South Asia river ﬂow projections and their implications for water resources

South Asia is a region with a large and rising population, a high dependence on water intense industries, such as agriculture and a highly variable climate. In recent years, fears over the changing Asian Summer Monsoon (ASM) and rapidly retreating glaciers together with increasing demands for water resources have caused concern over the 5 reliability of water resources and the potential impact on intensely irrigated crops in this region. Despite these concerns, there is a lack of climate simulations with a high enough resolution to capture the complex orography and water resource analysis is limited by a lack of observations of the water cycle for the region. In this paper we present the ﬁrst 25 km resolution regional climate projections of river ﬂow for the South Asia region. Two 10 global climate models (GCMs), which represent the ASM reasonably well are downscaled (1960–2100) using a regional climate model (RCM). In the absence of robust observations, ERA-interim reanalysis is also downscaled providing a constrained estimate of the water balance for the region for comparison against the GCMs (1990–2006). The RCM river ﬂow is routed using a river-routing model to allow analysis of present day and future river ﬂows 15 through comparison with available river gauge observations. We examine how useful these simulations are for understanding potential changes in water resources for the South Asia region. In general the downscaled GCMs capture the seasonality of the river ﬂows but overestimate the maximum river ﬂows compared to the observations probably due to a positive rainfall bias and a lack of abstraction in the model. The simulations suggest an increasing 20 trend in annual mean river ﬂows for some of the river gauges in this analysis, in some cases almost doubling by the end of the century. The future maximum river ﬂow rates still occur during the ASM period, with a magnitude in some cases, greater than the present day natural variability. Increases in river ﬂow could mean additional water resource for irrigation, the largest

global climate models (GCMs), which represent the ASM reasonably well are downscaled (1960-2100) using a regional climate model (RCM). In the absence of robust observations, ERA-interim reanalysis is also downscaled providing a constrained estimate of the water balance for the region for comparison against the GCMs (1990GCMs ( -2006. The RCM river flow is routed using a river-routing model to allow analysis of present day and future river flows 15 through comparison with available river gauge observations. We examine how useful these simulations are for understanding potential changes in water resources for the South Asia region. In general the downscaled GCMs capture the seasonality of the river flows but overestimate the maximum river flows compared to the observations probably due to a positive rainfall bias and a lack of abstraction in the model. The simulations suggest an increasing 20 trend in annual mean river flows for some of the river gauges in this analysis, in some cases almost doubling by the end of the century. The future maximum river flow rates still occur during the ASM period, with a magnitude in some cases, greater than the present day natural variability. Increases in river flow could mean additional water resource for irrigation, the largest usage of water in this region but has implications in terms of inundation risk. 25 These projected increases could be more than countered by changes in demand due to depleted groundwater, increases in domestic use or expansion of water intense industries.
Including missing hydrological processes in the model would make these projections more robust but could also change the sign of the projections.

Introduction
South Asia, the Indo-Gangetic plain in particular, is a region of rapid socio-economic change where both population growth and climate change is expected to have a large impact on 5 available water resource and food security. The region is home to almost 1.6 billion people and the population is forecast to increase to more than 2 billion by 2050 (United Nations, 2013). The economy of this region is rural and highly dependant on climate sensitive sectors such as the agricultural and horticultural industry, characterised by a large demand for water resources. As a result, over the coming decades, the demand for water from all sectors; 10 domestic, agricultural and industrial is likely to increase (Gupta and Deshpande, 2004;Kumar et al., 2005).
The climate of South Asia is dominated by the Asian Summer Monsoon (ASM), with much of the water resource across the region provided by this climatological phenomena during the months of June-September (Goswami and Xavier, 2005). The contribution from 15 glacial melt to water resources is less certain but likely to be important outside the ASM period during periods of low river flow . Glaciers and seasonal snowpacks are natural hydrological buffers releasing water during the drier periods such as spring and autumn when the flows of some catchments in this region are at their lowest. Similarly they may act to buffer inter-annual variability as well releasing water 20 during warmer drier years and accumulating during wetter colder years (Barnett et al., 2005). However, Kaser et al. (2010) show that the influence of glacial melt reduces with distance downstream, as other influences such as evaporation and precipitation increase in importance. Immerzeel et al. (2010) found that by the 2050s the main upstream water supply could decrease due to a reduction in snow and glacial melt (reductions of 8 % 25 for the upper Indus and more than 18 % for the Ganges and Brahmaputra). Meltwater plays an important role for the Indus and Brahmaputra particularly, accounting for a larger percentage of the downstream flow than the Ganges (where meltwater is approximately 10 % of the downstream flow). However Immerzeel et al. (2010) also show that these reductions in melt water are offset by an increase in precipitation in all three basins. Immerzeel et al. (2010) use coarse resolution general circulation models (GCMs) known to have difficulties in capturing monsoon precipitation and in estimating the relationship 5 between daily mean temperature and melting of snow and ice.
Recent studies have highlighted uncertainty in both glacier mass balance and ASM rainfall. Fujita and Nuimura (2011) show a negative mass balance for three benchmark glaciers in the Nepal Himalaya. Bolch et al. (2012) and Gardelle et al. (2013) highlight losses more generally from western, eastern and central Himalayan glaciers. These observed 10 changes in Himalayan glaciers can be attributed to the increase in temperature already experienced across the region, with warming more pronounced at higher elevations and during winter months (Shrestha and Aryal, 2011). There are however some glaciers in the Karakoram region showing increases in mass which has been attributed to a decrease in temperature for this region (Bolch et al., 2012;Gardelle et al., 2013). Projections of future 15 glacial change are challenging due to poor understanding of glacial processes, diversity in climate extremes and the complex orography of the region (Bolch et al., 2012). Complex orography contributes to other processes such as avalanching and therefore debris cover. The relationship between debris cover and melt is complex with a wide variety of responses across different glaciers across the Himalayan arc (Gardelle et al., 2013). The thickness of 20 debris cover is widely thought to significantly affect the response of the glacier to climate, with thick debris cover tending to slow down surface melting (Bolch et al., 2012;Scherler et al., 2011). However on the regional scale Kääb et al. (2012) found, using satellite data, similar thinning rates between clean and debris covered ice despite insulation by debris cover at some sites. Kääb et al. (2012) suggest that the insulating effect of debris layers 25 with thicknesses exceeding a few centimetres depends on the continuity of the coverage. Therefore changes in the thickness of debris across a glacier could change the melt rate on a local scale even across a single glacier tongue.

4
The ASM is also uncertain, Christensen et al. (2007) highlight two climate features that could influence the ASM, including a general weakening of monsoonal flows while enhanced moisture convergence could increase precipitation. Any reduction in water availability from either resource is likely to put more pressure on groundwater resources which is not sustainable in the longer term (Rodell et al., 2009). There is some disagreement 5 in the literature regarding the main effects of climate change on this region. Gregory et al. (2005) suggest that the availability and quality of ground water for irrigation could be more important factors influencing food security than the direct effects of climate change, particularly for India. However, Aggarwal et al. (2012) suggest that an increase in extremes (both temperature and precipitation) could lead to instability in food production and it is this 10 variability in food production that is potentially the most significant effect of climate change for the South Asia region.
Despite the general uncertainty in the reliability of water resources and the impacts of climate change for this region there are few simulations available with a high enough resolution for capturing the complex topography of the Himalayan region. The water 15 balance for the South Asia region as a whole is generally poorly understood with limited observing networks and data availability for both precipitation and river flows presenting a real challenge for validating models and estimates of water balance. This analysis seeks to use regional climate simulations to develop our understanding of the water cycle for the region in the context of the complete climate system, while acknowledging that more 20 needs to be done to address the missing hydrological processes in the model. Regional Climate Model (RCM) simulations are a widely used method across climate science for downscaling GCMs, including the regional IPCC assessment but are used in many other regional climate projects (Christensen et al., 2007;Murphy et al., 2009;Jacob et al., 2007). RCMs are based on the same physical equations as GCMs and therefore represent 25 the entire climate system including the carbon and water cycle. Though there are some limitations due to missing processes, their higher resolution allows a better representation of the regional scale processes; especially in regions of complex topography such as the Himalaya (Lucas-Picher et al., 2011). RCMs are designed to maintain the conservation of 5 water, mass, energy and momentum, essential for analysis on climate timescales. Lucas-Picher et al. (2011) conduct a comprehensive assessment of four RCMs run over South Asia demonstrating their ability to capture the monsoon; this analysis includes the RCM used here. Mathison et al. (2013) compare GCM and RCM outputs for temperature and precipitation specifically for the RCM used in this analysis. 5 Perhaps due to the lack of adequate resolution regional climate simulations available for this region there are relatively few studies that consider the value of downscaling using RCMs for hydrological applications for this region. However Akhtar et al. (2008) found that RCM data produced better results when used with a hydrological model than using poorquality observation data; this implies greater confidence in the RCM simulated meteorology 10 than available observational data for this region (Wiltshire, 2013). Therefore in the literature hydrological analysis is typically at the global scale using GCMs coupled with hydrological models (Milly et al., 2005;Hirabayashi et al., 2008;Falloon et al., 2011;Wiltshire et al., 2013a, b) or at the basin scale using stand-alone hydrological models (Singh and Kumar, 1997;Singh and Bengtsson, 2005;Singh et al., 2008;Seidel et al., 2000) such as the 15 Soil Water Assessment Tool (SWAT -Arnold et al., 1998). Weather data in SWAT is either simulated within the model using a weather generator or taken from observations of daily precipitation and maximum/minimum temperature (Nyeko, 2015). This approach may be appropriate for small domains within which there is consistency in rainfall patterns but may not be suitable for large domains in South Asia due to the high temporal and 20 spatial variability in precipitation across the region (Hijioka et al., 2014). Gosain et al. (2006) use the SWAT model with 50 km resolution daily RCM weather data to conduct a climate change impact assessment of the hydrology of several individual basins over India for two 20-year periods representing the present day (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) and future (2041-2060). Gosain et al. (2006) compare the differences between the two periods, rather than focussing 25 on absolute values, to find that climate change causes an increase in precipitation, river flow and evaporation for the Ganges basin. High variability across basins and sub-basins means that parts of the Ganges basin could experience seasonal or regular water-stressed conditions under climate change (Gosain et al., 2006) although it is not clear exactly which 6 climate change scenario has been used for these simulations. There are more examples of the application of RCMs for hydrological analysis for other regions such as the UK and Europe. Kay et al. (2006) use 25km RCM data in a catchment-based rainfall-runoff model to estimate the flood frequency of small UK river basins to good effect. Dankers et al. (2007) use an RCM to evaluate the benefits of using high spatial resolution climate information for 5 the Danube basin. Sampson et al. (2014) have also demonstrated the importance of the resolution of precipitation data for a region of Ireland for hydrological impact modelling.
The typical domain and resolution of RCM simulations enables the analysis of areas spanning multiple river basins covering a larger area than is usually possible with hydrological models. This means that there is consistent forcing across different basins. 10 The use of the RCM generated runoff within the hydrological model also preserves the consistency of the projections with atmospheric forcing which is not possible if the runoff is derived within a hydrological model. However there are few regional river flow analyses currently available, where these consistencies are maintained. Biemans et al. (2013) analyse the RCM projections used in this analysis in terms of water availability for food 15 production for selected river basins using a coupled hydrology and dynamic vegetation model, however so far no specific analysis of river flows has been done for these RCMs.
Therefore we present the first 25 km resolution regional climate projections of river flow for the South Asia region by using RCM generated runoff within a routing model to estimate river flow thereby enabling consistency to be maintained across basins and with the driving 20 climate scenario. This is a new application of the highest resolution RCM data currently available for this region, to enable analysis of the impacts of climate on river flows in conjunction with the strategic sampling of climate variability from selected GCMs. We use a novel approach to the consideration of variability of river flows through analysis of the upper and lower parts of the distribution, in addition to the mean flows. 25 The aim of this analysis is to examine how useful RCM simulations are for understanding how river flows could change in South Asia in the future. Irrigation is an important part of the agricultural industry for this part of the world, with the Indo-Gangetic plains traditionally providing the staple crops of rice and wheat (Aggarwal et al., 2000) for India and South Asia as a whole; the continued success of these crops is therefore important for the food and water security of the region. We discuss the potential implications of projected changes in the water resources needed to maintain yields of these crops in a changing climate. The models, observations and the analysis used are described in Sect. 2, while a brief evaluation of the driving data and the river flow analysis is presented in Sect. 3. The 5 implications of the potential changes in river flows on water resources and conclusions are discussed in Sects. 4 and 5 respectively.

Models
Figure 1 summarises the methodology described in this section in a flow chart, highlighting 10 the main stages in the generation of the presented river flow projections and the approximate resolution of the model data used.

GCM and RCM forcing
This analysis utilizes 25 km resolution regional climate modelling of the Indian subcontinent to provide simulations across the Hindu-Kush Karakoram Himalaya mountain 15 belt. These RCM simulations form part of the ensemble produced for the EU-HighNoon project (referred to hereafter as HNRCMs), for the whole of the Indian subcontinent (25 • N, 79 • E-32 • N, 88 • E), for the period 1960-2100. The other simulations in the HighNoon ensemble which used another RCM, the Regional Model from the Max Planck Institute for Meteorology (REMO -Jacob et al., 2007) were unavailable for use in this analysis. 20 Therefore one RCM, the HadRM3 RCM  is used to generate the river flow projections presented here. While the additional RCM would be useful here, analysis of the ERAint driven HadRM3 and REMO simulations over the Western Himalaya by Dimri et al. (2013) show that both models run at 25 km resolution over comparable domains have similar distributions of precipitation, temperature and interannual variablility, despite 25 8 having different representations of orography. Analysis of the complete HighNoon ensemble for the Ganges/Brahmaputra basin in Mathison et al. (2013) also indicate a small spread between HNRCMs for the 30-year mean climatologies of temperature. Precipitation is more variable for this basin, with a larger spread between HNRCMs. However the RCM uncertainty defined by these two models (REMO and HadRM3) is still smaller than the 5 climate uncertainty represented by the selected GCMs with the influence of the GCM on the projections of precipitation as great as the variability between RCMs . Therefore the most important contribution to the input uncertainty is from the GCM (Akhtar et al., 2008) and using two GCMs to provide boundary data to one RCM provides a better estimate of climate uncertainty than using a single GCM to drive two RCMs, which would 10 be the computational equivalent. On this basis we use HadRM3 driven by two carefully selected GCMs for this analysis. However other RCMs, not yet applied to this region could produce different projections.
In order to sample climate uncertainty, we use two GCM simulations that have been shown to capture a range of temperatures and variability in precipitation similar to the AR4 15 ensemble for Asia (Christensen et al., 2007). Although using just two ensemble members is unlikely to capture the full range of uncertainty of a larger ensemble, the two models used for these simulations have been shown to capture the main features of the large-scale circulation particularly the ASM Annamalai et al., 2007;Mathison et al., 2013) which is not true of all GCMs. The experimental design of the HighNoon ensemble 20 compromises between the need for higher resolution climate information for the region and the need for a number of ensemble members to provide a range of uncertainty. The length of the simulations needed and the limited number of GCMs that are able to simulate the ASM also affect the number of ensemble members. These factors are all important given the limited computational resources available. The GCMs; The Third version of the Met Office 25 Hadley Centre Climate Model (HadCM3 - Pope et al., 2000;Gordon et al., 2000, a version of the Met Office Unified Model) and ECHAM5 (3rd realization - Roeckner et al., 2003) are downscaled using the HadRM3 RCM . These two GCMs capture the uncertainty in the sign of the projected change in precipitation with one showing an 9 increase (HadCM3) and the other a decrease (ECHAM5). This feature is a key reason for the selection of these two GCMs. In addition to the GCMs, ERA-interim data (Simmons et al., 2007;Dee et al., 2011) is also downscaled using the HadRM3 RCM. ERA-Interim is reanalysis product that combines model and observations to provide a constrained estimate of the water balance of the region. The ERAinterim simulation has also been shown to 5 capture the role of steep topography on moisture transport fluxes and vertical flow for the western Himalayas (Dimri et al., 2013). Therefore, for this region, where there is a lack of robust observations, particularly of the water cycle (see Sects. 2.3.1 and 3.1), it provides a useful benchmark against which to compare the GCM driven simulations. A similar approach is described in a previous study by Wiltshire (2013). 10 These RCM simulations are currently the finest resolution climate modelling available for this region Moors et al., 2011;. HadRM3 has 19 atmospheric levels and the lateral atmospheric boundary conditions are updated 3 hourly and interpolated to a 150 s timestep. These simulations include a detailed representation of the land surface in the form of version 2.2 of the Met Office Surface Exchange Scheme 15 (MOSESv2.2, Essery et al., 2003) which includes a full physical energy-balance snow model (Lucas-Picher et al., 2011). MOSESv2.2 treats subgrid land-cover heterogeneity explicitly with separate surface temperatures, radiative fluxes (long wave and shortwave), heat fluxes (sensible, latent and ground), canopy moisture contents, snow masses and snowmelt rates computed for each surface type in a grid box (Essery et al., 2001). However 20 the air temperature, humidity and wind speed above the surface are treated as homogenous across the gridbox and precipitation is applied uniformly over the different surface types of each gridbox. The relationship between the precipitation and the generation of runoff is complicated, depending on not only the intensity, duration and distribution of the rainfall but also the characteristics of the surface. The infiltration capacity of the soil, the vegetation 25 cover, steepness of the orography within the catchment and the size of the catchment are important influencing factors on runoff generation (Linsley et al., 1982). In GCMs and even 25 km RCMs such as the one presented here, the resolution is often too coarse to explicitly model the large variations of soil moisture and runoff within a catchment and 10 therefore the major processes are parameterized (Gedney and Cox, 2003). The method used within MOSES2.2 for generating surface and subsurface runoff across a gridbox is through partitioning the precipitation into interception by vegetation canopies, throughfall, runoff and infiltration for each surface type (Essery et al., 2003). The Dolman and Gregory (1992) infiltration excess mechanism generates surface runoff; this assumes an exponential 5 distribution of point rainfall rate across the fraction of the catchment where it is raining (Clark and Gedney, 2008). Moisture fluxes are allowed between soil layers; these are calculated using the Darcy equation, with the water going into the top layer defined by the gridbox average and any excess removed by lateral flow (Essery et al., 2001). Excess moisture in the bottom soil layer drains from the bottom of the soil column at a rate equal 10 to the hydraulic conductivity of the bottom layer as subsurface runoff (Clark and Gedney, 2008). The performance of MOSESv2.2 is discussed in the context of a GCM in Essery et al. (2001), however no formal assessment of MOSESv2.2 and the runoff generation in particular has been done for the RCM. 15 In this analysis the simulated 25km gridbox runoff is converted into river flow using the 0.5 • Total Runoff Integrating Pathways river routing scheme (TRIP; Oki and Sud, 1998) as a post-processing step. TRIP is a simple model that moves water along a pre-defined 0.5 • river network; the Simulated Topological Network at 30 min resolution (STN-30p, version 6.01; Vörösmarty et al., 2000a, b;Fekete et al., 2001) in order to provide mean runoff per 20 unit area of the basin; this can be compared directly with river gauge observations. TRIP has been used previously in Falloon et al. (2011) which used GCM outputs directly to assess the skill of a global river-routing scheme. The TRIP model has been shown to agree well with observed river flow gauge data (Oki et al., 1999) and largely showed good skill when comparing runoff from several land surface models (Morse et al., 2009). Implementation 25 of TRIP in two GCMs; HadCM3 and HadGEM1 is described by Falloon et al. (2007) and was found to improve the seasonality of the river flows into the ocean for most of the major rivers. Using TRIP ensures the river flow forcing is consistent with the atmospheric forcing, 11 however it also assumes that all runoff is routed to the river network and as such there is no net aquifer recharge/discharge. This may not be the case in regions with significant ground water extraction which is subsequently lost though evaporation and transported out of the basin. These simulations do not include representation of extraction, reservoirs or dams. Many of the river gauges used in this analysis and described in Section 2.2 are 5 located at large dams along rivers in these basins and therefore the comparison between the simulations and the river gauges could be affected by these large features. Extraction, particularly for irrigation purposes is large in this region (Biemans et al., 2013); this means that the extraction-evaporation and subsequent recycling of water in a catchment (Harding et al., 2013;Tuinenburg et al., 2014) is not considered in this analysis. The routed runoff of the HNRCM simulations are generally referred to hereafter using only the global driving data abbreviations; ERAint, ECHAM5 and HadCM3 (except Sect. 3.1 where we refer to the HadCM3 GCM and ERAint datasets before downscaling).

Emission Scenario
These simulations use the SRES A1B scenario (Nakicenovic et al., 2000). The SRES 15 scenarios were devised according to the production of greenhouse gases and aerosol precursor emissions as part of the AR4 IPCC report (Christensen et al., 2007). The A1 storyline and scenario family represents a future world of very rapid economic growth, global population that peaks in mid-century and declines thereafter, and rapid introduction of new and more efficient technologies. The A1B scenario specifically, represents this future world 20 where there is balance across energy sources i.e. a mixture of fossil and non-fossil fuels (Nakicenovic et al., 2000). This scenario does not represent changes in landuse which remains fixed through the duration of these simulations. This is useful for understanding the effect of climate change in the absence of any adaptation.

Observations
This analysis uses observations of precipitation and river flow to assess the present day RCM hydrology. The precipitation observations are from the Asian Precipitation-Highly Resolved Observational Data Integration Towards the Evaluation of Water Resources (APHRODITE -Yatagai et al., 2012) dataset. APHRODITE is a daily, 0.25 • resolution 5 gridded dataset.
The river flow analysis focuses on a selection of river gauges from the Global Runoff Data Centre (GRDC, 2014) that are located within the three major river basins for South Asia; the Indus and the Ganges/Brahmaputra. These gauges provide observations which are used, in addition to downscaled ERA-interim river flows, to evaluate the downscaled 10 GCM river flows. The selection of these river gauges aims to illustrate from the perspective of river flows as modelled in an RCM, that the influence of the ASM on precipitation totals increases, from west to east and north to south across the Himalayan mountain range, while that of western disturbances reduces (Wiltshire, 2013;Dimri et al., 2013;Ridley et al., 2013;Collins et al., 2013). The differing influences across the Himalayan arc result in complex 15 regional differences in sensitivity to climate change; with western regions dominated by non-monsoonal winter precipitation and therefore potentially less susceptible to reductions in annual snowfall (Wiltshire, 2013;Kapnick et al., 2014). A brief geographical description of the rivers and the chosen gauges is given in this section, their locations are shown in Fig. 2 and listed in Table 1 (including the abbreviations shown in Fig. 2 and the gauge location in 20 terms of latitude and longitude).
The Indus, originates at an elevation of more than 5000 m in western Tibet on the northern slopes of the Himalayas, flowing through the mountainous regions of India and Pakistan to the west of the Himalayas. The upper part of the Indus basin is greatly influenced by western disturbances which contribute late winter snowfall to the largest 25 glaciers and snow fields outside the polar regions; the meltwaters from these have a crucial role in defining the water resource of the Indus basin (Wescoat Jr, 1991). In this analysis the Attock gauge is the furthest upstream and the Kotri gauge, located further downstream 13 provide observations on the main trunk of the Indus river. The Chenab river, located in the Panjnad basin and in this analysis represented by the Panjnad gauge, is a major eastern tributary of the Indus, originating in the Indian state of Himachal Pradesh. In the upper parts of the Chenab sub-basin western disturbances contribute considerably to precipitation while the foothills are also influenced by the ASM (Wescoat Jr, 1991). 5 The Ganges river originates on southern slopes of the Himalayas (Thenkabail et al., 2005) and traverses thousands of kilometres before joining with the Brahmaputra in Bangladesh and emptying into the Bay of Bengal (Mirza et al., 1998). The Ganges basin has a population density 10 times the global average making it the most populated river basin in the world (Johnston and Smakhtin, 2014), it covers 1.09 million km 2 with 79 % in 10 India, 13 % in Nepal, 4 % in Bangladesh and 4 % in China (Harding et al., 2013). The main trunk of the Ganges is represented in this analysis by the gauge at the Farakka barrage, located at the India-Bangladeshi border, to the East of the Himalayas. The Bhagirathi river, located in "the Upper Ganga basin", is one of the main head streams of the Ganges.
The Bhagirathi river originates from Gaumukh 3920 m a.s.l. at the terminus of the Gangotri 15 glacier in Uttarakhand, India (Bajracharya and Shrestha, 2011). The Tehri dam is located on this tributary, providing the most central data point on the Himalayan arc in this analysis (not a GRDC gauge).
The Karnali river (also known as Ghaghara), drains from the Himalaya originating in Nepal flowing across the border to India where it drains into the Ganges. The Karnali is 20 the largest river in Nepal and a major tributary of the Ganges (Bajracharya and Shrestha, 2011) accounting for approximately 11 % of the Ganges discharge, 5 % of its area and 12 % of its snowfall in the HNRCMs. Two of the river gauges in this analysis; the Benighat and the Chisapani are located on this river. Two other sub-catchments complete those covering the Ganges basin; the Narayani river (or Gandaki River, represented here by the Devghat 25 river gauge); reportedly very dependant on glaciers at low flow times of the year with over 1700 glaciers covering more than 2200 km 2 (Bajracharya and Shrestha, 2011). The Arun river, part of the Koshi river basin originates in Tibet, flows south through the Himalayas to 14 Nepal. The Arun, represented in this analysis by the Turkeghat gauge joins the Koshi river which flows in a southwest direction as a tributary of the Ganges.
The Brahmaputra originates from the glaciers of Mount Kailash at more than 5000 m a.s.l., on the northern side of the Himalayas in Tibet flowing into India, and Bangladesh before merging with the Padma in the Ganges Delta. The Brahmaputra is prone to flooding due 5 to its surrounding orography and the amount of rainfall the catchment receives (Dhar and Nandargi, 2000). The Brahmaputra is represented in this analysis by three gauges; Yangcun, the highest upstream gauge, Pandas in the middle and Bahadurabad furthest downstream but above the merge with the Padma.
There are no known observation errors for the GRDC observations (personal 10 communication, GRDC). Estimates of observation errors for river gauges vary in the literature with a recommendation in Falloon et al. (2011) for GCMs to be consistently within 20 % of the observations while Oki et al. (1999) suggest that errors of 5 % at the 95 % confidence interval might be expected. McMillan et al. (2010) propose a method for quantifying the uncertainty in river discharge measurements by defining confidence bounds. 15 In this analysis, these methods are hindered by the lack of observations concurrent with the model simulations. Therefore the method for approximating the inter-annual variability in this analysis is based on the model variability and is described in Sect. 2.3.

Methods
There are two stages to the analysis presented, comparison of the simulations with 20 observations (for both RCM precipitation and river flows) and analysis of future climate. The comparison against observations aims to assess if the RCM reproduces the regional hydrology in terms of precipitation and river flow compared with available observations. The objective of the analysis of future climate is to understand how these simulations compare against the present day high and low flows i.e. present day natural variability. 25 In this section we describe the methods used in each stage of the analysis; the comparison against observations is described in Sect. 2.3.1 and the analysis of future river flows in Sect. 2.3.2.

Comparison against observations
The total precipitation from each of the downscaled GCM simulations are compared against a downscaled ERAinterim simulation and APHRODITE observations. This comparison is on the basin scale, focussing on the basins included in the river flow analysis (see Sect. 2.2); the Indus and the Ganges/Brahmaputra. The TRIP model basin boundaries for each of 5 these basins are shown in Fig. 4. The Ganges and Brahmaputra catchments are considered together in this analysis as these rivers join together in the Ganges Delta and are not clearly delineated in TRIP (see Fig. 4b). The precipitation patterns for each basin are useful for understanding the changes in the river flows within the catchments although rain gauges in the APHRODITE dataset are particularly sparse at higher elevations (see Yatagai et al.,10 2012, Fig. 1). This leads to underestimation of the basin wide water budgets particularly for mountainous regions (Andermann et al., 2011). This is confirmed by Immerzeel et al. (2015) for the Indus basin where they find a high altitude precipitation of up to ten times higher than current gridded datasets is needed to close the water balance for this basin. We compare the observations and simulations in terms of their annual timeseries and the 15 climatology for each basin. The climatologies are calculated using the 1971-2000 period for HadCM3 andECHAM5 and1990-2006 for the ERAint simulation in order to capture a typical seasonal cycle for each simulation and basin. This analysis is repeated for river flows in Sect. 3.2 for each of the 12 gauges described in Sect. 2.2. We also calculate the 1.5 standard deviation (SD) over a 30 year period to 20 define the inter-annual variability. A value of plus 1.5 SD indicates an approx 1 in 10 year wet event, a value of minus 1.5 SD indicates a 1 in 10 year dry event. This approach is taken to indicate the possible impact of such a change under the hypothesis that current socio-economic levels of climate adaptation can cope with a 1 in 10 year events. The change driving mechanism could be anthropogenic climate or decadal variability. This 25 assumes that interannual variability is independent of climate change whether that is due to decadal variability or externally forced change. In this context it is indicative of the timing and magnitude of possible changes under the A1B emissions scenario. More work and ensemble members would be required to control for the role of decadal variability while the substantial computation expense in running high-resolution RCM experiments currently precludes the use of initial condition ensembles.

Future analysis
In Sect. 3.3 we use the annual timeseries of the whole simulation period to highlight 5 any trends in future precipitation, evaporation (at the basin scale) and river flows (for each gauge) over the century. We also calculate the climatologies for two future 30-year periods; 2040-2070 (referred to as the 2050s) and 2068-2098 (referred to as the 2080s). The monthly climatology for the two periods is compared against the 1971-2000 range of natural variability. The purpose of the climatology analysis is twofold. The first objective is to 10 establish if there is any change in the seasonality of the river flow. The second objective is to establish if there is any increase in the future 30-year mean river flows that is outside the present day variability, thereby indicating an increase in future events that are equivalent to the 1971-2000 1 in 10 year wet (dry) events (see Sect. 3.3.1).
Analysis of the 30 year mean is useful for understanding the general climatology of the 15 region but often it is the periods of high and low river flow that are critical in terms of water resources. Mathison et al. (2013) highlight the importance of potential changes in the seasonal maximum and minimum river flows for the agricultural sector. The analysis in Sect. 3.3.2 uses Kernal Density Estimation (KDE, (KDE, Scott, 2009;Silverman, 1986)) to calculate the probability density functions (pdfs) of the river flows for each river gauge and 20 30 year period. The main aim of this analysis is to establish if there is any change in the distribution of the highest and lowest river flows for the 2050s and 2080s compared with the 1971-2000 period (see Sect. 3.3.2). Given these distributions, we then attempt to quantify the changes in highest and lowest river flows for the two future periods by focussing on the changes in the lowest and highest 10 % of flows using two different approaches. In the first 25 approach in Sect. 3.3.3 we apply the upper and lower 10% of river flows for the 1971-2000 period as thresholds for the 2050s and 2080s. In Sect. 3.3.4, we take the principle of the threshold analysis one step further by calculating the 10th and 90th percentile threshold for each decade, simulation and gauge. The aim of this second approach is to establish if there is any systematic change in the upper and lower parts of the distribution through the century.

Results
The results are divided into three sections. Precipitation has a key influence on river flows, 5 therefore in Sect. 3.1 we consider the previous evaluation of the HNRCM simulations comparing the RCM precipitation for major South Asia basins with observations and ERAint. In Sect. 3.2 we focus on river flows themselves for 12 gauges within these basins distributed across the Himalayan arc. The methods used in Sect. 3.1 and 3.2 are described in Sect. 2.3.1). In Sect. 3.3 we analyse the future projections of precipitation, evaporation 10 and river flow to understand the water cycle of the region (see Sect. 2.3.2 for the methods used).

Comparison of present day driving data with observations
The HNRCM simulations have been evaluated in several previous publications. Lucas-Picher et al. (2011) evaluates the ability of RCMs to capture the ASM using ERA-15 40 data.  analysed the HNRCMs forced with ERA-Interim data. The GCM and HNRCM simulations are also evaluated against a range of observations for the Ganges/Brahmaputra river basin in Mathison et al. (2013). Figure 3 shows the observed spatial distribution of total precipitation for the monsoon period (June to September Goswami and Xavier, 2005) together with the HadCM3 and ERAint prior to 20 and post downscaling. The HNRCMs (Figure 3 d and e) improve the spatial distribution of precipitation and therefore compare well with the observations shown in Figure 3c. This is highlighted by the additional detail shown in the precipitation fields through comparison of the pre-downscaled datasets for the HadCM3 GCM ( Figure 3a) with those downscaled using HadRM3 (Figure 3e). This comparison is also possible for an improved precipitation representation compared with the pre-downscaled dataset ( Figure 3b). The higher resolution orography used in the 25km RCMs is more able than the much coarser resolution datasets to capture the particularly varied terrain of this region and the effects of this on the precipitation distribution. In general the HNRCM simulations capture the spatial characteristics of the ASM, successfully reproducing regions of high 5 convective precipitation, maximum land rainfall and the rain shadow over the east coast of India . Through adequately representing the spatial precipitation characteristics across the region, the areas of maximum and minimum precipitation can have a direct impact on the river flows for the appropriate basin. This is shown by the improvement in the timing and magnitude of the maximum precipitation for the RCM 10 (HadRM3) compared with the GCM (HadCM3) shown in Fig. 6c (Indus) and Fig. 6d (Ganges/Brahmaputra). The RCMs are also able to reproduce the inter-annual variability of the region although they underestimate the magnitude of the variation . The GCMs in the AR4 ensemble tend to exhibit cold and wet biases compared to observations both globally (Nohara et al., 2006) and for South Asia (Christensen et al., 15 2007). Although these are generally reduced in the RCM simulations there is a cold bias in the RCM that is probably carried over from the larger bias in the GCMs . The remaining analysis focuses on the downscaled simulations of HadCM3, ECHAM5 and ERAint using the HadRM3 RCM. The RCM simulations shown in Fig. 6 appear to 20 over estimate the seasonal cycle of total precipitation compared with the APHRODITE observations; this is highlighted by the annual mean of the total precipitation shown in Fig. 5. However given the limitations of the observations at high elevations discussed in Sect. 2.3.1 we compare HadCM3 and ECHAM5 against an ERAint simulation. The annual mean (Fig. 5) and the monthly climatology (Fig. 6) show that, for these catchments, the 25 ERAint simulation lies between the two HighNoon ensemble members for much of the year. However, during peak periods of precipitation the magnitude of total precipitation for ERAint is larger.
The seasonal cycles of total precipitation are distinctly different between the basins shown. The Indus basin (Fig. 6a), indicates two periods of precipitation; one smaller peak between January and May and another larger one between July and September. The timings of the largest peak compare well, however the smaller peak occurs later than both ERAint and APHRODITE for ECHAM5 and HadCM3. The magnitude of the 5 peaks in precipitation in the APHRODITE observations are consistently lower throughout the year than the simulations. The magnitude of the ERAint total precipitation is typically the largest while the ECHAM5 simulation is the lowest and closest to the APHRODITE observations. HadCM3 is between ECHAM5 and ERAint for most of the year. In contrast the Ganges/Brahmaputra catchment (Fig. 6b) has one strong peak between July and 10 September. In general this seasonal cycle is captured reasonably well by the simulations, both in terms of magnitude and timing of the highest period of precipitation. However there is a tendency for the simulations to overestimate rainfall between January and June compared to the observations, thus lengthening the wet season . Mathison et al. (2013) also show that in these simulations, the region of maximum precipitation along the 15 Himalayan foothills is displaced slightly to the north of that shown in the observations. One explanation for this could be that the peak in total precipitation is due to the distribution of observations already discussed. Alternatively it could be due to the model resolution, which may still be too coarse to adequately capture the influence of the orography on the region of maximum precipitation. The downscaled ERAint simulation also indicates 20 a higher total precipitation for January-May that is within the range of uncertainty of the GCM driven simulations. However for the remainder of the Monsoon period, ERAint has a higher total precipitation than the GCM driven simulations. Fig. 3d illustrates this, showing a slightly larger and more intense area of maximum rainfall over the eastern Himalayas for the downscaled ERAint simulation than shown in the other RCM simulations ( Fig. 3e and f) 25 or APHRODITE (Fig. 3c). 20

Present day modelled river flows
In this section we compare present day modelled river flows with observations and a downscaled ERAint simulation using annual average river flows (see Fig. 8) and monthly climatologies (see Fig. 9). It is clear from Fig. 8 that observed river flow data is generally limited which makes statistical analysis of the observations difficult. River flow data for 5 this region is considered sensitive and is therefore not readily available particularly for the present day. For each of the gauges shown here, there are generally several complete years of data but often the time the data was collected pre-dates the start of the model run. The ERAint simulation is also shown (cyan line-ERAint) to provide a benchmark in the absence of well-constrained observations (See Sect. 3.1). The comparison between the 10 model and observations shown in Figs. 8 and 9 is therefore to establish if the model and observations are comparable in terms of the average seasonal cycle and mean river flow rate without over-interpreting how well they replicate the observations. The Tehri Dam on the Bhagirathi river is not a GRDC gauge therefore observations are not shown. Observations for this gauge were received via personal communication from the Tehri Dam operator and 15 therefore could not be adequately referenced.
The Kotri gauge on the Indus (Fig. 9a) and the Yangcun gauge on the Brahmaputra (Fig. 9k) are the only two gauges where the modelled river flow is higher than the observations and not within the estimated variability (1.5 SD) of the region. The ERAint simulation is also outside the estimated variability (1.5 SD) for the Benighat gauge on the 20 Karnali river (Fig. 9e). The differences in these gauges are also reflected in the annual mean river flows (Fig. 8) for these river gauges which are higher than observed. The high bias in modelled river flow at the Kotri gauge could be due to the extraction of water which is not included in the model. The Indus has the largest irrigation scheme in the world and a semi-arid climate (Immerzeel et al., 2015) which means the extraction rate for this basin is 25 large (Biemans et al., 2013). This gauge is also located relatively close to the river mouth to the west of the Himalayas (see Fig. 2 and Table 1), therefore the river flows are less likely to be affected by the ASM and more likely to be affected by meltwater from winter precipitation. The Yangcun gauge is a more upstream gauge and the differences between the model and observations for this gauge are more likely to be related to the precipitation distribution. Fig. 3 shows a region of intense precipitation in the simulations (Fig. 3d, e and f) for the ASM period close to this gauge. The APHRODITE data (Fig. 3c) also shows a region of higher rainfall although this is not as large as that shown for the simulations. This 5 could be having a direct effect on the river flow.
The other two gauges on the Brahmaputra are located downstream of the Yangcun gauge; the Pandu (Fig. 9j) and Bahadurabad (Fig. 9l). At these two gauges, the seasonal cycle of river flow has a very broad peak particularly in the modelled river flows compared to the other gauges. In the simulations the snowfall climatology for the Ganges/Brahmaputra 10 basin (not shown) has a similar seasonal cycle to that of the river flow for the Bahadurabad and the Pandu gauges. It is therefore likely that the broad peak in river flow is related to the broad peak in snowfall and subsequent snowmelt. The Pandu gauge is also one of only two gauges where the modelled river flow is less than the observations for at least part of the year, the other being the Devghat gauge on the Narayani river (Fig. 9g). Both of these 15 gauges are located in the Himalayan foothills close to the region of simulated maximum total precipitation. If the simulations put the location of this maximum below these gauges this could cause the river flows at the gauges to be lower than observed. The river flow on the main trunk of the Ganges at the Farakka barrage (shown in Fig. 9i), is a reasonable approximation to the observations in terms of magnitude, however the timing of the peak 20 flow seems to be later in the models. It could be argued this also happens in some of the other gauges although it is more noticeable for the Farakka barrage. All the gauges shown here are for glacierized river basins. Snow fields and snow melt are represented in the simulations in this analysis and will therefore replicate some aspects of melt affecting river flow. However glacial melt is not explicitly represented in the RCM used for these 25 simulations. Including glacial processes specifically could act to reduce runoff because more snow is stored as ice or increase runoff where there is an increased melting (Bolch et al., 2012). Therefore including glacial processes could be important for the timing and magnitude of the maximum and minimum river flows for these catchments.

22
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

Future river flows
In this section we consider the future HNRCM simulations. Figure 5 highlights the variability in the future projections of total precipitation for South Asia between basins. In these simulations the Ganges/Brahmaputra catchment shows an increasing trend in total precipitation and there is considerable variation between the simulations (Fig. 5b). The 5 Indus basin (Fig. 5a), however, has a much flatter trajectory to 2100 and the simulations are more similar. The annual timeseries of evaporation (Fig. 7) over these catchments shows a similar picture, with an increasing trend for the Ganges/Brahmaputra basin (Fig. 7b) but no real trend for the Indus (Fig. 7a). The annual mean runoff efficiency (not shown), defined here as the ratio of annual runoff (streamflow per unit area) to annual precipitation, 10 shows no real trend for either basin. The trends in river flow (see Fig. 8) vary between gauges, although none indicate decreasing river flows. There is an upward trend in river flows at some of the gauges, in particular, the Narayani-Devghat (Fig. 8g), Arun-Turkeghat (Fig. 8, h) and Ganges-Farakka (Fig. 8i). These gauges suggest an upward trend toward the 2100s that actually represents a doubling of the river flow rate. The increase in river 15 flow for the Narayani-Devghat gauge (Fig. 8g) are consistent with analysis by Shrestha and Aryal (2011) using a hydrological model for the Narayani basin. Ganges-Farakka is the most downstream gauge in the Ganges/Brahmaputra basin in this river flow analysis therefore providing an approximation for the whole Ganges basin. These simulations show an increase in precipitation for the Ganges/Brahmaputra basin of approximately 20 % 20 (See Fig. 5) and an increase of approximately 10 % in evaporation (See Fig. 7), over the course of the century. This suggests the changes in runoff over the Ganges catchment are predominantly driven by precipitation on the annual scale. However regional analysis by Jhajharia et al. (2012) covering the humid northeastern part of India and a global analysis by McVicar et al. (2012) suggest there has been a decline in the evaporation caused by 25 lighter surface winds and reduced radiation. A future reduction in evaporation could also contribute to future increases in runoff. Analysis using a conceptual hydrological model by Singh and Bengtsson (2005) suggests that the type of precipitation being received at different elevations and the changes in melt and evaporation from snowpacks in a warmer climate could also be important for changes in runoff.

Climatology analysis
In this section we use climatologies to compare future river flows with the present day interannual variability (defined in Sect. 2.3.1). South Asia is a very variable region, yet these 5 models suggest the future mean river flow could lie outside the present day variability for peak flows for some of the gauges in this study. This could have important implications for water resources for the region. The gauges that show an increase in maximum river flows (see Fig. 10) are mainly those in the middle of the Himalayan arc (see Fig. 2). The seasonal cycle for the western most (Indus gauges) and the eastern most (Brahmaputra gauges) are 10 typically still within the range of present day variability. This could be due to the changes in the influence on river flow from west to east becoming more influenced by the ASM and less by western disturbances, with basins in the centre of the Himalayas and to the north influenced by both phenomena. Figure 10 also suggests that the maximum river flows still occur mainly during the ASM for many of the gauges shown.

High and low flow analysis
The analysis of the high and low flows is of particular importance to water resources and future availability, therefore in this section we calculate the distributions of the river flows for each of the gauges (see Sect. 2.3.1). These are shown in the form of probability density functions (pdfs) in Fig. 11 for the 1971-2000, 2050s and the 2080s. Figure 11 illustrates 20 how the lowest flows dominate the distributions for each of the three periods. In most of the gauges 1971-2000 period has the highest frequency of the lowest flows, the curves then tend to flatten in the middle of the distribution before tailing off toward zero for the highest flows. The two future periods also follow a similar trajectory, although in general there is a reduction in the frequency of the lowest flows and an increase in the magnitude of the 25 highest flows for all of the gauges and both simulations towards 2100.

24
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | The Yangcun gauge on the Brahmaputra (Fig. 11k) shows the least change of all the gauges between the 1971-2000 period, future periods and simulations. The distributions for the gauges downstream of Yangcun; the Pandu (Fig. 11l) and the Bahadurabad (Fig. 11j) are notable for their differences from all the other gauges. All the other gauges shown have a single peak toward the lower end of the river flow distribution. The Pandu and Bahadurabad 5 gauges have two distinct peaks in frequency with a second peak occurring toward the middle of the distribution, where the distribution for most other gauges flattens out. This is consistent with the broader peak in the 30-year mean seasonal cycle shown for these gauges in Fig. 10 and is probably similarly explained by snowmelt (see Sect. 3.2). In some of the other gauges there is a small increase in the middle of the river flow distribution but 10 this tends to be smaller and restricted to the two future periods e.g. the two Karnali river gauges ( Fig. 11e and f).

Threshold analysis
The pdfs shown in Fig. 11 and described in Sect. 3.3.2 suggest future changes in the lower and upper ends of the river flow distribution. In this section we consider these parts of the 15 distribution in order to confirm this pattern. We compare the two future periods (2050s and 2080s) against the 1971-2000 period explicitly using thresholds defined by the 10th and 90th percentiles for this present day period for each river gauge. Graphical examples from the results of this analysis are shown for all three periods (historical (a), 2050s (b), 2080s (c)) for the Farakka Barrage on the River Ganges in Fig. 12 and Fig. 13. In Fig. 12 the 20 number of months where river flow is below the present day 10th percentile reduces in each of the future decades. However for flows greater than the present day 90th percentile there is an increase in each of the future decades (Fig. 13). Table 2 illustrates that the patterns shown in Figs. 12 and 13 are generally true for almost every other gauge in the analysis. The Tehri Dam (Bhagirathi) is the only exception of the gauges shown in Table 2, showing 25 an increase of 12 % in the number of incidences where the river flow is less than the 1971-2000 10th percentile for the 2080s. This is mainly due to the ECHAM5 model which has a high number of incidences. The Yangcun gauge (Brahmaputra) is the only gauge where there is no change in the number of incidences where the river flow is less than the 10th percentile for 1971-2000 in either of the future periods. This is probably because the lowest river flows are already very low at this gauge.
At every gauge there is an increase in the number of incidences where river flows are greater than the 90th percentile for 1971-2000 for the two future periods. Several of the 5 gauges have increases in the number of events above the 90th percentile for the 1971-2000 period of more than 100 %. This confirms the conclusions drawn visually from the analysis in Fig. 11 that the general distributions move toward the higher flows for these gauges and simulations. 10 The annual timeseries shown in Fig. 8 is very variable and systematic changes throughout the century could be masked by this variability. On the basis that there are changes in the upper and lower parts of the future river flow distributions, the 10th and 90th percentiles for each decade and each simulation are calculated. At the lower end of the distribution, there is little change in the 10th percentile (not shown) for most of the gauges, probably 15 because of very low flows at the lowest times of the year. Only the Pandu and Bahadurabad gauges on the Brahmaputra and the Farakka gauge on the Ganges show a non-zero value for the lowest 10 % of river flows through to the 2100s. These three gauges indicate a slight increase for the 10th percentile for each decade through to 2100.

Decadal percentile analysis
The 90th percentile values (Fig. 14) are generally much more variable throughout the 20 century than those for the 10th percentile to the 2100s. We consider the gauges according to their location across the Himalayan arc from west to east (see Fig. 2). The HadCM3 simulation projects an increase in river flows for the most westerly gauges in this analysis; Attock and Kotri gauges located on the Indus (see Fig. 14a and b) and the Chenab-Panjnad gauge (see Fig. 14c). ECHAM5, on the other hand, shows a much flatter trajectory for these 25 gauges. This may be explained by the HadCM3 simulation depicting an increase in the occurance of western disturbances and an increase in total snowfall which is not simulated by ECHAM5 .
The gauges located toward the middle of the Himalayan arc generally show increases across the decades to 2100 in both models; these are the Bhagirathi-Tehri (Fig. 14d), both Karnali river gauges (Benighat - Fig. 14e and Chisapani- Fig. 14f), Narayani-Devghat and Arun-Turkeghat ( Fig. 14g and h). There is very close agreement between the two simulations for the Narayani-Devghat, Arun-Turkeghat ( Fig. 14g and h) and Bhagirathi-5 Tehri (Fig. 14d) gauges, with the former two showing less variability between decades than the others in the analysis. The Karnali-Benighat gauge (Fig. 14e) also has less variability between the decades, however there is a systematic difference between the two simulations that remains fairly constant across the decades. From the subset of gauges in this analysis that are the most central on the Himalayan arc, the Karnali-Chisapani gauge (Fig. 14f) has 10 the largest variability between simulations and decades. However this gauge still shows an increase overall in both simulations with a steeper increase for HadCM3 than ECHAM5. The closer agreement between simulations at these more central gauges may be due to the reducing influence of the western disturbances in the HadCM3 simulation from west to east across the Himalaya therefore resulting in smaller differences between the the two 15 simulations.
The Farakka-Ganges gauge (Fig. 14i) and two of the Brahmaputra gauges -Bahadurabad (Fig. 14j) and Pandu (Fig. 14l), represent three of the most easterly river gauges in the analysis. These gauges show an increase in both simulations through to the 2100s, in this case more pronounced in ECHAM5 than HadCM3 for these two Brahmaputra 20 gauges. There is much closer agreement between the two simulations at the Farakka-Ganges gauge (Fig. 14i) which is located slightly further west than the two Brahmaputra gauges. The other Brahmaputra gauge, the Yangcun (Fig. 14k) is very variable through the century, there is a period with consecutive decades of increasing river flows in the middle of the century but over the whole century neither model shows a consistent change. 25 This analysis shows that neither simulation is consistently showing a systematic increase in the 90th percentile of river flows across all the gauges. Instead it highlights the changing conditions and the different behaviour of the two simulations across the Himalayan arc.

Implications of changes in future river flows
In this section we consider the implications of the projected future changes in river flows for South Asia on water resources. We highlight the broader challenges facing the region and where the current RCMs need development to represent key processes for this region.
The key points from this discussion are summarised in Table 3. In the present day, water 5 resources in South Asia are complicated, precariously balanced between excess and shortage. Parts of South Asia receive some of the largest volumes of precipitation in the world and are therefore at frequent risk of flooding and yet others regularly endure water stress. The complexity is increased by the competition between states and countries for resources from rivers that flow large distances crossing state and country borders, each 10 with their own demands on resource. There is a considerable gap between the amount of water resource flowing through South Asia and the actual usable amount (Aggarwal et al., 2012), for example the total flow for the Brahmaputra basin is approximately 629 km 3 of which only 24 km 3 is usable (Kumar et al., 2005). There is therefore huge potential for improvements in the efficiency of systems for irrigation and domestic water supply that 15 could ease pressures on water resources, currently and predicted, as demand increases.
In the last 50 years there have already been efficiency improvements, such as development of irrigation systems and use of high yielding-water efficient crop varieties. These improvements have fuelled the rapid development in agriculture across South Asia making the region more self-sustained and alleviating poverty (Kumar et al., 2005). However 20 these advances have also had a large impact on the regions river ecosystems resulting in habitat loss, reduced biodiversity (Sarkar et al., 2012) and water pollution (Vörösmarty et al., 2010). Historically arbitrary thresholds based on a percentage of the annual mean flow have been used to estimate minimum flows, but these simplistic estimates do not take account of the flow variability that is crucial for sustaining river ecosystems (Arthington et al., 2006; 25  , referred to as environmental flows. Environmental flows are defined by Smakhtin and Anputhas (2006) as the ecologically acceptable flow regime designed to maintain a river in an agreed or predetermined state. The variability in river flows through the year have important ecological significance; for example low flows are important for algae control and therefore maintaining water quality. High flows are important for wetland flooding and preserving the river channel. When considering the implications of future changes in climate on river flows and therefore surface water resources, estimates of flow variability and minimum flows are an important consideration. However these are not easily quantified 5 in general terms with many methods requiring calibration for applications to different regions and basins. In our simulations there is an intensification of the seasonal cycle and therefore an increase in the flow variability and a reduction in the occurrance of the lowest flows. These changes could have implications for the biodiversity of these catchments.
In India the domestic requirement for water is the highest priority but is only 5 % of  (Murphy et al., 2004) and find an increase in water resources for South Asia at the annual 25 timescale due to climate change. The analysis shown here shows a similar result with increases in river flow, particularly the magnitudes of the higher river flows at these gauges, in some cases above the range of variability used for this analysis (1.5 SD). However, the analysis shown here on the monthly timescale, also highlights that these increases 29 in resource tend to occur during the ASM, when river flow is at its maximum. This could mean that the benefits of an increase in water resource may not be realised due to the timing of this increase within the year. Although these projected changes in river flow are not critical for water resources they could still be beneficial where there is the capacity to store the additional flow for use during periods of low flow. Additional water storage 5 capacity for example through rainwater harvesting, could greatly increase the useable water resource for the Ganges-Brahmaputra catchments (Kumar et al., 2005) and potentially alleviate the increased risk of flooding during the ASM when rainfall is most persistent and rivers are already at their peak flow. South Asia, even in the current climate, is particularly susceptible to flooding due to the high temporal and spatial variability of rainfall of the 10 region. It is estimated that approximately 20 % of Bangladesh floods annually (Mirza, 2002). Several studies have highlighted increases in both the extremes (Sharma, 2012;Rajeevan et al., 2008;Goswami et al., 2006;Joshi and Rajeevan, 2006) and the variability (Gupta et al., 2005) of precipitation in recent years that cause extreme rainfall events resulting in catastrophic levels of river flooding. Over 30 million people in India alone are affected by 15 floods and more than 1500 lives are lost each year (Gupta et al., 2003), the economic cost of flooding is also considerable with the cumulative flood related losses estimated to be of the order of USD 16 billion between 1978 and 2006 (Singh and Kumar, 2013).
The timing of the peak flows of major rivers in this region is also very important in terms of flooding. In 1998 the peak flows of the Ganges and the Brahmaputra rivers occurred 20 within 2 days of each other resulting in devastating flooding across the entire central region of Bangladesh. Approximately 70 % of the country was inundated, the flood waters then remained above danger levels for more than 60 days (Mirza, 2002). This event caused extensive loss of life and livelihood in terms of damaged crops, fisheries and property and the slow recedance of flood waters hindered the relief operation and recovery of the 25 region. This analysis does not suggest any change to the timing of the peak flows, only the magnitude. However given the high probability of two rivers in this region having coincident peak flows in any given year (Mirza, 2002) and the likelihood that severe flooding will result, an increase in the magnitude of the peak could still be significant. Flooding can have a large impact on crops, for example in Bangladesh over 30 % of the total flood related damages are due to the loss of crops. The estimated crop damage from the 1998 floods was estimated to be 3.0 million t (Gain et al., 2013). The slow receding of flood water can also mean the ground is not in a suitable condition to sow the next crop, restricting the growing time and potentially affecting crop yields for the following year. On the other hand a limited amount 5 of flooding could also be a benefit, particularly for rice crops. Inundation of clear water can benefit crop yield, due to the fertilization effect of nitrogen producing blue-green algae in the water (Mirza et al., 2003).
In our simulations the reduced occurrence of the lowest flows could translate into an increase in the surface water resource in this region especially during periods when the river 10 flows are traditionally very low. This could mean that the current and increasing pressure on ground water (Rodell et al., 2009) may be alleviated in future years. Alternatively increases in the lowest flows may enable adaptation to a changing climate and the modification of irrigation practises. Current projections of future climate suggest that temperatures could also increase for this region (Cruz et al., 2007). Increasing temperatures poses a threat 15 to crop yields of a different kind because this is a region where temperatures are already at a physiological maxima for some crops (Gornall et al., 2010). Rice yield, for example, is adversely affected by temperatures above 35 • C at the critical flowering stage of its development (Yoshida, 1981). Wheat yields could also be affected by rising temperatures, with estimated losses of 4-5 million tonnes per degree of temperature rise through the 20 growing period (Aggarwal et al., 2012). Additional water resource for irrigation at previously low flow times of the year could allow sowing to take place at a different time of the year in order to avoid the highest temperatures, thereby reducing the likelihood of crop failure. However with increasing variability and extremes, a potential feature of the future climate for this region (Hijioka et al., 2014), there is also the increased risk of longer periods with below 25 average rainfall and potentially more incidences of drought. This could lead to additional demand for water for irrigation to prevent crops becoming water stressed (Aggarwal et al., 2012). There may also be increases in demand from other sources other than agriculture, for example the increasing population (United Nations, 2013) or the reduced availability of ground water of an acceptable quality for domestic use (Gregory et al., 2005). Any of these factors, either individually or combined, could effectively cancel out any or all increases in resource from increased river flow due to climate change.
In addition there are a number of processes missing from the models used for these simulations that could change the sign of the projected changes. There is no irrigation 5 included in these simulations, which could be important particularly on the basin scale. The impacts of extensive irrigation on the atmosphere are complex but could have a positive impact on water availability (Harding et al., 2013) due to evaporation and water being recycled within the basin. Tuinenburg et al. (2014) estimate that up to 35 % of additional evaporation is recycled within the Ganges basin. Therefore this aspect of the regional water 10 cycle is not accounted for in these simulations. There is also no representation of glaciers which could act to increase or reduce river flows depending on the occurance of negative or positive mass balance respectively. In these simulations snowmelt is represented, however representing glacial processes as snowmelt could act to enhance the seasonal cycle in the simulated river flows for both present day and future projections as snow melts more readily 15 than ice. These simulations also do not explicitly include groundwater, primarily focusing on river flows. Groundwater is a highly exploited part of water resources for South Asia. Representation of this would give a more complete picture of the total water resources for this region. 20 We present the first 25 km resolution regional climate projections of river flow for the South Asia region. A sub-selection of the HNRCMs are used to provide runoff to a river routing model in order to provide river flow rate which can be compared directly with ERAInt and any available river gauge data for the South Asia region. This analysis focuses on the major South Asia river basins which originate in the glaciated Hindu-Kush Karakoram Himalaya; 25 the Ganges/Brahmaputra and the Indus. The aim of this analysis is firstly to understand the river flows in the ECHAM5 and HadCM3 simulations and secondly examine how useful they are for understanding the changes in water resources for South Asia. We also consider what the projected changes in river flow to the 2100s might mean for water resources across the Himalaya region.

Conclusions
The driving GCMs (ECHAM5 and HadCM3) have previously been shown to capture a range of temperatures and variability in precipitation similar to the AR4 ensemble for the 5 much larger domain of Asia (Christensen et al., 2007). However using just two ensemble members cannot capture the full range of these larger ensembles. In this analysis the seasonal cycle of precipitation, a key influence on river flows, is captured reasonably well for the downscaled GCMs compared to both observations and the downscaled ERAint simulation. Although observed precipitation is lower than in the model the underestimation 10 inherrent in precipitation observations at higher elevations is likely to be an important factor for this analysis, which includes the high Himalaya.
A number of GRDC gauge stations (GRDC, 2014), selected to capture the range of conditions across the Himalayan arc and sample the major river basins, provide observations of river flow for comparison against the HNRCM simulations. The lack of 15 recent river flow data limited the gauges that could be selected for analysis. In the absence of robust observations we use a downscaled ERAint simulation in addition to the available observations to provide a useful benchmark against which to compare the downscaled GCM simulations. In general there is a tendancy for overestimation of river flow rate across the selected gauges compared with GRDC observations, however comparison against the 20 ERAint simulation is more mixed with some gauges showing higher and others with lower river flows than ERAInt. In general most of the simulations broadly agree with observations and ERAint to within the range of natural variability (of 1.5 SD) and agree on the periods of highest and lowest river flow. Therefore indicating that the RCM is able to capture the main features of both the climate and hydrology of this region for the present day. 25 The future projections indicate an increase in surface water resources, with river flow rates at some of the gauges almost doubled by the end of the century. These increases in river flow occur for the gauges in the Ganges/Brahmaputra basin, which also shows an increasing trend in both evaporation and precipitation. Therefore the changes in river flow are likely to be mainly driven by precipitation on the annual scale which more than counters the evaporation caused by increasing temperatures in the model. This is consistent with other analyses of precipitation which also use the A1B climate scenario (Nepal and Shrestha, 2015), which is a useful result. The trajectories of the annual average river flow, evaporation and precipitation for the Indus are much flatter, showing little or no trend. 5 The increases in the annual mean river flows are reflected in the seasonal cycles of river flow for the two future periods (2050s and 2080s) which indicate that most of the changes occur during peak flow periods. Some of the gauges toward the middle of the Himalayan arc, show changes above the range of present day natural variability. This could be due to the increasing influence of the ASM and reducing influence of western disturbances from 10 west to east having an additive effect. The gauges located furthest west and east in this analysis lie within the present day natural variability. There were also differences between the two simulations across the Himalayan arc with HadCM3 suggesting increases in river flow at the upper end of the distribution for western gauges that was not evident in ECHAM5. The analysis shown here does not suggest a systematic change in the models for the timing 15 of the maximum and minimum river flows relative to the present day suggesting an over all increase in water resources at the top and bottom of the distribution. This has positive and negative implications with potentially more resource during usually water scarce periods. However there are also implications in terms of increased future flood risk during periods where the river flow is particularly high. Increases in maximum flows for rivers in this region 20 could be important in terms of loss of life,livelihoods, particularly agriculture and damage to infrastructure.
While this analysis suggests increasing surface water resources due to climate change, there are a number of other factors that could affect this result, both in terms of this analysis and uncertainties surrounding the region itself. The South Asia region is changing rapidly, 25 therefore other factors could have a large effect on water resources for this region. A rising population, expansion of industry (other than agriculture) and the continued depletion of ground water could change the demand for surface water resource from other parts of the South Asia economy. In addition increasing variability of an already changeable climate could lead to extended periods throughout the year of rainfall below the annual average, leading to an increase in demand for irrigation resource. In terms of this analysis, this is only one RCM and another RCM could produce a different result. Also there are missing hydrological processes in the RCM and river flow model that could impact the river flows directly. The RCM and river flow model do not include abstraction and irrigation, 5 groundwater recharge or explicitly include glacial processes and their contribution to river flow. Including glacial processes in the form of a glacier model together with river routing within the land-surface representation will be useful to establish if the contribution from glaciers changes the timing and/or magnitude of both the lowest and highest flows in these gauges. Likewise including representation of water extraction (both from rivers and 10 groundwater) particularly for irrigation, the biggest user of water in the region, will help to provide a more complete picture of the demand for water resources for the South Asia region. Including irrigation and therefore the associated evaporation will capture part of the water cycle not possible with the current model and maintain the regional water balance. Including representation of these processes in the RCM or river flow model would improve 15 the robustness of the future projections of water resources and further our understanding of the water balance for this region. These processes could have a large impact on the water balance in the model potentially changing the signal of the projected changes in river flow. Understanding the interactions between availability of water resources, irrigation and food production for this region by using a more integrated approach, such as that used in 20 Biemans et al. (2013) may also help with understanding how pressures on resources could change with time. In support of this work and others, there is also a need for good quality observations of both precipitation and river flow available for long enough time periods to conduct robust water resource assessments for this region. adaptation -A modeling study, Science of The Total Environment, 468-469, Supplement, S117-S131, doi:http://dx.doi.org /10.1016/j.scitotenv.2013.05.092, http://dx.doi.org/10.1016/j.scitotenv. 2013.05.092, 2013 The State and Fate of Himalayan Glaciers, 5 Science , 336, 310-314, doi:10.1126/science.1215828, http://0-www.sciencemag.org.wam.leeds. ac.uk/content/336/6079/310.abstract, 2012T., Laprise, R., na Rueda, V. M., Mearns, L., Meneńdez, C., Räisänen, J., Rinke, A., Sarr, A., and Whetton, P.  Cruz, R., Harasawa, H., Lal, M., Wu, S., Anokhin, Y., Punsalmaa, B., Honda, Y., Jafari, M., Li, C., and Huu Ninh, N Simmons,A. J.,Berrisford,P.,Poli,P.,Kobayashi,S.,Andrae,U.,30 Balmaseda, M. A., P., G. B., Bauer, Bechtold, P., Beljaars, A. C. M., van de Berg, L., Bidlot, J., Bormann, N., Delsol, C., Dragani, R., Fuentes, M., Geer, A. J., Haimberger, L., Healy, S. B., Hersbach, H., Hólm, E. V., Isaksen, L., Kållberg, P., Köhler, M., Matricardi, M., McNally, A. P., Monge-Sanz, B. M., Morcrette, J.-J., Park, B.-K., Peubey, C., de Rosnay, P., Tavolato, C., Thépaut, J.-N., and Vitart, F.: The ERA-Interim reanalysis: configuration and performance of the data assimilation system, Quarterly Journal of the Royal Meteorological Society, 137, 553-597, doi:10.1002/qj.828, http://dx.doi.org/10.1002/qj.828, 2011 A study of floods in the Brahmaputra basin in India, Interna-5 tional Journal of Climatology, 20, 771-781, doi:10.1002/1097 Resources Development, 21, 385-398, doi:10.1080/07900620500037818, http://dx.doi.org/10. 1080/07900620500037818, 2005 Water for India in 2050: first order assessment of available options, Current Science, 86, 1216-1224. Gupta, S., Javed, A., and Datt, D.: Economics of Flood Protection in India, Natural Hazards, 28, 5 199-210, 2003  Developing new crops takes time.
Changes in Increases in peak flows could be Improving river channel capacity. Flood protection levels peak flow -timing positive for irrigation and Diverting excess water to a different valley. do not match demographic and magnitude domestic supply but could increase Storing the excess water for low flow periods trends so vulnerability the risk of flooding. e.g. through rainwater harvesting. to flooding remains high Peak flows occurring later and/or Improving drainage and water recycling. in this region decreases in peak flows could reduce Adopting varieties of crops that grow (Gupta et al., 2003). availability of water for irrigation when water for irrigation is more Market development for at crucial crop development stages readily available new crops takes time negatively impacting yields.

Changes in low
Increases in the magnitude of the low Adaptations to avoid flooding during flows -timing flows could be positive for irrigation peak flow periods could provide resource and magnitude and domestic supply. during low flow periods. Decreases could mean less resource Development of water efficient, available for irrigation high yielding leading to reduced yields crop varieties      Comparison of the lowest 10 % of monthly river flows at the Farakka barrage on the Ganges river against the 10th percentile for the 1971-2000 period for 1971-2000 (top), 2050s (middle) and 2080s (bottom) for HadCM3 (red triangles) and ECHAM5 (blue stars). Each star or triangle represents a month within the 30 year period where the value is less than the 10th percentile of the 1971-2000 period with the total number for each of the simulations given in the top right corner of each plot. Figure 13. Comparison of the highest 10 % of monthly river flows at the Farakka barrage on the Ganges river against the 90th percentile for the 1971-2000 period for 1971-2000 (top), 2050s (middle) and 2080s (bottom) for HadCM3 (red triangles) and ECHAM5 (blue stars). Each star or triangle represents a month within the 30 year period where the value is greater than the 90th percentile of the 1971-2000 period with the total number for each of the simulations given in the top right corner of each plot.   Figure 14. The 90th percentile of river flow for each decade for HadCM3 (red triangles) and ECHAM5 (blue circles) for each river gauge.