Atmospheric teleconnection-based conditional streamflow distributions for the Han River and its sub-watersheds in Korea

Authors


Abstract

An improved understanding of hydroclimatic variability is central to efforts that seek to anticipate and adapt water resources management in a variable and changing climate. In Korea's Han River Basin, warm season (June–September) streamflow contributes nearly 74% of the annual water supply. In this study, a 40-year record of simulated daily unimpaired streamflow for the Han River Basin and its 24 sub-watersheds is analysed with a view to identify relationships with atmospheric teleconnection patterns. Tree-based classification methods identify East Atlantic-Western Russia (EA-WR) and East Atlantic (EA) as the two dominant atmospheric circulation patterns that influence seasonal streamflow variability in the Han River Basin. The classification method allows partitioning of predictor space into three sub-regions, wherein conditional streamflow statistics were computed based on the joint phases of EA-WR and EA teleconnection indices. On the basis of conditional hydrologic-teleconnection information, we estimated the 50-year return period high- and low-flow events (fitted to a lognormal distribution), thus allowing a simultaneous assessment of the joint effect of changes in the mean and variability of streamflow. Application of this approach for the 24 sub-watersheds clarified the spatial coherence of teleconnection-related impacts on streamflow. Despite the relatively modest sample size, the approach presented here illustrates the potential for identification of conditional streamflow information based on climatic precursors. This diagnostic study is a critical first step in developing seasonal streamflow estimates conditioned upon large-scale climatic state, particularly for multi-objective reservoir and watershed management. Copyright © 2011 Royal Meteorological Society

1. Introduction

The Han River Basin has been a major source of drinking and industrial water supply, fisheries, irrigation, hydropower generation, and recreational water use. As a result, changes in water quality and quantity are of great concern. Recent studies suggest that water quality in some tributaries have been deteriorating as a result of population growth and inadequate wastewater treatment facilities (Chang, 2008; Kim et al., 2008; Jung et al., 2009). Furthermore, rapid population growth and future development plans have raised concerns such as water allocation issues, deterioration of water quality and rising energy demands (Kim et al., 2008; Normile, 2010). In the Korean peninsula, warm season (June–September) hydrologic variability plays an important role in sustainable water supply for cities and irrigation, hydropower generation, as well as ecosystem services. During the warm season, the frequency and magnitude of precipitation are mainly determined by the intensity of monsoon and typhoon activities (Byun and Lee, 2002; Kim et al., 2006a; Sun et al., 2007; Wang et al., 2007; Kim et al., 2010b). Large-scale patterns of atmospheric circulation are responsible for tropical moisture transport into the Korean peninsula. As a result, atmospheric teleconnection patterns have significant impacts on the regional and river basin-scale hydrologic variability. Kim et al. (2010b) explored the modulation of summer season streamflow volumes by the East Atlantic-Western Russia teleconnection pattern using a linear modelling approach.

Several research studies have focused on identifying contemporaneous and lag relationship between large-scale atmospheric circulation patterns and hydroclimatic variables worldwide (Glantz et al., 1991; Thompson and Wallace, 2001; Choi et al., 2010). This study considered a suite of atmospheric teleconnection indices (available at from NOAA's Climate Prediction Center) as predictors to explore optimal classification using a tree model for Han River streamflow variability associated during the June–September season. As contrasted with the linear analysis of Kim et al. (2010b), tree-based approach presented here allows an exploratory assessment of the key teleconnection indices with very limited assumptions. Tree-based methods are known to be effective in isolating interesting structures and interactions that are often missed by linear approaches (De'ath and Fabricius, 2000). Furthermore, this study extends the analysis of climate-informed conditional streamflow to the 24 sub-watersheds of the Han River, of which some have critical multipurpose dams and reservoirs. In what follows, the study region, materials and methods are described in Section 2. The analysis results are presented in Section 3. Section 4 reports the summary and conclusions.

2. Materials and methods

2.1. Study region

The Han River Basin is located in the middle of the Korean peninsula, 36°30′–38°55′ north latitude and 126°24′–129°02′ east longitude (Figure 1). Drainage area (26 356 km2) for the river basin accounts for 23% of the South Korean territory. The precipitation regime in the Han River Basin shows a strong seasonal character, with the highest rainfall during July (307 mm) and lowest rainfall during December (21 mm). Figure 2 shows long-term mean daily discharge in the Han River Basin. The June–September season, on average, provides nearly three-fourths of annual rainfall, with a seasonal coefficient of variation (the ratio of seasonal standard deviation to the mean) of 0.42.

Figure 1.

The Han River Basin, located in the center of the Korean Peninsula (http://wamis.go.kr). A four-digit number represents the subwatershed number. The full name of stations can be found in the Supplementary Information Section. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

Figure 2.

Seasonality of streamflow in the Han River Basin based on modelled unimpaired flows for the 1966–2005 period. Boxplots summarize changes in timing of annual maximum flow (AMF), AMF-based peaks over threshold (POT), center of volume (COV), and 7-day low flow (LOW FLOW) using a boxplot (0.1, 0.25, 0.5, 0.75, 0.9 quantile levels are shown). The solid line shows long-term median flows

According to 2005 statistics compiled by the South Korean Government, the population in the Han-River Basin makes up more than 40% of the national total, which has almost tripled between 1966 and 2005, from about 7.5 million to about 19.1 million. Urban areas make up just 1% of the basin. Forests make up 78%, cultivated areas 16%, and grasslands and water bodies 5% (UNESCO, 2011).

In the Han River Basin, there are 22 dams in operation, including three multipurpose dams, two reservoirs for municipal and industrial water supply, nine hydroelectric dams, six agricultural dams (effective storage: more than 10 million m3), and two estuarine barrages. The total water supply by these 22 dams is 5150 million m3/year. This is almost 35% of the total water supply in South Korea. Average annual energy generation is approximately 1200 Giga Watt-hours (GWh) from 3 multipurpose dams and 3600 GWh with 9 hydroelectric dams (MLTM, 2010).

2.2. Data and methods

The long-term runoff data (1966–2005) is based on daily, unimpaired runoff records provided by the Water Resources Management Information System (WAMIS) http://wamis.go.kr/). The runoff data is generated through hydrologic model simulation and provided by WAMIS is an accomplishment of the national watershed research project in Korea. As there are only a few flow-gauging stations available in the Han River Basin, we used runoff data from physically based hydrologic modelling of the Han River Basin using the Precipitation Runoff Modelling System (PRMS) approach. The PRMS model, developed by the US Geological Survey (USGS), has been used for watershed studies in the several regions around the world including South Korea (Leavesley, 1983; Kim et al., 2005; Dressler, 2006; Bae et al., 2008). Bae et al. (2008) provide extensive details regarding model calibration and verification and the spatial and temporal variability of streamflow based on the PRMS modelling effort over the South Korea portion of the Korean peninsula region.

Figure 2 shows the simulated daily mean discharge for the 1966–2005 periods. On the basis of daily streamflow during the 1966–2005 periods, approximately 74% of the annual flow occurs during the summer monsoon period (June–September). Furthermore, the importance of the JJAS season goes beyond the total volume of runoff delivered during this time. Some of the key features of streamflow, of relevance to environmental flow management and ecological integrity (Postel and Richter, 2003), also occur during this warm season. For example, Figure 2 shows the distribution of four key metrics of ecologically relevant flow variability: (1) Annual maximum flows (AMF), with a median date of 31 July is commonly used for infrastructure design. It is also an important hydrologic extreme that is linked to important riverine and riparian processes, such as sediment deposition and redistribution to support species lifecycle stages. The empirical distribution of the Julian date of the annual maximum flow, summarized by the boxplot, also falls in the warm season. (2) Number of peaks exceeding the long-term mean AMF occurred from June to September. (3) Centre of volume (COV) date corresponds to the calendar when half the annual flow volume has occurred. COV has been used to investigate to diagnose the trends in the timing of streamflow (Maurer et al., 2007). The median time for the occurrence of COV is 29 July. (4) Low-flow period, however, occurred in the winter-spring months (January–March). A suite of these streamflow metrics is important to quantify the alteration of hydrologic regime for protecting freshwater biodiversity and ecosystem services (Poff et al., 1997; Postel and Richter, 2003).

As noted in Kim et al. (2010a), atmospheric teleconnections are associated with persistent meteorological anomalies (geopotential heights and associated moisture fluxes) that result in precipitation excess and deficits. However, the separability of effects of episodic typhoons and seasonal teleconnection patterns is a necessary step in efforts to understand and quantify current or future changes in the warm season hydroclimatology in this region. To this end, this study models the relationship between teleconnection patterns and streamflow by considering a spatial domain (120°E–138°E, 32°N–40°N) from the Korean Meteorological Administration (KMA. 1996; Oh and Moon, 2008) to enable a subjective separation of typhoon and non-typhoon seasonal precipitation and streamflow. Kim et al. (2010b) provide a detailed explanation of this approach. Their analysis suggests that, on average, 22.6% of the seasonal flow volume is linked to the episodic typhoons over the Korean peninsula (refer to Supplementary Information Figure 1).

While atmospheric teleconnections are known to influence regional precipitation and temperature patterns worldwide (Glantz et al., 1991), regional impacts (as seen in precipitation, temperature and streamflow) often result from nonlinear interactions between these patterns. In this study, a principal focus is that of ascertaining the combination of teleconnections patterns that modulate seasonal streamflow variability in the Han River Basin. As noted previously, tree models (Breiman et al., 1984; De'ath and Fabricius, 2000) are particularly suited to data set with very limited assumptions to isolate interesting structures and interactions, which are not often captured by linear approaches. We conducted a conditional partitioning analysis using a tree model to identify the influence of atmospheric circulation patterns on hydrologic variability. The tree-based approach is especially effective in partitioning the predictor space, iteratively selecting the most influential predictors. This classification approach has been successfully applied to environmental responses to anthropogenic stresses (De'ath and Fabricius, 2000; Vayssieres et al., 2000; Rovertson et al., 2001; Prasad et al., 2006). For tree-based methods, we used seasonal streamflow (June–September) as the response variables and atmospheric teleconnections (North Atlantic Oscillation (NAO), East Atlantic Pattern (EA), West Pacific Pattern (WP), East Pacific/North Pacific Pattern (EP-NP), Pacific/North American Pattern (PNA), East Atlantic/West Russia Pattern (EA-WR), Scandinavia Pattern (SCA), Polar/Eurasia Pattern (POL)) as potential predictor variables. On the basis of regional precipitation pattern associated with the teleconnection patterns (NOAA CPC: http://www.cpc.noaa.gov/data/teledoc/tele- contents.shtml), EA-WR and EA show distinct regional footprint on the Korean Peninsula, and this additional confirmatory information was used to decide the pruning procedure (complexity parameter). Furthermore, reduction in sample size (n) due to further branching would severely limit the reliability of conditional mean (µ) and standard deviation (σ) estimates. The optimal tree size is based on the smallest splits (number of split = 2) whose cross-validation error is the minimum. Tree-based modelling approach involves the following steps: an initial response variable is found for which local split is optimal. The data is divided into two groups, and then each sub-group is recursively partitioned until a splitting criterion is satisfied or until improvement cannot be made (Maindonald and Braun, 2007). We anticipate conditional partitioning analysis based on the tree-modelling approach to provide an improved description of the variability in summer streamflow using simultaneous signals of spatial loading patterns of atmospheric circulation patterns.

3. Analysis and results

3.1. Conditional hydrologic information

A nonparametric regression analysis provides useful information for understanding an empirical relationship between teleconnection patterns and the Han River streamflow during the JJAS season. On the basis of the linear analysis of Kim et al. (2010b), we examined the locally linear estimate of the nonparametric regression for EA-WR index and streamflow (Figure 3(a)); departures from linearity are evident. Furthermore, nonparametric regression between EA-WR index and typhoon-related streamflow estimates for the JJAS season and non-typhoon streamflow is shown in Figure 3(b) and (c). The results confirm that observed correlation stems from the teleconnection index relationship with non-typhoon flow, and that typhoon events have no significant relationship with the EA-WR index. To further investigate and model the potential relationship between teleconnection patterns and the Han River streamflow, a tree-based classification methodology was used, motivated by our interest in investigating conditional distributions of streamflow based on teleconnection indices. Partitioning the predictor space into coherent sub-spaces is an attractive approach to model the joint nature of climate–hydrologic linkages for the Han River. Here, we considered a suite of teleconnection indices as predictors to explore optimal classification. The result of the tree-based model is summarized in Table I. The optimal tree and distribution of flow volumes based on the resulting hydrologic-teleconnection model are presented in Figure 4(a) and (b). To evaluate the performance of this model, we examined surrogate splits for each category. Predicated on the optimal split (number of split = 2), the surrogate split examines the potential predictive value of the remainder of teleconnection indices. This is achieved by reapplying the partitioning that best reproduces the optimal split (in this case, EA-WR and EA patterns), though using other independent teleconnection patterns. In Table II, surrogate split variables are summarized for the optimal split. For example, at node 1, one surrogate split (EA < 1) shows a high degree of concordance (85.0%) with the optimal split. The strongest surrogate variable is shown in two surrogate splits (POL < 0.78) at the second node. Their agreement with the optimal split is 81.2%. This procedure confirms the optimal split selected by this analysis. The partitions determined by the tree-based model approach show systematic conditional relationships between atmospheric teleconnections and seasonal streamflow (June–September). In Figure 5, we note that the highest flow volumes occurred frequently during years when EA-WR < − 0.455. On the other hand, lowest flow volumes corresponded to EA-WR ≥ − 0.455 in combination with a positive EA (≥0.320). The observations of negative EA-WR (<− 0.455) indicate that 5 out of the 10 highest flows occurred in this predictor sub-space. Both the EA-WR phase (≥− 0.455) and the EA (<0.320) account for 4 out of the 10 highest flows and 3 out of the 10 lowest flows. However, the EA-WR (≥− 0.455) and EA (≥0.320) correspond to 7 out of 10 of the lowest flows and 1 out of 10 of the highest flows. The conditional distribution of streamflow associated with both the EA-WR and EA patterns is summarized in Figure 6, highlighting the separation of conditional distributions based on the selected teleconnection indices. Based on the tree-modelling methodology, we pursued an independent confirmatory analysis to assess the regional precipitation correspond to the three groups of teleconnection index phases using GPCP precipitation dataset. We construct seasonal composite anomalies (Figure S2 in the Supplementary Information). The results confirm the conditional departures (high and low) as noted for streamflow.

Figure 3.

Nonparametric estimates of the relationship between seasonal flow volume and EA-WR pattern. (a) June–September seasonal flow volume versus EA-WR. (b) Non-typhoon flow volume versus EA-WR. (c) Estimated typhoon-induced flow volume versus EA-WR. Here, a four-day time window after entering date of typhoon path within the restricted domain (120E–138E, 32N–40N) was used to calculate the contribution of typhoon to seasonal streamflow. Solid lines show a smoothed regression corresponding to EA-WR. Dotted lines indicate the standard error of the estimate around the fit

Figure 4.

Change in seasonal flow volume based on teleconnection condition. (a) Tree-based model. (b) Distribution for seasonal volume flows based on dominant teleconnection pattern. For each group, the variability in flow volume is summarized using a boxplot (0.1, 0.25, 0.5, 0.75, 0.9 quantile levels are shown)

Figure 5.

Joint influence of the EA-WR and EA patterns. Partitions of the predictor space are determined by the tree-based model approach. Solid triangles represent the 10 highest (upward pointing triangles) and the ten lowest (downward pointing triangles) flows in Han-River (1966–2005)

Figure 6.

Empirical probability density functions. The lines indicate distribution for seasonal volume flows determined by the tree-based model approach

Table I. Cross-validated error versus Complexity Parameter (CP), for the flow volume data in the Han-River Basin, Korea. (The relative error gives the relative error rate for predictions for the data generated the model. The column headed x error presents the cross-validated error)
NO.CPNumber of splitRelative errorSum-of-squared errorsStandard error
10.25501.0001.0430.298
20.08910.7451.0210.315
30.08020.6561.0080.278
Table II. Surrogate splits for split nodes shown in Figure 4(a) and their agreement with the optimal split. (The split threshold shows variables that would go to the left side or right side of the split, with codes for variable categories in parentheses)
NodeSurrogate splitDirectionAgreement
Node 1 (n = 40)EA < 1to the left0.850
 SCA < − 0.930to the right0.850
 NAO < − 0.800to the right0.825
 EP-NP < − 0.715to the right0.825
Node 2 (n = 32)POL < 0.780to the right0.812
 EP-NO < − 0.585to the left0.750
 PNA < − 1.070to the left0.750
 EA-WR < − 0.380to the left0.750
 NAO < − 0.260to the left0.719

3.2. Hydrologic variability at sub-watershed scale

In the previous section, the tree-based classification approach provided the joint model based on two dominant teleconnection patterns (EA-WR, EA). The threshold values corresponding to the partitioned predictor space (Figure 4(a)) were used to investigate the sub-watershed scale sensitivity of streamflow to teleconnection indices. This analysis allows a spatially refined view of changes in the conditional streamflow distributions for individual sub-watershed. In Figure 7, at the sub-watershed scale, the mean (µ) and the standard deviation (σ) change in streamflow are compared by groups conditioned upon the dominant phases of EA and EA-WR. As noted in the previous section, empirically separated non-typhoon streamflow estimates are used for modelling and estimating conditional distribution parameters. In Group I, the seasonal flow volumes (dominated by negative phase of EA-WR pattern) show consistent increases in conditional streamflow distribution, with a 65% increase in the mean and a 5% increase in the standard deviation. However, Group III (EA-WR ≥ − 0.455 and EA ≥ 0.320) shows large decreases in streamflow, with 45% decrease in the conditional mean and 60% decrease in the conditional standard deviation. Overall, substantial decreases in seasonal flow volume distributions (changes, µ: 35–54%, σ: 32–77%) depending on sub-watersheds are evident from results presented in Figure 7. A modest change in the streamflow can be found in Group II (changes, µ: 3.22%, σ: 39.48%). As noted in Section 2.1, sub-watersheds scale is most appropriate scale for reservoir operations based upon conditional streamflow information.

Figure 7.

Distribution of seasonal flow volume (June–September) for 24 sub-basins in the Han-River Basin (1966–2005). (a) Percent changes in the mean and standard deviation of seasonal flow volume. Groups determined by the tree-based model approach are identified with distinct symbols. (b) Conditional distribution of seasonal flow volume for 24 sub-basins. Changes in the mean and standard deviation based on joint influence of the EA-WR and EA are reported as a percent of long-term mean and standard deviation. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

In the Han River Basin, there are three multipurpose dams (Soyang River dam, Chungju dam, Hwengsung dam) in operation for flood control during the wet season and water storage during the dry season, as well as hydroelectric power generation for energy supply. The total water supply by these dams is 4712.5 million m3/year. Three dams have a capacity of 1125.5 million m3/year for flood control and a generating capacity of 1202.6 GWh/year for energy supply (MLTM, 2010). Dam operations to maximize the multiple objectives rely upon seasonal streamflow, and thus have the potential to benefit significantly from seasonal forecasts, as well as conditional risk of extremes, that may be included in adaptive management efforts within this hydrosystem. On the basis of results of our tree-based modelling study, we computed conditional extremes for the sub-watersheds most relevant for these three dams. Here, we used the total streamflow volumes to allow a proper assessment of the impact of conditional climate information for reservoir management. We estimated the 50-year return period high flow and low flow fitted to lognormal distribution based on three groups of conditional distribution, which shows a simultaneous assessment of the joint effect of potential changes in the mean and variability of streamflow. For Group I, the Soyang River shows an increase of 47% in mean and decrease of 14% in standard deviation. The Hwengsung Dam Basin also shows similar results with an increase of 63% in mean and decrease of 32% in standard deviation. In both cases, a 50-year return period flow by lognormal distribution shows modest changes (0.1, 10.2%). However, the estimates of the Chungju dam show a comparatively large increase (26% departure from a 40-year mean) in the magnitude of high-flow events. While the change in mean (39%) is smaller than other dams (Soyang River, Hwengsung), the standard deviation shows high variability that indicates an increase of a 50-year return period flow. In the case of Group III, three multipurpose dams show decreases in mean (Soyang River Dam: 33%, Hwengsung Dam: 37%, Chungju Dam: 28%) and standard deviation (Soyang River Dam: 29%, Hwengsung Dam: 43%, Chungju Dam: 22%). The estimates show substantial decreases in the 50-year return period high flow (Soyang River Dam: 37%, Hwengsung Dam: 45%, Chungju Dam: 31%) and 50-year return low flow (Soyang River Dam: 28%, Hwengsung Dam: 22%, Chungju Dam: 23%) and a departure from the long-term mean based on the period of record. A detailed report summarising percent changes in mean and standard deviation for sub-watersheds in Han River Basin is presented in the Supplementary Information Section. As the commingling effects of changes in the mean and standard deviation has been noted, the results partitioned by the tree model clearly show important role of the interaction between EA-WR and EA teleconnection patterns. While the sensitivity of streamflow distributions to teleconnection patterns is evident from the tree-based modelling, it is important to emphasize that sample size imposes limits on the reliability of the conditional estimates—a common concern when working with historical data. Large sample studies based on atmospheric modelling simulations conditioned upon sea surface temperature conditions are likely the most appropriate approach to further confirm the relationship elucidated in this study.

The availability of conditional streamflow information has the potential in being an important input in to adaptive reservoir management efforts, wherein depending on the prevailing atmospheric circulation patterns, the risk for unusually high and low seasonal flow can be estimated on a yearly basis. Furthermore, given the competing demands on available water, selected updating of risk can potentially afford both economic and ecological benefits (Kim et al., 2006b).

4. Summary and conclusions

This study investigated hydrologic variability in the Han River Basin within the context of contemporaneous atmospheric circulation patterns. Correlation results show that negative EA-WR is strongly correlated with flow volumes (June–September) in the Han River Basin (correlation coefficient = − 0.40; p-value = 0.01). However, as correlation analysis is restricted to the signal of two spatial loading patterns at a time, we conducted conditional partitioning analysis based on the tree-modelling approach for multiple teleconnection patterns that exert influences on moisture fluxes and storm track variability in the Han River Basin.

As contrasted with the linear analyses (Kim et al., 2010b), the tree-based method effectively isolated interesting features of the interaction with very limited assumptions (Figure S2 and S3 in Supplementary Information). On the basis of the resulting hydrologic-teleconnection model, the high flows tend to increase during the negative EA-WR phase (EA-WR < − 0.445), and the low flows corresponded to a positive EA-WR in combination with EA pattern (<0.32). Furthermore, this conditional partitioning analysis provides an improved hydroclimatic description of the variability in summer streamflow at the regional scales, as well as the sub-watershed scales. Results of frequency analysis for sub-watersheds correspond to the three multipurpose dams show significant changes in the mean and variability of streamflow caused by commingling effects of teleconnection patterns. The specificity of conditional hydroclimatic information at sub-watershed scales is particularly suitable for the increasingly complex reservoir operations in the Han River region. The Han River is critical source of drinking water, industrial water supply, irrigation, hydropower generation, as well as ecosystem services. An improved understanding of warm season hydroclimatic variability in the Han River Basin is central to efforts that seek to anticipate and adapt water resources management in a variable and changing climate. This diagnostic study is a critical first step in developing seasonal streamflow estimates conditioned upon large-scale climatic state, particularly for multi-objective reservoir and watershed management.

Ancillary

Advertisement