Multi-model ensemble projection of mean and extreme streamflow of Brahmaputra River Basin under the impact of climate change

The streamflow of Brahmaputra River Basin is vital for sustainable socioeconomic development of the Ganges delta. Frequent floods and droughts in the past decades indicate the susceptibility of the region to climate variability. Although there are multiple studies investigating the basin’s future water availability, most of those are based on limited climate change scenarios despite the wide range of uncertainties in different climate model projections. This study aims to provide a better estimation of projected future streamflow for a combination of 18 climate change scenarios. We develop a hydrologic model of the basin and simulate the future water availability based on these climate change scenarios. Our results show that the simulated mean annual, mean seasonal and annual maximum streamflow of the basin is expected to increase in future. By the end of the 21st century, the projected increase in mean annual, mean dry season, mean wet season, and annual maximum streamflow is about 25, 178, 11, and 22%, respectively. We also demonstrate that this projected streamflow can be expressed as a multivariate linear regression of projected changes in temperature and precipitation in the basin and would be very useful for policy makers to make informed decision regarding climate change adaptation.

. However, water shortage is prominent in the northwest part of Bangladesh because of low water availability in the dry seasons. These extreme variations in water scarcity and abundance that the basin has been experiencing in the past will raise concerns about the basin's resiliency under the future potential climate changes (Gain et al. ). Climate change will potentially increase intense rainfall in this region leading to increased flooding phenomenon (UNFCCC ). Moreover, snowpack in the mountains is susceptible to warming temperatures and may cause a shift in snow melting time.
Several studies have been conducted on the hydroclimatological impacts on the streamflow of BRB (e.g.  Seidel et al. () and Alam et al. () indicate that there will be about a 1.37% increase in mean annual streamflow for a 1% precipitation increase (in contrast to a 1.37% decrease for a 1 C temperature increase), where flood peaks are relatively more sensitive to climate change (30% increase in monsoon flood peak for a 10% increase in precipitation and a 1.5 C temperature increase). Similarly, Mohammed et al. (a, b) assessed the impacts of 1.5 and 2 C global warming on the extreme flows and water availability and found frequent flood flows and less frequent low flow events. Other than synthetic climate change studies, there are scenario-based studies conducted on BRB. Ghosh & Dutta () used downscaled climate data for the A2 SRES scenario from PRECIS (Providing Regional Climates for Impacts Studies) and concluded that the pre-monsoonal peak discharge will increase by 12% at The streamflow projections obtained from such an analysis provide an improved estimate of future changes expected in BRB.

RESEARCH APPROACH
Our methodology involves three key steps: (1) identifying representative climate scenarios, (2) hydrologic modeling for future streamflow projection, and (3) statistical modeling for future streamflow projection. Figure 2 shows a schematic of the research approach followed in this study. The first step is to analyze climate projections from multiple GCMs and RCPs to identify representative scenarios for streamflow investigation. The second step is to develop a hydrologic model and simulate the model for selected climate scenarios. The third step is to develop a statistical model using the outputs from steps 1 and 2 for predicting regional streamflow changes. In the following sections, we discuss each of these steps in further detail. HRU is the smallest spatial scale in the model that is created from unique combinations of land use, soil type and slope: Note, SWC o and SWC t are the initial and final (on day t) soil water content; precipitation, evapotranspiration, surface runoff, amount of water entering the vadose zone from the soil profile, and the amount of return flow on day t are denoted as P, E, R sup ,W a , and R sub , respectively.

Model setup, calibration and validation
This process involved three major steps: (1) [1978][1979][1980], and the validation period is 1996-2010. We conducted sensitivity analysis using regressing Latin Hypercube generated parameters against objective function values to identify parameters that show sensitivity to streamflow (Abbaspour ). We selected 11 parameters and found that streamflow is most sensitive to curve number (CN2). Table S1 in the supplementary information provides a detailed description of the parameters and the optimum values. Figure (2)- (5)): (4)    Although there are limitations of using Marksim, it is a      Table 2 along with the uncertainty analysis of such projections (95% confidence interval of median changes, and 95% bootstrap confidence interval of means). Figure 8 Table 2 along with the uncertainty analysis of such projections (95% confidence interval of median changes, and 95% bootstrap confidence interval of means). Because of the limited data points, it was necessary to run statistical significance tests to assess the performance of the regression models (Table 3). First, we analyzed the t-statistics, which is a measure of the likelihood that the actual value of the exponent and intercept of a regression equation is not zero (Islam & Seneka ). For the MVR of mean annual streamflow (Equation (6)), the t-statistics for the intercept, ΔT and ΔP, were found to be 20.34 (P-value: 2.46 × 10 -12 ), 1.33 (P-value: 2.03 × 10 -1 ), and 5.78 (P-value: 3.61 × 10 -5 ), respectively. The higher t-statistics and associated lower Pvalues indicate that it is less likely the actual value of the parameter could be zero. While performing the F-test, we noticed that the F-value and associated P-value for the MVR of mean annual streamflow are 25.94 and 1.35 × 10 -5 , respectively. The low P-value associated with the F-value indicates the validity of the regression equation in fitting the modelled results zero (Islam & Seneka ).
A statistical significant test for the other two models (MVR of mean dry season streamflow, MVR of the mean wet season streamflow) provides similar results (see Table 3 for details).
In order to assess the performance of these fitted regression equations, the mean annual streamflow and mean seasonal streamflow (at different temperature and precipitation changes for the six climate change scenarios selected for this study) estimated using Equations (6)- (8) were plotted against the SWAT simulated streamflows ( Figure 9). Also, statistical performance was checked through coefficient of determination (R 2 ) and the error bound as a form of 95% confidence interval of fitted values (e.g. intercept, ΔT and ΔP) (see Table 3 for details). It appears that the fitted equation for mean annual streamflow and mean wet season streamflow (Equations (6) and (8)) is in good

Maximum annual streamflow
The Ganges delta is one of the world's most flood-prone  Table 4 along with the uncertainty analysis of such projections (95% confidence interval of median changes, and 95% bootstrap confidence interval of means).
Except for a few scenarios (Coolest scenario, Driest scenario), all the other scenarios project an increase in annual maximum streamflow over the 21st century. The projected increase in annual maximum streamflow is  where ΔT and ΔP are the changes in basin scale mean temperature ( C) and basin scale mean precipitation (%) from the climate normal period , respectively and Q Max is the annual maximum streamflow of Brahmaputra at Bahadurabad for the future periods.
The results of the statistical significance test performed for Equation (9) are satisfactory. The t-statistic (and associated probability, or P-value) for the intercept, ΔT and ΔP, were found to be 34.24 (P-value: 1.17 × 10 -15 ), 4.73 (P-value: 2.68 × 10 -4 ), and 9.31 (P-value: 1.26 × 10 -7 ), respectively. The F-value and associated P-values for the MVR of annual maximum streamflow were found to be 91.29 and 4.01 × 10 -9 , respectively. In order to assess the performance of these fitted regression equations, annual maximum streamflow (at different temperature and precipitation changes for the six climate change scenarios) estimated using Equation (9) were plotted against the SWAT simulated annual maximum streamflow ( Figure 11). Also, statistical performance was checked through coefficient of determination (R 2 ) and the error bound as a form of 95% confidence interval of fitted values (e.g. intercept, ΔT and ΔP). It appears that the fitted equation is in good compliance with the simulated flow with higher R 2 (0.92) and less uncertainty (comparatively narrow bound of 95% confidence interval).

CONCLUSIONS
(1) The mean annual streamflow of the basin is expected to increase gradually over the 21st century. By the end of the 21st century, the projected increase in mean annual streamflow, with respect to the 1981-2010 climate normal period, ranges from about 14 to 47%, with a median and mean of about 24 and 25%, respectively.
However, the uncertainty associated with these projected changes grows as we progress to the distant future.
(2) The mean dry season streamflow of the basin is expected to increase at a much higher rate than the mean wet season streamflow. By the end of the 21st century, the median and mean projected increase in dry season streamflow is about 180 and 178%, respectively. However, the median and mean projected increase in wet season streamflow is only about 12 and 11%, respectively. This relatively high increase in dry season streamflow compared to the mean annual or mean wet season streamflow is presumably attributed to the enhanced snowmelt at the headwaters due to increase in temperature.
(3) By the end of the 21st century, the annual maximum streamflow is projected to increase by about 22% (ranges from about 6 to 64%).
(4) The projected mean annual, mean seasonal and annual maximum streamflow can be expressed as a multivariate linear regression of projected changes in basin scale mean temperature and basin scale mean precipitation.
The fitted equations for the annual maximum streamflow result in higher statistical goodness (R 2 ¼ 0.92) and lower uncertainty. The fitted equations for the mean annual streamflow (R 2 ¼ 0.78) and mean wet season streamflow (R 2 ¼ 0.72) also show satisfactory results. However, the fitted equations for the mean dry season streamflow demonstrate lower statistical goodness (R 2 ¼ 0.57) and higher uncertainty.