A Study on Assimilation of CYGNSS Wind Speed Data for Tropical Convection during 2018 January MJO

: The National Aeronautics and Space Administration (NASA) Cyclone Global Navigation Satellite System (CYGNSS) mission was launched in December 2016. CYGNSS provides ocean surface wind speed retrieval along specular reﬂection tracks at an interval resolution of approximately 25 km. With a median revisit time of 2.8 h covering a ± 35 ◦ latitude, CYGNSS can provide more frequent and accurate measurements of surface wind over the tropical oceans under heavy precipitation, especially within tropical cyclone cores and deep convection regions, than traditional scatterometers. In this study, CYGNSS v2.1 Level 2 wind speed data were assimilated into the Weather Research and Forecasting (WRF) model using the WRF Data Assimilation (WRFDA) system with hybrid 3- and 4-dimensional variational ensemble technology. Case studies were conducted to examine the impact of the CYGNSS data on forecasts of tropical cyclone (TC) Irving and a westerly wind burst (WWB) during the Madden–Julian oscillation (MJO) event over the Indian Ocean in early January 2018. The results indicate a positive impact of the CYGNSS data on the wind ﬁeld. However, the impact from the CYGNSS data decreases rapidly within 4 h after data assimilation. Also, the inﬂuence of CYGNSS data only on precipitation forecast is found to be limited. The assimilation of CYGNSS data was further explored with an additional experiment in which CYGNSS data was combined with Global Precipitation Mission (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG) hourly precipitation and Advanced Scatterometer (ASCAT) wind vector and were assimilated into the WRF model. A signiﬁcant positive impact was found on the tropical cyclone intensity and track forecasts. The short-term forecast of wind and precipitation ﬁelds were also improved for both TC Irving and the WWB event when the combined satellite data was assimilated.


Introduction
Owing to better global analyses and advances in numerical models and physics parameterizations, tropical cyclone (TC) forecasts have improved steadily in the past two decades (See, e.g., [1,2]). Despite the substantial amount of effort, however, accurate predictions of the structure and intensity of TCs and tropical convections remain challenging. The structure and intensity of TCs are not only influenced by the large-scale environment; the air-sea exchange also plays a vital role. The variations in structure and intensity also depend largely on the meso-and micro-scale in-storm dynamic, physics, and radiative processes, as well as the interaction of environmental flow and storm-scale processes [3,4]. An accurate description of the atmospheric state and the environment is essential for simulations of tropical cyclones and tropical convections.
Since tropical cyclones and maritime tropical convection spend most of their lifetime over oceans where conventional data and aircraft measurements are limited in space and time, the need for a more number of improvements to the data calibration, v2.1 is of much better quality than the previous version; this is especially true for the young seas/limited fetch (YSLF) product. Ruf et al. [41] calculated the root-mean-square (RMS) difference between CYGNSS v2.0 YSLF data and Stepped Frequency Microwave Radiometers (SFMRs) data onboard the NOAA P-3 aircraft in and near hurricanes. For wind retrievals above 20 m s −1 , the RMS was 5.01 m s −1 when the uncertainty of the SFMR measurements was removed. When CYGNSS v2.1 YSLF data were evaluated, Ruf et al. [40] indicated that this RMS value had been reduced to 3.2 m s −1 . In the current study, we examined the assimilation of both FDS and YSLF v2.1 wind speed products. Through the present study, our objectives are: (1) to evaluate the impact and reveal the challenge of utilizing the CYGNSS ocean surface wind speed data in a numerical model, and (2) to explore how to better utilize the CYGNSS data and improve short-term forecast skill.

Data
The main datasets used in this study include the Level 2 CYGNSS version 2.1 data, Global Precipitation Mission (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG) data, and Level 2 ASCAT data. The CYGNSS Level 2 data is obtainable from the Physical Oceanography Distributed Active Archive Center (PODAAC) (https://podaac-tools.jpl.nasa.gov/ drive/files/OceanWinds/cygnss/L2/v2.1). The GPM IMERG is a Level 3 gridded product for global rainfall estimates. Its spatial resolution is 0.1 • and temporal resolution is 0.5 h. The IMERG product incorporates precipitation estimates from GPM and partner satellite microwave and infrared (IR) sensors. Rain gauge data was used in calibration of IMERG [42]. GPM IMERG rainfall data is available at https://pmm.nasa.gov/data-access/downloads/gpm. ASCAT is one of several instruments on the Meteorological Operational (MetOp) polar-orbiting satellite. It is an active microwave sensor designed to provide surface wind observations over the global oceans. The ASCAT ocean surface wind vector data has a resolution of 25 km and is available twice a day over the model domain. ASCAT data can be accessed from the PODAAC (https://podaac-tools.jpl.nasa.gov/drive/files/OceanWinds/ascat).

The 2018 January MJO and CYGNSS Observation
The MJO is the dominant mode of intraseasonal variability (30-90 days), characterized by eastward propagation of enhanced deep convection at an average speed of 5 m s −1 during boreal winter [43]. The MJO convective signal typically initiates in the Indian Ocean and propagates to (and often across) the Maritime Continent to the western Pacific. The Real-time Multivariate MJO (RMM) index [44] is a commonly used index in MJO studies and operation forecasts as a measure of the phase and intensity of an MJO event. Figure 1 illustrates the RMM index provided by the Australian Bureau of Meteorology (data available at http://www.bom.gov.au/climate/mjo/graphics/rmm.74toRealtime. txt) from late December 2017 to early February 2018. RMM1 and RMM2 represent the first and second principal components of the combined empirical orthogonal functions (EOFs) of outgoing longwave radiation (OLR), as well as zonal wind at 200 and 850 hPa averaged between 15 • N and 15 • S. Points outside the center circle indicate a significant MJO event; the farther from the origin, the stronger the MJO amplitude. The phase triangles indicate the primary location of Indian Ocean, Maritime Continent, western Pacific, or western hemisphere and Africa. The 2018 MJO was one of the strongest events in the last few decades [45]. Barrett [45] showed that the MJO progressed eastward in January from the Indian Ocean to the Pacific Ocean in February, and then back to the Indian Ocean in March. The apparent counterclockwise motion shown in Figure 1a indicates the eastward propagation of the event. As in Figure 1a, the MJO was in phases 2 and 3 over the Indian Ocean from January 1 to 16, with maximum RMM of 2.1. The MJO intensified as it passed through Maritime Continent (Phases 4 and 5) from January 17 to 26. The MJO was strong in the western Pacific in phases 6 and 7 from January 27 to February 4, with a peak RMM close to 4. Hoover et al. [46] found that synthetic CYGNSS observations were able to measure outflow from convective storms in simulated MJO events from October and December 2011. Figure 1b shows the CYGNSS surface wind speed averaged over the domain 70 • -80 • E and 0 • -10 • S from 1 to 31 January 2018. This is within the core initiation region of the MJO in the Indian Ocean. The average CYGNSS wind speed showed a peak of 8 m s −1 on 7-8 January, which is in agreement with the maximum RMM value at the same time in phases 2 and 3. A trough in the average CYGNSS wind speed with a magnitude lower than 4 m s −1 occurred in 12-16 January. This is also in agreement with the low RMM index in 10-15 January shown in Figure 1a. Another period of large CYGNSS wind over 7 m s −1 occurred during 24-29 January 2018, which also coincides with strong convection and increasing RMM values during that time period as the MJO signal transitioned to the Maritime Continent. This indicates that, in general, CYGNSS data were able to detect strong winds related to this MJO event. However, it is noteworthy that the average CYGNSS wind speed calculated here is sensitive to WWBs during the MJO, and also to any weather system enhancing wind speeds within the domain; thus this metric should not be viewed as a definitive measure of MJO strength. The late January enhancement in CYGNSS wind speed appears to be similar to that of early January, with an enhanced near-Equator westerly flow associated with the presence of TC Cebile. were able to measure outflow from convective storms in simulated MJO events from October and December 2011. Figure 1b shows the CYGNSS surface wind speed averaged over the domain 70°-80°E and 0°-10°S from 1 to 31 January 2018. This is within the core initiation region of the MJO in the Indian Ocean. The average CYGNSS wind speed showed a peak of 8 m s -1 on 7-8 January, which is in agreement with the maximum RMM value at the same time in phases 2 and 3. A trough in the average CYGNSS wind speed with a magnitude lower than 4 m s -1 occurred in 12-16 January. This is also in agreement with the low RMM index in 10-15 January shown in Figure 1a. Another period of large CYGNSS wind over 7 m s -1 occurred during 24-29 January 2018, which also coincides with strong convection and increasing RMM values during that time period as the MJO signal transitioned to the Maritime Continent. This indicates that, in general, CYGNSS data were able to detect strong winds related to this MJO event. However, it is noteworthy that the average CYGNSS wind speed calculated here is sensitive to WWBs during the MJO, and also to any weather system enhancing wind speeds within the domain; thus this metric should not be viewed as a definitive measure of MJO strength. The late January enhancement in CYGNSS wind speed appears to be similar to that of early January, with an enhanced near-Equator westerly flow associated with the presence of TC Cebile. Previous studies found that CYGNSS measurements during TC overpasses can help in finding the correct center location, estimating the intensity, and constructing the wind distribution within the TC [40,47,48]. Hoover et al. [46] also showed that synthetic CYGNSS data may be used to monitor westerly wind burst (WWB) events during MJO events. In this study, we will examine the CYGNSS data for TC Irving and a WWB event during the 2018 January MJO, and evaluate the impact of the CYGNSS data on the short-term numerical prediction of these events.
TC Irving initiated from a low-pressure system in the southern Indian Ocean at ~1800 UTC on 4 January 2018 and developed into a tropical storm after 18 h. It continued to strengthen and became a tropical cyclone by 0600 UTC on 7 January, with the center minimum sea level pressure (MSLP) falling to 981 hPa and the maximum surface winds reaching 36 m s -1 . TC Irving gained further intensity as it moved southwestward, and reached peak intensity with a MSLP of 962 hPa and maximum surface wind of 49 m s -1 at 0000 UTC on 8 January. Thereafter, it started to weaken before it dissipated at around 0600 UTC on 11 January 2018. In this study, we focused on the intensification from 0000 UTC on 6-8 January, and the weakening afterwards until 0000 UTC on 9 January 2018.
In this study, a few quality control steps were applied to the Level 2 CYGNSS data v2.1. Data with the range corrected gain below 3 and the zenith sun angle lower than zero were removed to Previous studies found that CYGNSS measurements during TC overpasses can help in finding the correct center location, estimating the intensity, and constructing the wind distribution within the TC [40,47,48]. Hoover et al. [46] also showed that synthetic CYGNSS data may be used to monitor westerly wind burst (WWB) events during MJO events. In this study, we will examine the CYGNSS data for TC Irving and a WWB event during the 2018 January MJO, and evaluate the impact of the CYGNSS data on the short-term numerical prediction of these events.
TC Irving initiated from a low-pressure system in the southern Indian Ocean at~1800 UTC on 4 January 2018 and developed into a tropical storm after 18 h. It continued to strengthen and became a tropical cyclone by 0600 UTC on 7 January, with the center minimum sea level pressure (MSLP) falling to 981 hPa and the maximum surface winds reaching 36 m s −1 . TC Irving gained further intensity as it moved southwestward, and reached peak intensity with a MSLP of 962 hPa and maximum surface wind of 49 m s −1 at 0000 UTC on 8 January. Thereafter, it started to weaken before it dissipated at around 0600 UTC on 11 January 2018. In this study, we focused on the intensification from 0000 UTC on 6-8 January, and the weakening afterwards until 0000 UTC on 9 January 2018.
In this study, a few quality control steps were applied to the Level 2 CYGNSS data v2.1. Data with the range corrected gain below 3 and the zenith sun angle lower than zero were removed to filter out the unreliable observations. Also, the uncertainties in the Block II-F GPS transmitter power and antenna gain patterns can significantly degrade the quality of wind speed products [49]. Therefore, all Level 2 retrievals from Block II-F GPS satellites were excluded from the Level 2 product. Further quality control (e.g., gross check, vertical consistency, dry convective adjustment, etc.) was conducted during WRFDA data process before the CYGNSS wind speed observation was assimilated. Two different wind speed products retrieved by different geophysical model functions (GMFs) are provided in v2.1, i.e., the FDS and the YSLF. The FDS is appropriate for most conditions, while the YSLF data can be used typically for the inner core of TCs. YSLF GMF was developed for conditions when the long-wave portion of the sea state has not responded fully to the local surface winds [40,41]. Figure  filter out the unreliable observations. Also, the uncertainties in the Block II-F GPS transmitter power and antenna gain patterns can significantly degrade the quality of wind speed products [49]. Therefore, all Level 2 retrievals from Block II-F GPS satellites were excluded from the Level 2 product. Further quality control (e.g., gross check, vertical consistency, dry convective adjustment, etc.) was conducted during WRFDA data process before the CYGNSS wind speed observation was assimilated. Two different wind speed products retrieved by different geophysical model functions (GMFs) are provided in v2.1, i.e., the FDS and the YSLF. The FDS is appropriate for most conditions, while the YSLF data can be used typically for the inner core of TCs. YSLF GMF was developed for conditions when the long-wave portion of the sea state has not responded fully to the local surface winds [40,41]. Figure
The WRFDA system was used to conduct data assimilation with the hybrid ensemble variational technique. WRFDA is a unified community data assimilation system that has 3-dimensional variational (3DVAR) and 4-dimensional ensemble variational (4DVAR) capabilities, a hybrid variational ensemble algorithm (EnVar) and ensemble transform Kalman Filter technique. Being built in model-space, WRFDA is a flexible, state-of-the-art atmospheric data assimilation system that is portable and efficient for various platforms and for multiple models. WRFDA is suitable for the assimilation of a wide variety of observations, including conventional surface and upper-air soundings, ship, buoy, aircraft, satellite retrieved products, satellite radiance, radar, surface precipitation, etc. In this research, data assimilation is conducted with "cv option 5" which defines streamfunction, unbalanced velocity potential, unbalanced temperature, pseudo relative humidity, and unbalanced surface pressure as the control variables. The wind observations were assimilated using the wind speed and direction assimilation package based on Huang et al. [57].
Four different experiments were conducted. Table 1 lists the observation data assimilated in each experiment and the data assimilation analysis times. All experiments began at 0000 UTC on 6 January and ended at 0000 UTC on 9 January 2018. The control experiment (CTRL) was a WRF regional simulation. The initial and boundary conditions were interpolated from the 0.25 • resolution global NCEP Final Analysis (FNL). The impact of the CYGNSS data was firstly examined with two experiments-DA_cyg_fd and DA_cyg_lf, in which the 3-dimensional ensemble variational (3DEnVar) technique was adopted in the assimilation of only CYGNSS FDS and YSLF data, respectively. Ensemble WRF forecasts with 12 members were generated with randomly perturbed initial conditions created using the ensemble transform Kalman filter (ETKF) technique [58]. The ensemble forecast was used to produce a flow-dependent background error matrix. Cycled assimilation of the CYGNSS data was conducted at 0000, 1200, 1500, and 2100 UTC from 6 to 8 January, based on the data availability. There are more CYGNSS data available for TC Irving and the WWB event at 0000 and 1500 UTC than at other times. The observational error was set as 2 m s −1 for winds below 20 m s −1 and 10% for winds above 20 m s −1 based on the general guidance from Ruf et al. [40]. If there were multiple observations within a model grid box, a distance-weighted average of all observations was calculated and assimilated. The last experiment (DA_com) assimilated the combined CYGNSS wind speed, IMERG rainrate, and ASCAT vector wind data using the 4-dimensional ensemble variational (4DEnVar) technique. The 4DEnVar also used 12-member ensemble WRF forecasts to generate the flow dependent background error matrix. In this experiment, the CYGNSS data include YSLF data within 800 km from the center of TC Irving and FDS data elsewhere. Figure 3 illustrates the time windows of the 4DEnVAR process in DA_com.

CYGNSS Data Assimilation
To examine the impact of the CYGNSS wind speed data, the analysis fields and short-term forecasts from experiments with and without the assimilation of CYGNSS data are compared and analyzed in this section. Figure 4 plots the CYGNSS wind, background wind field, and data assimilation analysis increment ('analysis -background', or the difference between data assimilation analysis and the background) in the surface wind field at 1500 UTC on 6 January 2018. At this time, CYGNSS FDS ( Figure 4a) observed a maximum wind speed of 18 m s -1 , with most of the observed wind being slower than 15 m s -1 , while CYGNSS YSLF (Figure 4b) detected wind as strong as 47 m s -1 near TC Irving and many points with wind speed above 15 m s -1 . In the background field ( Figure  4e), an area of wind above 15 m s -1 was predicted over the eyewall region of TC Irving with the maximum wind speed of 20 m s -1 . Therefore, the wind speed observed by CYGNSS FDS is generally lower than the background. When the FDS data were assimilated, the analysis increments were mostly negative near TC Irving (Figure 4c). With the assimilation of high winds from CYGNSS YSLF data, the surface wind increased by up to 6 m s -1 near TC Irving and 2 m s -1 for the WWB event ( Figure  4d). However, the limitation of assimilating the CYGNSS data was also apparent. Due to the trackbased nature of CYGNSS observation, the analysis increment in DA_cyg_lf shows elongated shapes at several places. This very likely does not represent the correct characteristics of the wind field, and therefore, may cause inaccurate asymmetry in wind and, hence, mass fields.

CYGNSS Data Assimilation
To examine the impact of the CYGNSS wind speed data, the analysis fields and short-term forecasts from experiments with and without the assimilation of CYGNSS data are compared and analyzed in this section. Figure 4 plots the CYGNSS wind, background wind field, and data assimilation analysis increment ('analysis -background', or the difference between data assimilation analysis and the background) in the surface wind field at 1500 UTC on 6 January 2018.  (Figure 4e), an area of wind above 15 m s −1 was predicted over the eyewall region of TC Irving with the maximum wind speed of 20 m s −1 . Therefore, the wind speed observed by CYGNSS FDS is generally lower than the background. When the FDS data were assimilated, the analysis increments were mostly negative near TC Irving (Figure 4c). With the assimilation of high winds from CYGNSS YSLF data, the surface wind increased by up to 6 m s −1 near TC Irving and 2 m s −1 for the WWB event ( Figure 4d). However, the limitation of assimilating the CYGNSS data was also apparent. Due to the track-based nature of CYGNSS observation, the analysis increment in DA_cyg_lf shows elongated shapes at several places. This very likely does not represent the correct characteristics of the wind field, and therefore, may cause inaccurate asymmetry in wind and, hence, mass fields.
The RMS difference in surface wind speed is calculated between the data assimilation experiments (DA_cyg_lf and DA_cyg_fd) and the background field, and averaged over the model domain. Figure 5 shows the domain averaged RMS difference of surface wind speed within 9 h after the assimilation of CYGNSS data was complete at 1500 UTC on 6 January 2018. At the analysis time, the averaged RMS was 1.2 m s −1 for DA_cyg_lf and 0.86 m s −1 for DA_cyg_fd. It rapidly declined for both experiments within the first hour after data assimilation. After 4 h, the averaged RMS were only 0.41 m s −1 and 0.32 m s −1 for DA_cyg_lf and DA_cyg_fd, which were 30% and 37% of the original values, respectively. This indicates that the strongest impact of the CYGNSS data will be within the first~4 h of the forecast. The RMS difference in surface wind speed is calculated between the data assimilation experiments (DA_cyg_lf and DA_cyg_fd) and the background field, and averaged over the model domain. Figure 5 shows the domain averaged RMS difference of surface wind speed within 9 h after the assimilation of CYGNSS data was complete at 1500 UTC on 6 January 2018. At the analysis time, the averaged RMS was 1.2 m s -1 for DA_cyg_lf and 0.86 m s -1 for DA_cyg_fd. It rapidly declined for both experiments within the first hour after data assimilation. After 4 h, the averaged RMS were only 0.41 m s -1 and 0.32 m s -1 for DA_cyg_lf and DA_cyg_fd, which were 30% and 37% of the original values, respectively. This indicates that the strongest impact of the CYGNSS data will be within the first ~4 h of the forecast. With the assimilation of CYGNSS data in the initial conditions, positive impacts have been made in the simulations of TC Irving and the WWB event. Figure 6   The impact of CYGNSS data assimilation on the intensity and track forecasts of TC Irving was compared with the JTWC best track data in Figures 7 and 8. From 00 UTC on 6 January to 00 UTC on 8 January, TC Irving experienced an intensification, from a tropical storm to a category-2 tropical cyclone, with a peak intensity of 962 hPa in MSLP of and 49 m s -1 in maximum surface wind speed, followed by a weakening stage until 0000 UTC on 9 January with 973 hPa in MSLP of and 41 m s -1 in The impact of CYGNSS data assimilation on the intensity and track forecasts of TC Irving was compared with the JTWC best track data in Figures 7 and 8. From 00 UTC on 6 January to 00 UTC on 8 January, TC Irving experienced an intensification, from a tropical storm to a category-2 tropical cyclone, with a peak intensity of 962 hPa in MSLP of and 49 m s −1 in maximum surface wind speed, followed by a weakening stage until 0000 UTC on 9 January with 973 hPa in MSLP of and 41 m s −1 in maximum surface wind speed. From Figure 7, CTRL greatly underestimated the intensity of TC Irving. The forecast in CTRL shows a slow strengthening of Irving from 999 hPa in MSLP and 17 m s −1 in maximum surface wind speed at 0000 UTC on 6 to 990 hPa and 25 m s −1 at 0000 UTC on 9 January 2018. Improvement in intensity forecast by the assimilation of the CYGNSS data was found from 0000 UTC on 7 January to 0000 UTC on 8 January 2018. During this time period, both DA_cyg_lf and DA_cyg_fd outperformed CTRL with the strongest storm generated by DA_cyg_lf. CTRL missed the weakening of the storm from 0000 UTC on 8 to 0000 UTC on 9 January 2018. The intensity of the storms in DA_cyg_lf and DA_cyg_fd showed some fluctuations, but no apparent weakening was predicted in either experiment. It is noteworthy that the intensity forecast in data assimilation experiments was not always better than CTRL (e.g., from 1200 UTC on 8 January to 0000 UTC on 9 January 2018 in DA_cyg_fd).        The track forecasts from CTRL and data assimilation experiments were compared with the JTWC best track in Figure 8. As shown in Figure 8a, all experiments captured the southwestward path of TC Irving. CTRL produced the track to the south of the observed storm with a track error ranging from 35 km to 265 km. The assimilation of CYGNSS FDS data provided help with the tracking forecast. The track error in DA_cyg_fd was generally smaller than those of either CTRL or DA_cyg_lf, except at 1800 UTC on 7 January and in the last 6 h of the forecast period. For example, at 0000 UTC on 8 January, the center location of TC Irving in DA_cyg_fd (the track error of 100 km) was much closer to the observed location when compared with the one in CTRL (the track error of 175 km) and the one in DA_cyg_lf (the track error of 137 km). Figure 9 displays IMERG precipitation, CYGNSS FDS and YSLF wind speed, ASCAT wind speed and model simulated wind from DA_cyg_lf and CTRL over the region of the WWB event when ASCAT captured a portion of the event near 1500 UTC on 8 January 2018. IMERG (Figure 9a) shows the convection related to the WWB event produced a maximum precipitation over 10 mm h −1 . ASCAT (Figure 9b)  The overall impacts of the CYGNSS observation on surface wind and precipitation fields are shown in Tables 2 and 3. Table 2   The overall impacts of the CYGNSS observation on surface wind and precipitation fields are shown in Tables 2 and 3. Table 2 lists the domain averaged RMS of surface wind speed difference between different experiments and ASCAT wind speed within the model domain available around 0302 and 1448 UTC on 7 January and 0239 and 1436 UTC on 8 January 2018. As shown in the table, the assimilation of both CYGNSS FDS and YSLF data led to smaller RMS values at all times. At 1500 UTC on 8 January, the RMS reduced by up to 0.7 m s −1 in DA_cyg_lf when comparing with CTRL. Table 3 lists the threat scores (TS) of precipitation forecast with a threshold value of 2 mm h −1 from different experiments. The IMERG data were used as the reference dataset. The TS was calculated using the equation from Xiao et al. [59]: where C is the number of correct forecast events (points with precipitation exceeding the threshold value), R is the number of events in observation, and F is the number of events in forecast. As indicated in Table 3, the differences in TSs between CTRL and the data assimilation experiments DA_cyg_fd and DA_cyg_lf were as small as 0.01-0.02, which indicates that the influence from assimilation of CYGNSS data only was not significant. This could be attributed to the continuous correction of model dynamics, which commonly occurs when only surface data are assimilated. Another possible reason is the lack of dynamic support from the convective and thermodynamic fields for convection to grow. This was the motivation for the DA_com data assimilation experiment, as described next.

Assimilation of the Combined Satellite Datasets
The above results indicate the positive impacts of CYGNSS data, as well as the limitations of the CYGNSS data when being assimilated into the initial conditions of NWP simulations. Since the IMERG data can measure the precipitation of tropical convection, the ASCAT data can describe the cyclonic circulation of TCs, and the CYGNSS data can provide wind speed observations under heavy precipitation, it would be of great interest to examine how the combined datasets would help with TC and tropical convection intensity and structure.
The overall performance of DA_com is compared with other experiments in Tables 2 and 3. As shown in Table 2, DA_com produced the lowest RMS values at all times among all experiments. Specifically, the domain averaged RMS values in DA_com ranged from 1.47 to 2.08 m s −1 . When comparing to CTRL, the RMS values reduced by 1.09 to 1.59 m s −1 , indicating a 41-49% relative reduction. From Table 3, it is shown that the precipitation forecast skill in DA_com was apparently higher than the rest of the experiments. For example, the TS in CTRL was only 0.15 at 0000 UTC on 9 January 2018, while the value in DA_com was 0.33, which is a relative improvement of 120%. These results indicate large positive impacts with the assimilation of the combined datasets.
To illustrate the prediction of intensity and structure of TC Irving, Figure 10 shows a comparison of the sea level pressure (SLP) over-plotted on surface wind from DA_com and CTRL near TC Irving from around 0300 UTC on 7 January and 1500 UTC on 8 January 2018, when ASCAT observation was available. At both times, DA_com produced better intensity estimates than CTRL compared to the best track data. At 0300 UTC on 7 January, the best track data showed that TC Irving had a MSLP between 986 hPa (0000 UTC) and 981 hPa (0600 UTC) and a maximum surface wind speed between 31 m s −1 (0000 UTC) and 36 m s −1 (0600 UTC). ASCAT (Figure 10a) missed the strong wind in the core region due to contamination by heavy precipitation. Therefore, the wind asymmetry of TC Irving is unclear. CTRL (Figure 10b) Figure 10f) with a MSLP of 982 hPa and a maximum surface wind speed of 32 m s −1 and a smaller eye. From ASCAT, it seems that the strongest wind at this time appeared at the front right quadrant of TC Irving, and that there was no apparent wind asymmetry around the eyewall. In CTRL, the strongest wind was at the front to front left of TC Irving with a strong wind asymmetry. DA_com shows the strong wind occurring in both front left and front right quadrants. The asymmetry was not as apparent as the wind field in CTRL.
Remote Sens. 2020, 12, x FOR PEER REVIEW 14 of 21 between 986 hPa (0000 UTC) and 981 hPa (0600 UTC) and a maximum surface wind speed between 31 m s -1 (0000 UTC) and 36 m s -1 (0600 UTC). ASCAT (Figure 10a) missed the strong wind in the core region due to contamination by heavy precipitation. Therefore, the wind asymmetry of TC Irving is unclear. CTRL (Figure 10b)  From ASCAT, it seems that the strongest wind at this time appeared at the front right quadrant of TC Irving, and that there was no apparent wind asymmetry around the eyewall. In CTRL, the strongest wind was at the front to front left of TC Irving with a strong wind asymmetry. DA_com shows the strong wind occurring in both front left and front right quadrants. The asymmetry was not as apparent as the wind field in CTRL. Figure 11 shows the surface wind field for the WWB event around 1500 UTC on 7 January and 1600 UTC on 8 January 2018 when ASCAT captured a portion of the event. At both times, DA_com produced stronger wind field than CTRL. ASCAT (Figure 11a Figure 12 compares the precipitation predictions from CTRL and DA_com with IMERG data at 1500 UTC on 7 January and 2100 UTC on 8 January 2018. At both times, DA_com generated a much stronger and better organized precipitation over the core region of TC Irving than CTRL. At 1500 UTC on 7 January, DA_com (Figure 12c) produced maximum rainfall of 36 mm h -1 around the eyewall, which was pretty close to the observed 34 mm h -1 by IMERG (Figure 12a), and 12 mm h -1 higher than the one predicted in CTRL (Figure 12b). At 2100 UTC on 8 January 2018, both DA_com and CTRL underestimated the precipitation for TC Irving. IMERG (Figure 12d) estimated the maximum rainfall rate of 56 mm h -1 . CTRL (Figure 12e) predicted maximum rainfall of 20 mm h -1 . DA_com (Figure 12f) outperformed CTRL with a much stronger precipitation field with a maximum of 42 mm h -1 . Precipitation related to the WWB event was also better represented in DA_com ( Figure  13). At 1200 UTC on 7 January 2018, IMERG observed strong convection with the maximum precipitation rate of 32 mm h -1 at 75°-80°E and 5°-7°S. In CTRL, this storm was not predicted. With the assimilation of the combined datasets, DA_com produced heavy precipitation near the observed location with the maximum precipitation > 40 mm h -1 . A similar result was found at 2100 UTC on 8 January 2018. IMERG observed convection with maximum precipitation of 23 mm h -1 at 67°-78°E and 4°-8°S. CTRL predicted multiple strong convective storms, but the locations were far from the observed one. DA_com was able to generate the precipitation at 70°-78°E and 4°-8°S with maximum precipitation of 26 mm h -1 .  Figure 12 compares the precipitation predictions from CTRL and DA_com with IMERG data at 1500 UTC on 7 January and 2100 UTC on 8 January 2018. At both times, DA_com generated a much stronger and better organized precipitation over the core region of TC Irving than CTRL. At 1500 UTC on 7 January, DA_com (Figure 12c) produced maximum rainfall of 36 mm h −1 around the eyewall, which was pretty close to the observed 34 mm h −1 by IMERG (Figure 12a), and 12 mm h −1 higher than the one predicted in CTRL (Figure 12b). At 2100 UTC on 8 January 2018, both DA_com and CTRL underestimated the precipitation for TC Irving. IMERG (Figure 12d) estimated the maximum rainfall rate of 56 mm h −1 . CTRL (Figure 12e) predicted maximum rainfall of 20 mm h −1 . DA_com (Figure 12f) outperformed CTRL with a much stronger precipitation field with a maximum of 42 mm h −1 . Precipitation related to the WWB event was also better represented in DA_com (Figure 13). At 1200 UTC on 7 January 2018, IMERG observed strong convection with the maximum precipitation rate of 32 mm h −1 at 75 • -80 • E and 5 • -7 • S. In CTRL, this storm was not predicted. With the assimilation of the combined datasets, DA_com produced heavy precipitation near the observed location with the maximum precipitation > 40 mm h −1 . A similar result was found at 2100 UTC on 8 January 2018. IMERG observed convection with maximum precipitation of 23 mm h −1 at 67 • -78 • E and 4 • -8 • S. CTRL predicted multiple strong convective storms, but the locations were far from the observed one. DA_com was able to generate the precipitation at 70 • -78 • E and 4 • -8 • S with maximum precipitation of 26 mm h −1 .  The impacts of the combined satellite datasets on the intensity and track of TC Irving are displayed in Figures 7 and 8. As shown in Figure 7, DA_com also underestimated the intensity of TC Irving, but apparently performed better when comparing to CTRL, DA_cyg_fd, and DA_cyg_lf. An intensification was predicted from 999 hPa in MSLP and 17 m s -1 in maximum surface wind speed at 0000 UTC on 6 January to 985 hPa and 35 m s -1 at 1800 UTC on 7 January 2018. The weakening after 0600 UTC on 8 January was also captured. With the assimilation of the combined satellite datasets, the track forecast of TC Irving in DA_com was also improved. The center locations in DA_com were very close to the observed ones with the track error ranges from 11 km at 1200 UTC on 6 January 2018 to 88 km at 0000 UTC on 9 January 2018. In comparison, the track errors were 125, 93, and 94 km at  The impacts of the combined satellite datasets on the intensity and track of TC Irving are displayed in Figures 7 and 8. As shown in Figure 7, DA_com also underestimated the intensity of TC Irving, but apparently performed better when comparing to CTRL, DA_cyg_fd, and DA_cyg_lf. An intensification was predicted from 999 hPa in MSLP and 17 m s -1 in maximum surface wind speed at 0000 UTC on 6 January to 985 hPa and 35 m s -1 at 1800 UTC on 7 January 2018. The weakening after 0600 UTC on 8 January was also captured. With the assimilation of the combined satellite datasets, the track forecast of TC Irving in DA_com was also improved. The center locations in DA_com were very close to the observed ones with the track error ranges from 11 km at 1200 UTC on 6 January 2018 to 88 km at 0000 UTC on 9 January 2018. In comparison, the track errors were 125, 93, and 94 km at The impacts of the combined satellite datasets on the intensity and track of TC Irving are displayed in Figures 7 and 8. As shown in Figure 7, DA_com also underestimated the intensity of TC Irving, but apparently performed better when comparing to CTRL, DA_cyg_fd, and DA_cyg_lf. An intensification was predicted from 999 hPa in MSLP and 17 m s −1 in maximum surface wind speed at 0000 UTC on 6 January to 985 hPa and 35 m s −1 at 1800 UTC on 7 January 2018. The weakening after 0600 UTC on 8 January was also captured. With the assimilation of the combined satellite datasets, the track forecast of TC Irving in DA_com was also improved. The center locations in DA_com were very close to the observed ones with the track error ranges from 11 km at 1200 UTC on 6 January 2018 to 88 km at 0000 UTC on 9 January 2018. In comparison, the track errors were 125, 93, and 94 km at 1200 UTC on 6 January 2018, and 266, 259, 198 km at 0000 UTC on 9 January 2018 in CTRL, DA_cyg_fd, and DA_cyg_lf, respectively.

Discussion
In this study, the application of CYGNSS v2.1 Level 2 data was examined in a numerical simulation of TC Irving and a WWB event that occurred during the 2018 January MJO event. WRF model simulations and WRFDA 3DEnVar and 4DEnVar data assimilation experiments were conducted to investigate the impact of the CYGNSS data when it was assimilated individually, as well as combined with the IMERG rainfall and ASCAT ocean surface wind vector observations. The goal of this study was to report on the benefits and challenge in the assimilation of CYGNSS data.
The results indicate that CYGNSS data can capture the high winds and sharp wind gradient under heavily precipitating conditions within TC Irving and the WWB event. When CYGNSS data were assimilated individually, a positive impact was found on the wind field for both TC Irving and the WWB event. The intensity and track forecast of TC Irving were also positively affected. However, the main impact of the data appears to only last about 4 h. In addition, the influence of the CYGNSS winds on the precipitation field was modest at best. When CYGNSS winds were assimilated together with IMERG hourly rainfall and ASCAT ocean surface vector winds, significant impacts were found in the wind, pressure, and precipitation fields. The intensity and track forecast of TC Irving were largely improved. Specifically, the structure in wind and precipitation fields in the vicinity of the TC core region and across the WWB event was better represented with the assimilation of the combined satellite data. Unfortunately, one caveat of the current research is the lack of an experiment parallel to DA_com that assimilates IMERG and ASCAT data using 4DEnVar. Due to the expensive computational cost of 4DEnVar, we did not have the time or computational resource to conduct this experiment. Therefore, the sole impact of CYGNSS data is unclear when IMERG and ASCAT data are also assimilated. This is quite important for a thorough understanding on the overall benefit and limitation of CYGNSS observation. In future studies, we will conduct these experiments with more case studies to further explore the impact of CYGNSS data when it is combined with other datasets.

Conclusions
The results shown here are quite encouraging. However, the conclusions in this study are based on only one case study. More experiments with additional cases for a variety of tropical weather systems are needed to demonstrate a thorough evaluation of the consistent impacts of the CYGNSS observations to NWP forecasts. New versions of the CYGNSS data (v3.0 and Climate Data Record-CDR v1.0) are expected to be available soon; v3.0 includes a near-real-time corrections for variations in the transmit power of the GPS satellites, and is expected to improve the temporal and spatial sampling properties, as well as the accuracy. CDR data include track-wise debiasing that is expected to improve wind speed calibration. In both new data versions, a reduction is expected in the uncertainty of the wind speed products for both high and low wind speed [40]. Therefore, future research should be conducted with WRF simulations and data assimilation experiments for new versions of CYGNSS data to further explore the application of the latest products. In addition, the frequent revisit time of CYGNSS is one of the great advantages of this instrument. Hence, it would be of great interest to conduct continuous assimilation of CYGNSS data and to investigate the benefit it may bring to midrange to subseasonal forecasts of tropical convection and MJO events.