-
The WRF hybrid DA system is based on the 3DVAR framework by including the extended control variables a (Lorenc, 2003). The traditional 3DVAR is framed to provide an analysis increment
${{x}}{'}$ with the following cost function,where
${{{J}}_1}$ is associated with the static covariance matrix${{B}}$ and${{{J}}_{\rm{o}}}$ is the observation term that is associated with observation error covariance matrix${{R}}$ .${{{y}}'_{\rm{o}}} = {{{y}}_{\rm{o}}} - {{H}}\left({{{{x}}_{^{\rm{b}}}}} \right)$ is the innovation vector. Here${{{y}}_{\rm{o}}}$ and${{{x}}_{\rm{b}}}$ is the observation vector and the background state vector respectively.${{H}}$ is the observation operator, while${{H}}$ is the linearized observation operator. For the hybrid En3DVAR system, a sum of two terms is the final analysis increment${{x}}'$ , described asThe first term
${{x}}{'_1}$ in Eq. (2) represents the increment associated with the static BEC in 3DVAR and the second term of Eq. (2) is a linear combination of the extended control variable${{{a}}_k}$ ($k = 1, \cdots,K$ ) with the kth ensemble perturbations${({{{x}}_k})_{\rm{e}}}$ . Symbol. denotes the element-by-element product of the vectors. With the necessity of the ensemble covariance localization, the coefficients of${{{a}}_k}$ vary in space as a vector. Otherwise, the coefficients of${{{a}}_k}$ can be represented by scalars (Lorenc, 2003) in the absence of any localization. The above-mentioned increment${{x}}{'_1}$ and the extended control variable${{{a}}_k}$ are obtained by minimizing the following cost function for hybridwhere
${{{J}}_{\rm{e}}}$ is associated with the ensemble covariance that is used to constrain the extended control vector${{a}}$ . A is applied for the spatial correlation as the block diagonal matrix. The two coefficients,$\,{\beta _1}$ and$\,{\beta _2}$ , determine corresponding weights prescribed to the flow-dependent ensemble covariance and static covariance (Wang et al., 2008), with the constraint as, -
The GMI 1b radiance data are assimilated into the WRFDA system for both 3DVAR and hybrid methods in this study. GMI is a microwave radiometer with 13 channels, ranging from 10 GHz to 183 GHz (Table 1). The first 9 channels are standard microwave imager channels sensitive to precipitation and total column water vapor. Channel 8–9 at 89.0 GHZ are sensitive to convective rain areas. Channels 10–13 are responsible for detection of light precipitation and snowfall. In this study, only channels 3–7 are chosen to be assimilated carefully. It has been proven that raw radiance observations thinned to a grid with 2–6 times the model grid resolution are able to remove the potential error correlations between adjacent observations (Schwartz et al., 2012). A thinning mesh with 90 km is determined as an initial attempt to the assimilation of GMI radiances data.
Channel Frequency/GHz Polarisation Footprint/km 1,2 10.65 V, H 19.4×32.2 3,4 18.7 V, H 11.2×18.3 5 23.8 V 9.2×15.0 6,7 36.5 V, H 8.6×15.0 8,9 89.0 V, H 4.4×7.3 10,11 166 V, H 4.4×7.3 12 183±3 V 4.4×7.3 13 183±7 V 4.4×7.3 Table 1. GMI sensor characteristics
The Community Radiative Transfer Model (CRTM; Liu and Weng, 2006) coupled within the WRFDA was applied as the observation operator for GMI radiances. The temperature and humidity information from the model states are essential inputs for CRTM to calculate the simulated brightness temperature. The procedures of quality control and bias correction were conducted before data assimilation. For quality control: 1) Radiance data over mixed surfaces or with large bias were rejected. 2) Radiance observations were rejected if the retrieved level-2 cloud water liquid path (CLWP) exceeded the threshold listed in Table 2. The CLWP thresholds refers to those in Yang et al. (2016) and Kazumori et al. (2008). The systematic biases from the observed radiances were corrected before assimilation with 7 predictors (Liu et al., 2012; Xu et al., 2013) using the variational bias correction (VarBC) scheme. The applied predictors are the scan position, the square and cube of the scan position, the 200–50 hPa and 1000–300 hPa layer thicknesses, total column water vapor, and surface skin temperature. The quality control procedure works effectively for the criteria by checking the GMI observations after the quality control. In addition, the bias correction scheme was able to remove the systematic bias for the typhoon cases in our current study (not shown). The observation errors calculated offline are listed in Table 2 with GMI observations samples over 0000 UTC 1 July 2014 to 1200 UTC 21 July 2014. The statistics of the observation error is obtained by estimating the standard deviation between the observed and the simulated brightness temperature.
Channel Observation error
(Units: K)CLWP threshold
(Units: kg m−2)3 1.30 0.30 4 1.65 0.30 5 1.63 0.25 6 1.30 0.10 7 2.67 0.10 Table 2. Observation error and quality control thresholds
-
Four typhoon cases are employed in this study to validate the impact of GMI data assimilation with the hybrid method. The first case is Typhoon Matmo (2014) and the second case is Typhoon Chan-hom (2015). The other two cases are Meranti (2016) and Mangkhut (2018). The case Matmo (2014) is selected for the detailed comparison of the 3DVAR and the hybrid method. These typhoon cases are selected since they are effectively observed by the GMI radiance data.
From the record of the China Meteorological Administration (CMA), Matmo (2014) is the 10th typhoon, which occurred in the Western North Pacific Ocean. It made landfall in eastern Taiwan at 1600 UTC 22 July 2014 and then made its second landfall along the China coast near Fujian Province with the MSW reaching 30 m/s at 0700 UTC 23 July 2014. The landfall location was approximately 100 km away from Quanzhou Bay. Subsequently, Matmo (2014) passed through Fujian and Jiangxi Provinces, and continued northward to Shandong Province. Under the influence of Matmo (2014), heavy rainstorms occurred in northwest and southeast Quanzhou. Over its inland path, Matmo (2014) brought heavy precipitation, causing severe damage to 10 provinces in China.
Chan-hom (2015) was reported as the strongest TC landfall in Zhejiang Province since 1949. On 1 July, Chan-hom (2015) was clarified as a severe tropical storm. Early on 2 July, Chan-hom (2015) began to turn to the west-southwest with increasing intensity. Late on 9 July, Chan-hom (2015) reached its peak strength with estimated winds of 165 km/h and minimum sea level pressure of 935 hPa. Chan-hom (2015) made its landfall in Zhoushan, Zhejiang Province on 11 July around 0840 UTC.
Typhoon Meranti (2016) was one of the most powerful tropical cyclones on record and caused extensive damage to the Batanes in the Philippines, Taiwan, as well as Fujian Province in September 2016. Similarly, Typhoon Mangkhut (2018) was an extremely intense and catastrophic tropical cyclone that impacted Guam, the Philippines and South China in September 2018.
-
All experiments were conducted with the WRF (Skamarock et al., 2008), which is a compressible and non-hydrostatic atmospheric model in three dimensions. A single domain was applied with 57 vertical levels and a model top at 10 hPa for all experiments. The horizontal grid spacing was 15-km for all cases. For the physics parameterizations, the Kain-Fritsch cumulus parameterization (Kain and Fritsch, 1990; Kain, 2004) with a modified trigger function (Ma and Tan, 2009) and the WRF Single-Moment 6-Class microphysics scheme (Hong et al., 2004) were applied along with the Yonsei University (YSU) boundary layer scheme (Hong et al., 2006) and the 5-layer thermal diffusion model for land surface processes scheme. For the radiation scheme, the MM5 shortwave radiation scheme (Dudhia, 1989) and the Rapid Radiative Transfer Model (RRTM) longwave radiation scheme (Mlawer et al., 1997) were utilized.
-
For Typhoon Matmo (2014), three experiments were configured to evaluate the impact of assimilating GMI radiance data with the 3DVAR and the hybrid method on the subsequent forecasts in Table 3. The 3d-gts experiment assimilates only conventional observations from the operational Global Telecommunication System dataset in the National Centers for Environmental Prediction (NCEP) with the traditional 3DVAR method (Fig. 1a). The 3d-gmi experiment not only assimilates the conventional observations but also assimilates the GMI radiance data (Fig. 1b). Similar to the 3d-gmi experiment, h-gmi experiment employs the hybrid method with 40 ensemble members using the mean of the ensemble forecasts as the background.
Experiment Description 3d-gts GTS data using 3DVAR 3d-gmi GTS and GMI data using 3DVAR h-gmi GTS and GMI data using the hybrid method Table 3. List of experiments
Figure 1. (a) The distribution of observations from 1400 UTC 21 July to 1800 UTC 21 July. The numbers of each observation are marked on the right, (b) The GMI observations at 1600 UTC 21 July 2014. The red typhoon signals show the best track from 1800 UTC 21 July 2014 to 1200 UTC 24 July 2014 for Typhoon Matmo (2014).
Both 3DVAR and hybrid DA experiments were initialized using the NCEP operational 0.5º
$ \times $ 0.5º degree GFS analysis data as the initial and lateral boundary conditions. The initial conditions for Matmo (2014) are valid at 0600 UTC 21 July 2014. For 3DVAR, the background for DA is the 10 h spin-up forecast from 0600 UTC 21 July to 1600 UTC 21 July. Similarly, the initial ensemble members at 0600 UTC 21 July were generated by adding Gaussian random perturbations to the GFS analysis for the hybrid DA experiments. The Gaussian perturbations were drawn based on the static BECs (Torn et al., 2006). The h-gmi experiment employs the hybrid method using the ensemble mean as the background, and 10-h ensemble forecasts were launched to generate the ensemble members at 1600 UTC 21 July for the hybrid experiments. A 68-h deterministic forecast was launched at 1600 UTC 21 July by the analysis in 3DVAR and hybrid experiments, respectively.For the other three typhoons cases, only the two experiments 3d-gmi and h-gmi were conducted for each case. The analysis time for Chan-hom (2015) and Meranti (2016) are at 1800 UTC 9 July 2015 and at 0000 UTC 12 September 2016, respectively. For Mangkhut (2018), the valid time for the analysis is at 1800 UTC 15 September 2018.
With the limited ensemble members, horizontal and vertical localizations were applied to reduce spurious correlations caused by sampling error with a 750 km horizontal localization radius. The vertical localization scheme was based on an empirical function that considered the distance between two levels and the model height-dependent localization radius (Shen et al., 2017). The full 100% weight was prescribed to the ensemble-based BEC for the hybrid experiments. Observations within ±2 h were applied to the analysis time. The static BEC statistics used in the 3DVAR were derived based on the “NMC” method from the differences between 24-h and 12-h forecasts (Parrish and Derber, 1992) by using the WRFDA utility (Barker et al., 2012) for five control variables (velocity potential, stream function, unbalanced temperature, surface pressure and relative humidity).
Channel | Frequency/GHz | Polarisation | Footprint/km |
1,2 | 10.65 | V, H | 19.4×32.2 |
3,4 | 18.7 | V, H | 11.2×18.3 |
5 | 23.8 | V | 9.2×15.0 |
6,7 | 36.5 | V, H | 8.6×15.0 |
8,9 | 89.0 | V, H | 4.4×7.3 |
10,11 | 166 | V, H | 4.4×7.3 |
12 | 183±3 | V | 4.4×7.3 |
13 | 183±7 | V | 4.4×7.3 |