Data assimilation effect

Data Assimilation Effects

Data Assimilation Techniques

3D-Var

Even if the number of observations continues to increase, it has the advantage of having a low computational cost and creating a more balanced analysis, but there is a limitation of using homogeneous/isotropic/fixed background error covariance.

EnKF

The error statistics of the 3D-Var technique are static, isotropic, and almost homogeneous, misrepresenting the actual error statistics that are inherently flow-dependent across space and time. However, the EnKF data assimilation method is designed to provide flow-dependent background error covariance. Meanwhile, the 3D-Var technique is particularly suitable for operational purposes as it requires fewer computational resources and does not require the construction of ensembles like EnKF or the simulation of model trajectories like 4D-Var.

Hybrid (E3DVAR)

Comparing the performance of WRF-based 3DVAR, EnKF, and E3DVAR (hybrid) in East Asia, E3DVAR, a hybrid data assimilation method, outperformed 3DVAR and EnKF in East Asia.

Improving Short-term Precipitation Forecasting Effect of Data Assimilation

When 3D-Var data assimilation is applied every 1 hour, the RMSE of 2m temperature decreases by about 75%, and the RMSE of precipitation per hour decreases by about 57%. It was confirmed that frequent cycle data assimilation limits initial condition errors, captures medium-scale convective evolution, and improves spatial agreement with radar reflection observations.

AI-Based Short-Term Precipitation Prediction

Existing global precipitation prediction models have limitations in real-time response due to the vast amount of computation, and it was often difficult to accurately predict Korea's local precipitation patterns with complex topographical characteristics. Professor Hong Young-joon's team at Seoul National University developed an ultra-short precipitation prediction AI model that combines a local time-space attention mechanism and a new up-sampling structure.

Initial Condition Improvement Effect

The fundamental role of data assimilation is to calibrate the atmospheric conditions (initial conditions) from which the model departs close to actual observations.

The quality of the initial conditions is determined by the forecasting model, assimilated observation data, and data assimilation techniques. A more accurate forecasting model creates a more accurate first background, which allows more observations to be assimilated, and creates a virtuous cycle that further improves the short-term forecast (background) by increasing the accuracy of the analysis site.

In particular, correction of initial conditions in the underwater phase is key to predicting local precipitation. When the water vapor in the early stage of the model is restored by using the statistical relationship between radar reflectance, relative humidity, and water bodies, the water vapor content in the middle layer of the model increases in a direction that matches the convective region well after data assimilation, and the precipitation distribution simulation was successful.

Spin-up Troubleshooting Effect

Spin-up: A phenomenon in which the model's atmospheric conditions are unstable when it first starts up, resulting in poor predictions in the first few hours.

The prediction accuracy of the medium-sized NWP model is negatively affected by the error of the spin-up effect and the initial/side boundary conditions. It has been shown through end-of-life studies that assimilation of real-time observations including weather radar data can improve the reliability of precipitation prediction of the NWP model.

Applying data assimilation in the repeated cycling (RUC) method solves this problem more effectively. Since each cycle starts in the analysis field of the previous cycle, the hydrological and meteorological variables have sufficient time to spin-up, which has an important practical advantage over the re-initialization to cold-start in the global analysis field.

Effects of Data Assimilation Cycle

The graph compares how the temperature prediction error (RMSE) differs according to the data assimilation cycle .

DA cycle : Cycle1=1 hour, cycle3=3 hours, cycle6=6 hours, cycle12=12 hours

The temperature was lowest when applied with a 1-hour cycle, and as the cycle lengthened, the error increased, so that when no data assimilation was performed at all (Non-DA), it surged to over 1.2 degrees.

→ The shorter the data assimilation cycle, the more frequently the initial condition error is corrected, improving prediction accuracy.

Comparison of Spatial Distribution of Cumulative Precipitation

GPM satellite observations showed that precipitation was concentrated in western Bali, and the Cycle 1 experiment best reproduced this pattern.

Conversely, as the data assimilation period lengthened, the location of precipitation shifted eastward and western precipitation was underestimated, while the location of precipitation itself was significantly misaligned in the Non-DA experiment.

→ Data assimilation plays a key role in predicting the spatial location of precipitation.

Comparison of Wind Speed Bias

In all experiments, wind speed was underestimated, but the magnitude of this underestimation decreases as the data assimilation period shortens.

The bias for cycle 1 is approximately -1.25 m/s, whereas for Non-DA it reaches -4.5 m/s.

Wind speed is directly linked to low-level convergence; the better low-level convergence is represented, the more accurately convective organization is simulated.

→ This graph indirectly illustrates the dynamic mechanism by which data assimilation improves precipitation forecasting.

Radar Reflectivity Comparison

This figure compares the actual BMKG radar observed reflectances with the simulated reflectances from each experiment. Higher reflectances indicate stronger precipitation.

The actual radar observed strong convective zones exceeding 40 dBZ across western to central Bali.

Cycle 1 simulated reflectances of 20–40 dBZ in the same region, showing the closest results.

Cycles 3 and 6 showed weak reflectances of 10–25 dBZ, while Non-DA was only 5–10 dBZ, failing to capture the convective intensity itself.

The comparison of radar reflectances directly demonstrates that data assimilation improves both the strength and location of the convective system.