In addition, an analysis of the impact of the data assimilation component for 3DVAR was done. One standard deviation of the daily values is given in each row for (a) and (b).
Introduction
Background and Motivations
The data assimilation method was first adapted to the atmospheric field to generate reanalysis and forecast data, such as the National Centers for Environmental Prediction (NCEP) final reanalysis (FNL; for East Asia, Liu et al. 2011). adapted 3DVAR data assimilation with the GOCART aerosol module.
Research Objectives
Model Configuration
For the KORUS-AQ case, the Fire Inventory developed by NCAR (FINN; Wiedinmyer et al. 2011) was adapted. MEGAN is the biochemical model that can generate the biological emission inventory from leaf area index, weather and atmospheric chemical composition.
Data Assimilation Methods
Pf and Pa are the uncertainty of the prior and posterior values, respectively, and Q is the model error covariance. Due to the size of the matrix, an alternative approach was proposed using a sample of developed statistics, which is an ensemble.
Data
For the comparison of the impact of the data assimilation variants, a sensitivity test was performed. 𝐼𝑖=1𝑝(𝑦𝑖, 𝑜𝑖) represents the forecast proportion correct (i.e. the forecast accuracy of the actual forecasts), and ∑𝐼𝑖=1𝑝(𝑝(𝑦) The assimilation of satellite data helps to compare the values in the sea points and the downstream part improved, as in the northern part of the main island of Japan.
The correlation between DA and NoDA generally becomes equal after 24 hours, which shows that the uncertainty of the model itself is greater than the data assimilation effect. This pattern appears to be strongly related to the seasonal difference in the mean correlation of the NoDA experiment. In the case of the false alarm rate, both PM10 and PM2.5 were higher for the bias-corrected experiment.
By comparing EnKF_M and EnKF_N, the reliability of the surface PM data is significantly increased. From the relationship of RMSE and spread, the EnKF is difficult to generate high-quality data assimilation unless the ensemble spread problem is solved. In contrast to NoDA (Fig. 4-4f), assimilation of GOCI AOD occurs increased surface PM10.
This is not driven by satellite data assimilation (Figs. 4-5f), as it is outside the GOCI retrieval domain. This wind change is significant in the middle region of the peninsula, which is reliable for the overestimated PM regions. However, the variation of the model simulation is too wide, thus also reducing the data assimilation and prediction capabilities.
For each of the data assimilation methods, handling accurate covariance of background errors is a crucial part.
Experiments
Verification of Data Assimilation
- Case Verification for KORUS-AQ
- Case Verification for All-Season
The Asian dust event in case A is clearly shown in the observed PM10 concentration averaged over South Korea (Fig. 3-1a). As the concentration of PM10 includes PM2.5, surface PM2.5 contributes relatively more to PM10 in the long-range transport regime during the period 25 May – 1 June. This is again due to the direct use of surface observation PM2.5 in the data assimilation process, which tends to correct the systematic underestimation bias in NoDA as in PM10.
In South Korea, NoDA tends to simulate relatively high AOD values in the south, but not as high as in the GOCI observations. First, the overall overestimation of PM10 concentration in central and southern China and underestimation of PM10 concentration in northern China and the Korean Peninsula are reduced by the data assimilation. When comparing the scatter plots, the scatter plots from the NoDA experiment are dissimilar in the overall shape of both PM10 and PM2.5, and the models generally underestimate the PM concentrations than the observations.
In particular, in the case of correlation, PM10 is 0.34 and PM2.5 is 0.44, indicating a low agreement with the observed data. From these results, it can be seen that the particle concentration at the point of analysis was effectively improved by the data assimilation system. Errors are generally large in the lower atmosphere, as most aerosols originate from the surface.
Large concentration changes in the near-surface levels contribute to the model errors and forecast uncertainty. In the case of PM10, it is similar to PM2.5, but has the highest concentration in April, which is due to the Asian dust flying from the continent in spring (Jung et al.
Impacts on Air Quality Forecast
- Forecast Analysis for KORUS-AQ
- Forecast Analysis for All-Season
China region and southern sea did not show much difference between NoDA and the forecast, the data assimilation can fill the gap portion of the background model, which is not properly resolved by NoDA. The unique features of the local pollution case are that most of the highly polluted regions remained after the 24-h simulation. This means that the meteorological conditions of the current domain for this case are quite stable and the atmospheric transport stagnates, which can maximize the data assimilation effect.
It can be said that the correlation is calculated lower in the process of calculating the error balance. The special thing is that even in NoDA, the correlation increases with the passing of the forecast time, which is thought to be the uncertainty of the model itself. The results of Figure 3-17 are categorized into the grade predictions described in Chapter 2, and the data assimilation prediction results using data assimilation and bias correction for hit rate, false alarm rate, and HSS are shown in Figure 3-18. .
In the case of the DA experiment, the average hit rate decreased with the prediction duration and the false alarm rate gradually increased. In the case of the NoDA prediction, it was about 0.05, which was almost identical to the random reference prediction, and if data assimilation was done, PM10 increased to 0.37 and PM2.5 increased to 0.24. In the case of PM2.5, the increase in the prediction skill score due to data assimilation was very large and the bias correction effect was relatively small, but similar to the impact rate results, the amount of the score the ability to decrease followed by the duration of the prediction is reduced.
Data Assimilation Method
However, none of the EnKF results show better performance than 3DVAR, in which the 3DVAR produces better PM analysis data, at least for the assimilation period. The characteristics of the EnKF results are that mean bias is very small, compared to the 3DVAR. As a result, the RMSE for EnKF became smaller than the 3DVAR results, due to the overestimation of 3DVAR.
However, for the current situation, the EnKF method may be comparable to 3DVAR, but it cannot surpass the data assimilation ability of 3DVAR. In order to investigate the relationship between RMSE and spread of the EnKF method, a scatter plot between RMSE and spread was shown in Figure 4-3. In the EnKF, the RMSE indicates the overall bias of the analysis relative to the observation, and the range represents the standard deviation of the ensemble variables.
Both the EnKF experiments show a smaller spread than RMSE, which means that the EnKF is not in its optimal state. However, in the EnKF experiments, the comparison of EnKF_M and EnKF_N shows interesting differences in the relationships. The optimal condition is RMSE equal to dispersion, which means that the RMSE and dispersions have a positive relationship and the correlation will become 1 when the EnKF becomes an optimal condition.
Observing System Experiment
In addition, the surface PM data assimilation can affect not only the location of the surface station, but also the environment of the station. 4-5g and 4-5h indicate that the surface PM data assimilation helps suppress the overestimation bias in the continental interior by NoDA. When both the satellite AOD and surface PM observations are supplied to the data assimilation in DA3 (Fig.
Correlation and relative RMSE (rRMSE) are also shown for quantitative metrics to measure data assimilation performance. Data assimilation of surface PM observations provides the best fit to observed PM values (e.g., Fig. However, note that assimilation of satellite AOD data also improved surface PM analyzes (Fig. 4- 7b and f), especially in PM10.
When the initial conditions are determined by the data assimilation with both satellite AOD and surface PM observations (DA3), the forecast skill is best for both surface PM10 and PM2.5, except for the first few hours. The RMSE increases as the prediction time increases and the impact of data assimilation gradually disappears. Compared with DA3, DA2 exhibits significant skill, suggesting that surface PM data assimilation is one of the crucial components in improving the quality of surface PM forecasts.
Constraining Meteorology
Comparing surface PM10 concentrations between NF and FDDA, NF shows an overall higher concentration than FDDA, and the middle region of the Korean Peninsula shows the maximum difference than the surroundings. The peak of the decrease in wind speed is observed in the eastern and central part of the peninsula. The wind bias vector shows that the eastern coast of the peninsula shows easterly wind bias, whose eastward transport is more prohibited in the NF result.
From Chapter 4 onwards, one of the sensitivity tests compared the data assimilation methods and analyzed the data assimilation skills between 3DVAR and EnKF. A suitable EnKF can simulate the assimilation of aerosol data significantly better than the current simulation, the optimization of the EnKF system could be a suitable work for processing. 3DVAR and EnKF have their advantages and disadvantages of the method, and to combine the advantages of both methods, it was designed as a hybrid data assimilation.
There are several methods to perform hybrid data assimilation, and to start the research, the combination of 3DVAR and EnKF methods will be performed as En3DVAR. Quarterly Journal of the Royal Meteorological Society: Journal of Atmospheric Sciences, Applied Meteorology and Physical Oceanography. Woo, J.-H. and coauthors, 2020: Development of the CREATE inventory to support integrated climate and air quality modeling for Asia.