels as well as the number of AMSU-A and MHS observations before and after QC. Following the completion of all QC operations, the proportions of data utilized in the DA system were reduced by 30–50% for various channels.

Figure 2 shows the coverages of AMSU-A channel 6 (a and c) and MHS channel 5 (e) brightness temperatures data assimilated on 1200 UTC on 25 September 2016, following thinning and quality checking. Figure 2b, d, and f also shows the simulated brightness temperatures of the respective channel derived using the CRTM model.

Figure 2

Brightness temperatures (K) of AMSU-A channel 6 from NOAA15 (a) and NOAA18 (c) and MHS channel 5 from NOAA18 (e) at the analysis time of 1200 UTC on 25 September 2016 after QC. The brightness temperatures of the relevant channel estimated by CRTM are shown in the right column (b, d, and f).

Before direct radiances assimilation, satellite data must be simulated by modeling the background field. However, the simulated bright temperature frequently differs from the observed bright temperature due to systematic errors in the observations, in the background or analytical fields, and in the radiative transfer model. Radiance bias correction is required for satellite radiance DA. With prediction parameters such as air mass characteristics and scan location, the variational bias correction [44] scheme is utilized in this study. By introducing the coefficients of the bias revision forecast operator as control variables into the assimilation system, the predictor coefficients are dynamically updated in real time to dynamically estimate the revised observation bias based on satellite observations and unbiased conventional observations across the full assimilation interval. Scatter graphs of observed (OBS) against CRTM simulated brightness temperatures at 1200 UTC on 25 September 2016, for channel 6 of AMSU-A from NOAA-15, -18, and channel 5 of MHS from NOAA-18 are shown in Figure 3. The mean (MEAN), standard deviation (STDV), and root mean square error (RMS) are given in the lower right corner of the figure, respectively.

Figure 3

Scatter plots of the OBS versus simulated brightness temperatures (K) for channel 6 of AMSU-A from NOAA-15 (the first line), NOAA-18 (the second line), and channel 5 of MHS (the third line) from NOAA-18 at the analysis time 1200 UTC on 25 September 2016. (a, d, and g) BAK (background) versus OBS without bias correction (no BC), (b, e, and h) BAK versus OBS with bias correction (with BC), and (c, f, and i) ANA versus OBS with bias correction (with BC).

In comparison to the bias before correction (no BC) for channel 6 of AMSU-A from NOAA-15/-18, the RMSE of background (BAK) with bias correction (with BC) dropped from 3.869/0.878 to 0.342/0.316 (by around 91.2t/64.0%). With radiance DA, the RMSE of analysis (ANA) was further lowered to 0.280/0.185 and the STDV to 0.241/0.183. The RMSE and STDV of ANA for channel 5 of MHS were lowered by roughly 36.7 and 32.7%, respectively, as compared to the data of BAK following BC.

Figure 4 shows the histogram distribution that corresponds to Figure 3. The peak of observation minus background (OMB) is at −0.1 to 0.2 K, suggesting that the bias correction minimizes the difference between the simulated bright temperature of the background and the observation field. Because of the assimilation effect, the WRFDA analysis field in the third column derived after assimilation of the satellite observations is closer to the measured brightness temperature than the background field. The peak of observation minus analysis (OMA) values was closer to 0 K than OMB in the second column.

Figure 4

Histogram distributions of brightness temperature bias for channel 6 of AMSU-A from NOAA-15 (the first line), NOAA-18 (the second line), and channel 5 of MHS (the third line) from NOAA-18. The first column (a, d, and g) is the histogram of the frequency distribution of observed brightness temperature minus background brightness temperature (OMB) before the bias correction, the second column (b, e, and h) the distribution of OMB after the bias correction, and the third column (c, f, and i) is the observed brightness temperature minus the simulated brightness temperature from the analyses (OMA) after the bias correction.

4.2 Impact on Typhoon Megi (2016) track and intensity forecast

The performance of the 60 h Typhoon Megi (2016) path and intensity forecasts (without and with AMSU-A and MHS DA) was compared to the best available data (BST) from the China Meteorological Administration (CMA) TC database (http://tcdata.typhoon.org.cn) [45]. Every 6 h, the track record and intensity data of typhoons are obtained, including the typhoon’s center position (latitude and longitude), maximum wind speed (MSW) (10 min average), and central sea level pressure (SLP).

Figure 5 shows the forecast track and average location errors for Typhoon Megi (2016) based on the six trials. The characters “BST,” “CTL,” “A15,” “A18,” “A58,” “M18,” and “AM18” in the figure represent the best typhoon observation data from the CMA, the control experiment without assimilation, the assimilation of NOAA-15 AMSU-A, the assimilation of NOAA-18 AMSU-A, the assimilation of both NOAA-15 and -18 AMSU-A, the assimilation of NOAA-18 MHS, and the assimilation of combined AMSU-A and MHS onboard NOAA-18, respectively.

Figure 5

Forecast track (a) and mean track error (b) of Typhoon Megi (2016) simulated initial at 1200 UTC on 25 September 2016 for different experiments.

As can be observed in Figure 5a, all of these studies showed a similar trend in track prediction when compared to the best route of Typhoon Megi (2016). Overall, the tests A15, M18, and CTL have a south twist, whereas the other three have a north twist to some extent. During the early prediction periods, the Typhoon locations are quite near to the best track measurements. The track errors of all trials, on the other hand, grow with the projected lead time.

We calculated the track errors for the six assimilation experiments (Table 3) and plot them as a function of forecast hour in Figure 5b for a more thorough analysis. When compared to the CTL run (Bias of 103.76 km and RMSE of 123.95 km), the data demonstrate that virtually all DA tests helped to reduce prediction track errors. During the 60 h model integration, the projected track from the A15 experiment (Bias of 60.26 km and RMSE of 69.01 km) matches more closely with the best track data than the other five trials. However, the A18 and A58 experiments using NOAA-18 AMSU-A radiance assimilation simulated the typhoon’s first landfall better than the A15 trial. When comparing the assimilation results of AMSU-A and MHS onboard NOAA-18, the A18 simulations’ track and landfall position are closer to observation than M18 for the whole 60 h simulation. The typhoon track in M18 has the minimum error, with a Bias of 31.77 km and an RMSE of 41.84 km, before 30 h of model integration, followed by A15, which is closer to the optimal track than CTL. However, after 30 h of integration, M18’s track mistakes skyrocket, eventually overwhelming the other five studies. The predicted inaccuracy increases dramatically after 60 h.

In brief, MHS assimilation has a greater favorable influence in short-range prediction (30 h lead time) than in long-range prediction (beyond 60 h) and NOAA-15 AMSU-A assimilation leads to the best results for overall performance.

Similarly, in Figure 6, we depict the predicted minimum sea level pressure (MSLP) and MSW as a function of forecast lead time. The errors of forecast MSLP and MSW from various experiments over time in the 60 h are also shown in Figure 6 and Table 4. The MSLP and MSW were predicted to be of similar strength in all experiments.

Figure 6

(a) MSLP, (b) forecast MSLP error (unit: hPa), (c) maximum surface wind (MSW), and (d) forecast MSW error (unit: m/s).

For the MSLP forecast, all experiments with radiance assimilation except M18 are closer to the best track data than the CTL experiment. According to RMSE, A15 has the fewest mistakes, with a Bias of 6.76 hPa and an RMSE of 8.19 hPa, followed by AM18, A18, A58, and CTL. M18 is the only one with a larger error than CTL. However, the MSLP error of M18 is less than CTL in the first 18 h of model integration, demonstrating that MHS assimilation is beneficial for a short lead time.

A15 outperforms A18 and A58, demonstrating that assimilation satellite data from several platforms are not always superior than data from a single satellite. A15 has the least mistakes overall, with a Bias of 7.80 m/s and an RMSE of 8.34 m/s, according to the MSW prediction.

However, M18 is superior to A15 for the first 30 h integration, but this preference flips beyond that time. This refers to the fact that assimilation of MHS offers a short-term advantage for maximium wind speed forecasting, but has a long-term detrimental influence (the last 30 h). The uncertainty of the impact of MHS needs to be further investigated. Overall, both AMSU-A and MHS satellite radiances assimilations have a positive impact on the track and intensity forecast by improving initial temperature and moisture conditions [17,46–48].

4.3 Analysis of the impact of assimilation on meteorological elements

In order to further understand the mechanism of assimilation on the path and intensity of Typhoon Megi (2016), the adjustment of fundamental meteorological components before and after assimilation, as well as the incremental analysis of each physical quantity field, were examined by minusing the regional average value of the background fields at each height with the corresponding analysis fields (Figure 7).

Figure 7

Distribution of the mean vertical deviation of regional average temperature (a, in °C), relative humidity (b, in %), latitudinal wind (c, in m/s), and longitudinal winds (d, in m/s) after DA.

For initial temperature, humidity, latitudinal, and longitudinal winds, as shown in Figure 7, the magnitude of adjustment for each element after assimilation occurs mostly above 500 hPa. This could be related to the vertical weighting distribution of the satellite information. MHS had a higher influence on upper-level relative humidity than AMSUA. The NOAA15 experiments test indicated mostly cooling and humidity reduction in the lower layers below 700 hPa, whereas the NOAA18 test had minimal effect in the lower layers, mainly showing cooling and humidity increase above 500 hPa.

Because the overall large-scale regional average does not directly reflect how the basic elements affect typhoon movement and intensity changes, analysis of the physical quantity fields that have a large impact on typhoon movement and intensity changes is required to investigate assimilation mechanisms. The difference between the analysis field after assimilation (i.e., the initial field of the assimilation) and the background field before assimilation (i.e., the initial field of the CTL test) is known as the assimilation analysis increment. The SLP field, 500 hPa height and wind field, and 200 hPa wind field are chosen for the assimilation increment analysis.

Figure 8 shows the distributions of the SLP field (a) and the analysis increment distributions for assimilation experiments (b–f). The red dot represents the observed typhoon position, whereas the green dot represents CTL simulated typhoon position. As shown in the analysis increment distributions for assimilation experiments (b–f), there is a clear negative zone around the typhoon’s central location, and a circular positive zone around the typhoon’s eye, with a further circular positive zone beyond the positive zone, indicating further strengthening after the moment of assimilation analysis. The positive zone on the southeast side of the typhoon eye area has higher values than the positive zone on the west side, and the pressure gradient force points to the northwest, indicating that the typhoon will move in a northwesterly direction. The positive regions of A18, A58, and AM18 are substantially larger than those of A15 and M18, indicating that the former is heading northwesterly. This is consistent with the simulated in the previous description, suggesting that the assimilation test is a better simulation of typhoon intensity and intensity trends at the initial moment than the CTL test.

Figure 8

SLP field (a) and the analysis increment distributions for assimilation experiments (b–f) at analysis time.

The 500 hPa geopotential height and wind field and the analysis increment distributions for assimilation experiments at analysis time are given in Figure 9. The shaded colors represent the geopotential height field (in gpm) and the vector arrows represent wind field. As shown in Figure 9a, the typhoon is bordered to the north by subtropical high pressure, and travels westward due to the easterly wind belt. The initial wind field was altered in all sets of assimilation testing by changing the 500 hPa wind field structure. Comparing the analysis increments (Figure 9b–f) of the assimilation experiments shows that the geopotential height and wind fields around the typhoon area are adjusted significantly more in A18 than in the other sets of experiments, and the incremental distribution pattern after the combined assimilation of A15 or M18 is more or less the same as in A18, indicating that the contribution of AMSUA on NOAA18 to the assimilation is more pronounced.

Figure 9

The 500 hPa geopotential height and wind field (a) and the analysis increment distributions for assimilation experiments (b–f) at analysis time.

Figure 10

The 200 hPa wind field (a, m/s) and the analysis increment distributions for assimilation experiments (b–f) at analysis time. Black dot: observed typhoon position; red dot: CTL simulated typhoon position.

The higher level westerly jet is also strongly connected to typhoon intensity. In general, if the upper-level westerly jet is located to the north of the typhoon, the outflow of the typhoon’s upper-level dispersion is aided by the strengthening of the upper-level westerly jet, which increases the wind-pump effects and helps the typhoon to grow and intensify more. As seen in Figure 9a, the wind speed on the north side of the typhoon is substantially greater than on the south side, and the wind speed inside the eye area of the typhoon simulated in the CTL run is low, below 12 m/s, but beyond the eye area, there are areas of strong winds exceeding 24 m/s.

Meanwhile, north of 25°N, there exists a zone of fast westerly winds with a central wind speed of 24 m/s or greater. The high westerly jet promotes the dispersion and outflow of the typhoon’s top levels, implying that the typhoon’s strength in the CTL test would be increased. Comparing the incremental analysis of assimilation tests (Figure 9b–f).

As can be seen, each test has negative incremental zones to the east and west and positive incremental zones to the north and south, while the positive incremental zone to the north just corresponds to the strong outflow zone to the north of the typhoon in Figure 9a, indicating the intensification of the typhoon. The north side of A15 and M18 has a substantially larger intensity of the positive incremental zone than the south side, suggesting that the typhoon strength simulated by A15 and M18 is stronger than the CTL test. Moreover, the positive northern incremental zone in M18 is higher than that of A15, indicating that the simulated typhoon intensity of M18 is higher than that of A15 (Figure 10). This shows that M18 has the best assimilation impact and the closest correction to the measured upper 200 hPa wind field at the early simulation time period, which is consistent with the assimilation experiments in Section 4.2.