HTML
-
This study is based on the Rapid-refresh Multi-scale Analysis and Prediction System—Short Term (RMAPS-ST), which runs operationally at the Beijing Meteorological Bureau (Fan et al., 2013; Chen et al., 2014; He et al., 2019; Xie et al., 2019). In RMAPS-ST, the Advanced Research version of the WRF model, version 3.9, and the WRFDA system are employed. Assimilation is performed on a large domain with a grid spacing of 9 km and
$ 649\times 500 $ grid points, using 51 vertical levels laid out with an upper limit of 50 hPa.The experimental domain is shown in Fig. 7 with the distribution of the main conventional observations at 0000 UTC 1 November 2018. AMVs derived from the WV-High, WV-Mid and IR channels are assimilated in this study by merging with conventional observations. Both the initial and boundary conditions are obtained from the Global Forecast System produced by the National Centers for Environmental Prediction. The Thompson microphysics scheme (Thompson et al., 2008) and RRTMG long- and shortwave radiation scheme (Iacono et al., 2008) are used. In addition, the Kain−Fritsch cumulus parameterization scheme (Kain, 2004) is also adopted in this domain.
Figure 7. Domain of the model and the locations of the (a) main types of conventional observation data, (b) WV-High AMVs, (c) WV-Mid AMVs and (d) IR AMVs (SYNOP denotes surface station observations; SHIP denotes ship observations; BUOY denotes buoy observations, and SOUND denotes radiosonde observations).
-
Capability to assimilate AMVs derived from the three channels with new observation errors has been implemented within RMAPS-ST in this study. Five experiments (Table 1) are configured to investigate the added value from the assimilation of FY-4A AMVs on analyses and forecasts with new observation error. Firstly, the control experiment only assimilates the conventional observations (hereafter CTRL). Then, AMVs from WV-High (CONWVh), WV-Mid (CONWVm) and IR (CONIR) are assimilated together with conventional data, respectively. Moreover, an experiment with AMVs from the three channels assimilated simultaneously (CONALL) is added as well. Additionally, to better display the impact of the new observation error on analyses and forecasts, an experiment assimilating three-channel AMVs with default observation error in WRFDA is necessary, and is named CONALL-DEF.
Experiment Observation error Assimilation data CTRL Default value defined in WRFDA Conventional observations CONWVh New observation error Rwvh AMVs from WV-High Conventional observations CONWVm New observation error Rwvm AMVs from WV-Mid Conventional observations CONIR New observation error Rir AMVs from IR Conventional observations CONALL New observation error Rwvh, Rwvm and Rir for each channel Conventional observations AMVs from three channels CONALL-DEF Default observation error in WRFDA Conventional observations AMVs from three channels Table 1. Experiment design schemes.
In the six experiments, partial cycle operation (Hsiao et al., 2012) is adopted at 0000 UTC every day, which means that the initial field of 0000 UTC comes from a 6-h forecast beginning at 1800 UTC of the previous day. Then, the assimilation runs eight times at 0000, 0300, 0600, 0900, 1200, 1500, 1800 and 2100 UTC per day with 3-h cycling. At each analysis time, data within a 3-h window centered on the synoptic time will be ingested and a 24-h forecast will be obtained.
4.1. Model configuration
4.2. Experimental setup
-
Table 2 calculates the average proportion of the AMVs absorbed by the assimilation system to the total AMVs used in each assimilation process from experiments with new observation errors first. On average, more than 90% of the data are absorbed in each assimilation cycle, indicating the utilization rate of FY-4A AMVs from the three channels with QI greater than 80 is relatively high. Since observations from radiosondes are obtained twice a day at 0000 UTC and 1200 UTC, the time series of AMV observation minus background (OMB) and AMV observation minus analysis (OMA) of horizontal wind components at 1200 UTC are statistically analyzed in Fig. 8. It can be seen from the time series of OMB (Figs. 8a and c), the U component in FY-4A AMVs is smaller than the background, while the V component is larger than the background. However, the assimilation clearly reduces the deviation between the analysis and observation (Figs. 8b and d), with a more stable time series and a weakened amplitude.
Experiment Layers (hPa) 900−1000 800−900 800−700 700−600 600−400 400−300 300−200 200−100 100−50 CONWVh − − − 99.6 99.97 99.94 99.95 99.99 99.99 CONWVm − − − 99.99 99.98 99.96 99.97 99.99 99.99 CONIR 97.5 98.7 99.84 99.96 99.9 99.98 99.98 99.99 99.99 CONALL 97.5 98.7 99.79 99.89 99.93 99.96 99.96 99.97 99.99 Table 2. Average absorption ratios (%) of AMVs in different layers (hPa) from experiments assimilating with new observation errors.
Figure 8. Time series of the (a, c) OMB and (b, d) OMA statistics of the U and V components for assimilating FY-4A AMVs at 1200 UTC from the experiment assimilating FY-4A AMVs with new observation errors.
Moreover, the OMB sequence obtained by simultaneous assimilation of AMVs from the three channels is stable with no obvious outliers under the mutual restriction of data from the three channels. Also, the OMA is closer to 0 than that of the single-channel assimilation experiments, which means that the wind analysis obtained from the multi-channel assimilation experiment is better than that from the single-channel assimilation experiments.
-
The RMSEs for the analysis and forecast against sounding observations are calculated by the Model Evaluation Tools (MET) verification package (Kirby, 2004), which is designed to be a highly configurable, state-of-the-art suite of verification tools. MET provides a variety of verification techniques, including standard verification scores comparing gridded model data obtained from the WRF modeling system with point-based observations.
-
Since the AMV products mainly contain wind field information, only horizontal wind components (U and V) are discussed. Figure 9 shows the vertical profiles of the mean analysis RMSEs for horizontal winds (U, V) (solid lines) and the forecast RMSEs at t+6 h (dashed lines) and t+24 h (dotted lines) of CTRL and the experiments assimilating the single-channel AMVs. The experiments assimilating FY-4A AMVs achieve comparable analysis with CTRL, and even slightly inferior to CTRL, especially for the U component at 300 hPa, which may be related to the sparse resolution of the reference radiosonde observations.
Figure 9. Averaged vertical RMSE profiles for analyses and 6-h and 24-h forecasts in comparison with radiosonde data for CTRL and single-channel assimilation experiments.
After 6-h forecasts, the experiments assimilating FY-4A AMVs reduce the RMSEs in wind fields compared to CTRL. Positive effects mainly occur in the upper and middle layers, which is in accordance with the layer where FY-4A AMVs are distributed (Fig. 1d). Besides, among the three channels, the assimilation of WV-High AMVs produces more obvious improvement, while AMVs from the other two channels have a relatively weak effect on the 6-h forecast. After 24-h forecasts, the RMSEs of the horizontal wind components become larger, and the advantage of assimilating FY-4A AMVs gradually diminishes, but still appears in the U component below 500 hPa and V component above 500 hPa. Moreover, AMVs from the IR channel have a clearer positive impact on the 24-h forecast than that from the two water vapor channels.
At the same time, the mean RMSEs in the three layers for different experiments are calculated for various lead times. The results (Fig. 10) show that the assimilation of FY-4A AMVs reduces the RMSEs of the U and V components in the upper layer, and the improvement can even be seen in the 24-h forecast. Focusing in the middle and lower layers, AMVs from the two water vapor channels produce little positive effect because of no observations below 700 hPa. However, the assimilation of AMVs from IR improves the forecast of the V component in the middle layer, and improvement of the horizontal components in the lower layers in CONIR is gradually shown after 18 h of forecasting.
Figure 10. Average RMSEs at different lead times for the (a, b) upper layer (above 400 hPa), (c, d) middle layer (400−700 hPa) and (e, f) lower layer (below 700 hPa) for the CTRL and single-channel assimilation experiments.
In terms of the vertical distribution of FY-4A AMVs, a large number of AMVs from the three channels concentrate at 300 hPa, while only AMVs from IR are available below 700 hPa. Figure 11 displays the trend of RMSEs at 300 hPa and 700 hPa with lead time. It is clear that AMVs derived from the two water vapor channels have an advantage in improving the forecast in the upper layer, especially the data from WV-High, while AMVs from IR can produce noticeable improvement in the lower layer after 12 h of forecasting.
-
Several precipitation processes occur during the research period, so the impact of FY-4A AMVs on the precipitation forecasting skill is evaluated. The bias score (BS), threat score (TS), probability of detection (POD), and false alarm ratio (FAR) (formulae shown in Table 3) are used. The closer the BS, TS and POD are to 1, the higher the prediction accuracy of precipitation; and the smaller the FAR, the fewer false alarms in the precipitation forecast. The precipitation data for comparison are the hourly observations from surface stations provided by the CMA.
Index Formula Bias Score (BS) (hits + false alarms)/(hits + misses) Threat Score (TS) hits/(hits + false alarms + misses) Probability of Detection (POD) hits/(hits + misses) False Alarm Ratio (FAR) false alarms/(hits + false alarms) Table 3. Four statistical indices for rainfall forecast skill.
The average TS, BS, POD and FAR for the 6-h accumulated precipitation with thresholds of 0.1 mm, 1.0 mm, 5.0 mm 10.0 mm and 25.0 mm are shown in Fig. 12. Compared to CTRL, the assimilation of the FY-4A AMVs results in a decrease in BS for 6-h accumulated precipitation with intensity between 1.0 mm and 10.0 mm, and CONIR has the smallest values. Besides, the average TS and POD in CONWVm and CONIR show a slight improvement in precipitation with thresholds between 1.0 mm and 10.0 mm. In addition, the FAR of the experiments assimilating FY-4A AMVs decrease when compared to CTRL, and CONIR produces the lowest FAR value. However, owing to the actual intensity of the 6-h accumulated precipitation from these rainfall events being lower than 25.0 mm, the added AMVs lead to an under-prediction with the threshold of 25.0 mm, overall.
Figure 12. Average BS (a), TS (b), POD (c) and FAR (d) for the 6-h accumulated precipitation with different thresholds for the CTRL and single-channel assimilation experiments.
Generally, the results suggest that AMVs derived from the IR channel produce the largest improvement among the three channels. The reason may be that the precipitation intensity of these processes during the experiment is relatively weak and the vertical development relatively low. As a result, the wind of the lower layer in the IR channel has a positive impact on the precipitation.
-
Through the above discussion of the impact of assimilating single-channel AMVs on the analysis and forecasts, it can be concluded that the FY-4A AMVs from the water vapor channels mainly improve the wind field in the upper layer. Meanwhile, assimilating AMVs from IR has a slightly positive impact in the middle and lower layers. Next, to combine the advantages of the three channels, FY-4A AMVs from the three channels are simultaneously assimilated, and the impact on the forecast further analyzed.
Figure 13 displays the average RMSEs of the analysis and forecasts for the horizontal winds (U, V) of CTRL, CONALL and CONALL-DEF. Similar to Fig. 9, the assimilation of FY-4A AMVs derived from the three channels produces analyses a little inferior to that of CTRL, especially for the upper layer where a large number of AMVs are distributed, which is due to the rough spatial resolution of the radiosonde data. Also, the improvement from assimilating FY-4A AMVs will be seen after integrating for several hours. For example, the differences between the RMSEs of the experiments assimilating AMVs and that of CTRL decrease after six hours of model integration, and their RMSEs are relatively smaller than those of CTRL for the V component under 400 hPa. For the 24-h forecast, the RMSEs of U and V in CONALL and CONALL-DEF are smaller than those in CTRL at almost all levels. The main advantage of assimilating FY-4A AMVs in the U component gradually shifts from the upper layer to the middle and lower layer, while the improvement of the V component is still shown in the upper layer.
Figure 13. Averaged vertical RMSE profiles for analyses and 6-h and 24-h forecasts in comparison with radiosonde data for CTRL, CONALL and CONALL-DEF.
In addition, Fig. 13 also shows the impact of different observation errors on assimilation and prediction results. As can be seen from the RMSE profiles, when compared to CONALL-DEF (gray lines), assimilating FY-4A AMVs with new observation errors (purple lines) can clearly reduce the RMSEs of the horizontal wind components in almost all layers both in the analyses and 6-h and 24-h forecasts, which can serve as a reference for the significance of estimated observation error in new observation datasets.
Besides, according to the precipitation forecasting skill scores displayed in Fig. 14, except for the threshold of 0.1 mm, CONALL results in a decrease of BS and FAR when compared to CTRL. Moreover, the average TS and POD in CONALL show a slight improvement for precipitation with thresholds between 1.0 mm and 10.0 mm, which is similar to the result obtained in the single-channel assimilation.