The temporal and spatial structures of recent and future trends in extreme indices over Korea from a regional climate projection

Authors

  • E. S. Im,

    1. Earth System Physics, Abdus Salam International Centre for Theoretical Physics, Trieste, Italy
    Search for more papers by this author
  • I. W. Jung,

    Corresponding author
    1. Department of Civil and Environmental Engineering, Sejong University, 98 Kunja–Dong, Kwangjin–Gu, Seoul 143–747, Korea
    • Department of Civil and Environmental Engineering, Sejong University, 98 Kunja–Dong, Kwangjin–Gu, Seoul 143–747, Korea.
    Search for more papers by this author
  • D. H. Bae

    1. Department of Civil and Environmental Engineering, Sejong University, 98 Kunja–Dong, Kwangjin–Gu, Seoul 143–747, Korea
    Search for more papers by this author

Abstract

The temporal and spatial characteristics of trends in extreme indices over Korea between 1971 and 2100 are investigated using daily minimum (Tmin) and maximum (Tmax) temperature and precipitation data from a regional climate projection at 20 km grid spacing. Five temperature–based indices and five precipitation–based indices are selected to comprehensively consider the frequency, intensity, and persistence of extreme events. In addition, Mann–Kendall tests are used to detect the statistical significance of trends in these indices. For validation during the reference period (1971–2000), the model reasonably simulates the temporal and spatial pattern of the trend. The model captures observed direction and magnitude well in various types of extremes. Indices based on Tmin show a considerable change towards warmer climate conditions while indices based on Tmax do not reveal any distinct trend, implying an asymmetric response of Tmin and Tmax to global warming. Indices of the frequency and intensity of heavy precipitation show a significant increase, whereas the duration of dry and wet consecutive days shows no change. For future projections, the temperature–based indices exhibit a much more significant and consistent trend than the precipitation–based indices, with statistical significance at the 95% confidence level for all indices. The frequency and intensity of heavy precipitation are projected to increase in the 21st century, continuing the trend of the reference climate. Although the future projected changes in the duration of consecutive dry and wet days are not statistically significant, the signal becomes more pronounced with respect to the reference simulation. Copyright © 2010 Royal Meteorological Society

1. Introduction

The climate of Korea has experienced a gradual warming throughout the 20th century. The change of the mean climate state could lead to a change of climate extremes due to a shift in the temperature distribution. Negative impacts of climate change on society and ecosystems are mostly expected to arise from extreme events, which highlights the need to identify climate extremes (Sillmann and Roeckner, 2008). In fact, there has been discernible evidence of changes in climate extremes over Korea (Choi, 2004; Kwon, 2005; Chang and Kwon, 2007; Jung et al., 2009). As anthropogenic climate change is projected to accelerate in the future (IPCC, 2007), Korea will be vulnerable to changes in various types of extreme climate events. This concern is supported by climate change studies over Korea (Boo et al., 2006; Im et al., 2007a; Im et al., 2008b). Boo et al. (2006) and Im et al. (2008b) reported an increase in the number of the days of heavy precipitation as well as in the mean amount from a regional climate simulation for Korea. Im et al. (2007a) also found a significant increase (decrease) of hot (frost) spells from a regional climate projection over Korea under the Special Report on Emission Scenario (SRES) B2 emission forcing.

Although changes of climate extremes should be considered as critical factors of climate change impact, intensive examination of their characteristics is relatively limited so far. One of the main reasons is related to a lack of reliable and homogeneous long-term daily time series of both observations and simulation data. In particular, climate downscaling results with a focus on the Korean peninsula are rare. As climate extremes may have a complex spatial and temporal variation (Boo et al., 2006) and Korea is characterised by complicated mountainous terrain, high resolution information appropriate for analysis at the regional or local scale is needed to better understand climate extremes and their application to impact assessments over Korea (Bae et al., 2008b). To this end, we developed the RegCM3 one-way double-nested regional climate modelling system (Im et al., 2006). Im et al. (2006, 2007b) then demonstrated that the RegCM3 modelling system is a useful tool to provide reasonable fine-scale climate information over the Korean peninsula. Using the same modelling system, we carried out a downscaling of a global ECHAM5/MPI-OM A1B scenario simulation covering the period 1971–2100.

In this study, we focus on the temporal and spatial structure of trends of climate extremes of daily maximum and minimum temperature and precipitation from the nested domain covering the Korean peninsula at 20 km grid spacing. A continuous 130-year simulation allows us to derive meaningful statistics in response to either external forcing (e.g. greenhouse gas) or interdecadal modes of internal variability. Detailed insight into long-term trends has revealed changes in the mean climatic state as well as the pattern of evolution. We consider ten different types of indices based on three general categories of frequency, intensity, and persistence for defining a climatology of extremes (see Section 2.2). The behaviour of extremes can be quite nonlinear. For example, the rate of increase in the number of warm days and the corresponding magnitude of warming could have different trends even though they tend to evolve towards the same direction. Therefore, it is important to examine changes in frequency and magnitude of climate extremes as well as changes in the mean state. To investigate these changes, we first estimate the performance of our modelling system in capturing the spatial variation and temporal evolution of past trends in various types of climate extremes against a dense observational network during a reference period (1971–2000). We then investigate the future change of the temporal and spatial structure of the long-term trends in climate extremes under global warming (2001–2100). Note that this study is the first report of the global warming effect over Korea using a dynamically downscaled climatic change simulation based on the fourth Assessment Report (AR4) of the Intergovernmental Panel on Climate Change (IPCC). Furthermore, even though there are several references (Boo et al., 2006; Im et al., 2007a, 2008b) dealing with the extreme issues, most of these previous studies were not comprehensive, either considering intensity or frequency only at a certain future period (e.g. 2051–2080 or 2071–2100) with respect to present climate (e.g. 1971–2000). This study will provide a baseline for detecting the regional climate change over Korea caused by global warming.

2. Model and Method

2.1. Regional climate model

The regional climate model used in this study is the latest version of the International Centre for Theoretical Physics (ICTP) regional climate model, RegCM3. It is an upgraded version of the model originally developed by Giorgi et al. (1993a, 1993b) and improved as discussed by Giorgi and Mearns (1999) and Pal et al. (2007). The dynamic core of the RegCM3 is equivalent to the hydrostatic version of the National Center for Atmospheric Research (NCAR)/Pennsylvania State University mesoscale model MM5 (Grell et al., 1994). The physical parameterisations employed in this simulation include the comprehensive radiative transfer package of the NCAR Community Climate Model, version CCM3 (Kiehl et al., 1996), the non-local boundary layer scheme of Holtslag et al. (1990), the biosphere–atmosphere transfer scheme (BATS) land surface scheme (Dickinson et al., 1993), and MIT-Emanuel convection scheme (Emanuel, 1991).

The one-way double-nested technique and model domain configuration applied in the present study are the same as in Im et al. (2006, 2007b). The mother domain covers East Asia at 60 km grid spacing, while the nested domain focusses on the South Korean peninsula at 20 km grid spacing (Figure 1). Comparison of the two domains shows that the mountain ranges in the nested domain are much more realistic than in the mother domain, which emphasises the necessity of the double-nesting system. In this one-way nesting, relevant meteorological fields (wind, temperature, water vapour, and surface pressure) are interpolated from a coarse resolution mother domain onto the lateral buffer area of a high resolution nested domain in order to provide lateral boundary conditions for the nested domain simulations. Validation of the experiments using observation National Centers for Environmental Prediction (NCEP)/NCAR Reanalysis II, (Im et al., 2006) and GCM simulation (ECHO-G, Im et al., 2007b) as the initial and boundary condition demonstrated that the RegCM3 nesting system simulates both climatological and regional characteristics well. Moreover, the MIT-Emanuel convection scheme recently implemented in RegCM3 considerably improved the systematic cold and dry biases (Im et al., 2008a, 2008b).

Figure 1.

Model domain and topography (m) for the mother (a: 60 km resolution) and nested (b: 20 km resolution) simulations. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

In our experiment, the initial and time-dependent meteorological lateral boundary conditions for the mother domain simulation are interpolated at 6-hourly intervals from an ECHAM5/MPI-OM A1B scenario simulation. The integration spans the 130-year period 1971–2100, of which the first 30 years covering present-day conditions are used as a reference (1971–2000) and the remaining 100 years are used to simulate future climate conditions (2001–2100). The ECHAM5/MPI-OM is a coupled atmosphere–ocean global climate model that was developed by the Max-Planck-Institute for Meteorology (Jungclaus et al., 2006). The atmospheric component, ECHAM5, has a horizontal resolution of T63, corresponding to a grid resolution of 1.875° × 1.875°, and 31 vertical hybrid levels. The oceanic component (MPI-OM) of the coupled model is a primitive equation z-coordinate model with an integrated sea ice model. The ECHAM5/MPI-OM has simulated three members with different initial conditions under SRES A1B scenario. The third member is used here and identical to that used for the IPCC AR4 climate scenario integration (IPCC, 2007). The SRES A1B emission forcing is part of the A1 family that gives a total radiative forcing of approximately 6 W/m2 in year 2100 compared to pre-industrial times (IPCC, 2000). The A1B emission forcing lies in the AR4 as a middle-ground scenario and is the most commonly used scenario in AR4.

2.2. Extreme indices

Table I presents various types of extreme indices, including five temperature-based indices and five precipitation-based indices. To comprehensively investigate the characteristics of climate extremes, we select indices that consider frequency, intensity, and duration properties. All indices are calculated as annual values.

Table I. Abbreviation and definition of the indices of extremes based on maximum and minimum temperature and precipitation used in this study
VariableAbbreviation (unit)Definition of extreme indices
TemperatureFD (days)Number of FDs with Tmin below 0°
 HD (days)Number of HDs with Tmax above 30°
 TX95 ( °C)Averaged Tmax above 95th percentile
 TN5 ( °C)Averaged Tmin below 5th percentile
 HW (days)Maximum duration of consecutive hot days
PrecipitationPN80 (days)Number of days with precipitation above 80 mm intensity
 PPL95 (%)Percentage of total rainfall from events above long-term 95th percentile
 PX1D (mm)Greatest 1-day total precipitation
 MDRY (days)Maximum duration of consecutive dry days
 MWET (days)Maximum duration of consecutive wet days

In the case of temperature-based indices, the number of frost days (FD) is defined as the total number of days per year with minimum temperature (Tmin) below 0 °C, while the number of hot days (HD) is defined as the total number of days per year with maximum temperature (Tmax) above 30 °C. The number of days exceeding (or undershooting) an absolute threshold is the simplest way to reveal the change of frequency behaviour and interpret the shift of the distribution in Tmin and Tmax. To assess the change of magnitude of warm and cold days, we calculate the intensity that exceeds a percentile threshold. TX95 indicates the averaged Tmax above the 95th percentile while TN5 indicates the averaged Tmin below the 5th percentile. Changes in TX95 and TN5 show the modification of the upper and lower tail thickness in the Tmax and Tmin distributions, respectively. Finally, we consider the maximum duration of consecutive hot days (HW), which measures the severity of heat stress during the summer season.

As a precipitation index to measure the frequency of heavy precipitation, PN80 is defined as the number of days in which the daily precipitation is greater than 80 mm. The 80 mm/day value is the threshold used to issue severe weather alerts by the Korean Meteorological Administration (KMA) and has been used in previous studies (Chang and Kwon, 2007). We used the percentile of total rainfall exceeding the climatological 95th percentile precipitation (PPL95) and the greatest 1 day total precipitation (PX1D) in order to analyse the change of extreme precipitation intensity. For rainfall persistence characteristics, we consider the maximum duration of consecutive dry days (MDRY) and wet days (MWET). In this study, a wet day is defined as a day with precipitation accumulation greater than or equal to 1.0 mm, whereas a dry day represents a day with precipitation less than 1.0 mm. MDRY (MWET) is an indication of drought (flood) occurrence likelihood.

The selected indices in this study have a relatively short return period in order to obtain robust statistics. In other words, we deal with extreme events of which occurrence are typically large enough to allow meaningful trend analysis. These indices potentially include many aspects of changing climate conditions, and therefore the information derived from these indices should be useful for climate change impact studies.

2.3. Trend analysis method

We used the Mann–Kendall test (Mann, 1945; Kendall, 1975) to evaluate trends in the temperature and precipitation indices. The Mann–Kendall test is one of the widely used non-parametric tests to detect significant trends in time series data. The method is simple and robust, and it also has advantages of being able to deal with missing values and values below a detection limit. The null hypothesis H0 that Zc is not statistically significant or has no significant trend is accepted if − Z1−α/2ZcZ1−α/2, where ± Z1−α/2 are the standard normal deviates and α is the significance level for the test. Alternatively, it is accepted that H1 or Zc is statistically significant if Zc < − Z1−α/2 or if Zc > Z1−α/2 (Xu et al., 2005; Bae et al., 2008a). Kendall's statistic S is computed as follow:

equation image(1)

where, xk, xi are sequential data values; n is the length of the dataset; sgn(θ) = 1 if θ> 0, sgn(θ) = 0 if θ = 0, sgn(θ) = − 1 if θ< 0; m is the number of tied groups; and ei is the size of the ith tied group. S is expected to have normal distribution with mean 0 and variance S with the null hypothesis H0 that there is no trend displayed by the time series. The Mann–Kendall test statistic Zc is estimated as follows:

equation image(2)

where, Zc is a standard normal variable.

2.4. Verification strategy

For the validation of the reference simulation covering the period 1971–2000, we used the climate observations from 57 stations with long-term time series and few missing data. The quality control of the data was maintained by the KMA for the period from 1975 to 2004 throughout the southern part of Korea. These selected weather stations are relatively evenly distributed and located on diverse elevations (Jung et al., 2009). Spatial and interannual variability of precipitation extremes are ordinarily larger than for temperature ones. Chang and Kwon (2007) investigated the homogeneity of these data using Bayesian procedure in software AnClim (Stepanek, 2007). They demonstrated that using these quality controlled data has an advantage on study of precipitation extremes. In our study, daily Tmax, Tmin, and precipitation are compared to corresponding values from the inner nested domain at 20 km grid spacing. We compare simulated results with individual station values using the grid points closest to the stations. Comparing the simulated grid-point data of GCM with observed station data is difficult because the simulated data is spatially averaged and the spatial climate nature in subgrid scale is unclear (Skelly and Henderson-Sellers, 1996). However, recent studies (Fowler et al., 2005, Halenka et al., 2006, Im et al., 2008b) show that the relatively high model resolution justifies the comparison between the station data and the model data at the grid point closest to the station location. This dataset allows a first-order validation of the fine-scale structure of the nested domain simulation.

3. Results

3.1. Basic performance of daily Tmax, Tmin, and precipitation

Model inaccuracies in the representation of the daily characteristics of the daily frequency of meteorological events may cause systematic errors in the simulation of climate extremes. Therefore, it is important to assess whether the modelling system is capable of reproducing observed features of daily characteristics. This validation of the reference simulation adds to the reliability of projected changes in extreme indices. We thus begin our analysis with a validation of the frequency distribution of daily variables from the reference simulation in comparison with observations.

Figure 2 shows the probability density function (PDF) of Tmax and Tmin at all stations over Korea for both annual and seasonal (DJF and JJA) periods. We pooled together daily Tmax and Tmin dataset at the individual 57 grid points, not averaged over the 57 grid points (the same method for daily precipitation of Figure 3). It gives a measure of the mean and variance of the daily values. Overall, the simulated Tmax and Tmin distributions reasonably reproduce seasonal characteristics not only for the relative probability but also the variation range. The simulated means (values in parentheses in Figure 2) are also very close to observed one, differences mostly ranging within 1 °C (expect for DJF Tmin). Regarding the Tmax distribution, the shape of the annual pattern exhibits a bimodal structure reflecting the contribution of different PDF characteristics for the cold season (lower and broader) and the warm season (higher and narrower). Although the model follows the double-peaked structure shown by the observed pattern, the seasonal dispersion is slightly broader for winter (standard deviation: model = 5.6, OBS = 4.9) and narrower for summer (standard deviation: model = 3.5, OBS = 3.8) than observed. Moving to the Tmin distribution, we find that the summer pattern agrees better with observations than the winter one. The bias on the left side of the annual distribution may be partially attributed to an overestimation of winter Tmin. The upper tail of the winter distribution tends to be overestimated, with a shift towards warmer values. There is a noticeable difference in the shape of Tmax and Tmin distributions, with Tmax displaying a more asymmetric distribution. In spite of the model deficiency in variance and mean tendency, the model is able to capture the basic structure of each distribution realistically.

Figure 2.

PDF of the distribution for daily Tmax (upper panels) and Tmin (lower panels) in Korea. Here, left, middle, and right panel is annual, winter, and summer, respectively. Here, the numbers in parentheses are the mean values obtained from the simulated (MODEL) and observed (OBS) distribution

Figure 3.

PDF of the distribution for daily precipitation in Korea. (a), (b), and (c) is annual, winter, and summer, respectively. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 3 shows the same quantities for daily precipitation. As the climate of Korea is strongly governed by the monsoon system, the precipitation amounts show a pronounced seasonal variation. This characteristic is reflected in the seasonal frequency distribution. In winter, most precipitation intensities are less than 150 mm/day, whereas the summer distribution has a much longer tail at the high intensity range exceeding 500 mm/day. The simulated distribution during the summer season matches the observations remarkably well, except for a slight underestimation in the mid-range of 20–100 mm/day. Conversely, the model shows the tendency to overestimate winter precipitation above the 50 mm/day intensity level. Considering the fact that precipitation-based extreme events are mostly expected to occur in the summer season, which receives approximately two-thirds of the annual precipitation (Chang and Kwon, 2007), the ability to successfully capture the extremes at high intensity during the summer season is beneficial for this study.

In summary, Figures 2 and 3 show that the model performs reasonably well in reproducing daily Tmax, Tmin, and precipitation, indicating good fidelity of capturing the prominent distribution pattern and its seasonal dependency. In the next section, we turn our attention to the analysis of trends in extreme events derived from daily Tmax, Tmin, and precipitation.

3.2. Temporal structure of trends in extreme indices

Figure 4 presents the modelled time series of the five temperature-based indices averaged over 57 stations in Korea throughout the entire integration period. During the reference period (1971–2000), observed estimates are displayed as well. We first found that the trends of HD, TX95, and HW derived from Tmax are different from those of FD and TN5 derived from Tmin during the same period (1971–2000). The extreme indices based on Tmax remain constant or slightly increase, whereas the extreme indices based on Tmin show a significant trend (decreasing for FD and increasing for TN5) with a faster rate in both observations and simulations. This implies that the behaviour of Tmax and Tmin in response to global warming is nonlinear with a distinct seasonal pattern, as the extreme indices based on Tmax occur in the warm season, while the extreme indices based on Tmin occur in the cold season.

Figure 4.

Time series of temperature-based indices averaged over Korea for reference and future period. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

The model is able to realistically capture this asymmetric characteristic shown by observed climatological patterns despite systematic biases. These errors are reflected in the PDF of Tmin during the winter and Tmax during the summer as indicated by Figure 2. The upper tail underestimation of the Tmax summer distribution leads to an underestimation of the HD, TX95, and HW indices while the underestimation between 0° and − 10° of the Tmin winter distribution is the main cause of the FD underestimation and the TN5 overestimation. However, the qualitative aspects, in terms of the direction and slope of the trend are in good agreement with the observed estimate, enhancing the reliability in the trends projected for the future climate conditions.

In view of the statistical analysis (Table II), the Mann–Kendall test shows that all indices except for TN5 are not statistically significant at the 90% confidence level during the reference period. Even though these trends are not statistically significant, it does not necessarily mean that there is no trend of environmental significance (Bae et al., 2008a). For example, the relevant decline of FD is visible in both simulated and observed trends. But, their trend was not significance at the 90% confidence level because the natural variability exceeds that of the linear trend.

Table II. Summary of basic statistics derived from Figures 3 and 4
AbbreviationMeanStandardised coefficient of slopeMann–Kendall statistic
 ObservedReferenceFutureObservedReferenceFutureObservedReferenceFuture
  • a

    Significant at the 0.05 significance level.

  • b

    Significant at the 0.1 significance level.

FD98.783.265.3− 0.251− 0.305− 0.745− 1.231− 1.409− 7.883a
HD35.221.647.20.0710.1070.8830.5530.57110.378a
TX9532.631.233.20.033− 0.0010.8210.0710.1439.500a
TN5− 9.7− 8.8− 6.60.5320.4230.6732.622a2.765a7.010a
HW11.96.214.00.0650.0720.650− 0.2320.5357.853a
PN802.32.52.70.5190.3050.1942.837a1.803b1.954a
PPL9527.932.734.40.5120.3370.3362.784a2.266a3.344a
PX1D135.9156.9172.30.4500.4050.3742.373a2.301a3.615a
MDRY22.517.820.00.103− 0.1410.1510.624− 0.8921.415
MWET7.410.710.70.309− 0.041− 0.1591.4640.268− 1.176

In the future projection for the 21st century, the degree of warming is sharply accelerated. The standardised slope coefficients are higher than those of the reference simulation, in particular, for the Tmax-based indices. The trends of projected temperature-based indices are all significant at the 95% confidence level. Predominant increases of HD, TX95, and HW are projected and are tied to the change of the TMAX summer distribution in the future projection (not shown). At the same time, FD and TN5 noticeably decreases and increases, respectively. When viewed in the context of the mean value over the reference and future periods (Table II), TX95 and TN5 increase by approximately 2 °C. Consequently, HD increases by 26 days while FD decreases by 18 days. HW is also prolonged by more than a factor of 2 in the future with respect to the present-day climate. Note that future means from Table II are defined as the average over the whole future period of 2001–2100. The mean over the 30-year period at the end of the 21st century (2071–2100) is expected to have much higher values than the mean for the whole period. In the future, the climate of Korea is projected to undergo significant warming, which could be attributed to the combined effect of the frequency, intensity, and duration changes in extreme events.

Figure 5 shows the temporal evolution of the precipitation-based extreme indices. The structure of the temporal evolution depends on the index type. During the reference period, the indices related to the intensity and frequency of heavy precipitation (PN80, PPL95, and PX1D) exhibit statistical significance in both the simulated and observed trends, whereas the indices based on duration (MDRY and MWET) do not reveal any readily apparent trend. The PN80, PPL95, and PX1D show a consistently increasing trend in spite of the large interannual variability. In the case of PN80, the model captures fairly well the variation range as well as the trend. The model also shows a good qualitative agreement in the trend of PPL95 and PX1D with those observed but tends to overestimate these indices. The characteristics of heavy precipitation are tightly coupled with the distribution of summer precipitation because most of the heavy rain events over South Korea are associated with the development of the summer monsoon (Im et al., 2008a). As shown in Figure 3, the model distribution has a tendency to overestimate the very high intensity range, which could be attributed to the overestimation in the absolute value of PPL95 and PX1D. In the case of MDRY and MWET, it is very difficult to judge the positive or negative trend. According to the standardised coefficient, the trends in model and observations seem to have opposite signs. However, the magnitude of the slope is not statistically significant according to the Mann–Kendall test. The model tends to overestimate MWET and underestimate MDRY compared to observations. This is due to the model deficiency of producing excessive occurrences of very weak precipitation intensity, which is a commonly noted problem in precipitation produced by climate models.

Figure 5.

Time series of precipitation-based indices averaged over Korea for reference and future period. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

The future behaviour of precipitation-based indices is less distinct compared to those of temperature-based indices, which somewhat monotonically respond to the emission forcing (e.g. CO2 concentration). In general, the future tendency of the five indices continuously follows the characteristics of the reference climate. Indices related to heavy precipitation show gradually increasing trends with substantial variability. These trends are significant at the 95% confidence level. On the other hand, a well-defined regular trend is not certain in the MDRY and MWET time series. However, the enhancement of variability is discernible, including anomalous peak episodes, which could be an indication of increased likelihood of severe drought and flood events under global warming.

In Figures 4 and 5, the trends between the future and reference period seems to be rather discontinuous. This is not related to data discontinuity, but to the chosen period of trends. If different periods are selected, different slope of trends (in spite of same direction) could be drawn. Trends are strongly dependent on the chosen period, and long-term datasets include both interannual and interdecadal oscillations, which will affect the trend terms of the long-term period's sub-period (Bae et al., 2008a; Jung et al.2009).

3.3. Spatial structure of trends in extreme indices

In this section, we move our attention to the spatial structure of trends in the extreme indices. Figure 6 presents the spatial distribution of the trend in temperature-based indices for observations, and both reference and future simulations. First, we tried to examine the consistency between the temporal and spatial trend pattern for the past. The stations showing the statistical significance at the 90% confidence level are restricted to a small number of locations for the Tmax-based indices. In addition, the upward and downward patterns are more mixed. It is difficult to determine consistent tendencies in the temporal evolution. On the other hand, Tmin-based indices contain many stations with high statistical confidence covering a relatively large area. This pattern appears to agree with the temporal aspect explained in Figure 4.

Figure 6.

Spatial distribution of trend in temperature-based indices over Korea. Here, left, middle, and right panels indicate observation, reference simulation, and future projection, respectively. Shaded area indicates the downward trend

An unexpected aspect is seen in FD distribution. Observations show a positive trend over some of the central and southern regions. It seems to be a rather conflicting feature considering a report of increasing winter temperatures over Korea (Choi, 2004). It might partially be due to the absolute threshold (0 °C) which is used for determining the FD. Year-to-year variability in frost-day counts is related to the variability of temperature during the cold season, and it exhibits regional dependence (Klein Tank and Konnen, 2003). This implies that the threshold to represent the climatological cold day should vary from region to region. This is the reason, when using the values of the 5 percentile for the FD threshold, the trend in the number of days below the threshold for Tmin decreases in all regions (not shown).

Regarding the change of the Tmin magnitude, there are mostly significant positive trends in both simulations and observations throughout the entire region. The model successfully captures the general pattern of the trend, but it shows less significance at several sites compared to the observed estimates, which contribute to the lower gradient of temporal evolution of the TN5 averaged over Korea as shown in Figure 4.

While recent 30-year trends derived from observations and the reference simulation show mixed features in terms of direction and significance, the projected trends in the future indeed show strong and consistent pattern. All stations' trends are statistically significant at the 90% confidence level and indicate the same sign in correspondence of the temporal structure. This suggests a robustness in temperature-based indices for warmer climates in the future projections, at least for the A1B scenario and the GCM–RCM combination considered in this study.

Figure 7 shows the same quantities as Figure 6 for precipitation-based indices. Regarding the indices measuring the frequency and intensity of heavy precipitation (PN80, PPL95, PX1D), both the simulated and observed trends show an overall increasing pattern except for randomly scattered small areas. The model follows fairly well the direction of observed trends even though there are differences of statistically significant areas between the model and observations. A different situation is found for the indices based on duration (MDRY and MWET). MDRY shows spatial complexity of opposing signals and no statistical significance, thus it is difficult to draw any meaningful interpretation. In case of MWET, the observed trend shows an increase while the model tends to decrease across large areas, but these signals are mostly not statistically significant at 90% confidence level.

Figure 7.

Spatial distribution of trend in precipitation-based indices over Korea. Here, left, middle, and right panels indicate observation, reference simulation, and future projection, respectively. Shaded area indicates the downward trend

The precipitation-based indices also exhibit a more significant and consistent trend in the future projection, in line with those found in the temperature-based indices. A pronounced enhancement of PN80, PPL95, and PX1D is found in the southern part of Korea. As the increase of PPL95 demonstrates the increase of the contribution of extreme precipitation exceeding 95th percentile to the total annual precipitation, it is correlated with decreasing moderate rainfall. Therefore, less frequent but heavier precipitation events could lead to both an increased vulnerability to flood as well as drought episodes. Based on the normalised distribution of daily precipitation change, precipitation with intensity above the 50th percentile tends to increase its contribution to total precipitation, while precipitation of lower intensity yields a reduced contribution in the future climate (not shown). This hypothesis is well reflected in MDRY and MWET, with MDRY (MWET) projected to increase (decrease), even though the statistically significant region is rather restricted.

4. Uncertainty and Limitation of this Study

Climate change impact assessment on extreme indices should be attributed to various uncertainties such as different greenhouse gas emission scenarios, and Global Climate Model (GCM) and Regional Climate Model (RCM) configurations. The uncertainties based on GCM and RCM primarily stem from climate process description and approximations, parameterisation of subgrid-scale phenomena, and initial condition (IPCC, 2007). Previous studies (Frei et al., 2006; Goubanova and Li, 2007; Kyselý and Beranová, 2009) demonstrated that the results of climate change on precipitation extremes were strongly model dependent, particularly regarding parameterisation of convective processes. Kay et al. (2009) showed that GCM initial conditions also affected flood frequency regarding precipitation extremes. The present study is based on a single realisation—an A1B emission scenario, a RegCM3, and an ECHAM5/MPI-OM boundary forcing. This means that using a different realisation may result in a different conclusion.

In order to investigate uncertainty occurring due to the internal variability of GCM simulation used as the initial and boundary conditions for the RegCM3 downscaling system, monthly GCM precipitation and temperature are considered over East Asia. Figure 8 displays area-averaged (100E–150E, 20N–50N) time series in precipitation and temperature according to three members of ECHAM5/MPI-OM with A1B scenarios for 1971–2100, which have different initial conditions. The trends of seasonal precipitation and temperature and annual ones display identical directions, though their inter- and multi-decadal variations are different among each member. As a result of the Mann–Kendall test, most of the seasonal and the annual simulations of the three members except winter precipitation show significant increasing trends at the 99% confidence interval. Winter precipitation show slight increases for all members but they are not significant. Table III shows correlations of monthly precipitation and temperature among each member. The differences in temperature monthly variation among each member are negligible. The correlations of these members for precipitation are less than for temperature, but they are still high enough to conclude that all ensemble members with different initial conditions are highly correlated. Ideally, it would be desirable to carry out a downscaling of all three members of ECAHM5/MPI-OM simulations and analyse extreme indices from the ensemble mean. However, it would require tremendous computational cost without data sharing throughout systematic collaboration (e.g. ENSEMBLE project). Based on the comparison between the three members (Figure 8), there is not much difference between them. Therefore, a downscaling result of one member may not introduce a significant discrepancy from the ensemble mean. Figure 9 shows changes in seasonal mean temperature and precipitation in both future periods (2011–2040 and 2071–2100) with respect to the reference period (1971–2000) from 19 IPCC AR4 GCM projections with A1B emission forcing over East Asia. It gives an idea of the reliability of the ECHMA5/MPI-OM projection via assessing inter-model variability. Changes in temperature for both seasons revealed coherent increase. All models project warmer climate and increasing rate enhanced in the late 21st century in response to Greenhouse gas (GHG) concentration with no exception. Conversely, changes in precipitation show a mixed pattern and are much more model dependent. In spite of large spread (especially winter precipitation for 2020s), ECHAM5/MPI-OM provides projection information over East Asia (including the Korean peninsula) that is not markedly different from other simulations. All three members are included in the boundary area which is formed by the scatters of other simulations, regardless of seasons (DJF and JJA) and variables (temperature and precipitation). Changes in winter precipitation of ECHAM5/MPI-OM member 3 (used in this study, symbol 3) are located at rather marginal area, but winter precipitation mostly not affects the precipitation-based extreme index used in this study. Hence, these results can support the use of ECHAM5/MPI-OM simulation for downscaling.

Figure 8.

Area-averaged time series in summer, winter, and annual mean precipitation and temperature for 1971–2100 over East Asia. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 9.

Change in seasonal temperature and precipitation of 19 GCM simulations used in 2007 IPCC AR4, for 2020s (2011–2040) and 2080s (2071–2100) relative to 1971–2000. Symbol 1, 2, and 3 indicates three members of ECHAM5/MPI-OM with A1B scenario. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Table III. Pearson's correlation coefficient among 3 members of ECHAM5/MPI-OM
SimulationsReference period (1901–2000)Future period (2001–2100)
 Member 1Member 2Member 3Member 1Member 2Member 3
PrecipitationMember 11.000.850.851.000.830.82
 Member 20.851.000.850.831.000.83
 Member 30.850.851.000.820.831.00
TemperatureMember 11.000.990.991.000.990.99
 Member 20.991.000.990.991.000.99
 Member 30.990.991.000.990.991.00

The realism of any SRES scenario, GCM, and RCM simulation can be questioned due to various uncertainty sources. Therefore, interpretation of the projected future climate conditions based on modelling should be done with caution. To reduce the uncertainty range, there is a need for further ensemble experiments that will provide more constrained climate change information.

5. Conclusions

In this study, we presented an analysis of the trends in various extreme indices based on temperature and precipitation. For a fine-scale climate change scenario focussing on the Korean peninsula, we performed the dynamical downscaling of the ECHAM5/MPI-OM A1B simulation using a double-nested regional climate modelling system (1971–2100: 130-year). We focussed on the temporal and spatial structure of trends in five temperature-based extreme indices and five precipitation-based extreme indices derived from daily Tmax, Tmin, and precipitation from a nested domain simulation at 20 km grid spacing. The detection of trends in extreme indices and its statistical significance are estimated using the Mann–Kendall non-parametric test.

Based on the validation of a reference simulation against a dense network of station observations, the model reproduces the PDF of daily Tmin, Tmax, and precipitation distribution reasonably well. Therefore, extreme indices derived from daily Tmin, Tmax, and precipitation show overall good agreement with those from observations. Possible effects of global warming are mainly described by an increase in both the frequency and intensity of temperature-based extremes in warmer climate conditions, though an asymmetric response of Tmin and Tmax was found. Indices measuring the frequency and intensity of heavy precipitation exhibit significant increasing trends while MDRY and MWET based on duration show no distinct tendency. The model successfully captures the key characteristics in terms of direction and magnitude of trend. However, several quantitative differences are revealed. Simulated HD, TX95, and HW tend to be too low due to an underestimation of the upper tail in the summer Tmax distribution. On the other hand, underestimation of FD and overestimation of TN5 is due to a lower frequency of simulated temperatures below 0 °C in winter. PN80, PPL95, and PX1D also tend to be overestimated due to the longer tail of the summer distribution of precipitation produced by the model.

Regarding future projections, all indices based on Tmin and Tmax show a stronger trend with respect to the reference simulation, being statistically significant at the 95% confidence interval. The temporal evaluation of the regional averages over all station locations satisfies the statistical significance tests at the 95% confidence interval. The spatial coherence of trend patterns adds confidence to the projected climate change. It is evident from our analysis that extremely hot events occur more frequently with intensified magnitude, while extremely cold events occurring in the present climate would cease under global warming. Heavy precipitation is also projected to increase in frequency as well as magnitude. Even though statistical significance is restricted to several station locations, MDRY (MWET) tends to increase (decrease), implying less frequent and heavier precipitation events in the future projection.

In spite of some deficiencies of the reference simulation described above, the quality of the ECHAM5/MPI-OM downscaling results presented here demonstrates considerable improvement compared to previous GCM–RCM simulations over our study region (Hirakuchi and Giorgi, 1995; Kato et al., 2001). These results are even superior to the statistics reported by Im et al. (2007b), despite using the same regional climate modelling system. The main differences could originate from using a different GCM and a newly added convection scheme in the RegCM3. Therefore, the fine-scale long-term scenario simulation in this study provides useful information for comprehensive climate change impact assessment studies over the Korean peninsula. In particular, the quality of precipitation-based indices is encouraging, considering that summer precipitation over Korea is strongly determined by relatively less predictable convective processes, such as the occurrence of mesoscale convective systems and tropical storms (Im et al., 2006).

As all interpretations in this study are based on a single realisation, it is impossible to quantify various sources of uncertainty. However, the RegCM3 double-nesting system adopted in this study for dynamic downscaling has already been extensively evaluated. It shows reasonable performance in simulating both climatological and regional characteristics over Korea, regardless of lateral boundary forcing, such as in experiments using perfect boundary conditions (e.g. NCEP/NCAR reanalysis: Im et al., 2006, 2008a) and GCM boundary conditions (e.g. ECHO-G global climate model: Im et al., 2007b). Moreover, the majority of the future climate change aspects presented here are consistent with those found in downscaled ECHO-G projections reported by Im et al. (2007a, 2008b), except for several aspects accounting for the use of a different emission scenario. As downscaling scenarios with a focus on the Korean peninsular are lacking, this work could help to extend understanding of future climate change due to global warming over Korea.

Acknowledgements

The authors wish to thank the two anonymous reviewers whose valuable comments and suggestions greatly improved the quality of this paper. We extend our thanks to Dr Heejun Chang and Matthew Wood of Portland State University for carefully proofreading our manuscript. The downscaled dataset was produced by the first author with technical and computing support of Earth System Physics Section, the Abdus Salam ICTP. The second and third authors were supported by a grant (code# 1-9-3) from Sustainable Water Resources Research Center of the 21st Century Frontier Research Program in Korea.

Ancillary