Hydrological assessment of the Tungabhadra River Basin based on CMIP6 GCMs and multiple hydrological models

Global climate change has altered the intensity and frequency of hydro-climatological extremes (Yuan et al. 2015; Piras et al. 2016). Accordingly, reliable streamflow prediction under climate change is imperative for better water resources planning and management (Steinschneider et al. 2012; Singh et al. 2016). A typical approach for this purpose is to generate future streamflow by forcing future projections obtained from climate models under varying climate scenarios into a calibrated hydrological model. However, the obtained projections suffer from inherent biases in both climate and hydrological models (HMs) (Her et al. 2019; Wang et al. 2020). It is essential to account for and quantify these uncertainties to improve the skill of future hydrologic projections. As projections from a single climate model cannot address the non-linear interactions researchers employed multi-model projections (Georgakakos et al. 2004; Harding et al. 2012; Hagemann et al. 2013; Tegegne et al. 2017; Keteklahijani et al. 2019; Adib et al. 2020).

Future climate projections are generally obtained from the outputs of state-of-the-art tools called global climate models (GCMs), which represent the atmospheric processes through distinct mathematical expressions (Gouda et al. 2018). The Coupled Model Intercomparison Project Phase 6 (CMIP6) initiated by the Intergovernmental Panel on Climate Change (IPCC) has coordinated and compiled the simulation from climate model experiments obtained from various modelling teams worldwide. The simulations in CMIP6-based climate models were generated under a set of scenarios using combinations of various radiative forcing levels and shared socioeconomic pathways (Meinshausen et al. 2020). The uncertainties in the future streamflow projections generated with the help of climate model outputs propagate from model structures, sub-grid parameterizations and their simplification of processes, various forcing scenarios, initial conditions, and choice of downscaling methods (Zhang et al. 2015b; Anil et al. 2021). Many of the studies reported so far address the uncertainty in the future streamflow projections due to uncertainties in model structures, initial conditions, scenarios, and downscaling methods (Knutti et al. 2010; Chen et al. 2011; Wang et al. 2019; Alam et al. 2021). The uncertainty corresponding to the climate models has been reduced in the latest CMIP6 bundle of GCMs due to the advancement in understanding the underlying physics of atmospheric processes over the last couple of decades (Gusain et al. 2020; Priestley et al. 2020). However, the climate model uncertainty is considered a major contributor to the future streamflow projections’ uncertainty (Kwon et al. 2012). Moreover, it is essential to note that the hydrological model structure uncertainty has received comparatively less emphasis so far than the former (Poulin et al. 2011; Steinschneider et al. 2012). The uncertainties associated with hydrological modelling include uncertainty in model structure, the spatial scale of the model, initial hydrologic conditions, parameter equifinality, and uncertainty in the calibration data (Jakeman & Hornberger 1993; Beven 2006; Kirchner 2006; Ludwig et al. 2009; Her & Chaubey 2015).

Basin-level hydrological studies assist policymakers and end-users in implementing better adaptation strategies for improved water resources management under climate change. In this article, the uncertainty in the future ensemble streamflow projections of the Tungabhadra Basin is studied considering the projected hydrological alterations and low water use efficiency. The future projections of the Tungabhadra Basin, a tributary to Krishna River, predicted severe drought and alterations in hydrological variables such as streamflow, soil moisture, and evapotranspiration (Gosain et al. 2006; ACIWRM & WRD 2012; Chanapathi et al. 2020). Singh et al. (2016) assessed the parametric uncertainty in simulating the extreme peak and low flows in the Tungabhadra Basin using Soil Water Assessment Tool (SWAT). The results suggested the reliability of SWAT performance in simulating the observed peak and low flows. Meenu et al. (2013) calibrated Hydrologic Engineering Center Hydrologic Modeling System version 3.4 (HEC-HMS 3.4) and forced future climate projections obtained from a single GCM to assess the impact of climate change on streamflow. The results of the study suggest increasing precipitation and runoff, with decreasing evapotranspiration. In contrast, a recent study by Facer-Childs et al. (2021) shows a reduction in the streamflow projections in the Tungabhadra Basin because of the increased frequency and severity of droughts due to climate change. Their study addressed the necessity to adapt ensemble streamflow prediction strategies to increase the reliability of the design and operation of water resources systems. Moreover, the water usage in the Tungabhadra Basin is low leading to unreliable allocation to all the users. Furthermore, the groundwater levels in Vedavati sub-basin are depleting due to over-exploitation attributable to scanty rainfall and surface water availability (WA). Bejagam et al. (2020) reported that precipitation from CMIP5-based models projects a decline of about 1.31–14.57% in the Tungabhadra Basin. Hence, understanding the future streamflow projections from recent CMIP6-based climate models is quintessential for hydrological applications such as reservoir operation, hydropower production, and irrigation in the Tungabhadra Basin. To the best of the authors' knowledge, very few studies have been reported so far on the hydrological modelling of the Tungabhadra Basin and assessing the uncertainty in the ensemble streamflow projections (Meenu et al. 2013; Singh et al. 2013, 2016; Goyal & Khan 2017). However, no study has been reported on assessing the uncertainty in the future streamflow at the Tungabhadra Basin using multiple GCMs and HMs.

The main objective of this study is to account for the uncertainties in the future streamflow projections of the Tungabhadra Basin generated from multiple GCMs and multiple HMs for different future scenarios. For this purpose, five CMIP6-based GCMs were selected based on their performance in replicating the historical statistics. A detailed description of the methodology used for the selection of best-performing GCMs is mentioned in the succeeding sections. The climate projections obtained from the five selected GCMs for four different scenarios were forced into four HMs (with varying model structures) to generate ensemble streamflow. Consequently, this study also shows how the uncertainty in future precipitation and temperature projections are translated into hydrological variables such as streamflow, potential evapotranspiration (PET), and WA. The total uncertainty in the generated ensemble streamflows is further decomposed using ANOVA to estimate the contributions of three different sources (GCMs, HMs, and SSPs). An ensemble post-processing method is also employed to correct the ensemble spread of the simulated flows during the future period.

The gridded daily precipitation, maximum temperature, and minimum temperature data are obtained from the Indian Meteorological Department (IMD). The daily precipitation dataset is available at a spatial resolution of 0.25° × 0.25° (Pai et al. 2014), and the daily maximum and daily minimum temperature data are available at a resolution of 1° × 1° (Srivastava et al. 2009). The gridded precipitation and temperature data were collected for the study area from 1975 to 2010 to set up HMs. The observed daily streamflow of the Tungabhadra Basin at Haralahalli gauge station was collected from 1975 to 2010, from India Water Resources Information System (India-WRIS).

The future projections of precipitation and daily maximum and minimum temperature were obtained from the latest bundle of CMIP6-based GCMs. A new set of five scenarios were considered in CMIP6-based GCMs, namely ‘shared socioeconomic pathways’ (SSPs). These scenarios were developed based on socioeconomic factors such as urbanization, economic growth, population, education, and the rate of technological development (Riahi et al. 2017). However, the raw GCM projections are simulated at coarser spatial resolutions (>200 km²) that cannot be used in basin-scale climate change impact studies. Moreover, they are prone to inherent biases that should be corrected. Hence, bias-corrected projections are used in this study to generate future ensemble streamflow. Mishra et al. (2020) developed a bias-corrected dataset for six countries in South Asia using empirical quantile mapping (EQM) at a spatial resolution of 0.25° for the daily time step. In this study, future climate projections (precipitation, maximum, and minimum temperature) obtained from CMIP6-based, bias-corrected 13 GCMs for all 4 scenarios (i.e., SSP-126, SSP-245, SSP-370, and SSP-585) for the future period of 2015–2100 and historical projections for the period 1975–2010 were also extracted for the study area (Mishra et al. 2020). The selected CMIP6 GCMs used in this study and their originating institute, spatial resolutions, and country are tabulated in Table 1.

To consider the uncertainty induced by hydrological modelling in the streamflow projections, three lumped conceptual and physics-based semi-distributed HMs were used in this study. They are Hydrologiska Byråns Vattenbalansavdelning (HBV), simple hydrology (SIMHYD); identification of unit hydrographs and component flows from rainfall, evaporation, and streamflow data (IHACRES), and soil and water assessment tool (SWAT). The varying complexities of these models in terms of model structures and spatial scale appraise the hydrologic predictive uncertainty. These models were calibrated and validated for 1975–2010 at daily time steps to simulate the future flows in the basin for 2015–2100. A brief description of the selected HMs is presented in the succeeding sections.

HBV is a lumped conceptual rainfall-runoff model, developed by the Swedish Meteorological and Hydrological Institute (SMHI) for runoff simulation and hydrological forecasting in the early 1970s (Bergstrom 1975). The runoff generation process is controlled through a set of 15 parameters, out of which 5 parameters address the snow routine and the remaining parameters account for soil moisture, response, and routing routines (Mendez & Calvo-Valverde 2016). This model has been widely used around the world attributable to its strong scientific foundation (Chen et al. 2012), and the model requires minimal inputs such as basin-averaged precipitation, PET, and temperature (Ashagrie et al. 2006; Jin et al. 2009; Chen et al. 2012; Seibert & Vis 2012).

SIMHYD is a simple lumped conceptual rainfall-runoff model and different hydrologists have tested the applicability of the model on various catchments (Chiew et al. 2009; Vaze et al. 2010). It consists of seven parameters and three storages to account for interception losses, soil moisture, and groundwater. The simulated runoff in SIMHYD consists of three components: surface runoff, interflow, and base flow. The surface runoff is generated through the infiltration excess mechanism (simulated using the negative exponential function of soil wetness), whereas the interflow is generated based on the saturation excess mechanism. The surface runoff is estimated using a negative exponential function of soil wetness, and the interflow and base flow are estimated as linear functions of soil wetness and groundwater storage, respectively.

IHACRES is a six-parameter lumped conceptual rainfall-runoff model based on the principle of unit hydrograph, developed by the Integrated Catchment Assessment and Management Centre (iCAM), Australian National University, and the Institute for Hydrology, Wallingford, UK (Jakeman et al. 1990; Chiew et al. 1993; Post & Jakeman 1996). A different version of this model has been widely used in multiple applications across the globe (Croke & Jakeman 2004). Basically, the model identifies unit hydrograph and flow components from rainfall, evaporation, and streamflow. The unit hydrographs corresponding to quick and slow flows are defined to generate total runoff. The runoff generation mechanism consists of a non-linear and linear module, where the former converts rainfall to effective rainfall, and the latter routes the effective rainfall. A detailed description of the underlying equations in the IHACRES model can be found in Post & Jakeman (1996).

All three lumped models were calibrated using the genetic algorithm and the SWAT model was calibrated with the help of SWAT-CUP software using the Sequential Uncertainty Fitting (SUFI-2) algorithm. The parameters of the HMs are fine-tuned to match the observed discharge during the period 1975–1993. The calibrated parameters were validated for the period 1994–2010. A diagnostic evaluation of the performance of all the selected HMs is conducted using a set of metrics. The metrics were chosen such that they emphasize distinct features of hydrograph such as low flows, high flows, and overall water balance (Manikanta & Vema 2022). The selected metrics along with their equation ranges and ideal values are mentioned in Table 2. The calibrated parameters were used to generate the future projections during 2015–2100, using the best-performing GCMs under four climate scenarios.

Table 2

Performance evaluation metrics along with equation, range, and ideal values

Sl No .	Performance evaluation metric .	Equation .	Range .	Ideal value .
1	Percent Bias (PBIAS)		−∞ to +∞	0
2	Nash–Sutcliffe Coefficient of Efficiency (NSE)		0 to ∞	1
3	Logarithmic Nash–Sutcliffe Coefficient of Efficiency (logNSE)		−∞ to 1	1
4	Mean Absolute Error (MAE)		0 to ∞	0
5	Fourth Root Mean Quadrupled Error (R4MS4E)		0 to ∞	0
6	Skill Score (SS)		0 to 1	1

Sl No .	Performance evaluation metric .	Equation .	Range .	Ideal value .
1	Percent Bias (PBIAS)		−∞ to +∞	0
2	Nash–Sutcliffe Coefficient of Efficiency (NSE)		0 to ∞	1
3	Logarithmic Nash–Sutcliffe Coefficient of Efficiency (logNSE)		−∞ to 1	1
4	Mean Absolute Error (MAE)		0 to ∞	0
5	Fourth Root Mean Quadrupled Error (R4MS4E)		0 to ∞	0
6	Skill Score (SS)		0 to 1	1

where and represent the observed and simulated discharge values, respectively, and is the mean observed discharge, n represents the total sample size of data, represent the frequency of simulated and observed streamflows for a particular bin. bn represents the total number of bins selected to analyse the match between PDFs between simulated and observed flows.

View Large

The ranking is given to the GCMs at each grid and the ranking pattern in each grid is plausibly different. Hence, to facilitate impact assessment studies for a region, a group decision-making approach is used to extract the suitable set of GCMs for the entire study area. A detailed description of group decision-making methodology can be found in Srinivasa Raju et al. (2017). The group decision-making in this study is performed over all three variables at all grids to obtain five best-performing GCMs for carrying out further study. From the analysis, it was found that models CanESM5, INM-CM5-0, MPI-ESM1-2-HR, MPI-ESM1-2-LR, and MRI-ESM2-0 were found to be the best-performing models. The complete analysis is carried out using these selected five models in the subsequent section.

The difference between maximum and minimum values (ranges) of projected climate variables (precipitation and temperature) and hydrological components (QQ, DR, SS, etc.) was calculated as a measure of the amount of uncertainty contained in ensemble predictions made using multiple GCMs and multiple HMs. Uncertainty in ensemble hydrological components obtained from the multiple GCMs and multiple HMs were first quantified. Then, the uncertainty amount was quantified for two different sources, multiple GCMs, and multiple HMs.

One of the popular approaches to overcome the uncertainty in model prediction is to pool the ensemble members from all the models, also known as equal weighting. It is challenging to assign weights to individual GCMs based on past performance and create a weighted ensemble mean. Assigning weights based on the representation of history requires a comprehensive knowledge of the model's performance (Curry & Webster 2011). However, the weights will highly depend on the evaluation procedure, ground truth data, region of interest, and parameters evaluated. Hence, in this study, a simple ensemble mean is employed to maintain the information about the spread of the ensemble. Uncertainty in ensemble projections due to each GCM is quantified by estimating the range (difference between maximum and minimum) and rank given to individual GCMs in the contribution of overall uncertainty among the selected GCMs. GCM with higher uncertainty given the first rank and GCM with the least uncertainty given the last rank. Then, flow duration curves (FDCs) were drawn from the ensemble streamflows obtained to quantify the uncertainty in multiple HMs. In addition, the relationships between the quantities of uncertainty in the ensemble projections of the climate variables and hydrological components were then investigated to see which climate variables (precipitation, maximum, and minimum temperature) exerted the most significant influence on the hydrological prediction uncertainty.

The future streamflows generated using multiple GCMs, multiple HMs, and multiple scenarios are highly uncertain with a high ensemble spread. However, it is worth noting that the ensemble spread plays a vital role in decision-making under uncertainty. Hence, the correction of ensemble spread to reduce the uncertainty bands assists the policymakers in better decision-making (Zhang et al. 2022; Manikanta et al. 2023). For this purpose, QRF, a non-parametric method, is used to post-process the ensemble streamflows (Meinshausen 2017). In QRF, conditional quantiles and median values of the response variable are estimated for high-dimensional predictors. The conditional quantiles are estimated using binary regression trees known as classification and regression trees (CARTs). Sequentially, in each tree, iterative binary splitting is employed to determine a new vector of predictors and its associated leaf. This iterative binary splitting between predictors aggregates the observations according to their forecasts. Therefore, for every generated streamflow event, an ensemble of observations is restored to create an empirical cumulative distribution function (CDF). The final CDF is estimated as the weighted mean of the CDFs obtained from all the trees, from which the predictive quantiles are determined. For more details on QRF methodology, the readers are encouraged to refer to Taillardat et al. (2016). In this study, 20 ensemble members generated by 5 GCMs and 4 HMs were forced as predictors for post-processing the spread of historical ensemble streamflows under each SSP. The hyper-parameters such as the maximum terminal node size are set to 20 with a tree size of 1,000. The model established during the historical period is used to correct the ensemble spread of generated future streamflows. The post-processing of future ensemble streamflows using QRF is performed with the help of an R package called ‘quantregForest’ (Meinshausen 2017).

The model structural uncertainty plays a key role in a hydrological prediction system (Moriasi et al. 2007). Hence, using multiple model structures is essential to account for the uncertainties in future streamflow projection. In this study, four different HMs (SWAT, HBV, IHACRES, and SIMHYD) were calibrated and validated in simulating the observed streamflow. The ranges of parameters and their fitted values are presented in Table 3. Nash–Sutcliffe Efficiency (NSE) is used as an objective function to calibrate the selected HMs. The NSE values for HBV, IHACRES, SIMHYD, and SWAT were 0.71, 0.65, 0.59, and 0.73, respectively, during the calibration period (1975–1993) and 0.76, 0.72, 0.71, and 0.73, respectively, during the validation period (1994–2010). The performance of selected models during both calibration and validation periods in terms of NSE was reasonably acceptable (Moriasi et al. 2007, 2015). However, it is necessary to understand the performance of the simulated streamflow in capturing different flow segments of the observed hydrograph. The performance metrics such as fourth root mean quadrupled error (R4MS4E), Logarithmic Nash–Sutcliffe coefficient of efficiency (logNSE), percentage bias (PBIAS), SS, and mean absolute error (MAE) are estimated to evaluate the performance of simulated flows from the calibrated HMs.

For instance, the indicator R4MS4E puts more emphasis on peak flows, whereas logNSE is sensitive to low flows. The volumetric error is measured using PBIAS and the match between observed and simulated frequencies is measured using SS, and MAE yields the unbiased error estimate. The computed performance evaluation metrics for the calibration and validation periods are tabulated in Table 4. From the R4MS4E values, it is clear that the SWAT and HBV models perform well in capturing high flows in both calibration and validation periods. The logNSE values show that the HBV model performance in simulating the low flows is best in both calibration and validation periods. The PBIAS value of SWAT is lower (between ±5%) than that of other models, whereas the bias in HBV simulated flow is within acceptable limits. In terms of SS, all four models exhibit similar performance during both calibration and validation periods, with the HBV model exhibiting slightly better performance. The performance of HBV and SWAT in terms of MAE is found to be good during calibration and validation. The SIMHYD simulated flows are highly underestimating the observed flows leading to declined performance. SWAT being a physics-based semi-distributed model, and the inclusiveness of spatial variability in rainfall and physical characteristics of the catchment, is the reason behind its better performance (Singh et al. 2013). The process description of the HBV model structure is the plausible reason for its better performance.

The ensemble streamflows generated by forcing the GCM outputs for the calibration period (1975–1993) are post-processed using QRF to reduce the ensemble spread and to capture the observed streamflow. The model fitted during the calibration period is extrapolated to post-process the ensemble streamflows during the validation period. Figure (10) represents the FDC plotted using the ensemble streamflows before post-processing (hereby referred to as EnSF) and after post-processing (QRF-EnSF) during the validation period (1994–2010). In summary, the ensemble spread is highly optimized in the QRF-EnSF, where the ensemble spread captures observed FDC between the exceedance probabilities of 0.05 and 0.8 efficiently with minimal spread. It can also be noticed that the spread in EnSF is comparatively high. Moreover, the FDC by EnSF tends to overpredict the high flows and underpredict the low flows. The performance of both EnSF and QRF-EnSF is not satisfactory in capturing the streamflow extremes (extremely low and high values of streamflows). The results of the analysis suggest that QRF post-processed flows efficiently capture the majority of the flows and can be used to reduce the uncertainty in the generated streamflows.

The Tungabhadra River serves as an essential source of irrigation, hydropower generation, and water supply in major portions of Karnataka regions. Hence, assessing the WA under changing climate scenarios is a crucial issue in this basin. Limited studies have been reported so far assessing the impact of climate change on the streamflow of the Tungabhadra Basin (Meenu et al. 2013; Singh et al. 2013, 2016; Janga Reddy & Nagesh Kumar 2020; Venkatesh et al. 2020b). Moreover, no study was reported so far considering the hydrological model uncertainty along with GCM uncertainty in the future streamflow projections in the Tungabhadra Basin. Hence, the present study comprehensively measured the influence of uncertainty in climate projections and hydrological model structure uncertainty on different hydrological variables for the Tungabhadra Basin. The total uncertainty in the future streamflows is decompounded as the contribution of individual sources (GCMs, HMs, and SSPs) and their interactions using the ANOVA technique. Furthermore, a statistical post-processing method is used to calibrate the ensemble spread intended to reduce the uncertainty in the streamflows. From previous studies, it was suggested that the selection of suitable GCMs is based on the past performance (Srinivasa Raju & Nagesh Kumar 2015; Anil et al. 2021), which is also followed in the present study using a widely accepted approach such as CP (Srinivasa Raju et al. 2017). Four HMs (SWAT, HBV, SIMHYD, and IHACRES) were calibrated and validated in simulating the observed streamflow at the Haralahalli gauge point. The performance of all the calibrated models was found to be satisfactory. SWAT is s physics-based semi-distributed model and the previous studies supported the results for the hydrological assessment of the basin (Singh et al. 2013, 2016; Venkatesh et al. 2020a), proved its efficiency again with better performance in the present study. From Section 4.1, conceptual model HBV performance is equally good with SWAT in simulating future streamflow. Models IHACRES and SIMHYD performance might not match well with the other models mentioned before but their performance in simulating future streamflow cannot be ignored owing to the fact that model structure is one of the major sources of uncertainty.

The analysis shows that all the streamflow projections project an increase in the higher streamflow values and a decrease in the lower streamflow values which is in line with the previous studies (Zhang et al. 2015a). The results obtained from the ANOVA method demonstrated that the contribution of interaction between GCMs and SSPs to the total uncertainty was found to be significantly higher (between 44.19 and 58.62% in the near future and 40.24 and 59.37% in the far future). Despite having lower individual contributions, i.e., GCMs (5.04–9.48% in the near future and 5.06–9.26% in the far future) and SSPs (0.57–0.82% in the near future and 0.57–1.03% in the far future), their interaction has a major share in total uncertainty. It is also important to note that the uncertainty in the future ensemble streamflows is also due to the temporal uncertainty in the GCMs. It implies that the daily projections obtained from the GCMs vary drastically leading to high variations in the projected streamflow values. The uncertainty in the climate projections obtained from GCMs varies over different months, recommending improved GCM simulations for reliable basin-level impact assessment studies (Her et al. 2019). However, the current study using multiple HMs yields better insights into the uncertainties in the streamflow projections (Her et al. 2019; Wang et al. 2020). Since the inherent complexity is present among various parameters in the modelling chain, it is difficult to generalize a single hydrological model for a catchment. From higher QQ projections, it is understood that more intense precipitation is expected in the future with fewer rainy days (Meenu et al. 2013). The QRF post-processing technique was applied to reduce the uncertainty bands in the ensemble streamflow projections. The post-processed ensemble streamflows were found to be efficiently capturing the observed streamflows during the validation period with lower ensemble spread. The post-processed future streamflows also project increment in the future streamflows, but the increment is significantly lower than the ensemble streamflows before post-processing. The results of the present study are useful in framing adaptation strategies such as enhancing irrigation efficiency, managing water demand, and installing water harvesting schemes considering the future WA in the Tungabhadra Basin.

This study focuses on quantifying the uncertainty in the ensemble streamflow projections introduced by multiple GCMs and multiple HMs. The conclusions are as follows:

• The performance of selected HMs was found to be satisfactory with an acceptable NSE value (>0.6) during both calibration and validation periods and diagnostic evaluation of model performance indicates that SWAT and HBV performed better in simulating low flows.

• The ability of SWAT to account for rainfall spatial variability and spatial heterogeneity of catchment characteristics is attributable to its better performance, which agrees with previous studies (Singh et al. 2013). The performance of HBV, being a lumped model, was equally good as SWAT in simulating streamflows.

• Uncertainty in climate projections during the monsoon period was appraised and results show that the overall increment in PP projections ranges from 10.43 to 222.5% and QQ projections range from 34.50 to 377.7% in the monsoon season. Both Tmax and Tmin were found to be increasing with respect to scenarios. It was found that the changes in projected Tmax range from −3.0 to +3.4 °C, and Tmin ranges from −1.8 to +1.8 °C.

• • FDCs obtained from the models HBV, IHACRES, and SIMHYD show that the frequency of high flows is likely to be increased under all SSPs, and the frequency of low flows is expected to decrease in the future scenarios for SIMHYD and IHACRES models. The above results indicate that the flashiness of the catchment (slope of flow duration curve) may increase indicating higher variability of flows in future scenarios.

• Uncertainty decomposition using 3-way ANOVA suggested that the interaction between SSPs and GCMs is the highest contributor among all other sources ranging from 44.19 to 58.62% in the near future and 40.24 to 59.37% in the far future. The second highest contributor corresponds to the combined effect of HMs, SSPs, and GCMs with a share ranging between 29.18 and 38.79% in the near future and 30.14 and 40.85% in the far future. SSPs and HMs individual contributions are negligible.

• The mean flows after post-processing are likely to be reduced up to 7% in June and increase up to 70% in September under all scenarios. The increment in the mean flows was found to be highest in October in un-post-processed flows, whereas the results from QRF post-processed flows show that August is likely to have the highest increment. The ensemble spread has been reduced after the QRF post-processing application.

All relevant data are available from an online repository or repositories: https://www.imdpune.gov.in/cmpg/Griddata/Rainfall_1_NetCDF.html - For gridded precipitation data, https://www.imdpune.gov.in/cmpg/Griddata/Max_1_Bin.html - For gridded temperature data, https://indiawris.gov.in/wris/#/ - For observed streamflow data, https://swat.tamu.edu/data/ - India dataset for SWAT model, https://esgf-node.llnl.gov/projects/cmip6 - future climate dataset.

The authors declare there is no conflict.