Abstract

The spatial and temporal resolution of readily available climate change projections from general circulation models (GCM) has limited applicability. Consequently, several downscaling methods have been developed. These methods predominantly focus on a single meteorological series at specific sites. Spatial and temporal correlation of the precipitation and temperature fields is important for hydrologic applications. This research uses a nearest neighbor–genetic algorithm (NN–GA) method to analyze the Namhan River basin in the Korean Peninsula. Using the simulation results of the CNRM-CM for the RCP 8.5 climate change scenario, archived in the fifth phase of the Coupled Model Intercomparison Project (CMIP5), the GCM projections are downscaled through the NN–GA. The NN–GA simulations reproduce the features of the observed series in terms of site statistics as well as across variables and sites.

1. Introduction

Climate change may affect the spatial and temporal distribution of water resources, as well as the intensities and frequencies of extreme hydrological events (Huntington 2006). According to the IPCC (2013), which periodically publishes assessment reports on world climate change research, the global sea temperature is expected to rise by approximately 0.3°–4.8°C, with the sea level rising by about 26–82 cm by the late twenty-first century. As the temporal and spatial imbalance of global water resources is expected to worsen in the future, efforts are being made to quantitatively assess the effect of climate change on basin or regional scales.

Since general circulation model (GCM) (representing physical processes in the atmosphere, ocean, cryosphere, and land surface; www.ipcc-data.org/guidelines/pages/gcm_guide.html) outputs have low temporal and spatial resolution, downscaling is used at a regional scale. There are dynamic downscaling methods that use numerical, model-based regional climate models, which provide weather and climate variables and flux through meteorological equations. Furthermore, there are statistical downscaling methods that calculate the statistical dependence between GCM outputs and regional observations (Wilby et al. 2004; Goodess et al. 2005; Benestad et al. 2009). Each downscaling method has its own advantages and disadvantages (Hay and Clark 2003; Schmidli et al. 2007; Spak et al. 2007; Gutmann et al. 2012). Statistical downscaling is more widely adopted in hydrological studies, because it is less computationally demanding (Chen et al. 2010), and the final product can be used directly with hydrologic models without needing further bias correction. Statistical downscaling methods can be largely classified into three categories: weather typing, regression, and weather generators (Giorgi et al. 2001). Weather typing is a method of downscaling using hydrometeorological variable distribution in case certain meteorological phenomena were to occur at the surface of a specific region (Goodess and Palutikof 1998; Fowler et al. 2000; Bardossy et al. 2002, 2005). The frequency distributions of local or regional climate are then derived by weighting the local climate states with the relative frequencies of the weather classes (IPCC 2001). Kyung et al. (2011) applied the k method to examine the Korean Peninsula. Stochastic weather generators are statistical models designed to provide realistic random sequences of atmospheric variables, such as precipitation, temperature, and wind speeds. Kilsby et al. (2007) suggested a combination model, capable of generating rainfall models by connecting a Neyman–Scott rectangular pulses model with the weather factors presented by Watts et al. (2004). This model is capable of more accurately reproducing variability and extreme precipitation than the existing Markov chain model. One example of a regression model is to utilize a transfer function, which is to find a direct relationship between large-scale predictors and local predictands.

There have been a lot of studies using the above methodologies. For example, a number of researchers (e.g., Zorita and von Storch 1999; Yuval and Hsieh 2003; Dibike and Coulibaly 2006; Pasini and Langone 2010, 2012; Pasini and Modugno 2013) have used artificial neural networks for nonlinear prediction. Others (Karl et al. 1990; Wigley et al. 1990; von Storch et al. 1993; Busuioc et al. 2001; Huth 2002, Tomozeiu et al. 2006; Frias et al. 2006) have used canonical correlation analysis. The Statistical Downscaling Model, which was developed by Wilby et al. (2002) in combination with a stochastic model, has also been widely used. Recently, nonhomogeneous hidden Markov models were applied to downscale regional seasonal climate data for daily rainfall at various gauging sites (Robertson et al. 2007; Khalil et al. 2010).

Although many downscaling methods have been developed, producing actual reproductions of the hydrological process using these methods continues to be a challenge. Runoff in a basin can be accurately assessed only when the temporal correlation between the weather variables is preserved at each site, which then must be analyzed in relation with the spatial correlation between each site in the basin. The method presented in this study aims to preserve the correlations between weather variables, such as precipitation, temperature, humidity, wind velocity, and solar radiation. This downscaling method considers spatial dependence between each site for climate change projections. To reproduce these correlations, we introduce the nearest neighbor–genetic algorithm (NN–GA) method, which simulates the future daily weather series, operating under the assumption that the codependence relationships across the variables at a weather station are consistent with past events (i.e., at the microscale, the physics is unchanged, but the frequency of different types of events changes as a result of anthropogenic factors). Consistent with other statistical downscaling methods, we assumed that the potential changes in the larger-scale precipitation and temperature fields due to radiative effects, circulation changes, and atmospheric moisture holding capacity are reproduced by the GCM. The method was used to simulate a future daily weather series using the GCM data in the Namhan River basin of the Korean Peninsula.

2. Data used

a. Observed climate data

We used data from the Namhan River basin (Fig. 1), since it best represents the topographical and meteorological variations of the Korean Peninsula. In addition, there is a sufficient period of high-quality data available for this region. The total area of the basin is 12 577 km2, and the total length of the river is 375 km. The upstream area comprises highlands with elevations between 500 and 1550 m and the downstream area is hilly, with an average height of approximately 500 and m (see the DEM in Fig. 1). We selected six weather stations (the yellow circles in Fig. 1), which have collected data for at least 30 years, from the Korea Meteorological Administration. Geographic information for the weather stations and the monthly and daily data (minimum temperature, maximum temperature, precipitation, relative humidity, wind speed, and solar radiation) collecting periods are shown in Table 1.

Fig. 1.

Study area and location of stations.

Fig. 1.

Study area and location of stations.

Table 1.

Weather stations used in study.

Weather stations used in study.
Weather stations used in study.

b. NCEP–NCAR reanalysis data

The National Oceanic and Atmospheric Administration/ESRL Physical Sciences Division provides global climate observation data, including National Centers for Environmental Prediction (NCEP) reanalysis data. The NCEP–NCAR reanalysis, which is mainly used in climate change–related studies, provides data on temperature, relative humidity, specific humidity, sea level pressure, evapotranspiration, U-wind, V-wind, and so on, on a 4-h, daily, and monthly basis from 1 January 1948 to the present. This study used monthly NCEP data for the Korean Peninsula for the period of 1973–2008.

c. Climate model and scenario

This study uses data from the Centre National de Recherches Météorologiques Coupled Global Climate Model (CNRM-CM) outputs, obtained from the fifth phase of the Coupled Model Intercomparison Project (CMIP5) data archive (http://cmip-pcmdi.llnl.gov/cmip5/index.html). CNRM-CM has been developed jointly by the CNRM-Groupe d’Études de l’Atmosphere Météorologique (CNRM-GAME) with the Centre Européen de Recherche et de Formation Avancée en Calcul Scientifique (CERFACS) contributing to CMIP5. A full description and a basic evaluation of the system can be found in Voldoire et al. (2013). We analyzed the monthly mean data from the historical and future period using the radiative concentration pathway (RCP) 8.5 experiment from these CMIP5 models (Taylor et al. 2012).

d. Used data and periods

This study forecasted daily weather series data through downscaling. The input data are summarized in Table 2. The input climate variables were composed of the mean temperature, specific humidity, wind speed, and sea level pressure as provided by NCEP and the GCM from Kyoung et al.’s (2011) study. The training is performed for 36 yr (1973–2008) using the NCEP data, and is divided into a calibration period (1973–2003) and a verification period (2004–08). The forecasting is performed for 90 yr (2011–2100) from CNRM-CM data.

Table 2.

Input data of NN.

Input data of NN.
Input data of NN.

3. NN–GA for downscaling

a. Nearest neighbor

The nearest neighbor method has roots in nonlinear time series analysis in the state-space dimension and uses spatial correlation to improve nonlinear short-term forecasts. This NN method, which was proposed by Casdagli (1992), and Casdagli and Weigend (1994), is a method primarily applied to chaos time series under the assumption that past time series patterns would be repeated in the future after converting the time series data of a single variable into a set of vectors with the consideration of its delay time and embedding dimension.

This study extracted data from the meteorological variables provided and the month-to-month weather variables observed at each site by applying the simplest NN. This was based on the assumption that, under the condition of specific meteorological variables, the same weather characteristics can occur. Therefore, vectors were formed by using weather data series instead of using a single time series. Delay time was excluded, because it is insignificant, given that the monthly mean from averaging original time series data is used. The embedding dimension was assumed to have four dimensions of meteorological variables (temperature, specific humidity, wind speed, and sea level pressure) from the NCEP data in this study.

Above all, for application, a set of vectors needs to be constructed with the variables capable of affecting the weather variables, which are intended to be downscaled, as follows:

 
formula

Here, X is a meteorological variable, n is the number of meteorological variables affecting the weather variables to be downscaled, and t shows that the variables are time series data. To investigate the similarity between the training set and prediction set , this study calculated the distance between them, or . When the most similar with is identified through the process, the precipitation which occurred under the is inferred to occur again in .

b. Genetic algorithm

The genetic algorithm is based on C. Darwin’s “survival of the fittest” theory. The GA was proposed by Holland (1975) as a search algorithm that applied the principles outlined in Darwin’s assessment of the natural selection of organisms to mechanical learning areas. It has been applied to various applications, such as pattern recognition, including optimization, machine learning, robot engineering, and traveling-salesman problems. The GA is an organic evolution model derived from the natural world. It is a stochastic optimization method with excellent applicability in real-world problems, as it simulates a process where, among a group of individuals forming a generation, individuals with high environmental adaptability have higher survival rates in subsequent generations (survival of fittest), go through crossover and mutation, and form the next generation. In hydrology, GA was used as a methodology to overcome the local optimization of parameters (Ragab et al. 2001; Bhattacharjya 2004; Chang et al. 2006). In this research, we implemented a GA to a downscaling method, rather than an optimization method.

The general procedure for applying the genetic algorithm involves three major operators: selection, crossover, and mutation (as shown in Fig. 2). Selection is the process of selecting the fittest members in a population, which are preferred to be selected to reproduce promising new offspring. The purpose of crossover is exchanging important building blocks of two parent members to generate new offspring. For each dimension of the offspring member, a random number in the range [0, 1] is generated in the step of mutation.

Fig. 2.

Execution steps of the genetic algorithm.

Fig. 2.

Execution steps of the genetic algorithm.

c. Development of downscaling method through NN–GA

1) The method to generate daily weather series, considering the temporal correlation between the meteorological series and the spatial correlation of multiple sites

A method is necessary to maintain a temporal correlation of downscaled variables and spatial correlation between the downscaled variables. The NN method applied in this study has both advantages and disadvantages in overcoming this problem. Each site’s data of past events show the best description of the temporal correlation and spatial correlation. Thus, we can derive major events that occurred in the past by calculating the Euclidean distance to find nearest neighbors from all variables at the sites. Data series of each site have maintained the temporal correlation and spatial correlation. However, it is necessary to devise a method capable of effectively integrating each of the different weather variables showing past data in terms of Euclidean distance for each site. In addition, to predict an uncertain future, estimates show representative values based on sufficient numbers of data series, rather than predictions using a single series. This necessitates the concept of the GA, the basic assumption of which is that the derived weather variables of the past can be substituted for each other and that weather characteristics of one site can be reproduced in other sites. This study, therefore, presents a method for generating daily weather series that takes into consideration the temporal correlation between the weather variables and the spatial correlation between the sites (Fig. 3). Datasets existing in Euclidean distance D from the GCM were selected for each site (S1, S2, S3, and S4). The genetic algorithm was applied in order to downscale temporally the monthly datasets to a daily weather series. We used an adaptive genetic algorithm, which changes according to the population size, the probability of crossover, or the mutation probability while the genetic algorithm is running. After the genes composed of each site’s weather variables (D1, D2, and D3) derived in state space were crossed over, they could be substituted for the weather variables data of different years and different sites. The reconstituted gene in each site (S1′, S2′, S3′, and S4′) was subsequently evaluated as an objective function in state space, and the GA was realized using a repetitive selection of an excellent gene. The sum of squares error (SSE) between the historical data and the derived data from the GCM was used as an objective function in this study. The estimation was done by minimizing the SSE between the historical data and the derived data from the GCM as an objective function. In this study, a roulette wheel algorithm was applied to select genes for the crossover and mutation operations (Reca and Martinez 2006).

Fig. 3.

The process for applying the genetic algorithm for generating weather variables: 1) find historical datasets existing in Euclidean distance D from the predicted point [the red point (X1, X2) in state space]; 2) the genes are composed in each site; 3) crossover and mutation are repeated for selecting the best genes using an objective function.

Fig. 3.

The process for applying the genetic algorithm for generating weather variables: 1) find historical datasets existing in Euclidean distance D from the predicted point [the red point (X1, X2) in state space]; 2) the genes are composed in each site; 3) crossover and mutation are repeated for selecting the best genes using an objective function.

For precipitation, it is necessary to address changes in terms of both amount and occurrence. For example, in the future, the number of wet days may increase because of climate change. Therefore, we considered the frequency change to be a nonstationarity of precipitation in this study. A precipitation nonstationarity should be able to forecast a change in the number of wet days, as well as a change in the amount of precipitation. A method to generate weather variables beyond the range of prior observations and its pattern changes is needed since much stronger events are expected under climate change in the future. For this purpose, we developed the concept of mutation through a genetic algorithm, using the assumption that the daily data for each month can be considered to be a gene (Fig. 4). The daily precipitation for each of the relevant months can be estimated as the ratio to the monthly precipitation (assumed to be 100 mm in Fig. 4), after which it can be classified into wet and dry days. If a mutation of wet days occurs, a wet day (O) changes into a dry day (×) and vice versa. The quantity—ratio of daily precipitation—is therefore updated to generate a reinterpreted precipitation series (i.e., the gene that reflects the mutation, by multiplying the ratio of the daily precipitation amount by the total monthly precipitation). As a result, the maximum precipitation for our study site was updated from 30 to 33.3 mm, and the number of wet days was updated from 6 to 5 days (Fig. 4).

Fig. 4.

Method to apply mutation to the number of wet days.

Fig. 4.

Method to apply mutation to the number of wet days.

2) Generation of daily weather series through NN–GA

The NN–GA developed by this research can be modeled, as shown in Fig. 5. The contents of each stage can be summarized as follows.

  1. Find similar datasets (D1, D2, …, Dk) of the past using the NN method. The k in the NN method is determined by the number that leads to an optimal result while it is changing the number of similar training sets with prediction set in order of its similarity.

  2. Weather data of each weather station for the past relevant period of D1 are composed of one gene per month (D2, D3, …, Dk; follow the same method).

  3. One gene herein forms a pair of genes including the daily weather series (in this research, it forms six units of data: minimum temperature, maximum temperature, monthly precipitation, relative humidity, wind velocity, and solar radiation).

  4. Perform a crossover according to the crossover principle of the GA for genes of the weather stations (S1, S2, …, Sn) in each dataset.

  5. Together apply a principle of mutation when performing a crossover (in this research, the probability of mutation changes with the lapse of time).

  6. Set the selection conditions by organizing the objective function (the relationship between the monthly series by NN and the integrated monthly series from the simulated daily data by GA) and then select the satisfying and eliminate the unsatisfying data.

  7. Develop the algorithm by applying this process repeatedly.

Fig. 5.

Modeling mimetic diagram of the NN–GA method.

Fig. 5.

Modeling mimetic diagram of the NN–GA method.

4. Application result

a. Downscaling of climate data through the NN–GA method

1) Analysis and application method

The aim of the NN method is to determine the characteristics of the data series that should be forecasted using historical data. The best-performing k values were selected for each station: 11 for Daegoanrung, 15 for Wonju, 13 for Chungju, 13 for Yangpyeong, 18 for Icheon, and 9 for Jecheon. The embedding dimension m was selected to be four, since four kinds of climatic data were considered. To apply the NN method to show future climate change, we first had to determine a conceptual approach of nonstationarity. As the meteorological characteristics, including temperature and precipitation, have changed because of the impact of climate change, it was necessary to determine the characteristics of this change so that they could be reflected in the downscaling method for future forecasting. Both simple increases and decreases in precipitation as well as changes in wet days act as important factors showing changes in the precipitation characteristics, reflecting the occurrence of floods and droughts and representing the runoff systems in a basin. Accordingly, it is necessary to review the patterns of change in wet days so that these characteristics can be reflected in the downscaling method. This study solved this problem using the NN method and the concept of mutation. Generally, there is a large positive correlation between monthly precipitation and the number of wet days. This study reviewed the characteristics of long-term change in the number of wet days itself by eliminating the impact of monthly precipitation on the number of wet days. In other words, we reviewed the impact of other factors on wet days by using the calculated residuals from a linear regression between monthly precipitation and wet days. This study estimated the trend of monthly wet days (Fig. 6). The monthly estimated rate of increase is shown in Table 3.

Fig. 6.

The characteristics of monthly change in wet days: (left) relationship between monthly rainfall and wet days; (right) trends for wet days for (a) February and (b) August.

Fig. 6.

The characteristics of monthly change in wet days: (left) relationship between monthly rainfall and wet days; (right) trends for wet days for (a) February and (b) August.

Table 3.

Monthly rate of increase in wet days.

Monthly rate of increase in wet days.
Monthly rate of increase in wet days.

Analyzing past trends of monthly wet days indicated that the wet season (July–September) has increased precipitation and the dry season (October–April) has decreased precipitation. A probability of mutation was selected from the monthly trend of the calculated residuals from the linear regression in each month. For example, if the increased rate of a specific month is 0.01, the probability of mutation for the relevant month in the following year is 0.01; however, the probability of mutation for the relevant month after 10 yr is 0.1. Wet day trends can be accurately reflected using this concept of mutation. For the case of August 100 yr later, the daily precipitation series can be estimated, showing that the wet days increased by approximately 7.3 days.

2) Verification of NN–GA method

To verify the NN–GA method developed by this study, we analyzed the period of 2004–08 (5 yr), comparing the estimated data with the actual observation data in terms of the minimum temperature, maximum temperature, precipitation, relative humidity, wind speed, and solar radiation. Generally, weather generators, including NN–GA in this study, can be used to create synthetic meteorological data for given climate scenarios and operate on a daily time step. The daily time series are randomly resampled to represent the interest day’s weather. This means the time series of weather variables are statistically similar to observations. Therefore, it is not possible to perform a direct comparison between simulated and observed weather variables. Rather, the values of weather variables can be compared against the range of values of the same variable or the integrated monthly data. We performed a check to determine whether the simulated daily data were integrated into the monthly data with the range, and were well reflected in the features of the observed values. From the integrated weather series, the minimum temperature, maximum temperature, relative humidity, wind speed, and solar radiation were measured in terms of the monthly mean values, and precipitation was shown as monthly precipitation.

As shown in Fig. 7, we regarded the boundary of the top–bottom of the data series, produced using the NN–GA method, as the range of uncertainty. We designated the median values as the representative values and subsequently compared them with the observed values. We reviewed our simulation results using each downscaled-site weather series through the correlation coefficient and the Nash–Sutcliffe model efficiency coefficient (Table 4). A correlation coefficient exceeding 0.6 is highlighted in bold. In the case of minimum temperature, maximum temperature, precipitation, and relative humidity, the simulation results showed good results. For wind speed and solar radiation, the correlations showed comparatively lower results.

Fig. 7.

Verification of 2004–2009 weather data using the NN–GA method (Daegoanrung): (a) minimum and (b) maximum temperature, (c) monthly precipitation, (d) relative humidity, (e) wind speed, and (f) solar radiation.

Fig. 7.

Verification of 2004–2009 weather data using the NN–GA method (Daegoanrung): (a) minimum and (b) maximum temperature, (c) monthly precipitation, (d) relative humidity, (e) wind speed, and (f) solar radiation.

Table 4.

Comparison of correlation coefficient between observation data and simulation data. Correlation coefficients > 0.6 are boldface.

Comparison of correlation coefficient between observation data and simulation data. Correlation coefficients > 0.6 are boldface.
Comparison of correlation coefficient between observation data and simulation data. Correlation coefficients > 0.6 are boldface.

This study also compared the daily data between each observation value and simulation value using the correlation coefficients (as shown in Table 5) to determine whether the correlation between the weather variables that was downscaled to each site reflected the features of the observed data. Largely, there were positive correlations for the minimum temperature, maximum temperature, daily precipitation, and relative humidity. The average wind speed and solar radiation showed negative correlations. The difference in the correlation coefficient between the weather variables of each site for the observation value and simulation value were classified into three standards: 0–0.1, 0.1–0.2, and 0.2 or more (Table 5). Although each site shows slight variations, as a whole, solar radiation had the weakest reproducibility when compared with the other weather variables. However, in light of the fact that the maximum difference in the correlation coefficient was approximately 0.22 [wind speed and solar radiation of Daegoanrung (DG)], the simulation series was generally able to fairly accurately reproduce the correlations of the observation series (strong correlation vs weak correlation or positive correlation vs negative correlation).

Table 5.

Comparison of the observed series with series simulated using correlation coefficients between the weather variables for each site. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.

Comparison of the observed series with series simulated using correlation coefficients between the weather variables for each site. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.
Comparison of the observed series with series simulated using correlation coefficients between the weather variables for each site. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.

The results from the correlation coefficients for each site, comparing the daily data observation values and the simulation values of each of the weather variables characteristics are shown in Table 6. Overall, there was a good correlation between the sites of the weather variables. Conversely, DG showed a comparatively weaker correlation with other places, which may have been caused by geographical characteristics, as this site is located the farthest away from the center of the basin and at the highest altitude. We also differentiated between the observation value and simulation value for each of the weather variables characteristics (Table 6). For some of the sites, the difference in the correlation coefficient for the relative humidity, average wind speed, and solar radiation exceeded 0.1. However, most of the weather variables reproduced the spatial correlation comparatively well, with correlation coefficient differences of less than 0.2.

Table 6.

Comparison between the observed series and the simulated series for correlation coefficients between the sites for each weather variable. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.

Comparison between the observed series and the simulated series for correlation coefficients between the sites for each weather variable. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.
Comparison between the observed series and the simulated series for correlation coefficients between the sites for each weather variable. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.

b. Assessment of future weather variables using the NN–GA method

To assess the impact of climate change on future weather variables data, this study utilized CNRM-CM data for an RCP 8.5 climate change scenario and applied it to the NN–GA method that was developed in this study. The simulation period was 90 yr (2011–2100). We performed a simulation for each site series, and the analysis used the daily data mean. All of the coefficients, except relative humidity, showed a trend to increase in the first-order linear regression model, with the minimum temperature and maximum temperature demonstrating apparent increasing trends (Fig. 8). Comparing the weather variables for the year 2100 with those from 2011 showed increases in the minimum temperature and maximum temperature by approximately 4.5°C, with precipitation amounts increasing by 1.7 mm day−1, which is equivalent to an increase of about 616 mm yr−1 (Table 7). In the case of the Namhan River basin, where the average annual precipitation is about 1905 mm, the rainfall is estimated to increase by 32.3% in the future. In contrast, the relative humidity showed a decreasing trend, which is considered to be a result of the dry season increasing. Thus, extreme precipitation is expected to occur frequently because of the changes in the precipitation patterns, such as an increase in the precipitation amount and a decrease in wet days.

Fig. 8.

Results of future (2011–3000) weather variables assessment for (a) minimum and (b) maximum temperature, (c) daily precipitation, (d) relative humidity, (e) wind speed, and (f) solar radiation.

Fig. 8.

Results of future (2011–3000) weather variables assessment for (a) minimum and (b) maximum temperature, (c) daily precipitation, (d) relative humidity, (e) wind speed, and (f) solar radiation.

Table 7.

Comparison of estimated average increase amount.

Comparison of estimated average increase amount.
Comparison of estimated average increase amount.

The NN–GA method developed in this study focused on the following three purposes in its downscaling method application: 1) to reproduce the nonstationarity of a data series, 2) to reproduce spatial correlation between sites, and 3) to reproduce temporal correlation between data series. To determine the nonstationarity of the data series, we evaluated the precipitation series. This factor is the most important in the field of water resources, because there are many limitations when analyzing changes in precipitation amounts based on the simple concept of a mean value. Therefore, we examined changes in the precipitation patterns, based on the concept of extreme value. In addition to quantity, we also analyzed changes in precipitation occurrence. For this, the World Meteorological Organization (Klein Tank et al. 2009) suggested the use of the Expert Team on Climate Change Detection and Indices (ETCCDI) for the evaluation of extreme values. This study evaluated changes in the precipitation characteristics using four precipitation-related indices (Table 8).

Table 8.

ETCCDI used in evaluation of precipitation. Here RRij is the daily precipitation amount on day i in period j; RRwj is the daily precipitation amount on a wet day w (RR ≥ 1.0 mm) in period j; RRwn99 is the 99th percentile of precipitation on wet days.

ETCCDI used in evaluation of precipitation. Here RRij is the daily precipitation amount on day i in period j; RRwj is the daily precipitation amount on a wet day w (RR ≥ 1.0 mm) in period j; RRwn99 is the 99th percentile of precipitation on wet days.
ETCCDI used in evaluation of precipitation. Here RRij is the daily precipitation amount on day i in period j; RRwj is the daily precipitation amount on a wet day w (RR ≥ 1.0 mm) in period j; RRwn99 is the 99th percentile of precipitation on wet days.

Our evaluation results showed increasing tendencies for CDD, CWD, PRCPTOT, and R99pTOT (Fig. 9). The fact that CDD increased, despite the increase in the annual precipitation, indicates that the precipitation concentration occurs more frequently in the wet season, which is evidenced by the increasing trend toward extreme precipitation (Fig. 9d). This pattern was determined through changes in recent precipitation patterns in the Korean Peninsula; therefore, the NN–GA method has proven its applicability in reproducing the characteristics of change in various data series, including extreme values.

Fig. 9.

Results for 2010–3000 of evaluation using ETCCDI: (a) CDD, (b) CWD, (c) PRCTOT, and (d) R99TOT.

Fig. 9.

Results for 2010–3000 of evaluation using ETCCDI: (a) CDD, (b) CWD, (c) PRCTOT, and (d) R99TOT.

5. Discussion

In this research, we developed a downscaling method that utilizes GCM data in the field of water resources. This study focused on reproducing the nonstationarity of a data series and the correlation between the series and the sites. Hence, this study used the GA method as an NN method to derive data from the state space. In addition, the GA was used to consider the nonstationarity of data series and correlations between the series and the sites. The singularity of this study is that it utilized the GA method, which has mainly been used as an optimization method in hydrologic fields, by applying it to a downscaling method. This novel GA crossover method not only considers uncertainty by extending the data series while maintaining spatial correlation between the sites but also reproduces the changes in weather patterns while maintaining the conceptual mutation of the GA and the temporal correlation between the weather variables in order to interpret the conceptual nonstationarity of climate change. The following contents were ascertained based on the derived study results and the application of this method. As a result of analyzing the correlation between the sites and weather variables for daily weather series during the validation periods, the simulation series was able to accurately reproduce the correlation of the observation series, with a correlation coefficient error of less than 0.2, which validates this method’s applicability.

The results of this study were compared with the results of MarkSim as a Web-based weather generator tool (http://gismap.ciat.cgiar.org/MarkSimGCM). MarkSim estimates the daily maximum and minimum air temperatures and the daily solar radiation values from monthly means of these variables using the methods devised by Richardson (1981). The monthly solar radiation values are estimated from temperature, longitude, and latitude using the model from Donatelli and Campbell (1997). MarkSim is described in detail in Jones and Thornton (2013). The CNRM-CM GCM for the A2 SRES emission scenario in the Coupled Model Intercomparison Project, phase 3 (CMIP3), which is provided online, was selected for use, as it is the most similar model and scenario to this study. A comparison of the NN–GA with MarkSim for the correlation coefficient between the Daegoanrung and Wonju sites for 2011–20 is shown in Fig. 10. Only Daegoanrung and Wonju were selected because the spatial resolution of the MarkSim is 0.5° × 0.5°, and the grid at the Wonju site includes most of the other sites, excluding the Daegoanrung site. The NN–GA was able to better express the observed characteristics than MarkSim for all four of the variables: minimum temperature, maximum temperature, daily precipitation, and solar radiation (Fig. 10). The most difficult variable to estimate was solar radiation, which showed a relatively poor correlation for both the MarkSim and NN–GA methods; however, the NN–GA was able to fairly accurately reproduce a correlation between the observation series (Table 9). First of all, the results depend on the application of different methods and spatial resolution in each method. MarkSim uses 720 classes of weather around the world, and the basic algorithm is a daily rainfall simulator that uses a third-order Markov process to predict the occurrence of a rainy day. The spatial resolution is 30 arc-min (0.5°) on a pixel-by-pixel basis. On the other hand, the NN–GA method is based on the nearest neighbor linked with nonlinear genetic algorithm. The NN–GA downscales the weather directly from a GCM to a particular site using observed data at the site. Therefore, the use of observed data instead of averaged data shows better results in the NN–GA method. The strength of the NN–GA method is its capability to generate weather variables considering temporal and spatial correlation using observed data, and thus it is better suited for downscaling over a local area. Furthermore, it is very flexible when selecting input and output variables. However, it shows limitation when applied to an ungauged area or a large area with insufficient hydrological homogeneity. By contrast, the MarkSim can be applied to any region, including a large area or an ungauged area, since it was developed for the global area. But it offers only a few weather variables as output. It seems that the correlations are overestimated between solar radiation and temperatures (see Table 9). The MarkSim uses the method devised by Richardson (1981), but even his study showed that maximum temperature and solar radiation are positively correlated with a correlation coefficient of about 0.25 [(0.19–0.32)], minimum temperature and solar radiation are negatively correlated with a correlation coefficient of about −0.2 [−( 0.153–0.248)] at the Temple, Texas, Atlanta, Georgia, and Spokane, Washington, sites. We believe that the issue arises because of the integration process from sites to a grid.

Fig. 10.

Comparison of the observations vs NN–GA and MarkSim for the correlation coefficient between the DG and WJ sites for 2011–20.

Fig. 10.

Comparison of the observations vs NN–GA and MarkSim for the correlation coefficient between the DG and WJ sites for 2011–20.

Table 9.

Comparison of the correlation coefficients for NN–GA and MarkSim for the the simulated weather variables for the DG and WJ sites. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.

Comparison of the correlation coefficients for NN–GA and MarkSim for the the simulated weather variables for the DG and WJ sites. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.
Comparison of the correlation coefficients for NN–GA and MarkSim for the the simulated weather variables for the DG and WJ sites. Differences between observation and simulation are marked as follows: 0–0.1 (normal font), 0.1–0.2 (boldface font), and 0.2 and above (boldface italic font). Here temp indicates temperature.

Conversely, there were several potential issues and limitations. This study developed an NN–GA method, which focused on reproducing the correlation between the series and the sites by downscaling six weather variables to six sites. However, the results of mean wind speed and solar radiation were, comparatively, not accurate enough to show the temporal correlation and spatial correlation between the sites. We reviewed the correlation between the weather series and NCEP data to verify this. We selected WJ as a representative site and estimated the correlation coefficient between the weather variables and the NCEP data (average temperature, sea level pressure, specific humidity, and wind speed) adopted in this study, then applied the downscaling method. The mean wind speed and solar radiation had a weak correlation with the NCEP data, compared to other weather variables (shown in Fig. 11). This may influence continuously the application of the downscaling method. To downscale many weather variables, including the mean wind speed and solar radiation series, attention should be paid to the selection of variables for the NCEP and GCM. Future research should focus on developing a method capable of addressing these variables more efficiently.

Fig. 11.

Correlation between the NECP values and weather variables.

Fig. 11.

Correlation between the NECP values and weather variables.

This study calculated the monthly trend (increase rate) for wet days, based on a one-dimension regression that first eliminated the correlation between monthly precipitation and wet days, which is not suitable for long-term prediction. Consequently, we studied an additional method that could be used to link the change in wet days with the outside factors of the NCEP or GCM. Finally, the NN–GA method presented in this study was developed for utilization in the field of water resources; therefore, this method should be applied to other sites located in the same basin to maintain hydrologic homogeneity. Keeping these limitations in mind, this method can be utilized in the downscaling of GCM data to assess the effects of climate change in the field of water resources.

6. Conclusions

As the existing methods for downscaling a single weather variable to a single site using climate data, such as from a GCM, in the field of water resources are limited, this study developed the NN–GA method. The NN–GA method works by linking the nearest neighbor with a genetic algorithm to devise a downscaling method that is capable of targeting various weather variables and of taking into account the correlations between the weather variables and the downscaled sites. To determine the applicability of the developed NN–GA method, this study was verified using data from the Namhan River basin. The results of the application of six weather variables (minimum temperature, maximum temperature, precipitation, relative humidity, mean wind speed, and solar radiation) for six weather stations in the Namhan River basin showed that the NN–GA method produced a comparatively good reproduction of correlation between the weather variables and the downscaled sites. Accordingly, this method was used to estimate the future daily weather series in the Namhan River basin using the results of the CNRM-CM GCM, according to the RCP 8.5 climate change scenario, determining whether it was able to accurately reproduce the nonstationarity of the data series, the correlation between the weather variables, and the spatial correlation between the sites. Reviewing the change in patterns of extreme precipitation together with the change in daily weather series showed that the data series was able to fairly accurately reproduce the conceptual nonstationarity. This method was capable of very accurately reproducing the correlation between the weather variables and the sites.

Although many methods have been developed to utilize GCM data in water resources, there are many difficulties that still exist when directly utilizing the results of the downscaling method on a basin scale. The NN–GA method that was presented in this study is a downscaling method that can be used in the utilization of GCM data.

Acknowledgments

This work was supported by a National Research Foundation of Korea (NRF) grant, funded by the Korea government (MEST; 2011-0028564).

REFERENCES

REFERENCES
Bardossy
,
A.
,
J.
Stehlík
, and
H. J.
Caspary
,
2002
:
Automated objective classification of daily circulation patterns for precipitation and temperature downscaling based on optimized fuzzy rules
.
Climate Res.
,
23
,
11
22
, doi:.
Bardossy
,
A.
,
I.
Bogardi
, and
I.
Matyasovszky
,
2005
:
Fuzzy rule-based downscaling of precipitation
.
Theor. Appl. Climatol.
,
82
,
119
129
, doi:.
Benestad
,
R. E.
,
I.
Hanssen-Bauer
, and
D.
Chen
,
2009
: Empirical–Statistical Downscaling. World Scientific, 228 pp.
Bhattacharjya
,
R. K.
,
2004
:
Optimal design of unit hydrographs using probability distribution and genetic algorithms
.
Sadhana
,
29
,
499
508
, doi:.
Busuioc
,
A.
,
D.
Chen
, and
C.
Hellstrom
,
2001
:
Temporal and spatial variability of precipitation in Sweden and its link with large-scale circulation
.
Tellus
,
53A
,
348
367
, doi:.
Casdagli
,
M.
,
1992
: Dynamical systems approach to modeling input–output systems. Nonlinear Modeling and Forecasting, M. Casdagli and S. Eubank, Eds., Santa Fe Institute Series, Book 12, AddisonWesley, 265–281.
Casdagli
,
M.
, and
A.
Weigend
,
1994
: Exploring the continuum between deterministic and stochastic modeling. Forecasting the Future and Understanding the Past, A.S. Weigend and N.A. Gershenfeld, Eds., Santa Fe Institute Series, Book 15, Addison–Wesley, 347–366.
Chang
,
C. L.
,
S. L.
Lo
, and
S. L.
Yu
,
2006
:
The parameter optimization in the inverse distance method by genetic algorithm for estimating precipitation
.
Environ. Monit. Assess.
,
117
,
145
155
, doi:.
Chen
,
T. S.
,
P. S.
Yu
, and
Y. H.
Tang
,
2010
:
Statistical downscaling of daily precipitation using support vector machines and multivariate analysis
.
J. Hydrol.
,
385
,
13
22
, doi:.
Dibike
,
Y. B.
, and
P.
Coulibaly
,
2006
:
Temporal neural network for downscaling climate variability and extremes
.
Neural Networks
,
19
,
135
144
, doi:.
Donatelli
,
M.
, and
G. S.
Campbell
,
1997
: A simple model to estimate global solar radiation. Research Institute for Industrial Crops PANDA Project Res. Pap. 26, 3 pp.
Fowler
,
H. J.
,
C. G.
Kilsby
, and
P. E.
O’Connell
,
2000
:
A stochastic rainfall model for the assessment of regional water resource systems under changed climatic conditions
.
Hydrol. Earth Syst. Sci.
,
4
,
263
281
, doi:.
Frias
,
M. D.
,
E.
Zorita
,
J.
Fernandez
, and
C.
Rodriguez-Puebla
,
2006
:
Testing statistical downscaling methods in simulated climates
.
Geophys. Res. Lett.
,
33
,
L19807
, doi:.
Giorgi
,
F.
, and Coauthors
,
2001
: Regional climate information—Evaluation and projections. Climate Change 2001: The Scientific Basis, J. T. Houghton et al., Eds., Cambridge University Press, 583–638.
Goodess
,
C. M.
, and
J. P.
Palutikof
,
1998
:
Development of daily rainfall scenarios for southeast Spain using a circulation-type approach to downscaling
.
Int. J. Climatol.
,
18
,
1051
1083
, doi:.
Goodess
,
C. M.
, and Coauthors
,
2005
:
An intercomparison of statistical downscaling methods for Europe and European regions—Assessing their performance with respect to extreme temperature and precipitation events
. Climatic Research Unit Tech. Rep. CRU RP 11, 69 pp. [Available online at http://www.cru.uea.ac.uk/cru/pubs/crurp/CRU_RP11.pdf.]
Gutmann
,
E. D.
,
R. M.
Rasmussen
,
C.
Liu
,
K.
Ikeda
,
D. J.
Gochis
,
M. P.
Clark
,
J.
Dudhia
, and
G.
Thompson
,
2012
:
A comparison of statistical and dynamical downscaling of winter precipitation over complex terrain
.
J. Climate
,
25
,
262
281
, doi:.
Hay
,
L. E.
, and
M. P.
Clark
,
2003
:
Use of statistically and dynamically downscaled atmospheric model output for hydrologic simulations in three mountainous basins in the western United States
.
J. Hydrol.
,
282
,
56
75
, doi:.
Holland
,
J. H.
,
1975
: Adaptation in Natural and Artificial Systems. University of Michigan Press, 211 pp.
Huntington
,
T. G.
,
2006
:
Evidence for intensification of the global water cycle: Review and synthesis
.
J. Hydrol.
,
319
,
83
95
, doi:.
Huth
,
R.
,
2002
:
Statistical downscaling of daily temperature in central Europe
.
J. Climate
,
15
,
1731
1742
, doi:.
IPCC
,
2001
: Climate Change 2001: The Scientific Basis. Cambridge University Press, 881 pp.
IPCC
,
2013
: Climate Change 2013: The Physical Science Basis. Cambridge University Press, 1535 pp., doi:.
Jones
,
P. G.
, and
P. K.
Thornton
,
2013
:
Generating downscaled weather data from a suite of climate models for agricultural modelling applications
.
Agric. Syst.
,
114
,
1
5
, doi:.
Karl
,
T. R.
,
W. C.
Wang
,
M. E.
Schlesinger
,
R. W.
Knight
, and
D.
Portman
,
1990
:
A method of relating general circulation model simulated climate to the observed local climate. Part I: Seasonal statistics
.
J. Climate
,
3
,
1053
1079
, doi:.
Khalil
,
A. F.
,
H. H.
Kwon
,
U.
Lall
, and
Y. H.
Kaheil
,
2010
:
Predictive downscaling based on non-homogeneous hidden Markov models
.
Hydrol. Sci. J.
,
55
,
333
350
, doi:.
Kilsby
,
C. G.
, and Coauthors
,
2007
:
A daily weather generator for use in climate change studies
.
Environ. Modell. Software
,
22
,
1705
1719
, doi:.
Klein Tank
,
A. M. G.
,
F. W.
Zwiers
, and
X.
Zhang
,
2009
: Guidelines on analysis of extremes in a changing climate in support of informed decisions for adaptation. WMO Tech. Doc. WMO/TD-1500, 53 pp.
Kyoung
,
M.
,
H. S.
Kim
,
B.
Sivakumar
,
V. P.
Singh
, and
K. S.
Ahn
,
2011
:
Dynamic characteristics of monthly rainfall in the Korean peninsula under climate change
.
Stochastic Environ. Res. Risk Assess.
,
25
,
613
625
, doi:.
Pasini
,
A.
, and
R.
Langone
,
2010
:
Attribution of precipitation changes on a regional scale by neural network modeling: A case study
.
Water
,
2
,
321
332
, doi:.
Pasini
,
A.
, and
R.
Langone
,
2012
:
Influence of circulation patterns on temperature behavior at the regional scale: A case study investigated via neural network modeling
.
J. Climate
,
25
,
2123
2128
, doi:.
Pasini
,
A.
, and
G.
Modugno
,
2013
:
Climatic attribution at the regional scale: A case study on the role of circulation patterns and external forcings
.
Atmos. Sci. Lett.
,
14
,
301
305
, doi:.
Ragab
,
R.
,
B.
Austin
, and
D.
Moidinis
,
2001
:
The HYDROMED model and its application to semi-arid Mediterranean catchments with hill reservoirs 1: The rainfall-runoff model using a genetic algorithm for optimisation
.
Hydrol. Earth Syst. Sci.
,
5
,
543
553
.
Reca
,
J.
, and
J.
Martinez
,
2006
:
Genetic algorithms for the design of looped irrigation water distribution networks
.
Water Resour. Res.
,
42
,
W05416
, doi:.
Richardson
,
C. W.
,
1981
:
Stochastic simulation of daily precipitation, temperature and solar radiation
.
Water Resour. Res.
,
17
,
182
190
, doi:.
Robertson
,
A. W.
,
V. M. I.
Amor
, and
W. H.
James
,
2007
:
Downscaling of seasonal precipitation for crop simulation
.
J. Appl. Meteor. Climatol.
,
46
,
677
693
, doi:.
Schmidli
,
J.
,
C. M.
Goodess
,
C.
Frei
,
M. R.
Haylock
,
Y.
Hundecha
,
J.
Ribalaygua
, and
T.
Schmith
,
2007
:
Statistical and dynamical downscaling of precipitation: An evaluation and comparison of scenarios for the European Alps
.
J. Geophys. Res.
,
112
,
D04105
, doi:.
Spak
,
S.
,
T.
Holloway
,
B.
Lynn
, and
R.
Goldberg
,
2007
:
A comparison of statistical and dynamical downscaling for surface temperature in North America
.
J. Geophys. Res.
,
112
,
D08101
, doi:.
Taylor
,
K. E.
,
R. J.
Stouffer
, and
G. A.
Meehl
,
2012
:
An overview of CMIP5 and the experiment design
.
Bull. Amer. Meteor. Soc.
,
93
,
485
498
, doi:.
Tomozeiu
,
R.
,
C.
Cacciamani
,
V.
Pavan
,
A.
Morgillo
, and
A.
Busuioc
,
2006
:
Climate change scenarios of surface temperature in Emilia-Romagna (Italy) obtained using statistical downscaling models
.
Theor. Appl. Climatol.
,
90
,
25
47
, doi:.
Voldoire
,
A.
, and Coauthors
,
2013
:
The CNRM-CM5.1 global climate model: Description and basic evaluation
.
Climate Dyn.
,
40
,
2091
2121
, doi:.
von Storch
,
H.
,
E.
Zorita
, and
U.
Cubash
,
1993
:
Downscaling of global climate change estimates to regional scales: An application to Iberian rainfall in wintertime
.
J. Climate
,
6
,
1161
1171
, doi:.
Watts
,
M.
,
Goodess
C. M.
, and
Jones
P. D.
,
2004
: The CRU daily weather generator. Built Environment: Weather Scenarios for Investigation of Impacts and Extremes Tech. Note 1, version 2, 7 pp. [Available online at http://www.cru.uea.ac.uk/cru/projects/betwixt/documents/BETWIXT_TBN_1_v22.pdf.]
Wigley
T. M. L.
,
P. D.
Jones
,
K. R.
Briffa
, and
G.
Smith
,
1990
:
Obtaining sub-grid-scale information from coarse-resolution general circulation model output
.
J. Geophys. Res.
,
95
,
1943
1953
, doi:.
Wilby
,
R. L.
,
C. W.
Dawson
, and
E. M.
Barrow
,
2002
:
SDSM—A decision support tool for the assessment of regional climate change impacts
.
Environ. Modell. Software
,
17
,
145
157
, doi:.
Wilby
,
R. L.
,
S. P.
Charles
,
E.
Zorita
,
B.
Timbal
,
P.
Whetton
, and
L. O.
Mearns
,
2004
: Guidelines for use of climate scenarios developed from statistical downscaling methods. IPCC Tech. Rep., 27 pp. [Available online at http://www.ipcc-data.org/guidelines/dgm_no2_v1_09_2004.pdf.]
Yuval
, and
W. W.
Hsieh
,
2003
:
An adaptive nonlinear scheme for precipitation forecasts using neural networks
.
Wea. Forecasting
,
18
,
303
310
, doi:.
Zorita
,
E.
, and
H.
von Storch
,
1999
:
The analogue method as a simple statistical downscaling technique: Comparison with more complicated methods
.
J. Climate
,
12
,
2474
2489
, doi:.

Footnotes

*

Current affiliation: K-water Institute, Yuseong-gu, Daejeon, South Korea.