Sensitivity of Summer Precipitation over Korea to Convective Parameterizations in the RegCM4: An Updated Assessment

*is study investigates the performance of the latest version of RegCM4 in simulating summer precipitation over South Korea, comparing nine sensitivity experiments with different combinations of convective parameterization schemes (CPSs) between land and ocean. In addition to the gross pattern of seasonal and monthly mean precipitation, the northward propagation of the intense precipitation band and statistics from extreme daily precipitation are thoroughly evaluated against gridded and in situ station observations. *e comparative analysis of 10-year simulations demonstrates that no CPS shows superiority in both quantitative and qualitative aspects. Furthermore, a nontrivial discrepancy among the different observation datasets makes a robust assessment of model performance difficult. Regardless of the CPS over the ocean, the simulations with the Kain–Fritsch scheme over land show a severe dry bias, whereas the simulations with the Tiedtke scheme over land suffer from a limited accuracy in reproducing spatial distributions due to the excessive orographic precipitation. In general, the simulations with the Emanuel scheme over land are better at capturing the major characteristics of summer precipitation over South Korea, despite not all statistical metrics showing the best performance. When applying the Emanuel scheme to both land and the ocean, precipitation tends to be slightly overestimated. *is deficiency can be alleviated by using either the Tiedtke or Kain–Fritsch schemes over the ocean instead. As few studies have applied and evaluated the Tiedtke and Kain–Fritsch schemes to the Korean region within the RegCM framework, and this study introduces the potential of these new CPSs compared with the more frequently selected Emanuel scheme, which is particularly beneficial to RegCM users.


Introduction
During recent decades, substantial efforts have been made to improve the performance in simulating the detailed characteristics of regional to local climate over the Korean Peninsula. Dynamical downscaling using regional climate models (RCMs), in particular, has offered the potential to resolve the physical processes modulated by complicated geographical features [1][2][3][4], eventually resulting in improved climate statistics compared with those from largescale forcings (e.g., global climate models or reanalysis data). Despite advances in modeling, summer precipitation remains a great challenge due to distinct temporal and regional variations and its stochastic nature [5][6][7][8].
Not only approximately 70% of the annual precipitation is concentrated in the summer season [9] but also its interannual and intraseasonal variations are significant. In addition, summer precipitation over Korea is shaped by many different mechanisms, including the East Asian summer monsoon (EASM) frontal systems, mesoscale convective systems, and typhoon activities [8,[10][11][12]. e complex interactions between them make it difficult to reproduce both quantitative and qualitative aspects of summer precipitation. Although RCMs with higher resolution are able to employ the refined surface forcing, the mechanisms mentioned above are not greatly dominated by stationary forcings (e.g., topography); instead, they are dominated by stochastic factors such as convective instability. As a result, most RCMs face many difficulties capturing the temporal and spatial behaviors of summer precipitation at regional to local scales with a high level of accuracy. Specifically, the main difficulties include (1) characteristics of daily precipitation (i.e., overestimation of frequency and underestimation of intensity) [3,4,7]; (2) intensity and position of some localized extreme precipitation [7]; (3) phase and amplitude of the diurnal variation of precipitation [8]; and (4) occurrence and extent of typhoons [13].
Precipitation simulation is sensitive to many different experimental configurations (e.g., the lateral boundary conditions, horizontal resolution, and parametrization schemes) [14]. Because of the complex formation process of precipitation, there is no single method that can bring a dramatic improvement in simulating precipitation with respect to all aspects. However, selecting a suitable convective parameterization scheme (CPS) is most directly related to the precipitation formation in the model, especially in summer when precipitation is predominantly convective in nature [15][16][17][18]. erefore, finding the optimal CPS over the target region is essential for determining the model configuration if there is flexibility in the choice of multiple schemes under the same modeling framework.
e different versions of the Abdus Salam International Center for eoretical Physics (ICTP) Regional Climate Model (RegCM) is one of the most widely used RCMs applied to the Asian region, including the Korean peninsula. Two noticeable improvements related to CPS in RegCM version 4 (RegCM4) are running different CPSs over land and ocean, referred to as "mixed convection schemes (MCS)" [19], and including new CPSs, the Tiedtke scheme (added from RegCM4.3) and Kain-Fritsch scheme (added from RegCM4.5). Further options and increased complexity require additional tests for appropriate practical application of the upgraded version. MCS have been widely tested in a significant body of research over different regions [20][21][22][23][24][25], but the advantages and necessity of MCS are highly dependent on the domain and other configurations of the simulation. e evaluation of the Tiedtke and Kain-Fritsch schemes, newly incorporated in RegCM4, is relatively limited because these versions have only recently been released to the public. While Ali et al. [20] demonstrated that the Tiedtke scheme shows a better performance in simulating summer precipitation over China, Fuentes-Franco et al. [26] suggested the benefits of using the Kain-Fritsch scheme in simulations of tropical cyclones over Central America. From these exemplary studies, the implementation of MCS and the new CPSs are likely to benefit the regional climate simulation over the selected regions. However, good performance of a CPS in one region does not guarantee its success in other regions, because their performance is greatly influenced by region, seasonality, and their combinations with other schemes. For the regional climate simulation focusing on South Korea, the Emanuel scheme is the most commonly selected CPS within the RegCM configuration [4,6,7,27]. Any sensitivity tests for MCS and new CPSs have not been evaluated and documented yet, providing significant room for a systematic investigation and updated assessment.
In this study, we investigate the relative performance of MCS based on comparative analysis of the nine sensitivity experiments with different combinations of three CPSs (Emanuel, Tiedtke, and Kain-Fritsch) using the latest version RegCM4.7. We focus on the CPSs' sensitivity to summer precipitation over South Korea. e statistics, on monthly and daily time scales, are evaluated against both gridded and station observations. It helps understand the structural error from CPSs and relative strength and weakness of individual CPS. erefore, the updated assessment of the new CPSs in the latest version RegCM4 can provide an opportunity to augment previous findings and add to baseline information for the usage of RegCM4.  [32], and Kain-Fritsch scheme [33]. In our study, we exclude the Kuo and Grell schemes based on previous studies. It has been demonstrated that the Kuo scheme generally provides poor performance of precipitation simulation, particularly in high resolution (<30 km) [25,28], and the Grell scheme tends to underestimate summer precipitation over the Korean region, showing poorer performance than the Emanuel scheme [27,34]. erefore, we have designed nine experiments that include all combinations of the three CPSs (Emanuel, Tiedtke, Kain-Fritsch) over land and ocean as described in Table 1. When naming the individual experiments, the bold letters (E, T, and K) denote the three different CPSs by the first letter of the CPS names, and land and ocean are denoted by the small italic letters l and o, respectively. is experimental design makes it possible to quantify the sensitivity of these MCSs and estimate the relative performance of Tiedtke and Kain-Fritsch schemes, which have not yet been applied to the Korean region. e relative performances of these schemes are compared with that of the Emanuel scheme, which is the CPS most frequently selected for the Korean climate simulation, using RegCM. Other physics parameterizations employed are the National Center for Atmospheric Research (NCAR) Community Climate Model version 3 (CCM3) radiation scheme [35], the Holtslag planetary boundary layer scheme [36], the Subgrid Explicit Moisture Scheme (SUBEX) [37] for gridscale cloud process, and the Zeng ocean flux scheme [38]. All of the simulations are coupled with the land surface model Community Land Model version 4.5 (CLM4.5) [39]. Figure 1 presents the domain and topography used in our experiment. e domain covers northeastern Asia centered on the Korean Peninsula and adjacent seas with e initial and boundary conditions are obtained from the ERA-Interim reanalysis with a resolution of 0.75°× 0.75°at 6-hour intervals [40]. e simulations span from December 1997 to December 2007. e first month is considered as the spin-up period, and the 10-year period from 1998-2007 is used for analysis.
Since CRU, APHRODITE, and TRMM have a different grid system, they are interpolated to achieve the same resolution as the model using the inverse-distance-weighting method, which facilitates to calculate statistics such as the spatial correlation between the observation datasets and the model output. In addition to the multiple gridded observation datasets, we also use in-situ station observations for in-depth daily analysis. As well documented by Juneng et al. [46], a large discrepancy is found among CRU, APHRO-DITE, and TRMM, indicating that there is uncertainty embedded in the observation data. erefore, comparing the simulation with both multiple gridded and in-situ observations will enhance the reliability of the evaluation.

Evaluation of the Monthly and Seasonal Mean
Precipitation. In this section, the analysis focuses on the comparison between simulated and observed precipitation characteristics, at monthly or seasonal mean timescales. Figure 2 presents the annual cycle of monthly mean precipitation averaged over South Korea. For the 10-year climatology, the precipitation pattern shows a strong seasonal variation. is is because the Korean climate is characterized by cold dry winters and warm wet summers due to the influence of the East Asian monsoon system. In accordance with the amount of precipitation, the sensitivity of each CPS varies from season to season. While the difference in the performance among the CPSs seems to be negligible in the winter, performance deviation among different CPSs is significant during the summer. Similarly, the deviation of the CPS simulation is more pronounced in wet years than dry years. All simulations are reasonably good at simulating the dominant annual cycle of precipitation with a maximum peak in July or August. However, the CPS is able to derive a different performance mainly in the quantitative aspect, differentiating their performance in terms of the amount of summer precipitation. A severe underestimation of summer precipitation occurs in the simulation with the Kain-Fritsch scheme over land ("lK"), regardless of the CPS over ocean. On the other hand, the simulations with the Emanuel and Tiedtke scheme over land ("lE" and "lT") show a mixed performance with both underestimation and overestimation patterns. Furthermore, the discrepancy among the different observation datasets is not trivial, which makes it difficult to provide an unambiguous assessment of model performance. e best performing simulation varies according to which the observation dataset is used for the comparison. Figure 3 presents the spatial distribution of precipitation averaged over the warm season. Since approximately 70% of the annual total precipitation is concentrated within only several months (June-July-August-September, hereinafter as JJAS), we focus on the different behaviors of precipitation simulated during JJAS hereinafter. First, the spatial details across different observations are different, reinforcing the uncertainty embedded in observation data. While the maximum in the southern coastal region is prominent, the shape and magnitude of other maxima in the northern part of South Korea differ noticeably from one another. For example, APHRODITE shows two separated, localized maxima, but precipitation from CRU and TRMM forms a broad band. For the quantitative aspect, the difference between TRMM and APHRODITE reaches more than 1 mm/ day. Such a limited consistency may stem from different      Table 3), despite having the best spatial correlation with the observations (see correlation coefficient in Table 3). For the simulations with lE and lT (the first and third sets of 3 rows in Table 3), the spatial-averaged precipitation amount is more reasonable compared with those with lK. However, the simulations with lE and lT show some problematic behavior in the positioning of precipitation maxima. e excessive orographic precipitation occurs in the high-elevated mountains, particularly along the east coast in the simulations with lT. is is responsible for the lower spatial correlation coefficient presented in Table 3. is may be caused by over-sensitivity of parameterization of rainfall processes to topography. Table 3 provides a statistical comparison of the overall performance of the nine sensitivity experiments against the gridded observations of APHRO-DITE, TRMM, and CRU. All experiments show mixed performance with both strengths and weaknesses when compared with the observation datasets. Moreover, the simulation with the best performance varies according to which observation data is used. Clearly, there is no single CPS that shows a superior performance across all comparisons. Nevertheless, lE might be considered the most reasonable configuration, taking both quantitative and qualitative behaviors into considerations. e CPS over ocean, although not comparable with CPS over land, also affects the precipitation pattern over land through the advection of the convection system. e Tiedtke and Kain-Fritsch scheme over ocean (oT and oK) tend to reduce the amount of precipitation over land.
To further quantitatively evaluate how the simulations deviate from observed patterns in terms of the probability across different intensities of monthly mean precipitation, the empirical cumulative distribution functions (ECDFs) of the simulations and the gridded observations are compared, and the Kolmogorov-Smirnov (KS) distance between them is calculated (Figure 4). KS distance measures the maximum distance between two distributions, and thus it can be a good indicator for estimating the level of similarity between the simulated and observed distribution. Consistent with the  results obtained from the analysis of spatial distribution seen in Figure 3, lK measures the largest KS distance away from the observed distribution due to the severe underestimation. e cumulative probability reaches the 50th percentile at a daily precipitation less than 5 mm/day. When comprehensively compared with the three observation datasets, lE − o K, lE − o T, and lT −o E are well matched in the light to heavy precipitation.
In summary, the different combinations of CPSs employed in the latest version of RegCM4 show the different performance in simulating the statistics of seasonal and monthly mean precipitation. e three major findings are (1) the influence of CPS over land overwhelms the CPS over ocean. (2) Simulation with lK is characterized by copious dry biases suggesting that it is not appropriate to apply this scheme over land in South Korea using the RegCM4. (3) Although there is no CPS that shows the best performance in both quantitative and qualitative comparisons, lE is regarded as the most reasonable configuration for the simulation of summer precipitation over South Korea.

Evaluation of the Daily Precipitation.
On the gross pattern of monthly and seasonal mean precipitation against gridded and station observations, this section focuses on the performance of daily precipitation that can help in understanding the characteristics of the frequency, intensity, and extremes. For the simulation, we consider the case when a daily precipitation is higher than 0.1 mm. Figure 5 presents the time-latitude cross section of daily precipitation that is zonally averaged from 126°E to 129.5°E. Because summer precipitation in East Asia, including the Korean Peninsula, propagates towards the north along the monsoon front, capturing its evolution plays an important role in improving the accuracy of model simulation. Both the observation datasets (i.e., APHRODITE and TRMM) clearly reveal that heavy precipitation, exceeding 16 mm/ day, moves progressively towards the north from mid-June to early August. ere are some localized stationary precipitation bands after the peak is reached (early August), but there is no relevant propagation at that time. A similar behavior is observed in all simulations; however, different CPSs yield the sensitivity of magnitude of the intense precipitation band. Again, the simulations with lK suffer from a severe dry bias, despite capturing the shape of the evolutionary pattern and the peak in early August. Other simulations with lE and lT deliver reasonable overall performance when the timing and amplitude of the precipitation band are simulated.
Given that nontrivial discrepancy exists in the different gridded observation datasets (i.e., APH and TRMM), the analysis presented below is based on the comparison with insitu observation data obtained from 56 KMA stations. For this, gridded model simulations are converted into 56 stations, as presented in Figure 1. Figure 6 presents the temporal evolution of daily mean precipitation averaged over the 56 stations during the four months (JJAS) throughout the 10-year period. In general, RegCM4 delivers a reasonable performance in simulating interannual and intraseasonal precipitation variations. e simulated and observed daily  Table 3: Statistical metrics for comparison of the June-July-August-September (JJAS) mean (1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007) precipitation between model simulations and multiple observations. Blue and red colors indicate the best and worst performance, respectively. 1 ME refers to mean error, and the unit is mm/day. 2 RMSE refers to root-mean-square error, and the unit is mm/day. 3 Corr. refers to spatial correlation. 6 Advances in Meteorology fluctuations are well correlated (the temporal correlation coefficient exceeds 0.7). Additionally, the simulations effectively capture the variations in daily mean precipitation occurring each year, including the timing of the precipitation peaks in different years. For example, the model successfully captures the extreme cases observed in 1998 and 2002, except for the simulations with lK. In particular, lE − oE outperforms the others based on the temporal correlation and the precipitation-peak intensity.
To further investigate the precipitation characteristics of each simulation, the analyses of the frequency and intensity of daily precipitation are presented in Figure 7. For the individual experiments and the in-situ observation, a data sample is created by pooling the 10-year JJAS daily values from all the 56 stations (122 days × 10 years × 56 stations) together. Using this sample, Figure 7(a) presents the frequency distribution of precipitation with respect to the intensity of precipitation with 10 mm/day intervals, whereas Figures 7(b) and 7(c) display the boxplot of daily precipitation and extreme precipitation to emphasize the difference at the tail parts that are not easily distinguished in Figure 7(a). Firstly, the models exhibit the ability of capturing the gamma distribution of precipitation with similarity to the observed pattern. However, the performance variations increase as the intensity of precipitation increases. e lK simulations clearly underestimate the occurrence of precipitation with intensity exceeding 40 mm/day, whereas lE and lT simulations are considerably closer to the observation. Such discrepancy is also observed in the boxplot of extreme daily precipitation above the 95 th percentile. Almost the entire box from lK simulations does not overlap with the observed one, which means that the statistics (e.g., median and interquartile range) between the simulation and observation are fairly different owing to the severe dry bias observed in the lK simulations.
To compare far clearer shape of extreme behavior, Figure 8 presents the boxplots with the annual maximum consecutive 5-day accumulative precipitation and the annual number of days when precipitation exceeds 20 mm/day in 56 individual stations (10years × 56 stations) in case of each simulation. Consistently, the simulations with lK show the worst performance, underestimating both the mean and  extremes. In contrast, lE − oE and lT −oE deliver the best performances, particularly reproducing extreme consecutive precipitation events over 500 mm/5-day. A comparison between Figures 7 and 8 indicates that the sensitivity of intense precipitation to CPS tends to be higher than that of weak precipitation to CPS. In addition, the effect of CPS on the ocean is evident in the case of extreme precipitation (Figure 8), whereas it is negligible in the case of mean precipitation. For example, the performance of oE is better than those of oT or oK in simulating extreme precipitation when the same CPS is employed for the land. Irrespective of the CPS that is applied, all the simulations fail to capture the very extreme cases when the precipitation amount for 5 days exceeds 550 mm, which is possibly associated with a typhoon. As demonstrated in previous studies, most of the RCMs show a low accuracy when simulating heavy precipitation accompanied by a typhoon (e.g., Im et al. 2006 [5]). In general, the analysis of daily precipitation provides more in-depth insights into the performance of each CPS. It  Advances in Meteorology reveals that although the simulations can capture the temporal and spatial variabilities of the daily precipitation, they still unavoidably overestimate the frequency of light precipitation and underestimate the frequency and intensity of the mid-to-high precipitation. Overall, lE − oE and lT −o E offer an advantage over the other CPSs in terms of the extreme daily precipitation.

e Physical Mechanism behind the Precipitation Pattern.
To understand the physical processes underlying the different performance across the CPSs, two cases, l E − o E and l K − o K, which represent the best and worst performances, respectively, are selected, and in-depth analyses of the thermodynamic and vertical structures are performed. First, we assess the relative contribution of convective precipitation to the total precipitation, given that CPS primarily controls the formation of convective precipitation. Figure 9 presents the spatial distribution of convective precipitation and its ratio with respect to total precipitation derived from l E − o E and l K − o K simulations. In line with the amount of total precipitation (See Figures 3(a), 3(e)), the amount of convective precipitation is considerably higher in l E − o E simulation than in l K − o K simulation. It can be inferred from the spatial similarity in the maximum position between convective and total precipitations that convective precipitation derived from different CPSs can be a major factor in determining the performance of simulated precipitation in terms of magnitude and localized maxima. In contrast to the amount, interestingly, the ratio of convective precipitation to total precipitation exhibits an opposite pattern, yielding the higher values in l K − o K simulation than in l E − o E simulation. e l K − o K produces convective precipitation exceeding 70% in large areas in the Korean Peninsula. However, it was reported that stratiform large-scale precipitation rather than convective precipitation is the 10 Advances in Meteorology dominant precipitation type in Korea [47]. For l K − o K simulation, a small amount of convective precipitation constitutes large parts of the total precipitation, and it leads to substantial dry biases in total precipitation. In this regard, l E − o E shows a relatively better ability of capturing a realistic balance between the convective and large-scale precipitations. Note that although CPS plays a critical role in triggering and developing convective clouds at an unresolved subgrid scale, the modified vertical profiles of thermodynamic variables through the parameterization of convective precipitation can modulate the processes for resolved grid-scale precipitation. e partitioning of convective and large-scale precipitations is a result achieved from the active communication of both the processes within a model. erefore, appropriate partitioning of convective precipitation plays a key role in the successful simulation of precipitation characteristics [47,48].
Convective inhibition (CIN) may provide a clue for explaining the different amounts of convective precipitation derived from the two CPSs. Figure 10 presents the time-latitude cross-section of climatological (1998-2007: 10-year) daily CIN, zonally averaged from 126°E to 129.5°E for the month of June through September. Because CIN indicates the amount of energy inhibiting convection, a larger CIN indicates a more unfavorable condition for convection [49]. Consistent with the precipitation pattern appeared in Figure 5, the CIN calculated using the output from lK − oK simulation is much higher than that from lE − oE simulation. In particular, lK − oK simulation exhibits a very high CIN at the timing when lE − oE exhibits an intense precipitation band, explaining why lK − oK simulation suffers from a substantial dry bias.
To minimize the smoothing effect arising from averaging in the temporal and spatial dimensions, daily CIN values from 56 stations in the 10-year JJAS period are pooled together and are counted into corresponding bins in the frequency distribution ( Figure 11). e frequencies of CIN, which indicate atmospheric environments with weak and moderate inhibitions (CIN <100 J/kg), are mostly the same in lE − oE and lK − oK simulations. However, their differences are more evident at the range with very high CIN because lK − oK simulation has a longer and thicker tail. It means that CIN from lK − oK simulation is often too high to trigger the convection, which directly results in the suppression of convective precipitation.
We also check the vertical structures of moisture content and omega simulated from lE − oE and lK − oK to further explain the different circumstance potentially responsible for the formation of precipitation. Figure 12 presents the vertical cross section of specific humidity and omega along the latitudes at 127.6°E. A notable difference between the two simulations can be observed for low level moisture. e lK − oK simulation tends to produce drier conditions in the level across entire latitude of the Korean Peninsula. erefore, in spite of a stronger rising motion, the formation of clouds is fairly limited owing to a lack of moisture source.
e higher values of CIN seen in Figure 11 are typically generated by the flow of dry air [49]. e thermodynamic and vertical structures described above collectively differentiate the performances of precipitation when employing different CPSs such as lE − oE and lK − oK.

Discussion and Conclusion
e potential of the earlier version of RegCM to simulate the climate in Korea has been intensively examined [4,6,27,50].
is study provides an updated assessment of the latest version of RegCM4 by applying the newly implemented CPSs (i.e., the introduction of mixed convection schemes and the availability of Kain-Fritsch and Tiedtke schemes). We focus on the comparison of performances of nine sensitivity experiments, which have the different combinations of Emanuel, Kain-Fritsch, and Tiedtke schemes over land and ocean, in simulating the summer precipitation in South Korea. A comprehensive assessment, including the seasonal, monthly, and daily statistics against gridded and in situ observation datasets, is conducted to evaluate the model's performance with different CPS. e monthly and seasonal climatological mean precipitations clearly indicate that the determination of spatial pattern and amount of precipitation is highly dependent on the CPS that is applied over land. Irrespective of the CPS that is applied over the ocean, the Kain-Fritsch scheme over land tends to suffer the most from severe dry bias, although it provides a high spatial correlation with the observations. It is revealed that the thermodynamic and vertical structures in lK − oK simulation are unfavorable for triggering and developing the convection because the CIN is too high and moisture in low level is not sufficient, compared to lE − oE  that delivers the most reasonable performance in this study.
On the other hand, the Tiedtke scheme over land produces excessive orographic precipitation that ruins its reproduction of the spatial pattern. From this perspective, the Emanuel scheme over land is clearly superior in capturing the major characteristics of summer precipitation in South Korea. However, when applying the Emanuel scheme over both land and ocean, the precipitation tends to be slightly overestimated. Replacing it with the Tiedtke or Kain-Fritsch schemes over the ocean may alleviate this problem. For daily statistics, the Emanuel scheme over the ocean may benefit the simulation of extreme precipitation. Comparing with the Tiedtke or Kain-Fritsch schemes over the ocean, the results show good agreement with the observation in terms of heavy precipitation, especially when precipitation exceeds 500 mm/day. An analysis of the 10-year sensitivity tests with different CPSs indicates that for the simulation of summer  precipitation in South Korea using RegCM4, no CPS configuration exhibits an overall superiority across all the quantitative and qualitative aspects. e discrepancy among different observation datasets also makes it more difficult to derive a robust conclusion. Generally, the Emanuel scheme, which is the most frequently employed CPS in the Korea region, is still the best scheme over land and can capture the major characteristics of summer precipitation in this region. is study also demonstrates the merits of Tiedtke and Kain-Fritsch schemes, which are new additions to the recent versions of RegCM and have rarely been applied before, when employed when applied over the ocean in this region. erefore, model users should take these findings into careful consideration when determining the CPS that best aligns with the purpose of their simulations.
Data Availability e data that support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.