Climate impact emergence and flood peak synchronization projections in the Ganges, Brahmaputra and Meghna basins under CMIP5 and CMIP6 scenarios

The densely populated delta of the three river systems of the Ganges, Brahmaputra and Meghna is highly prone to floods. Potential climate change-related increases in flood intensity are therefore of major societal concern as more than 40 million people live in flood-prone areas in downstream Bangladesh. Here we report on new flood projections using a hydrological model forced by bias-adjusted ensembles of the latest-generation global climate models of CMIP6 (SSP5-8.5/SSP1-2.6) in comparison to CMIP5 (RCP8.5/RCP2.6). Results suggest increases in peak flow magnitude of 36% (16%) on average under SSP5-8.5 (SSP1-2.6), compared to 60% (17%) under RCP8.5 (RCP2.6) by 2070–2099 relative to 1971–2000. Under RCP8.5/SSP5-8.5 (2070–2099), the largest increase in flood risk is projected for the Ganges watershed, where higher flood peaks become the ‘new norm’ as early as mid-2030 implying a relatively short time window for adaptation. In the Brahmaputra and Meghna rivers, the climate impact signal on peak flow emerges after 2070 (CMIP5 and CMIP6 projections). Flood peak synchronization, when annual peak flow occurs simultaneously at (at least) two rivers leading to large flooding events within Bangladesh, show a consistent increase under both projections. While the variability across the ensemble remains high, the increases in flood magnitude are robust in the study basins. Our findings emphasize the need of stringent climate mitigation policies to reduce the climate change impact on peak flows (as presented using SSP1-2.6/RCP2.6) and to subsequently minimize adverse socioeconomic impacts and adaptation costs. Considering Bangladesh’s high overall vulnerability to climate change and its downstream location, synergies between climate change adaptation and mitigation and transboundary cooperation will need to be strengthened to improve overall climate resilience and achieve sustainable development.


Introduction
The majority of Bangladesh lies in one of the largest and most populous deltas of the world, downstream of the three river systems of the Ganges, Brahmaputra and Meghna (GBM). Floods are common in the GBM basins and are primarily caused by heavy rainfalls during the summer monsoon season (South/ Southeast Asian monsoon region). Every year, onequarter to one-third of Bangladesh is inundated (Islam et al 2010), impacting approximately 40 million people (>20% of the population) and often leading to a large number of fatalities and human displacements (Mehta and Kumar 2019). Economically, flooding results in large financial losses in the agriculture, housing, industry, and transportation sectors  (Mirza 2011). The effects of floods on public health are also a concern, with poor households being particularly vulnerable (Vinke et al 2017, Lee et al 2021. Flood duration and recession within Bangladesh depend on the upstream inflow of water, rainfall within Bangladesh, and tidal activity. River floods are particularly severe in Bangladesh if (i) peak flows are of high magnitude and (ii) there is a temporal match of the flood peaks of at least two of the three major rivers. A synchronization of flood peaks at the confluence often results in an increase in flood magnitude and an accelerated flood wave (Guse et al 2020). For Bangladesh, flood synchronization has led to catastrophic flood events, such as in 1988 and 1998 (figure 1), when peak flows of the Ganges and Brahmaputra (GB) occurred with a lag of a few days only (Mirza 2002), and the majority of the country was inundated.
Future climate projections based on the Coupled Model Intercomparison Project phase 5 (CMIP5; Taylor et al 2012) have indicated increasing air temperature and summer monsoon precipitation over South Asia (Srivastava andDelSole 2014, Uhe et al 2019). The evaluation of CMIP5 simulated precipitation in relation to observations has, however, revealed considerable uncertainties in capturing South Asian summer monsoon onset and precipitation amounts, especially for convective precipitation events (Chen and Zhou 2015, Sabeerali et al 2015, Ashfaq et al 2017. The latest CMIP6 global climate models (GCMs) differ from those of CMIP5 as a result of improved model parameterization and process representation (Eyring et al 2016), a later starting year of the future scenarios (2015 compared to 2006 in CMIP5), updated emission scenarios (shared socioeconomic pathways (SSPs) extending the representative concentration pathways; Meinshausen et al 2020), inclusion of new processes (mostly related to atmospheric chemistry and aerosols), higher spatial resolution, and updated land-use scenarios (Gidden et al 2019). The starting year of the CMIP6 future projections was updated to 2015 as observational datasets could be extended, especially CO 2 budgets. The 2007-2014 emissions have been notably higher than expected. Therefore the SSP1-2.6 scenario starts, for example, with higher CO 2 concentrations and the emission reductions are more rapid until 2100 compared to RCP2.6. The Indian summer monsoon is better captured by CMIP6 in comparison to CMIP5 models (Gusain et al 2020) as a results of smaller sea surface temperature biases over the Northwestern Pacific Ocean (Xin et al 2020). CMIP6 simulations exhibit a larger climate sensitivity to anthropogenic greenhouse gas emissions, which is most likely caused by cloud feedbacks and cloud-aerosol interactions (Meehl et al 2020), and resulting in overall higher warming over South Asia (Almazroui et al 2020). Similar to CMIP5, the CMIP6 models project an increase in total annual precipitation over South Asia during the twenty-first century (Chen et al 2020, Wang et al 2020, particularly over the northwestern parts of India, Nepal, Bhutan, and Bangladesh (Almazroui et al 2020). The differences between CMIP5 and CMIP6 precipitation and air temperature projections are most pronounced toward the end of the twenty-first century.
Previous studies using hydrological models forced by the simulated meteorological output of CMIP5 GCMs suggest an increase in floods risk, even when limiting global warming to 1.5 • C in accordance with the Paris Agreement (Uhe et al 2019).  showed, using a macroscale hydrological model (H08) forced by five CMIP5 GCMs under a high emission scenario (RCP8.5), that the Meghna basin will experience the highest increase in precipitation (+30%) and runoff (+40%) by the end of the century (2075-2099) compared to 1979-2003. Similarly, Mohammed et al (2018 highlighted that under different global warming scenarios (1.5 • C, 2.0 • C, and 4.0 • C), peak flows associated with a return interval of 100 years will increase in the GBM basins (Ganges +27%-54%, Brahmaputra +8%-63%), particularly in the Meghna river (+15%-81%), and especially under high warming (4.0 • C). Studies focusing solely on the Brahmaputra River have also found consistent increases in discharge and high flows when using CMIP5 global (Alam et al 2021) and CORDEX regional (Islam et al 2017) future climate projections. Using the latest generation of climate models (CMIP6), Hirabayashi et al (2021) showed that, globally, CMIP6 based flood projections result in similar flood risk compared to CMIP5. Projected flood risk differences between CMIP5 and CMIP6 climate forcing remain unknown in the GBM basins.
Moving beyond climate change impact assessments focusing on differences over time, the time of climate impact emergence (TCIE) is an important risk management indicator (Hawkins and Sutton 2012). TCIE provides stakeholders and policy makers an estimate of the time left for adaptation planningor in other words when will an extreme flood today be regarded as the 'new normal' in the future. The 'time of emergence' concept has mostly been applied to climatic variables such as mean temperature and precipitation (Diffenbaugh andScherer 2011, Hawkins andSutton 2012). More recently, the concept was extended to impacts assessing hydrological regime shifts (Leng et al 2016, Muelchi et al 2021 and future agricultural risks (Jägermeyr et al 2021).
The objective of this paper is to assess how the improvements in simulating summer monsoon precipitation of the latest generation of climate projections (CMIP6) translate into high and peak flows in Bangladesh in comparison to the previous generation of climate projections (CMIP5). The study focuses on three main aspects:

Overview of study basins
The watersheds of the GBM rivers cover together an area of ∼1.6 million km 2 . The Ganges basin (966 000 km 2 ) is the largest, followed by the Brahmaputra (532 000 km 2 ) and the Meghna (80 000 km 2 ). Elevation ranges from sea level to more than 8000 m in the Himalayan Mountains for the Ganges and the Brahmaputra basins and to up to 3000 m in the Manipur Hills for the Meghna basin. The majority the GBM delta in Bangladesh is less than 12 m above sea level. The GBM basins are shared among five countries: India (64%), China (18%), Nepal (8%), Bangladesh (7%), and Bhutan (3%) (FAO 2011). The individual country share in each basin is detailed in table 1. The climate is extremely diverse across the GBM basins (FAO 2011). Total annual (summer monsoon season June-September) precipitation amounts are presented in table 1 for the different watersheds. Topography plays a strong role on the spatial distribution of precipitation in the GB watersheds (figure S1). Precipitation is the lowest across the Tibetan Plateau (table 1 (China)) and the upper northwestern region of the Ganges watershed. In the Brahmaputra basin, large precipitation amounts occur across the Himalaya belt and the lowlands. The Meghna watershed receives the largest amount of precipitation per km 2 , with the world's highest precipitated area (Cherrapunji-Mawsynram region) receiving average annual precipitation of around 12 000 mm in the hilly northeastern state of Meghalaya (India)  . Within Bangladesh, a humid subtropical climate dominates, with mean annual precipitation ranging from 1500 mm in the west to about 4200 mm in the northeastern parts of the country (Shahid 2011, Mohammed et al 2018. About 80% of the total annual rainfall occurs in the summer monsoon season (Mirza 2002, Islam et al 2010. The remaining precipitation is distributed between the pre-monsoon (March-May) and post-monsoon (October-November) seasons. The annual maximum (mean) discharge amounts to ∼55 000 (∼16 000) m 3 s −1 in the Ganges river at Hardinge Bridge , ∼82 000 (∼24 000) m 3 s −1 in the Brahmaputra river at Bahadurabad (1980Bahadurabad ( -2012 and ∼13 000 (∼5500) m 3 s −1 in the Meghna river at Bhairab Bazar (1984-2012) on average based on discharge data provided by the Bangladesh Water Development Board. For better inter-basin comparison, runoff records (mm) are shown in table 1.
The land use distribution also differs between the watersheds (figure S2). The Indo-Gangetic Plain is largely used for agriculture ('South Asia's bread basket'), the Himalaya belt is forested, while savanna and grasslands dominate the Tibetan plateau. Glaciers, located at high altitudes, cover 0.02% of the Ganges and 0.85% of the Brahmaputra basins (RGI 2017). There are no documented glaciers nor permanent snowfields in the Meghna basin due to topographic and climatic conditions. Upstream regions in the Brahmaputra and the Ganges watershed are highly dependent on the seasonal snow and ice-melt for agriculture and hydropower generation (Immerzeel et al 2010, Ray et al 2015, Biemans et al 2019. For Bangladesh, the glacier melt contribution to total discharge is negligible, although glacier melt and extreme monsoon precipitation tend to temporally coincide. Based on the degree-day approach (which is also used in the hydrological model Soil and Water Integrated Model (SWIM) applied in this study) and upper-range values for the melt parameter, the glacier meltwater contribution to the mean annual maximum discharge (Q max ) was approximately 1.4% (∼0.15% in the Ganges and 2.5% in the Brahmaputra basins) and considerably less for the periods of more severe floods. With a projected glacier recession, this percentage is likely to decrease further in the future (Lutz et al 2014).

Data
All data used (table S1) (table S2). Here, we refer to these stations as outlet stations. Monthly discharge data from nine stations located in Nepal, in the upstream parts of the Ganges basin, were obtained from the Global Runoff Data Center (table S1). Table S2 summarizes the discharge data's location, temporal coverage, and quality.
Observation-based climate data (table S1) was obtained from two datasets: EWEMBI (Lange 2016) and W5E5 version 2 (Lange et al 2021). Both datasets are available globally at 0.5 • horizontal spatial resolution and at daily time step. The EWEMBI dataset, covers the time period 1979-2013 and W5E5 the period 1979-2019. Data sources of EWEMBI are ERA-Interim reanalysis data (Dee et al 2011), WATCH methodology applied to ERA-Interim reanalysis data (Weedon et al 2014), eartH2Observe (E2OBS) forcing data (Calton et al 2016) and NASA/GEWEX Surface Radiation Budget data (Stackhouse et al 2011). The Surface Radiation Budget data were used to bias-correct E2OBS shortwave and longwave radiation (Lange 2018). W5E5 combines version 2.0 of the WATCH Forcing Data methodology applied to ERA5 data (Weedon et al 2014, Cucchi et al 2020 over land with ERA5 reanalysis data  over the ocean. Precipitation over the land is bias-adjusted using observations from the Global Precipitation Climatology Centre (GPCC; Schneider et al 2011, version 6) and over the ocean using the Global Precipitation Climatology Project (Adler et al 2003, version 2.3).
Meteorological outputs (precipitation, air temperature (mean, minimum, maximum), humidity, downward shortwave radiation) of four CMIP5 and ten CMIP6 GCMs (table S3) were bias-adjusted at daily temporal resolution and statistically downscaled to 0.5 • spatial resolution within the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP) phases 2b (for CMIP5) and 3b (for CMIP6). The CMIP5 and CMIP6 GCMs used here were selected by ISIMIP based on benchmark performance during the historical period, equilibrium climate sensitivity, structural independence, output availability at the time of selection, and representation of the entire CMIP ensemble (details see in Frieler et al 2017). The bias-adjusted climate models form the basis for impact studies across many different sectors (www.isimip.org/). The ten CMIP6 GCMs were bias-adjusted to the W5E5 dataset, employing a quantile mapping approach that preserves trends in all quantiles of the distribution of simulated daily climate model outputs (Lange 2019). The bias-adjustment of the four CMIP5 models is based on the EWEMBI data set (1979-2013) based on a slightly advanced version of the method by Hempel et al (2013) as detailed in Frieler et al (2017) and Lange (2017). The approach is based on transfer functions that correct the multi-year monthly mean value followed by an adjustment of the intramonthly day-to-day variability. This method only preserves long-term mean-value trends in the climate variables. For the sake of brevity, in the following we refer to the bias-adjusted and statistically downscaled climate inputs from ISIMIP as CMIP5 and CMIP6 climate forcing, although our datasets presents only a subset of the available CMIP5/CMIP6 models. The comparison of the long-term mean spatial patterns of total annual and monsoon season precipitation (figure S1) and precipitation climatology (figure S3) shows a good agreement between the observation-based datasets (W5E5, EWEMBI) and CMIP5, CMIP6 and also in comparison to GPCC (table S4).

Hydrological model, set-up and evaluation
We used the process-based, semi-distributed SWIM (Krysanova et al 1998 to simulate the hydrological cycle of the GBM basins at a daily time step. SWIM was developed based on SWAT'93 (Arnold et al 1993) and MATSALU (Krysanova et al 1989) and has been applied to basins of similar hydro-climatic conditions , Wortmann et al 2018. Table  S5 provides an overview of the most relevant hydrological processes for this study and their representation in SWIM. SWIM is based on a three-level spatial disaggregation of the basin (catchment, subbasins, hydrotopes). The hydrotopes within a subbasin were constructed based on common land use, soil type, and elevation contours (300 meters). Data sources for elevation, land use and soils are shown in table S1. Glacier cover is considered based on the Randolph Glacier Inventory (RGI 2017), and glacier thickness was estimated by slope inversion and uniform shear stress of 10 5 Pa (Marshall et al 2011). Glacier extent was kept constant throughout the simulation periods as we estimated glacier contributions to floods to be negligible (see section 2). Similarly, the land cover was held constant within SWIM as our focus was on evaluating the climate change impact. Due to lack of data, the operation of reservoirs (75 in the Ganges, 6 in the Brahmaputra) is not included in the model parameterization.
Climate data needed to force SWIM include daily values of minimum, mean and maximum air temperature, total precipitation, solar radiation, and relative humidity. In this study, two SWIM set ups were used: • EWEMBI-SWIM was calibrated using the meteorological forcing of the observational dataset EWEMBI. • W5E5-SWIM was calibrated using the meteorological forcing of the observational dataset W5E5.
The calibration (1980-1995) and validation (1996-2012) of SWIM was based, as much as possible, on the enhanced approach by Krysanova et al (2018) using multiple statistical performance criteria (table S6). A multi-objective evolutionary algorithm (S-metric selection evolutionary multi-objective optimization algorithm, SMS-EMOA (Beume et al 2007)) was used to optimize the SWIM model parameters followed by manual fine-tuning for both model set-ups (EWEMBI and W5E5 climate forcing). The calibrated parameters and their value ranges (used in this study) are displayed in table S9. Calibration on daily discharge was restricted to the outlet stations due to data availability constraints. In the Ganges basin, additional upstream gauging stations, located in Nepal, with monthly discharge records were available for model validation (table  S2), which refined the model parameterization to represent internal basin processes reliably.

Indicator Description
Qmax (a proxy for floods) Annual maximum (daily mean) discharge (m 3 s −1 ) Q10 (a proxy for high flows) The magnitude of daily discharge that is exceeded 10% of the time in the daily time series of 30 years t(Qmax) Timing of Qmax (day and month) Flood synchronization Two criteria need to be fulfilled:

Indicators to assess climate change impact
SWIM was forced by the four CMIP5 (using calibrated/validated EWEMBI-SWIM) and ten CMIP6 (using calibrated/validated W5E5-SWIM) climate projections under the RCP2.6/RCP8.5 and SSP1-2.6/SSP5-8.5 scenarios, respectively. The radiative forcing levels of the chosen global warming scenarios are similar for RCP2.6/SSP1-2.6 and RCP8.5/SSP5-8.5 by 2100. The low-end global warming scenarios RCP2.6/SSP1-2.6 present the emission pathways limiting global warming to 2.0 • C above the preindustrial levels. RCP8.5/SSP5-8.5, on the contrary, represent the high-end global emission scenarios. In total, the model ensemble consisted of 28 simulations. We defined 1971-2000 to represent the reference period, 2031-2060 the near and 2070-2099 the far future. The impact of climate change on floods in Bangladesh was evaluated using different flood indicators, which are detailed in table 2. The indicators include a proxy for floods (Q max ) and high flows (Q 10 ), timing and synchronization of Q max and TCIE. Flood indicators were calculated at the outlet discharge stations of the three rivers located in Bangladesh (figure 2). We applied a Wilcoxon ranksum test to assess the statistical significance of the changes.
We evaluated flood synchronization across the three river basins for each river basin combination (GB, Brahmaputra and Meghna (BM), and Ganges and Meghna (GM), and all three river basins (GBM)) within a time period of 10, 20, and 30 d. We only evaluated events when the combined Q max of the two or three rivers exceeds a threshold value (ΣQ 0.1 ), which is based on the discharge magnitude that was exceeded on only 0.1% of all days of the observations (Q 0.1 , table S7). We counted the number of occurrences of such synchronous floods for each model ensemble member within each time period.
We also computed the TCIE following the definition and approach described in Jägermeyr et al (2021). TCIE defines the year when the average climate change impact occurs outside the envelope of the historical Q max flow variability. More specifically, the TCIE is the year in which the 30 year moving-average of simulated Q max (ensemble median) exceeds either the 95th (positive TCIE) or 5th (negative TCIE) percentile of the simulated Q max (ensemble median) of the reference period .

Model calibration and validation
The performance of SWIM was evaluated for both mean and high flow conditions. Results are largely similar for the two model versions (EWEMBI-SWIM, W5E5-SWIM). The comparison of daily simulated and measured discharge (figures S4(a), (c) and (e)) at the outlet stations shows high scores of the statistical performance criteria (table 3). Model performance is, on average, slightly lower during the model validation period compared to the calibration period. In the Ganges basin, good model performance was also achieved at most upstream gauges, which indicates that internal watershed processes are captured sufficiently well (figure S5). In the Brahmaputra basin, simulations and observations show higher differences compared to other two basins, which can be directly linked to data availability (no upstream discharge records) to constrain model parameterization. The seasonal discharge cycle is quite well reproduced by the SWIM simulations (figures S4(b), (d) and (f)). Larger differences in simulating seasonality exist based on W5E5 in the Brahmaputra (underestimation especially from July to September) and in the Ganges (overestimation from mid-August to September).
The recurrence intervals of Q max show a good agreement between observations and SWIM simulations forced by the observation-based datasets (EWEMBI, W5E5) as well as the historical CMIP simulations (figure 3). Difference between SWIM forced by EWEMBI or W5E5 are minor when assessing daily peak flows. At Hardinge Bridge (Ganges), simulated Q max by EWEMBI-SWIM (W5E5-SWIM) slightly underestimate (overestimate) the magnitude of frequent high flow events (return period of up to 5 years), low frequency events (return period >5 years) are generally underestimated by the simulations. At Bahadurabad (Brahmaputra), Q max of high frequency are well captured, while events of lower frequency tend to be underestimated by the simulations. At Bhairab Bazar (Meghna), higher (lower) frequency events are underestimated (well captured). Differences of Q max recurrence intervals between the CMIP historical climate forcing and the observationbased dataset used for the bias-adjustment, are low, except in the Brahmaputra basin and especially for CMIP5 versus EWEMBI.

Climate change impact assessment
The climate model ensembles consistently project increases in median summer monsoon season air temperature and precipitation across the GBM basins and global warming scenarios, despite large variability across the model ensembles (figure 4, table S8). The air temperature increase is largest under RCP8.5/SSP8.5 in the far future, especially in the Brahmaputra basin (>4 • C). The temperature increase is significant (p < 0.05) across all scenarios and is comparable between CMIP6 and CMIP5 (table S8). Considering the changes in the multi-model median, stringent climate mitigation policies limit warming to an increase of less than 2 • C (RCP2.6/SSP1-2.6) and summer precipitation increase to less than 10% in the GBM basins (except in the Ganges (CMIP6) and Meghna (CMIP5)). The CMIP5 models simulate the largest increase in summer monsoon season precipitation in the Meghna basin (28 ± 13%) under RCP8.5 in the far future (ensemble median ± interquartile range (difference between 75th and 25th percentiles)), while the CMIP6 models simulate the largest increase in the Ganges basin (27 ± 22% under SSP5-8.5 in the far future) (table S8).
Under all climate change scenarios, a consistent increase in high flows (Q 10 ) and flood peaks (Q max ) is simulated in the near and far future ( figure 4, table S8). The largest percent increase in Q 10 and Q max is projected in the Ganges basin under RCP8.5/SSP5-8.5 in the far future. The increase in Q 10 and Q max amplifies until the end of the century under RCP8.5/SSP5-8.5, while remaining relatively stable under RCP2.6/SSP1-2.6. Stringent emission reductions even result in a decrease of Q max under the SSP1-2.6 scenario in the far compared to the near future. In the Meghna watershed, however, Q 10 and Q max increase from the near to the  (1980-1995) and validation (1996-2012) based on daily discharge at the gauging stations Bhairab Bazar (Meghna), Hardinge Bridge (Ganges) and Bahadurabad (Brahmaputra), which are located in Bangladesh prior to their confluence. SWIM was calibrated using meteorological forcing of EWEMBI (basis for bias-adjustment of CMIP5) and W5E5 (basis for bias-adjustment of CMIP6). The statistical performance criteria include: Nash-Sutcliffe efficiency (NSE) (Nash and Sutcliffe 1970), percent bias (PBIAS) (%), mean error (ME) (m 3 s −1 ), Coefficient of determination (R 2 ), Index of Agreement (IoA) and the ratio of standard deviation (rSD  far future under the RCP2.6. When comparing the CMIP phases, a strong increase in Q max (ensemble median) of 60% is projected under RCP8.5, compared to 36% under SSP5-8.5 in the far future relative to the reference when averaged across the GBM basins (weighted by basin area). For Q 10 , the increase is comparable between the CMIP phases (RCP8.5: 42%, SSP5-8.5: 44%). Comparing the CMIP5 and CMIP6 forcing for the separate basins shows that: In the BM basins, percent increase in Q 10 and Q max is larger for CMIP5 compared to CMIP6 based projections across all scenarios and periods (except precipitation (RCP2.6, near future) and Q 10 (RCP8.5, far future) in the Brahmaputra). In the Ganges basin, the percent increase of Q 10 is larger under CMIP6 across all scenarios and periods, while CMIP5 based simulations result in larger Q max (except for SSP5-8.5 in the near future).
In the reference period, Q max (ensemble median) occurs first in the Brahmaputra (late July), followed by the Meghna (mid to late August) and Ganges (late August to early September) rivers (figure 5). Q 10 also occurs first (late-July) in the Brahmaputra river, and at similar time (late July to early August) at the GM rivers ( figure S7, reference period). In all basins, Q max occurs earlier under CMIP5 compared to CMIP6 simulations (∼1-2 weeks) in the reference period. In the Ganges basin, the timing of Q max occurrence remains unchanged under the future projections (compared to the reference) under CMIP6, while for CMIP5 Q max occurs later (earlier) under RCP2.6 (RCP8.5). For the Brahmaputra river, a tendency toward earlier occurrence of Q max (more pronounced under CMIP5 compared to CMIP6) is simulated when considering the ensemble median under all future scenarios, except in the near-future under the CMIP6 SSP scenarios. In the Meghna river, Q max occurs earlier in the near future and later in the far future under CMIP6 future projections compared to the reference. Under CMIP5, there is tendency towards earlier occurrence of Q max , except under RCP8.5 in the near future when no change compared to the reference is simulated. Q 10 occurs consistently earlier across all scenarios and in all river basins (figure S7).
The likelihood of simultaneously occurring Q max (ensemble median) at the confluence of the rivers increases under all scenarios when considering all three rivers, except in three combinations (10 d, SSP5-8.5 near future, SSP1-2.6 and RCP8.5 far future, figure 6) when no change is simulated. Such events have a magnitude of at least 200 000 m 3 s −1 and occur between zero (10 d window) to six times (30 d window) within a 30 year period under RCP8.5/SSP5-8.5 (ensemble median). The likelihood of simultaneously occurring Q max from two out of three river systems increases in most scenarios (89 out of 96 combinations (rivers, ensemble, scenario)). Exceptions are a decrease only for Brahmaputra-Meghna in the near future relative to the reference when forced by RCP2.6 (CMIP5). For all river combinations, higher chances of synchronized Q max is simulated during the reference period when SWIM is forced by CMIP5 climate data except for the combination Ganges-Meghna. Projections under CMIP5 RCP8.5 forcing show more frequently occurring simultaneous Q max for all combinations except Ganges-Meghna relative to CMIP6 SSP5-8.5 forcing.
The TCIE is positive, which means that the simulated 30 year moving average of Q max (ensemble median) exceeds historical Q max variability (expressed by the 95th percentile) of the reference period , in all basins until 2100 (figure 7). In the Ganges basin, TCIE occurs between the years 2034 and 2042 (all scenarios), which is considerably earlier compared to the other basins. TCIE is comparable between CMIP5 and CMIP6 based projections in the Ganges. The main difference between the CMIP phases is that under the SSP1-2.6 scenario, the simulated 30 year moving average of Q max remains outside of the envelope of the historic Q max variability after 2040, while under RCP2.6, mitigation efforts result in returning to the historical Q max flow variability after ∼2060. Projections show positive TCIE after the year 2070 in the BM basins under the scenarios SSP5-8.5/RCP8.5, ranging between 2073 (CMIP5) and 2078 (CMIP6) in the Brahmaputra and between 2073 (CMIP5) and 2095 (CMIP6) in the Meghna basin. This implies that the TCIE occurs earlier (5 years the Brahmaputra, 22 years Meghna) under CMIP5 compared to CMIP6 forcing. Q max stays below the 95th percentile threshold (and therefore within the envelope of the reference Q max variability) under RCP2.6/SSP1-2.6 in the BM rivers.

Discussion and conclusion
Our flood simulations for the Ganges-Brahmaputra-Meghna basins using the latest climate projections (CMIP6) indicate increases in high flows (Q 10 ) and floods (Q max ) in the 21st century, which is in line with previous estimates based on CMIP5 , Mohammed et al 2018, Uhe et al 2019. The projections show consistently that the Ganges basin is most prone to increased flooding, both in terms of magnitude and TCIE, under future climate change. In the BM basins, projections for high flows and floods are consistently lower (magnitude and TCEI) based on CMIP6 compared to CMIP5 scenarios. In the Meghna basin, the differences are most pronounced (e.g. TCEI occurs 20 years earlier under CMIP5 versus CMIP6 projections). While the overall variability across the ensemble remains high, our findings suggest robust increases in flood characteristics in the GBM basins.
High flow (Q 10 ) and flood (Q max ) projections under RCP2.6/SSP1-2.6 and RCP8.5/SSP5-8.5 are largely similar up to the middle of the century (∼2050), suggesting that climate mitigation efforts will only become effective in the second half of the century. Stringent climate mitigation policies have the potential to limit the climate impact on floods to within the envelope of historical variability, as suggested by our TCIE analysis (figure 7). This would considerably reduce adaptation costs. Continuing high greenhouse gas emissions will result in considerable increases in high flows and floods in the second part of the century. The TCIE analysis highlights that major shifts in flooding due to climate change are projected to occur after 2070 in the BM basins but likely as early as ∼2040 in the Ganges basin under RCP8.5/SSP5-8.5. The strong increase in high flows and floods under RCP8.5/SSP5-8.5 will translate into increased flood risk and likely devastating socioeconomic impacts. Our results indicate that existing flood-protection measures and infrastructure may not be effective and/or sufficient in the future, and that adaptations are needed already in the near future.
Despite the robust signal of increasing floods under climate change, we acknowledge that our approach neglects the uncertainties associated with the hydrological model, as only one hydrological model was used. However, most studies agree that the largest share of uncertainty in climate change impact assessments on water resources stems from the forcing climate models (Gädeke et al 2014 and projecting summer monsoon season precipitation is particularly challenging (Lee et al 2018). In addition, our hydrological model setup could be refined by additionally considering climate-induced changes in glacierized areas (kept constant in our setup) and water management/ regulation measures, such as water withdrawals, retention, and storage which play a role in the Ganges basin. However, as floods are primarily caused by (summer) monsoon precipitation and cyclone events in the GBM basins (e.g. Lutz et al 2014), glacier melt only plays a minor role during extreme events. Water management/regulation measures are not included due to unavailability of observations/lack of data but are also expected to only play a minor role during extreme flooding. Mohammed et al (2018) have pointed out that most water management and regulation measures are designed to be used for water storage rather than flood protection. For example, the Farakka barrage, located in India about 18 km upstream of the border to Bangladesh, diverts water during the dry season and is kept open during the monsoon season to avoid flooding in India.
The climate forcing data poses another uncertainty source. We used the meteorological output of the EWEMBI and W5E5 datasets during the calibration and validation of SWIM. The EWEMBI dataset also served as a basis for bias-adjusting the CMIP5 forcing, while W5E5 was used for bias-adjusting the CMIP6 forcing. In the GM basins, the differences in Q max based on CMIP5-SWIM and CMIP6-SWIM forcings are small, while larger differences exist in the Brahmaputra basin (figures 3 and S6). The biasadjustment and downscaling method used also introduces differences between the CMIP generations. For CMIP6, the improved bias-adjustment and the enhanced downscaling method were found to result in better representation of extreme events and spatial variability (Lange 2019), compared to the approach used for the CMIP5 climate data.
In conclusion, our study provides flood projections using the newest generation of climate projections bias-adjusted for impact assessments. Our results stress the need for considering a potential increase in high flows (up to +44%) and floods (up to +60%) from the three major rivers draining to Bangladesh as well as an increased risk of flood peak synchronization when adapting water management measures in Bangladesh and its upstream neighbors. High emission scenarios suggest a change in flood regimes as early as 2030 in the Ganges implying that flood management will face fundamentally changed risk profiles. This will require substantial investments and, ideally, transparent transboundary coordination of flood risk management considering climate change impacts. The results also indicate how important it is to mitigate climate change and what can be gained when following ambitious reduction targets, whereas failing in this task will imply devastating effects on one of the world's most vulnerable region to climate change in the world.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files).