Hydrological response to future climate change in the Agano River basin, Japan

To evaluate the impact of climate change on snowfall in Japan, a hydrological simulation was made in the Agano River basin by using a regional climate model’s output. A hindcast experiment was carried out for the two decades from 1980 to 1999. The average correlation coefficient of 0.79 for the monthly mean discharge in the winter season showed that the interannual variation of the river discharge could be reproduced and that the method can be used for climate change study. The future hydrological response to global warming in the 2070s was investigated using a pseudo-global-warming method. In comparison to data from the 1990s, the monthly mean discharge for the 2070s was projected to increase by approximately 43% in January and 55% in February, but to decrease by approximately 38% in April and 32% in May. The flood peak in the hydrograph was moved forward by approximately one month, changing from April in the 1990s to March in the 2070s. Furthermore, the projection for the 10-year average snowfall amount was projected to be approximately 49.5% lower in the 2070s than in the 1990s.


INTRODUCTION
Climate change is having a conspicuous effect on Japan.In particular, the Japan Meteorological Agency (2002) has reported that snowfall amounts have fallen sharply along Japan's eastern seaboard since the mid-1980s, resulting in a noticeable decrease in river discharge in this region in spring.Table I lists decadally averaged April discharge rates (m 3 /s) for the main rivers entering the Sea of Japan, with differences between the earlier decades and the 1990s given as percentages of the 1990s values.There are no negative change values in the table, indicating that all of the rivers have experienced a decrease in their mean April discharge rate.Indeed, April discharge rates in the 1990s were 10% to 36% smaller than those in the 1970s and 1980s, which also provides supporting evidence for climate change in this region.
To clarify the possible effects of climate change on water resources, quantitative analysis is required.Such analysis requires information on changes in river discharge determined from both observation and simulation.Hydrological simulations of rivers generally use data obtained from ground-based meteorological stations.The distribution and density of meteorological stations affect the accuracy of such hydrological modeling, especially for large river basins.Japan has approximately 110 manned meteorological stations, which measure wind speed, precipitation amount, air temperature, humidity, atmospheric pressure, and other variables.Nevertheless, this density does not meet the recommendations of the World Meteorological Organization (WMO, 1994).In addition, because Japan is a mountainous country, most stations are located in valley bottoms or other flat areas.
In recent years, a downscaling method that links atmospheric and hydrological models has been developed for hydrological simulations.Kite and Haberlandt (1999) tested hydrological simulations of the Mackenzie and upper Columbia rivers and showed that the coupling of atmospheric and hydrological models was useful in understanding the macro-scale hydrological cycle.Many other authors have also used the downscaling method to investigate changes in river hydrology (e.g.Wood et al., 2004).Fujihara et al. (2006) examined the influence of global warming on the water resources of the Tone River basin using the Table I.Mean decadal April discharge rates (m 3 /s) in the 1990s for the main rivers flowing into the Sea of Japan in the Hokuriku and Tohoku regions and the differences between 1990 and two earlier decades, given as percentages of 1990s values.
* 1 1970, 1994 and 1997  conducted a flood risk assessment of global warming using a regional climate model and high resolution GCM over Japan and indicated that the risk will increase in most areas.
As mentioned by Tachikawa et al. (2009), the bias must be removed from the GCM output data in order to do hydrological modeling because the statistical properties of the GCM output data do not necessarily match up to that of observed data.There is no guarantee that obtained technique, a bias correction technique from the current climate experimental data and observed data, is applicable to the future data.In this study, we examined the projections of a regional climate model used to investigate long-term hydrological responses using a hydrological model.We used the Weather Research and Forecasting (WRF) model with dynamic downscaling to simulate input variables for the Soil-Vegetation-Atmosphere Transfer and Hydrological Cycle (SVAT&HYCY) model.The study site was the Agano River basin, an area in Japan that receives heavy snowfall, located in the Hokuriku region.Model runs over a 20-year period were conducted, and the output hydrographs were compared with observational data.We then examined the effects of global warming on the hydrological processes of the Agano River basin using the pseudo-global-warming method.

STUDY AREA
The Agano River (Figure 1) drains into the Sea of Japan and is the second largest river in Japan with respect to annual discharge (12.9 billion m 3 , http://www.hrr.mlit.go.jp/agano/).The Agano River drains an area of 7,710 km 2 making it the eighth largest river basin in Japan.The two major tributaries of the Agano, the Tadami and Aga rivers, are both located in Fukushima Prefecture.The majority of the Agano River discharge is caused by the Tadami River, which drains an area receiving heavy winter snowfall.In its lower reaches, the Agano River flows through the Echigo Plain in Niigata Prefecture.The Echigo Plain is a major rice production area, ranking second in Japan in total rice output, and water is an important factor in determining the quantity and quality of the rice.In addition, there are many hydroelectric dams because of the abundant water resources of the basin.Therefore, the projection of future climate change in the Agano River is significant for the region's economy and society.

MODELS AND SETTING
To examine river discharge, we used the WRF regional climate model and the SVAT&HYCY model.

The WRF Model
The WRF model is a mesoscale numerical weather projection system designed to serve both operational forecasting and atmospheric research needs.The model uses full-compressible, non-hydrostatic equations.We used the Advanced Research WRF (ARW) core version 2.2 (Skamarock et al., 2005) with a two-way nesting technique and the WRF single-moment 6-class microphysics scheme.The parent domain was set to a wide area, between 28-46°N and 124-148°E on a 20-km grid.The inner domain was located in a smaller area between 35.5-39.3°Nand 135.7-141.5°Ewith a 5-km grid (Figure 2a).
We conducted two numerical experiments.One was the hindcast run (CTL) used to reproduce past hydrological events of the 1980s and 1990s.The other was a pseudoglobal-warming run (PGW) used to project the hydrological response in the 2070s.The National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR) 6-hourly reanalysis data (Kalnay et al., 1996) was used as the lateral boundary condition for the CTL.For the PGW, the lateral boundary condition was  adopted, following the method of Hara et al. (2008).A global warming component was added to the NCEP/NCAR reanalysis data for the 1990s.Global warming components were estimated as the monthly average difference between the 10-year average of the 21st century projection, based on scenario A2 of the Special Report on Emissions Scenarios (SRES-A2) (Nozawa et al., 2007), from 2071 to 2080, and the 20th century simulation from 1991 to 2000, from version 3.2 of the Model for Interdisciplinary Research on Climate (MIROC) (medres, approximately 300 km, T42), an atmosphere-ocean coupled general circulation model.The PGW method allows for the comparison of climate in the present year and that in a PGW year that is similar to the control year in terms of interannual variation while including future climatology.Therefore, by the PGW method, we could evaluate the river discharge under a future climate.
The output variables from the WRF model were (1) downward longwave flux at the ground surface, (2) downward shortwave flux at the ground surface, (3) surface air pressure, (4) water vapor mixing ratio at 2 m height, (5) accumulated total grid-scale precipitation, (6) surface air temperature (SAT), and ( 7) wind speed at 10 m height.

The SVAT&HYCY model
The SVAT&HYCY model (Ma et al., 2000) is a onedimensional hydrological model that includes a soilvegetation-atmosphere transfer scheme, runoff formation, and river routing.The model has a number of features applicable to understanding the water-energy cycle in a river basin.In this study, only the forest vegetation type was used.The precipitation was assumed to be snowfall when the SAT was less than 2°C according to the result of Ma et al. (1999).Evapotranspiration, snowmelt, surface runoff, infiltration, and base flow were simulated by the SVAT&HYCY model using the WRF output data listed above.The river network was constructed using the GTOPO30 dataset (http:// edc.usgs.gov/products/elevation/gtopo30/gtopo30.html) with 0.05° resolution (Figure 2b).The estimated basin area was 7,834 km 2 , which is close to the reported area of 7,710 km 2 .The velocity of the river flow was set at 0.7 m/s, in consideration of the mountainous topography.
The amount of discharge measured at Maoroshi (6,997 km 2 ) until 2003 is available from the Japan River Association.To confirm the performance of the hydrological model, we ran it for a 20-year experiment period from July 1979 to June 1999.The WRF model was run per year, from June 28 of this year to July 1 of the next year.We did not consider the effect of dams in this study because many dams along the river are used to operate with high water levels for hydroelectric power generation.We only focused on the relative evaluation of interannual variation, which is greatly influenced by climate factors in the monthly mean discharge.

RESULTS AND DISCUSSION
To avoid snow pack, the "water year" was set from July of the previous year to June of the current year.The simulated amount of precipitation was compared with 10-station data in the 20-year period.The result showed that the annual mean correlation coefficient and bias were 0.63 and 592 mm in overestimation (with a standard deviation of 663 mm) (Table S1 and Figure S1 in Supplement), respectively.Figure 3 shows the annual runoff from 1980 to 1999.Simulated annual runoff roughly corresponded to observed values throughout the two decades with the exception of 1996.The simulated interannual variation of annual mean runoff correlated well with the observed data, though the estimated precipitation was overestimated compared to some of the station data.The time series correlation was significant, with a significant level of 95% (0.62 in correlation coefficient).
Figure 4 shows the monthly mean discharge in winter from November to May.Although there was no significant difference between the simulation and observation, it is more important to examine the correlation rather than the comparison in absolute values because of human impact.For example, the water controlled by dams was not considered in this study.The correlation coefficient for the monthly mean discharge in average was 0.79 with a range from 0.66 in December to 0.84 in February.The correlation coefficient was 0.79, 0.75 and 0.83 in March, April and May, respectively.The average of these time series did not have a significant difference, with a significance level of 5%.It suggests that interannual variation is greatly influenced by the climate factor.The calculated monthly mean discharge in April in the 1990s decreased 16.6% compared with that of the 1980s, which was consistent with  the data and approximately 5% smaller than that shown in Table I.
In addition, The Nash-Sutcliffe efficiency coefficient (NSC) (Nash and Sutcliffe, 1970) was adopted to evaluate the reproducibility of the hydrological model.The NSC values for the monthly mean discharge were 0.474 for the 1980s and 0.476 for the 1990s.As Figure 4 demonstrated with the high NSC, the seasonal variation of river runoffs during the period is well reproduced in the hydrological model without dam operations, confirming that the effect of the dam operation is minor during the simulated season.Therefore, it is considered that the model is appropriate for studying climate change impact, especially during the winter season.
The monthly mean discharge in the 1990s and that under the PGW in the 2070s are shown in Figure 5.The figure presents a clear increase in discharge from 340 to 486 m 3 /s (43%) in January and from 365 to 565 m 3 /s (55%) in February.However, discharge decreases are shown in both April (from 674 to 417 m 3 /s, 38%) and May (from 427 to 292 m 3 /s, 32%).In the 2070s, the flood peak is projected to occur one month earlier, in March rather than April.There was little change between July and November, and in general, the annual mean changes of runoff between the 1990s and 2070s runs were small.This suggests that the total precipitation change between the 1990s and 2070s runs was small.Under global warming, the discharge change is mainly caused by the phase change of precipitation (Hara et al., 2008) and earlier snowmelt due to increased SAT.The characteristics of hydrological response for future climate change obtained in this study are consistent with the result of Tachikawa et al. (2009) for the Mogami River basin.
The 10-year average annual precipitation of 2,243 mm in the 1990s run increased by approximately 80 mm (3.6%) to 2,323 mm in the 2070s run.However, the amount of snowfall showed a marked decrease.The 10-year average annual amount of snowfall in the 1990s run was 507 mm, which decreased to 251 mm in the 2070s run, representing a basin-wide reduction of 49.5%.Figure 6 shows the distribution of snowfall decrease across the basin.The snowfall decreased by more than 20% (ranging from 23% to 95%).Particularly large snowfall reductions of more than 80% are shown in the downstream region, mostly because the PGW model projected no snow there.The changes of all grids over the basin were significant by t-test with the significance level of 5%.
The PGW components by AOGCM that we used in this study are obtained from MIROC (Nozawa et al., 2007).The future change of the annual precipitation around Japan is around zero by not only MIROC and but also the ensemble mean of the CMIP3 GCMs (Chapter 11 of IPCC, 2007).The difference of the 10-year-averaged global mean SAT of CMIP3 GCMs between 1990s and 2070s is 2.19°C in the ensemble mean (with a standard deviation of 0.43°C), and 2.61°C in MIROC, respectively.The difference of SAT around Japan (130E-140E, 30N-45N) between the 1990s and the 2070s is 2.53°C in the ensemble mean (with a standard deviation of 0.70°C), and 3.36°C in MIROC, respectively.The future global change of SAT in the MIROC is larger than the ensemble mean (Figure S2 in Supplement).Therefore, this projection reflects a climate condition under the relatively high SAT increase in the 2070s between the CMIP3 GCMs.The ratio of snowfall in total precipitation can be greater and the snowmelt season could come later when we use the multi-model ensemble mean as a global warming component.

CONCLUSION
To understand the impact of climate change on hydrological processes in snowy areas of Japan, a hydrological model was applied to the Agano River basin by using dynamic downscaling data.The model performance was checked using a 20-year hindcast for the period 1980 to 1999.A high correlation coefficient of 0.79 on average was obtained for the monthly mean discharge in the winter season and indicated that the method is suitable for future climate change studies.
The global warming numerical experiment suggested that the monthly mean hydrograph for the 2070s will differ from that observed in the 1990s.The monthly mean discharge was projected to increase by approximately 43% in January and 55% in February, compared to the respective  values for the 1990s.On the other hand, discharge will decrease by approximately 38% in April and 32% in May, respectively.Few changes in discharge were found for the period from July to December and in March.The 2070s flood peak was projected to occur one month earlier, in March, compared to the April peak in the 1990s.It seems that the impact of rising temperatures due to the change in precipitation is limited, considering the lack of major changes in river discharge characteristics.
The PGW results also showed a marked change in snowfall.The PGW snowfall amounts were 20% lower than those in the CTL-1990s, with a basin-wide average reduction of approximately 49.5% and a range from 23 to 95%.

Figure 2 .
Figure 2. Domain of the WRF model (a): ① center of the parent domain (20-km grid); ② center of the inner domain (5-km grid).River network of the Agano River basin (b).

Figure 1 .
Figure 1.Location of the Agano River basin.

Figure 4 .
Figure 4. Comparison of 10-year average monthly mean discharge between the simulation (CTL) and observations (OBS) in the 1980s and 1990s at Maoroshi.

Figure 3 .
Figure 3.Comparison of annual runoff between the simulation (CTL) and observations (OBS) at Maoroshi.

Figure 5 .
Figure 5. Same as in Figure 4 but for simulations in the 1990s (CTL) and the 2070s (PGW).

Figure 6 .
Figure 6.Distribution of the decrease ratio (percentage) of the 10-year average snowfall amount in the PGW run compared with the CTL run over the basin.