Impact assessment of multiple uncertainty sources on high flows under climate change

Extreme hydrological events like floods are one of the most severe natural disasters which have caused great economic losses and casualties in many places in the world (Kundzewicz et al. 2010; Sena et al. 2012a). Floods are closely related to the climate system and are usually very localized and sensitive to even modest changes of climate equivalent or smaller than changes expected from potential global warming (Knox 2000). Previous studies show that the frequency and intensity of the extreme climatic and hydrological events have been increasing and are likely to continue to increase due to global warming (Easterling et al. 2000; Bell et al. 2004). In China, over the past 40 years about 50% of summer precipitation in the lower and middle Yangtze River is in the form of rainstorms which brings an increasing risk of summer flood (Jiang et al. 2007). Similarly, in the Pearl River basin extreme flood events have significantly increased since 1980 (Wu et al. 2013). In coastal regions like Zhejiang Province, where the study basin is located, climate change has exerted impacts on the hydrologic cycle and affected river runoff. Typhoons usually bring much precipitation and cause floods almost every year, and there has been a significant increasing tendency in the past five decades (Tian et al. 2012; Xu et al. 2012, 2013).

Many studies on the impact of climate change on regional water resources focused on the long term average flows, but it is acknowledged that climate change may severely affect extreme events in hydrology on a regional scale (Sena et al. 2012b; Zakaria et al. 2012). Extreme events are of most concern for water management and hazard prevention (Qi et al. 2013). However, the ability to predict extreme events is much less compared to average flow events (Hirsch 2011; Tian et al. 2013). There is less confidence about changes in extremes than that in average flows (Alliance 2009; Tian et al. 2014). Besides, the uncertainty of future extreme events may increase with future climate variability and change, which requires us to develop a broader understanding of the range of possible hydrologic futures that we may face (Rolf et al. 2010).

In impact studies of climate change on hydrology, commonly there are four steps involved (Horton et al. 2006; Wilby & Harris 2006; Camici et al. 2014). First, the scenarios are set up to assist in understanding the possible future developments of the climate. Secondly, the global climate models (GCMs) are applied to project the future climate under different climate scenarios. Then, because of the mismatch between the resolution of GCMs and hydrologic models, downscaling of GCM projections is implemented to refine the GCM data so that the resolution could be compatible with the hydrologic models. Lastly, the downscaled GCM outputs are used as input to drive the hydrologic models to simulate for instance river discharge at regional or local scales.

However, the impact assessment of future climate change is subject to a series of uncertainties. Owing to limited knowledge of the reality and unpredictable future, uncertainties are inevitable in every step of the modeling process. There are many different classifications of uncertainty sources for various research purposes (Van der Keur et al. 2008; Warmink et al. 2011). In terms of the location, level and nature of uncertainties, Walker et al. (2003) defined several uncertainties that should be taken into account in uncertainty analysis, including model context, input, model structure and parameters. Xu et al. (2005) discussed the challenges and problems in studying hydrological consequences of climate change and pointed out that there are large uncertainties in climate scenarios, GCM output, downscaling methods and hydrologic models. Previous studies have analyzed the uncertainty from GCMs, scenarios, hydrological model input and hydrological models (Wilby & Harris 2006; Chen et al. 2013b), however, only limited studies on the uncertainty of downscaling methods compared to other uncertainty sources could be found. In this study, we did not aim to investigate the uncertainty caused by different downscaling methods. This has been done by several studies (Wilby et al. 1998; Chen et al. 2013a; Tofiq & Guven 2014). This study focused on the uncertainty of a statistical downscaling method, and made a comparative analysis to other uncertainty sources including GCMs, scenarios and hydrological models. The assessment of uncertainty along the whole modeling process is crucial because it provides more informative projections of climate change impacts on the river discharge and it is fundamental to improve the existing modeling techniques. Besides, it provides scientific support for policymakers in water resources management to make a robust decision under climate change.

Climate scenarios are possible images of the future. They are developed to describe how the future might be with respect to socio-economy, technology and emission of greenhouse gases. In the 1992 Supplementary Report to the IPCC Assessment, six alternative IPCC scenarios (IS92) were published which embodied a series of assumptions that have impacts on future greenhouse gas emissions without considering climate policies (Houghton et al. 1992). Based on IS92, in 2000 a set of IPCC scenarios which estimated the uncertainty of greenhouse gas emissions was published in the Special Report on Emission Scenarios (SRES) (Nakicenovic et al. 2000). Similarly as IS92, the SRES scenarios assumed no policy actions in climate change mitigation and adaptation either. Recently, the SRES scenarios were updated and expanded for the purposes that firstly more detailed information is needed in running the current generation of climate models, and secondly there is an increasing interest in exploring the impact of different climate policies (Van Vuuren et al. 2011a). The newly developed scenarios are called representative concentration pathways (RCPs), which have provided a new basis for climate modeling. Up to now, there have been some studies on the assessment of simulation results for the different RCPs from the ensemble of Coupled Model Inter-comparison Phase Five (CMIP5), and comparison between CMIP5 and CMIP3 (Moss et al. 2010; Knutti & Sedláček 2012). However, only few studies on extreme hydrology including quantification of all uncertainty sources under the new RCPs have been carried out.

This study investigates the uncertainties in extreme flows originating from RCPs, GCMs, downscaling method and hydrologic models. The major objective is to assess the impact of climate change on extreme discharges and to quantify the uncertainty of extreme flows stemming from different sources. The paper is organized as follows. First, the study area and the observed data are presented in the following section. Subsequently, the scenarios, GCMs, downscaling method and hydrologic models applied in this study are described. Then, the results are presented followed by the discussion section. Conclusions are drawn are in the final section.

Jinhua River has a drainage area of approximately 6,000 km². The Jinhua River basin is a subbasin of Qiantang River basin which is the largest and most important water resource for Zhejiang Province in eastern China. Two major tributaries, Dongyang River and Wuyi River, form the Jinhua River. The Jinhua River basin has a typical sub-monsoon climate with large inter-seasonal variability in temperature and precipitation. The annual mean precipitation from 1953 to 2008 was 1,386 mm. The annual maximum discharge at Jinhua station varied between 600 and 4,900 m³/s in the period 1961–1990. Almost every year floods occur in summer when typhoons bring much precipitation, and floods often cause great property loss and casualties (Tian et al. 2012).

The latest emission scenarios called Representative Concentration Pathways were developed to update and expand the scope of the old scenarios (Moss et al. 2010; Van Vuuren et al. 2011b). The four RCPs (RCP2.6, RCP4.5, RCP6.0 and RCP8.5) contain information on emissions, concentrations and radioactive forcing, which are used as input for climate model experiments. RCP2.6 represents a mitigation scenario leading to a low forcing level. RCP4.5 and RCP6.5 represent two medium stabilization scenarios. RCP 8.5 represents a high baseline emission scenario (Van Vuuren et al. 2011a). These four RCPs cover the full range of radioactive forcing values from 2.6 to 8.5 W/m² and greenhouse gas emissions in the literature.

The general circulation model is widely used to simulate climate changes as input for impact assessment studies (Shackley et al. 1998; Murphy et al. 2004; Lobell et al. 2008). In this study, the GCM outputs are derived from a multi-model data set of the World Climate Research Programme's Coupled CMIP5. CMIP5 has been carried out by a broad climate research community and aimed to provide a multi-model context for the assessment of the uncertainties of emission scenarios, model structures and initial conditions. It involves simulations from more than 20 modeling groups using more than 50 models. To make a comparison and analyze the uncertainties of different GCMs, three GCMs are selected. These three GCMs are different in many ways like grids division, boundary conditions and atmospheric compositions. The information of the three GCMs BCC-CSM 1.1 (Wu et al. 2008; Wu 2012), HadGEM2-ES (Jones et al. 2011) and GISS-E2-R applied in this study is given in Table 2.

LARS-WG is a stochastic weather generator which is commonly used for generating long time series of weather data and producing local daily climate change scenarios for a single site (Lazzarotto et al. 2010; Xu et al. 2012; Calanca & Semenov 2013). It has been tested by many researchers and shows generally a good performance in representing extreme events (Qian et al. 2008; Semenov 2008; Roshan et al. 2013).

Firstly, the statistical features of the observed weather data (in this study daily precipitation, maximum temperature and minimum temperature) are determined. Secondly, the monthly changes in equivalent variables from combinations of the four RCPs and three GCMs were calculated. At last, the synthetic weather data with the same statistical characteristics are generated on a day-to-day basis. The simulation of precipitation occurrence in LARS-WG utilizes a semi-empirical distribution for the length of wet and dry day series and daily precipitation. For a wet day, the value of the precipitation is generated from a semi-empirical distribution of precipitation for the particular month independent of the length of the wet series and the amount of precipitation on a previous day. Daily maximum and minimum temperature are regarded as stochastic processes with daily means and standard deviation conditioned on the wet or dry days (Semenov & Barrow 2002).

Ten sets of 30-year daily meteorological series are produced by varying random seeds which are predefined in LARS-WG. Random seeds control the stochastic component of LARS-WG and make it possible to generate a number of various realizations of the weather time series. These series are used as the input for the hydrological models to study the uncertainty of the downscaling method.

Three hydrological models are employed to study the uncertainty of model structure in this paper: GR4J, HBV and Xinanjiang. Xinanjiang model with 15 parameters is the most complicated model among the three, and GR4J model with four parameters is the simplest model among the three. The inputs of these models are precipitation and potential evapotranspiration. Potential evapotranspiration is estimated with the Hargreaves equation which uses daily maximum and minimum temperature as inputs (Hargreaves & Samani 1982). These hydrological models are widely used and they were developed based on different interpretations of the rainfall–runoff processes. The detailed description of the models is listed in the following.

The GR4J model with four parameters is a widely used conceptual model and was developed by the French institute CEMAGREF (Perrin et al. 2003). The GR4J model consists of four parameters: the maximum capacity of the production store (X1), the ground water exchange coefficient (X2), the 1 day ahead capacity of the routing store (X3) and the time base of the unit hydrograph (X4). The input of the GR4J model is daily precipitation and potential evapotranspiration, and the output of the model is daily discharge.

The HBV model was developed by the Swedish Meteorological and Hydrological Institute (Bergström 1992). There are eight parameters in the HBV model. The model is composed of a precipitation and snow accumulation routine, a soil moisture routine, a quick runoff routine, a baseflow routine and a transformation function. The inputs of the HBV model are daily precipitation and potential evapotranspiration, and the output is daily discharge. The maximum soil moisture storage (FC) controls the amount of water contributing to the runoff generating routing. The capillary flow from the upper response box to the zone of soil water is described by parameter CFLUX.

The Xinanjiang model was developed by Zhao (1992) for humid and semi-humid regions. There are 15 parameters in the Xinanjiang model. It consists of an evapotranspiration component, a runoff generation component, a runoff production component and a flow routing component. The inputs of the Xinanjiang model are daily precipitation and potential evapotranspiration, and the output is daily discharge.

The uncertainties from RCPs, GCMs, downscaling method and hydrologic models are addressed in this study. The extreme high flows are represented by annual maximum daily discharges. To compare the uncertainty range of each source, the uncertainties in simulated high flows from each source are assessed individually in this study. The uncertainty ranges of RCPs are estimated given the average value of design flows for GCMs, hydrologic models and downscaling. In a similar way, the uncertainty ranges are estimated for GCMs, hydrological models and downscaling methods. The design flows with 90% confidence intervals caused by different sources for 10-, 20- and 30-year return periods are estimated to analyze the uncertainty from different sources.

Uncertainties in precipitation changes due to different GCMs and climate scenarios not only present different magnitudes of changes, but also different directions of change. For BCC-CSM 1.1, there is an increase for most of the time from May to September, and in October there is a decrease. Under RCP2.6, RCP4.5 and RCP6.0, the increase of annual mean precipitation is 12 mm, 4 mm and 3 mm, respectively. Most of the largest increase for each month is detected under RCP2.6. For HadGEM2-ES, there is a decrease in June, July, October and November and an increase in August and September. Opposite to the annual mean precipitation for BCC-CSM 1.1, the annual mean precipitation decreases under four RCPs and the largest decrease occurs in RCP4.5 with a 13 mm drop. For GISS-E2-R, there is no consistent increase for the four RCPs in each month, and in March, April, August and October the precipitation decreased under all RCPs. The highest increase happens in September under RCP8.5 and for RCP4.5 while the largest decreases occur for 6 months out of a year.

Figure 7 shows observed extreme high flows in the baseline period and the extreme high flows under the four RCPs in future. The annual maximum discharges are fitted to the GEV distribution and the parameters in the GEV distribution are estimated by the Maximum Likelihood Estimation method. Compared to the baseline period, there is a decrease for design high flows under four RCPs. The differences of high flows between RCP 2.6, RCP4.5, RCP 6.0 and RCP8.5 are not significant, indicating a small uncertainty range from RCPs.

Figure 9 shows the extreme high flows for observations in the baseline period and the average extreme high flows for the three GCMs in the future. The future extreme high flows would decrease for the three GCMs compared to the baseline. Generally, the BCC-CSM 1.1 model simulates larger high flows than HadGEM2-ES and GISS. The GISS model predicts the smallest high flows. The uncertainty range of extreme high flows is the largest for HadGEM2-ES followed by BCC-CSM 1.1, and GISS-E2-R.

In this study, four emission scenarios (RCP2.6, RCP4.5, RCP 6.0 and RCP 8.5) were applied to estimate high flows in the near future. The results showed that the uncertainty ranges of high flows from GCMs were larger than those from emission scenarios. Precipitation and temperature were two major inputs to drive the hydrologic models. The average magnitude of the uncertainties of precipitation, maximum and minimum temperature from GCMs and emission scenarios was calculated and compared in Table 4. The bold numbers indicated the larger number between GCMs and emission scenarios. The uncertainties of precipitation, maximum and minimum temperature from the emission scenarios were smaller than those from GCMs in most months, especially for precipitation in spring and summer when heavy rainfall usually occurred and caused floods. These uncertainties in precipitation and maximum and minimum temperature were propagated into the next step of the modeling process, i.e. simulating the discharges. Hence a larger uncertainty of discharges from GCMs was detected than that from emission scenarios.

Model uncertainty was mainly from incomplete and biased knowledge of the rainfall–runoff processes. Studies had been carried out on comparing the uncertainty from different sources in simulating discharges, for instance input error, parameter uncertainty and model structures (Walker et al. 2003; Refsgaard et al. 2006). Results suggested that there were important benefits in exploring different model structures as part of the overall modeling approach because the model performance was strongly dependent on model structure and different model structures could capture different aspects of the catchment response (Butts et al. 2004). In this paper, the GR4J model and Xinanjiang model were more consistent and have a smaller range of uncertainty. The HBV model had the largest uncertainty range and was likely to be more dependent on the input data. Uncertainty ranges of the high flows due to different sources were compared and the results showed that the uncertainty from hydrologic models were the largest compared to GCMs, emission scenarios and the downscaling method, which also confirmed the point that the uncertainty from the hydrological model structure was important and should not be ignored.

In this study, a limited number of GCMs and hydrologic models were applied. It was possible that if more GCMs and hydrologic models were used, the uncertainty ranges from both might change. Besides, it should be noted that the small uncertainty range from emission scenarios was obtained based on the near future period. It was consistent with the fact that the differences of the RCP emission scenarios were small in the beginning and became larger in the second half the 21st century. In Prudhomme's (2003) study, seven GCMs and four scenarios were applied to estimate the floods in UK catchments. Results showed that there was more uncertainty due to the different GCMs than to the emission scenarios, which was in agreement with the result in this paper. In Bae's (2011) study, 13 GCMs, three emission scenarios and three hydrological models were applied to assess water resources in Korea. The results indicated that GCM structures were the main source of uncertainty for summer and autumn in their study area. Although different results could be obtained depending on the methods used and study areas, it should be noted that there were some consistent results stating that the uncertainties from GCMs were large. In this paper, not only the uncertainties from GCMs were noticeable but also the hydrologic models were found with large uncertainties. It was crucial to improve the performance of GCMs and hydrological models to reduce the uncertainty in precipitation and river discharge simulation, particularly high flow simulation, in climate change impact studies.

It was widely acknowledged that the projection of future high flows was important for risk assessment and flood prevention. In this case, the Jinhua River basin was located in the upper reaches of the Qiantang River basin and there was relatively less effect from human activities compared to lower reaches. However, it was also crucial to consider the impact of human activities on the changes of river flows. There were many studies on separating the effects of climate variability and human activities (e.g. Zhang et al. 2012, 2013, 2014). The assessment of the contribution of each factor to changes in high flows would be investigated in future study.

In the meantime, considering the large uncertainty existing in climate change impact analysis, quantification of uncertainties from different sources was important to identify the major uncertainty sources and analyze the possible impacts of each uncertainty source on future discharges. To be supportive for policy makers to make robust and resilient decisions under climate change, it would be more reliable and wholesome if more aspects about uncertainty were involved and we would like to explore it in the next step.

The uncertainty assessment of high flows due to different uncertainty sources under climate change has been carried out for the Jinhua River basin, eastern China. The climate output from three GCMs under four emission scenarios were downscaled by a statistical downscaling method (LARS-WG) with 10 sets of random seeds. Then, the downscaled outputs were used to drive three lumped hydrologic models. Changes of monthly precipitation, maximum and minimum temperature were compared between the baseline (1961–1990) and future periods (2011–2040). The daily discharges were estimated and extreme high flows represented by T-year flows were calculated for the same periods. The uncertainty ranges of high flows were compared systematically among different sources, including GCMs, RCPs, hydrological models and parameter sets in downscaling.

The results indicated that the uncertainties in precipitation were present in terms of not only the different ranges of changes, but also the directions of change. For the minimum temperature, increases all year round in the future period were found and for the maximum temperature increases were estimated in most cases except for the GISS-E2-R model under RCP 6.0 from April to July. Annual maximum discharges were expected to decrease under four emission scenarios in the near future compared to the baseline period. The magnitudes of decrease of high flows were directly related to the choice of different hydrological models and the RCPs.

Among different uncertainty sources, the analysis showed that hydrologic models and GCMs presented larger uncertainty ranges than parameter sets in downscaling methods and emission scenarios. There was a larger uncertainty range in high flows from the HadGEM2-ES model than that from GISS-E2-R and BCC-CSM 1.1. For hydrologic models, the uncertainty ranges of high flows simulated by the HBV model were larger than those by the GR4J and Xinanjiang models. The changes of average annual maximum discharges were inconsistent, but considering the uncertainty ranges, both the magnitudes and directions of changes would be uncertain. However, the results show that there would be a larger possibility that extreme high flows would decrease, which could indicate a release of pressure on flood prevention. For emission scenarios, RCP8.5 showed the smallest uncertainty compared to other emission scenarios.

This study was partially supported by the National Natural Science Foundation of China (51379183), Zhejiang Provincial Natural Science Foundation of China (LR14E090001), International Science & Technology Cooperation Program of China (2010DFA24320), Zhejiang Province Postdoc Foundation (BSH1401024) and China Postdoc Foundation (2014M561760). The National Climate Center of China and Bureau of Hydrology, Zhejiang Province are greatly acknowledged for providing weather and flow data in Qiantang River basin.