Impact of Assimilating Doppler Radar Data on Short-Term Numerical Weather Forecasting at Different Spatial Scales

Highlights

What are the main findings?

• Doppler radar data assimilation significantly improves the initial analysis of mesoscale systems and enhances the accuracy of short-term (0–3 h) quantitative precipitation forecasts, especially for heavy rainfall (>15 mm), compared to using only wind profiler radar data.
• The Barnes filter analysis reveals that forecast improvements from radar data assimilation are most pronounced at the mesoscale but are inherently limited to the first 2–3 h, after which skill rapidly decays. In contrast, large-scale systems show greater stability and predictability over a 6-h forecast period.

What is the implication of the main finding?

• This study underscores the critical value of high-resolution Doppler radar data for initializing and predicting mesoscale convective events, directly leading to more accurate and reliable short-term weather warnings for heavy rainfall.
• The findings highlight a fundamental predictability limit for mesoscale systems and suggest that future research should prioritize techniques to extend the useful forecast window of such rapidly evolving weather phenomena.

Abstract

This study explores the impact of assimilating Doppler radar data on short-term numerical weather forecasting for a heavy rainfall event in Southern China, focusing on different spatial scales. Results show that radar data assimilation significantly improves the initial analysis and enhances the accuracy of hourly precipitation forecasts by providing more detailed mesoscale system information, compared to assimilating only wind profiler radar data. The Barnes filter analysis reveals that radar data assimilation has a more pronounced effect on mesoscale systems, with improvements primarily concentrated in the first 2 h of the forecast. However, this improvement diminishes rapidly beyond the 2 h lead time, indicating the inherent predictability limits of mesoscale systems. In contrast, large-scale systems exhibit a greater stability and predictability, with radar data assimilation having a relatively smaller but still positive impact. The study emphasizes the importance of radar data assimilation for short-term forecasts at different spatial scales and suggests that future work prioritize extending mesoscale predictability.

1. Introduction

Numerical forecasting has rapidly advanced in recent years, leading to improved accuracy in weather prediction models. However, there still remains significant uncertainty in numerical forecasts at short time scales for extreme and high-impact weather events [1,2,3]. One important source of forecast error in numerical prediction often stems from initial errors [4,5,6,7,8]. Therefore, it is of crucial importance to enhance the accuracy of near-term numerical forecasts for heavy rain by incorporating additional observational data to refine the initial conditions.

Compared with conventional observations, Doppler weather radar delivers three-dimensional clouds and precipitation information at a higher spatio-temporal resolution and simultaneously depicts the structure and evolution of atmospheric flow [9,10,11]. It is commonly served as an important tool for monitoring mesoscale convective systems [12,13]. In many studies, Doppler weather radar observations are assimilated into NWP models to evaluate their impact on the forecasts of various weather systems, including the tornadic thunderstorm [14], supercell [15], squall line [16], Meiyu frontal rainfall [17], and tropical cyclone [18,19,20,21,22].

Diverse assimilation techniques, including the three-dimensional variational scheme (3DVar), the four-dimensional variational scheme (4DVar), the ensemble Kalman filter (EnKF), and the hybrid technique combining EnKF and 3DVar, which have been widely employed to assimilate Doppler radar data. Most studies on the radar data assimilation of radar data have been focused on using the 3DVar method because of its stability and lower computational cost [10,23,24,25,26]. 4DVar, capable of accurately retrieving both wind and thermodynamic fields, is a more sophisticated algorithm and has been integrated into radar assimilation systems such as the variational Doppler radar analysis system (VDRAS) [27], Japan Meteorological Agency’s 4DVar [28], China Meteorological Administration’s global forecast system [29], and the Weather Research and Forecasting (WRF) Model [30,31]. EnKF is also widely used for radar data assimilation (DA) due to its ability to provide flow-dependent covariance using an ensemble of forecasts. It is sometimes hybridized with 3DVar [32].

The enhancement of numerical forecasts by assimilating radial velocity and by assimilating radar reflectivity may be different. Xiao and Sun (2007) [10] assimilated radar reflectivity directly as a model variable using WRF-3DVar and indicated that reflectivity assimilation significantly modified the initial humidity and cloud water mixing ratio, whereas radial velocity assimilation has a larger impact on the initial wind. Tian et al. (2017) [33] showed that radar reflectivity DA improved storm event precipitation forecasting more significantly than radial velocity DA. Tsai et al. (2014) [34] suggested that radar reflectivity DA results in a faster response in quantitative precipitation nowcasting compared to radial velocity DA. Pu et al. (2009) [18] found that radar reflectivity DA can improve Hurricane Dennis’s thermodynamic and hydrological composition of the initial vortex, leading to more accurate precipitation forecasts than radar velocity DA. Conversely, radar velocity DA has a significant influence on the storm track and intensity. Similar results can also be seen in the analysis of Typhoon Morakot (2009) [35], Typhoon Jangmi (2008) [21], and Hurricane Ike (2012) [19].

Although the performance of different radar variables and DA methods varies across hindcast experiments, studies commonly show that radar DA significantly improves short-term precipitation forecasts by better resolving small-scale dynamic and thermodynamic structures in the initial conditions [25,36,37,38]. These studies mainly focus on evaluating the impact of radar DA on short-term forecasts at different validity times. Sun and Wang (2013) [30] investigated that radar DA by 3DVar had a positive effect on 0–6 h quantitative precipitation forecasts. Tsai et al. (2014) [34] indicated that the rainfall intensity and distribution for the first 6 h were improved with radar DA. Wang et al. (2016) [39] showed that the 24 h forecast of the typhoon best track and precipitation had the highest prediction skills after assimilating radar data. Bachmann et al. (2019) [40] compared the precipitation skill of 0–6 h forecasts with and without radar DA. All evidence suggest that radar DA is particularly effective for short-term numerical forecasts, but few studies have investigated the impact of radar DA on the short-term forecasts at different spatial scales.

In this paper, the impact of assimilating radar DA on the short-term numerical forecast of a heavy rainfall event in Southern China is investigated by separating the analyses and forecasts into different spatial scales using the Barnes filter method [41]. Then the influences of radar DA in meso- and large-scale structures of analyses and forecasts are explored. The weather case, quality control of observational data, model configuration, and experimental design are described in Section 2. In Section 3, the impacts of DA with and without radar data on the analysis and forecast of precipitation greater than 0.1 and 15 mm are compared. The relevant mechanisms are diagnosed and examined. Finally, the summary and conclusions are given in Section 4.

2. Weather Case, Model, and Experimental Design

2.1. Weather Case

An extreme precipitation event occurred in the Fujian Province of China during 14 to 17 May 2019. Figure 1a shows the observed 24 h accumulated precipitation from 1200 UTC 15 May to 1200 UTC 16 May.

It is noteworthy that the maximum 24 h accumulated precipitation in west-central Fujian was 308.9 mm, which broke the local historical record since 1961. Each maximum 6-hour accumulation (0−6, 6−12, 12−18, and 18−24 h, starting at 1200 UTC 15 May 2019) exceeded 70 mm. The 6−12 UTC period recorded a peak value of 193.3 mm (Figure 1b). Most global numerical weather prediction (NWP) models were limited in their ability to forecast this extreme precipitation event. Both the intensity and distribution of 24 h accumulated precipitation forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP) were not similar, based on the observations (Figure 1c,d).

Although the forecasts of precipitation location and intensity from mesoscale models were more accurate than that of the global models, the forecasted maximum 24 h precipitation was less than the observations, substantially underestimating the event and thereby degrading forecast utility. This exceptionally rare rainfall event served as a case study to evaluate the impact of radar DA at different spatial scales.

2.2. Model Configuration and Data

The Weather Research and Forecasting Model (Version 3.7) [42] with the 3DVar data assimilation scheme [43] is used to perform the experiments in this study. Same as in [44], the model domains are nested with 141 × 131 coarse grid points in the parent domain (D01) and 241 × 241 fine grid points in the inner domain (D02) (Figure 2). The horizontal resolutions of the two domains are 18 and 6 km, respectively, and the model is vertically divided into 53 layers. The physical parameterizations are the same as in the experiments carried out in [44], including the WRF single-moment five-class microphysics scheme [45], the rapid radiative transfer model longwave radiation scheme [46], the Goddard shortwave radiation scheme [47], the Grell 3D cumulus parameterization scheme [48], and the MYJ planetary boundary layer scheme [49]. Global forecasts from the global forecast system (GFS) at NCEP provide the background fields for data assimilation and free model forecasts.

The observations of the wind profiler radar (WPR), automatic weather station (AWS), and S-band weather radar are collected by the China Meteorological Administration (CMA) (Figure 2). There are approximately 42 WPR stations in D02, and all the observations are used for the DA experiment. The 850-hPa wind observations reveal that low-level convergence collocated with the extreme precipitation site, confirming that WPR winds are consistent with the precipitation. Both reflectivity and velocity observations from the S-band weather radar (Z9598) closest to the precipitation center have been assimilated in the sensitivity DA experiments, with reflectivity being assimilated directly as a model variable. The radar detection range and resolution are 230 km and 1°*1 km, respectively. Observational data was assimilated solely in D02 with the 3DVar data assimilation scheme. AWS observations are withheld from the DA experiments and used solely for the independent verification of forecast accuracy.

2.3. Radar Data Preprocessing

Radar data are preprocessed using the Python ARM Radar Toolkit (Py-ART, V1.11.0) [50]. Quality control includes reflectivity despeckling, velocity dealiasing, and velocity texture filtering. Reflectivity values below 10 DBZ are discarded to focus on the severe convection cores. Then, the quality-controlled radar reflectivity and velocity data are interpolated from spherical to Cartesian coordinates. Observation errors for radar reflectivity and velocity are estimated as the nine-point standard deviation around each central grid point [23]. After quality control, the radar reflectivity and velocity data are preprocessed by a typical thinning step. They are smoothed and interpolated into a grid with a horizontal resolution of 6 km that is the same as the resolution in the WRF model (Figure 3).

2.4. Experimental Design

Focusing on the 6 h peak precipitation window (1800 UTC 15 May to 0000 UTC 16 May 2019), this study examines how the Doppler radar DA affects forecasts of this event. Hindcast experiments were conducted within a cycling DA and forecast framework based on the WRF-3DVar (Figure 4). The WRF model was cold started at 1500 UTC 15 May, with first-guess (FG) and boundary conditions downscaled from the NCEP GFS global forecast. A 3 h forecast without DA was performed to provide the FG for the first DA cycle, initialized at 1800 UTC 15 May. Observations were then assimilated during the first DA cycle to produce the analysis that initialized a 6 h forecast with hourly output. Afterward, hourly assimilation was performed over the next six DA cycles, with FG fields taken from the preceding 6 h forecast. In the sensitivity experiment (RAD), both WPR and S-band radar data were assimilated, whereas the control run (CTL) used only WPR observations (

Table 1. Sensitivity experiments and the observations assimilated in the experiments.

Experiments	Observations Assimilated in the Experiment
CTL	Wind profiler radar
RAD	Both wind radar profiler and S-band Doppler radar

).

2.5. Methods

2.5.1. Barnes Filtering Method

The Barnes filtering method is a widely used technique in atmospheric sciences for spatial interpolation and data smoothing [41], which is particularly effective in separating mesoscale features from larger-scale atmospheric data. The method applies low-pass filtering to smooth the data and employs weighted averaging to update field estimates, with weights decreasing as the distance to the target point increases. By iteratively refining the initial estimate, first correction, and final correction, Barnes filtering produces a smooth and accurate spatial field.

For a target point $\left(x,y\right)$, mesoscale features are calculated by removing the large-scale component from the original field:

$$F_M\left(x,y\right)=F_O\left(x,y\right)-F_L\left(x,y\right)$$

(1)

where $F_L\left(x,y\right)$ indicates the large-scale component (formulated in Appendix A). $F_O\left(x,y\right)$ and $F_M\left(x,y\right)$ are the original and mesoscale atmospheric fields, respectively. $x$ and $y$ represent the longitude and latitude.

2.5.2. Threat Score and Bias Score

Surface precipitation forecasts were verified against automatic weather station (AWS) observations using the Threat Score ($TS$) and Bias Score ($BIAS$):

$$TS= {\displaystyle \frac{C}{F+NO-C }}$$

(2)

where $F$ is the number of forecasted events, $NO$ the number of observed events, and $C$ is the number of correct forecasts. $TS$ ranges from 0 to 1, with higher values indicating better skill.

$BIAS$ is defined as

$$BIAS={\displaystyle \frac{F}{NO}}$$

(3)

$BIAS$ ranges from 0 to positive infinity. A value ≈ 1 indicates an unbiased forecast, while a value > 1 (<1) signifies overestimation (underestimation).

2.5.3. Pattern Correlation Method

The pattern correlation method is widely used to quantify the similarity between datasets or patterns, with the Pearson correlation coefficient (PCC) being a common approach. It measures the linear relationship between two variables, ranging from −1 (perfect negative correlation) to +1 (perfect positive correlation), with 0 indicating no linear relationship. This method is applied in fields such as image recognition, time series analysis, signal processing, and data clustering, which is sensitive to outliers and primarily captures linear relationships.

The PCC is a measure of the linear relationship between two variables. Its formula is

$$\mathrm{P}\mathrm{C}\mathrm{C}={\displaystyle \frac{\sum _{i=0}^N\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{{\sum }_{i=0}^N{\left(X_i-\overline{X}\right)}^2}\sqrt{{\sum }_{i=0}^N{\left(Y_i-\overline{Y}\right)}^2}}}$$

(4)

where $X_i$ and $Y_i$ are the ith data points of the two variables, and $\overline{X}$ and $\overline{Y}$ are their means. $N$ is the sample number of the data points.

3. Results

3.1. Improvement in the Initial Analysis by Assimilating Radar Data

An effective data assimilation system can incorporate observational data into the analysis, resulting in an analysis that better aligns with the observations compared to the background [44]. To evaluate the agreement between radar observations (“O”) and both the background (“B”) and analysis (“A”), the probability density function (PDFs) of the observation-minus-background (OMB) and observation-minus-analysis (OMA) differences for the RV and REF, averaged over all cases, are assessed. As shown in Figure 5, the PDFs of both OMB and OMA for RV and REF present a Gaussian-like distribution with an approximately zero mean (Figure 5), implying that the fitting of both the analysis and background is nearly unbiased (i.e., exhibit approximately zero skewness). Quantitatively, the standard deviation of OMA is smaller than that of OMB by about 15% for the RV (2.85 vs. 3.37 m s⁻¹) and 9% for the REF (3.64 vs. 4.00 DBZ), suggesting that the analysis fits the observations better than the background.

Figure 6 shows 850-hPa meridional wind and wind vector analyses at 2000 UTC 15 May 2019 for CTL (Figure 6a), RAD (Figure 6b), and their analysis increment (i.e., analysis minus FG) (Figure 6c,d). Near 26°N, 117.5°E, where the heavy rainfall event occurred (Figure 1a), the RAD analysis exhibits stronger northerly winds than the CTL analysis. The southerly–northerly wind convergence in the RAD analysis (Figure 6b) is more favorable for heavy rainfall than the purely southerly wind in the CTL analysis (Figure 6a). The RAD analysis increment reveals a clearer convergent southerly–northerly couplet (Figure 6d), implying that radar observations resolve finer meso- and small-scale structures than WPR observations and thus enhance localized short-term heavy-rain forecasts. Similar features also appear at 925 and 700 hPa. Radar data assimilating intensifies the low-level southerly–northerly convergence near 26°N, 117.5°E. Conversely, radar data assimilation enhances the upper-level southerly–northerly divergence at 300 hPa over the same area. This low-level convergence coupled with the upper-level divergence favors heavy precipitation development.

3.2. Improvement in the Forecast by Assimilating Radar Data

The TS and BIAS for hourly precipitation forecasts from CTL and RAD, averaged over all cases, are shown in Figure 7. Both CTL and RAD exhibit a greater forecast skill for precipitation thresholds exceeding 0.1 mm (Figure 7a). Although the TS declines from 0.19 to 0.15 with the lead time, RAD consistently outperforms CTL in the hourly precipitation forecast, especially within the first 5 h. The maximum TS improvement of RAD over CTL reaches about 11% for the 0.1 mm thresholds at the 1 h lead time. For precipitation exceeding 15 mm, RAD demonstrates a significantly higher forecast skill than CTL. The TS of RAD is nearly three times that of CTL at 1h. It possibly implies that the WPR observation assimilated in the CTL are sparser and thus cannot provide as detailed information on mesoscale systems as radar observations. Compared with radar assimilation, the lack of mesoscale detailed in the WPR-based analyses may yield a less accurate precipitation forecast.

BIAS values of hourly precipitation exceeding 0.1 mm are similar between the CTL and RAD (Figure 7c). But for precipitation exceeding 15 mm, the CTL shows a significantly lower BIAS than RAD (Figure 7d), suggesting that the forecast with less detailed mesoscale information in the analysis may underestimate the area of heavy hourly rainfall. By resolving detailed mesoscale structures, radar assimilation enlarges the heavy-rain area and thus yields a higher forecast skill for intense precipitation.

Since the 6 h maximum accumulated precipitation peaked at 193.3 mm between 1800 UTC 15 May and 0000 UTC 16 May 2019 (Figure 1b), the hourly precipitation skill of CTL and RAD, initialized at 2000 UTC 15 May, is examined (Figure 8). Observations (Figure 8c,f,i) reveal a pronounced precipitation maximum over central Fujian, with hourly amounts exceeding 15 mm. Noticeable underestimation occurs in CTL (Figure 8a,d,g), where both the precipitation area and intensity are markedly smaller than observed. Meanwhile, the RAD better captures the intensity of the maximum precipitation center than CTL (Figure 8b,e,f). This improvement likely reflects radar assimilation refining the structure and intensity of mesoscale systems in the initial conditions (see Figure 6). Moreover, the forecasting skill of both RAD and CTL deteriorated markedly after the third hour, suggesting that the predictability limits of such mesoscale systems may be approximately three hours. To address this question, the CTL and RAD analyses and forecasts are decomposed into large-scale and mesoscale components using the Barnes filter, and the predictability of each scale is subsequently examined.

3.3. Impact of Radar DA on the Predictions at Different Spatial Scales

3.3.1. Selection of Weight Constant in the Response Function

The Barnes filter is a widely used spatial filtering technique that applies Gaussian weights to interpolate each grid point, yielding a low-pass filter that removes high-frequency fluctuations. By tuning the filtering constant C, low-frequency fluctuations at various scales are retained, enabling the separation of large-scale and mesoscale systems. Figure 9 illustrates the response functions for different filtering constants C. For a smaller C (e.g., 2000), the filter function converges rapidly at shorter wavelengths, and the response function quickly reaches its maximum value. This results in the filtering of small-scale wave systems with wavelengths less than 200 or 300 km. Conversely, for larger values of C (e.g., 40,000), the filter function converges at longer wavelengths, and the response function approaches its maximum more gradually. Thus, the filter predominantly affects larger-scale waves, removing systems with wavelengths smaller than 1000 km. In this study, a value of C = 7000 is chosen to separate systems with wavelengths smaller than 500 km from larger-scale features.

3.3.2. Impact of Radar DA on the Analysis at the Meso- and Large-Scale

Figure 10 compares meso- and large-scale analysis increments of 850-hPa divergence and wind vectors between the CTL and RAD experiments. In CTL, which assimilates only wind profiler radar data with a coarse spatial resolution of ~80 km, the resulting mesoscale divergence field is spatially homogeneous and weakly convergent. Wind vectors are scattered, reflecting the wind profiler radar’s inability to resolve fine-scale atmospheric structures (Figure 10a). At the large-scale, the convergence is weak, and the wind vectors lack necessary details (Figure 10c).

In contrast, the RAD experiment, which assimilates both the wind profiler and Doppler weather radar data, shows notable improvements. The Doppler weather radar, featuring a high spatial resolution of 1 km, enables RAD to resolve more concentrated and intense convergence zones at the mesoscale (Figure 10b). Correspondingly, wind vectors accordingly show a clear tendency to converge. The large-scale RAD field reveals local convergence characteristics and distinct wind vector circulation patterns (Figure 10d). Evidently, Doppler radar assimilation markedly enhances the accuracy of analyzing these key atmospheric parameters.

3.4. Temporal Evolution of Meso- and Large-Scale Systems in the Analysis

Figure 11 shows 850-hPa wind and divergence analyses at different spatial scales from 1800 UTC 15 May to 0000 UTC 16 May 2019. At the mesoscale, the RAD experiment exhibits well-defined strong convergence zones. RAD wind vectors also show a concentrated and organized convergence pattern, confirming its superiority in resolving mesoscale wind field convergence.

In the large-scale analysis, RAD (Figure 11(d1–d7)) provides a more finer-scale convergence field and wind vector circulation patterns than CTL (Figure 11(c1–c7)). Likewise, in the original analysis, RAD (Figure 11(f1–f7)) captures details more effectively than the CTL (Figure 11(e1–e7)). In summary, assimilating Doppler radar data markedly improves the depiction of mesoscale wind field convergence and the multi-scale temporal evolution of weather systems.

Because the RAD analysis resolves finer observational detail, the averaged Pearson correlation coefficients (PCCs) of 850-hPa zonal (U) and meridional (V) winds between the RAD analyses across all cases are calculated (Figure 12). For both zonal (U) and meridional (V) winds at the large-scale, the PCCs between the current and late-hour wind analysis show a gradual, steady decline, reflecting the slow evolution and persistence of the large-scale circulation systems. Compared with large-scale wind fields, the PCCs of mesoscale zonal and meridional components decline rapidly within the first three hours. After 3 h, the PCC index saturates at a low value, indicating that mesoscale wind fields evolve fastest and are the least predictable.

The comparison shows that U-wind PCCs remain consistently higher and decay more slowly than V-wind PCCs, indicating the more rapid evolution of V-wind. This may be attributed to the fact that U-winds are more prominently dominated by planetary-scale circulation. Especially at the mid to high latitudes, the westerly belt imposes a stable zonal wind pattern that is easier to predict. The V-wind field, in addition to being affected by large-scale circulation, is perturbed by complex factors such as frontal activities and topographic forcing, rendering its evolution more complex and less predictable.

3.5. Predictability in the Forecast at the Meso- and Large-Scale

Average PCCs of 850-hPa wind forecasts against the corresponding analysis are compared between the CTL and RAD at 1–6 h lead times for different spatial scales (Figure 13). The results indicate that large-scale information exhibits good stability and reliability, retaining high PCCs even at the 6-hour forecast time. In contrast, the mesoscale accuracy declines rapidly, with PCCs falling 50–60% between 1 and 6 h in both the RAD and CTL. This underscores the inherent complexity and variability of mesoscale weather systems. When comparing the two experiments, RAD generally exhibited higher PCCs than CTL, especially for mesoscale forecasts within the first 3 h, demonstrating the short-term advantages of radar assimilation. Weather radar assimilation markedly improves 850-hPa wind speed forecasts, particularly for mesoscale fields. However, these improvements diminish by the 6th hour, especially for the V-wind component, indicating the short-term benefits of weather radar data assimilation. While large-scale information demonstrates greater stability in forecasts, the impact of radar assimilation on large-scale forecasts is relatively limited.

3.6. Impact of Different Filter Parameters

Figure 14 illustrates how filter parameters (varying wavelengths) affect 850-hPa U- and V-wind field forecast improvements. Regardless of filter parameters, radar assimilation improves mesoscale winds only within the first 2 h and decays rapidly (Figure 14a,b). This consistent result indicates that radar assimilation offers a clear short-term forecast advantage but contributes little over a longer lead time.

Figure 14c,d further examine how filter parameters influence the improvement from radar assimilation. The results indicate that, as the filter parameter increases (i.e., wavelength lengthens), the enhancement effect weakens, suggesting that shorter wavelengths (mesoscale information) benefit more from radar assimilation. Moreover, the red–blue bars difference shows that shorter wavelengths yield a faster loss of both improvement and forecast skills. This highlights the greater effectiveness of radar assimilation for short-term forecasts at the mesoscale, although its effectiveness decreases rapidly over time.

4. Summary and Conclusions

This study investigated the impact of Doppler radar assimilation on short-term numerical weather forecasting for a Southern China heavy-rain event across different spatial scales. The following are the key findings and conclusions:

Improvement in Initial Analysis: Radar assimilation significantly improved the initial analysis compared to assimilating only the wind profiler data. The analysis increments showed more detailed and refined mesoscale features, particularly in the 850-hPa wind field. This improvement is crucial for enhancing the accuracy of short-term heavy precipitation forecasts.

Enhancement in Precipitation Forecasting: Radar assimilation leads to notable improvements in hourly precipitation forecasts, especially for precipitation exceeding 0.1 mm. The Threat Score and Bias Score verified that radar assimilation consistently outperformed wind profiler data assimilation, especially within the first 3 h. For precipitation over 15 mm, the improvement was even more significant, confirming that radar data resolves the mesoscale detail essential for heavy-rain accuracy.

Spatial Scale Analysis: The Barnes filter decomposition revealed that radar assimilation has a more pronounced effect on mesoscale weather systems. Improvement is primarily concentrated in the first 2 h and decays rapidly thereafter. This highlights the short-term benefits of radar assimilation and suggests that the predictability of mesoscale systems may be limited to about 3 h. This limitation may arise from the poor predictability of strongly nonlinear, rapidly evolving mesoscale weather systems.

Predictability of Weather Systems: Large-scale weather systems exhibit greater stability and predictability, maintaining relatively high Pearson correlation coefficients (PCCs), even at the validity time of 6 h. In contrast, the mesoscale systems showed a rapid decline in PCCs, indicating their inherent complexity and variability. Radar assimilation played a crucial role in enhancing the forecast accuracy for mesoscale systems in the short term but had limited impact on long-term forecasts.

Limitations and Future Work: Despite the significant improvements from radar data assimilation, the study also identified limitations. The rapid decline in forecast skill after the first few hours suggests that the predictability of mesoscale systems is inherently limited. Sparse observations and data assimilation techniques cannot fully resolve convective cells or boundary layer structures, inevitably introducing initial uncertainties. Future work should focus on refining numerical models and exploring advanced data assimilation techniques to extend the predictability of these systems. More radar data assimilation experiments for diverse weather cases (e.g., tropical cyclones, squall lines) will also be conducted.