Future changes in precipitation and impacts on extreme streamflow over Amazonian sub-basins

Many questions arise about the impact of climate change on water resources, which are fundamental for ecosystems and societies. Such questions are especially relevant in the Amazon River basin where severe hydroclimatic (precipitation, evapotranspiration (ET)) and anthropogenic (deforestation, land-use change...) changes are already threatening the region. Runoff changes in the basin have been characterized by more severe extreme values on the main stream for about forty years (Callede et al 2004, Espinoza et al 2009a) due to decadal modes of variability at regional scale (Marengo et al 1998, Labat et al 2004, Marengo 2004, Labat 2005, Espinoza et al 2006, 2009b) and to an interannual variability (Richey et al 1989, Marengo 1992, Guyot et al 1998, Uvo and Graham 1998, Ronchail et al 2005a, 2005b, 2006, Zeng et al 2008, Espinoza et al 2009a) leading to severe droughts such as the 1998, 2005 and 2010 events (Marengo et al 2008, Zeng et al 2008, Yoon and Zeng 2010, Espinoza et al 2011, Marengo et al 2011) and floods as experienced in 1999, 2009 (Chen et al 2010, Marengo et al 2010, 2012) and 2012. These events alter the livelihood of riverside populations (Drapeau et al 2011), the ecosystems and the vegetation functioning (Phillips et al 2009). Drought has been suggested to increase tree mortality and decrease photosynthetic activity in 2010 (Lewis et al 2011, Xu et al 2011). Furthermore the impact of such events on water resources (Bates et al 2008, Cox et al 2008), on fire and vegetation (Nepstad et al 2007, Cook et al 2012) might increase in the future. Are present-time extreme discharges precursors of future conditions?

Despite the uncertainties of future climate projections, related to CO₂ emission scenarios and to General and Regional Circulation Models (GCMs and RCMs respectively) (Li et al 2006, Vera et al 2006, Boulanger et al 2007, Marengo 2009), some consensus is found regarding climate change in the Amazon River basin. Indeed, a 4 °C warming is expected in tropical South America (Boulanger et al 2006) and a consensus is found about a precipitation decrease in austral winter (Vera et al 2006), a longer dry season and more severe droughts (Li et al 2006, 2008). Moreover, annual precipitation is expected to decrease in the eastern Amazon River basin and, in contrast, to increase and be more intense over the western Amazon (Meehl et al 2007, Alves and Marengo 2010).

Past studies on climate change and its impact on hydrology at global scale do not show any robust agreement about future runoff trends in the Amazon River basin. Some authors simulate a discharge reduction on the main stem of the Amazon River (Arora and Boer 2001) while others find a 5% increase in discharge at the end of the 21th century (Nohara et al 2006). Salati et al (2009), using 15 AR4 (4th Assessment Report) coupled models and the RCM HadRM3P (Hadley Centre Regional Model version 3P), project a 7–30% discharge decrease on the main stem of the Amazon River depending on the scenario (SRESB2 and SRESA2), the climate model and the period. At the regional scale, literature on future runoff projections in the Amazon River basin remains scarce because of hydrological modeling issues (Coe et al 2007, Decharme et al 2008, Beighley et al 2009, Trigg et al 2009, Getirana et al 2010, Paiva et al 2011a, Yamashima et al 2011, Guimberteau et al 2012, Paiva et al 2012). Milly et al (2005), using ensembles from 12 GCMs for the middle of the 21st century under the SRESA1B scenario, obtained a 5–10% runoff decrease in northeastern Amazon, contrasted to a 10–40% increase in western Amazon. Along the eastern edge of the Amazon River basin, Tomasella et al (2009) found a 30% (60%) mean annual (low-flow) decrease of the Rio Tocantins discharge at the end of the 21th century under the SRESA1B scenario. In the upper Amazon, Lavado Casimiro et al (2011) use two hydrological models, the climatic data from 3 GCMs under 2 SRES scenarios, and found both decreasing (4 basins) and increasing (3 basins) discharge trends in the Peruvian Amazon Andes basins. As a consequence of precipitation and runoff change, flooded areas and flood duration are found to increase in about one third of the basin, especially in Western Amazonia (Langerwisch et al 2012).

The above differences in future extreme flow estimations (low and high flows) for the Amazon River are due to the climatic forcing, model uncertainties but also to the lack of simulations at sub-basin scale. Will annual extreme flows be more severe during the 21st century than the present ones? Will there be differences within sub-basins? To try to answer these questions, our study focuses on sub-basin scale annual extreme values in the context of a changing climate. This was permitted by the availability of the Land Surface Model (LSM) ORCHIDEE (ORganising Carbon and Hydrology In Dynamic EcosystEms, section 2.1) which is able to accurately simulate the present-time streamflow in many stations over the basin (Guimberteau et al 2012). We build plausible scenarios of hydrological conditions that populations and ecosystems might have to cope with during the 21st century, based on a multi-scenarios approach from the IPCC's (Intergovernmental Panel on Climate Change) AR4 (Solomon et al 2007). Our methodology is first described (section 2). We detail the method based on anomalies to build the future climate forcings (section 2.2) and the hydrological signatures chosen to analyze the relationship between precipitation and streamflow (section 2.3). The hydrological signature results for present time are briefly given in section 3. The results from future simulations under SRESA1B at the middle of the 21st century are presented in section 4 and organized according to the different sub-basins. Section 5 summarizes the results obtained at the end of the century and under other scenarios.

2.1. The land surface model ORCHIDEE

2.2. Construction of climate change forcings

GCMs are known to poorly simulate precipitation patterns, even more at sub-basin scale (Solomon et al 2007). Moreover, because of the importance of local precipitation for hydrology, downscaling methods have been developed for hydrological studies and lead to improved results at basin scale (Wilby et al 1999). We applied the delta downscaling method approach to produce climate change forcings from GCMs results. This method is based on the addition of the monthly anomalies between two climatologies (simulated climate change and current climate results from a given GCM) to a baseline meteorological forcing. It appears to be a downscaling method well adapted for hydrological studies (Fowler et al 2007). In fact, if we assume that relationships between variables in the present-time climate are likely to be maintained in the future, the delta method incorporates from the present-time forcing more realistic spatio-temporal patterns in precipitation which are crucial for future streamflow simulation. However, this method prevents the study of future interannual variability as future monthly anomalies of climatologies are applied to the present-time forcing whose variability is maintained.

Climate change anomalies are derived from up to 8 AR4 GCMs (see table s1 available at stacks.iop.org/ERL/8/014035/mmedia), 2 different 20 yr running mean periods (2046–2065 and 2079–2098) and 3 emission scenarios (SRESB1, A1B and A2). We mainly focus on the results coming from SRESA1B, which projects a very rapid economic growth, a low population growth, a rapid introduction of new and more efficient technology, leading to a 720 ppm CO₂ concentration by the year 2100. This choice was driven by the better availability of the 8 atmospheric variables required to run ORCHIDEE (see table s2 available at stacks.iop.org/ERL/8/014035/mmedia) under SRESA1B scenario compared to SRESA2. The global meteorological dataset NCC (NCEP/NCAR Corrected by CRU data, Ngo-Duc et al 2005) is used as the baseline for this study. The 21 yr period representing the current climate in NCC dataset is the period 1980–2000, during which the precipitation dataset was corrected by ORE (Environmental Research Observatory) HYBAM (Geodynamical, hydrological and biogeochemical control of erosion/alteration and material transport in the Amazon River basin) (Cochonneau et al 2006) observations that have improved the present-time streamflow simulation (called 'ORCH4') by ORCHIDEE (Guimberteau et al 2012).

In each GCM grid-cell i, monthly anomalies (ANO(X_GCM,i,tGCM)) have been calculated from climatologies of the atmospheric variables outputs derived from each GCM (X_GCM,i,tGCM) (figure 1 and equation (1)). Linear spatial interpolation has been applied to the GCM outputs whose resolution was coarser than NCC (1.0°). The resulting monthly anomalies of a given variable of a GCM were applied to the corresponding NCC variable (X_NCC,i,tNCC) at each NCC time-step (t_NCC = 6 h) for the corresponding month (t_GCM = 1 month) of each year of the 1980–2000 period (figure 1 and equation (2)). Thus, simulations are performed with ORCHIDEE forced by these new climate change forcings starting from the same state of equilibrium of the model (5 yr spin-up over the 1975–1979 period).

$$\begin{eqnarray} &&\fl \mathrm{ANO}({X}_{\mathrm{GCM},i,{t}_{\mathrm{GCM}}})\nonumber\\ &&=\left (\frac{({X}^{\mathrm{SRES}\alpha }-{X}^{2 0 C 3 M})}{{X}^{2 0 C 3 M}}\right )_{\mathrm{GCM},i,{t}_{\mathrm{GCM}}} \end{eqnarray} \tag{ 1 }$$

$$\begin{eqnarray} &&\fl {X}_{{\mathrm{NCC}}_{\mathrm{futur}},i,{t}_{\mathrm{NCC}}}\nonumber\\ &&={X}_{\mathrm{NCC},i,{t}_{\mathrm{NCC}}}+(\mathrm{ANO}({X}_{\mathrm{GCM},i,{t}_{\mathrm{GCM}}})\ast {X}_{\mathrm{NCC},i,{t}_{\mathrm{NCC}}}). \end{eqnarray} \tag{ 2 }$$

Zoom In Zoom Out Reset image size

2.3. Hydrological signatures

The variations of three hydrological signatures (runoff coefficient, precipitation elasticity of streamflow and first/last deciles of streamflow) are studied at sub-basin scale using 16 ORE HYBAM gauge stations (table 1) to characterize the hydrological response to future precipitation change. In the stations where it accounts, the backwater effect (Meade et al 1991) that is not represented in ORCHIDEE, is eliminated in the measurements (Espinoza et al 2009a). For each signature, differences are performed between one simulated year in climate change condition and the baseline ORCH4 climatology. Thus we estimate for each year the signature variation between future simulation and average present-time condition. The results are presented from an ensemble which is an average of 168 differences (8 models *21 yearly differences) between future and present conditions.

Table 1.  List of ORE HYBAM gauge stations over the Amazon River basin (see figure 2 for their localization on a map). For each station, mean annual streamflow (Q_mean), runoff coefficients (R_coeff) and streamflow elasticity to precipitation (ε_p) are computed from observations (ORE HYBAM database) and present-time simulation ORCH4, on average over the 1980–2000 period (except for the observed streamflow at the river mouth whose value is from Callede et al (2010), for the 1972–2003 period average). The colors indicate the localization of the stations in the basin (black: mainstream; red: southeastern/southern region; green: western region; blue: northwestern/northern/northeastern region).

2.3.1. Annual runoff coefficient.

The annual average runoff coefficient (R_coeff, equation (3)) varying from 0 to 1, is the proportion of the long-term average precipitation ($\bar {P}$ in kg m⁻² s⁻¹) that runs off into streams (long-term average discharge $\bar {Q}$ in kg m⁻² s⁻¹), assuming no net change in storage. A high (low) runoff coefficient points out that a large amount of water leaves the basin as streamflow (as ET).

$$\begin{eqnarray} &&{R}_{\mathrm{coeff}}=\frac{\bar {Q}}{\bar {P}}. \end{eqnarray} \tag{ 3 }$$

Three runoff coefficients are computed for each sub-basin:

• ${R}_{\mathrm{coeff}}^{\mathrm{OBS}}$ is computed from ORE HYBAM discharge data (except for the river mouth discharge whose value is from Callede et al (2010), for the 1972–2003 period average) and from the precipitation dataset used to correct NCC forcing (see section 2.2), on average over the 1980–2000 period.
• ${R}_{\mathrm{coeff}}^{\mathrm{ORCH4}}$ is computed from simulated precipitation and streamflow values of the present-time simulation, ORCH4.
• ${R}_{\mathrm{coeff}}^{\mathrm{ENSEMBLE}}$, is computed from the averaged simulated precipitation and streamflow values of the different future simulations, for each scenario and time horizon.

2.3.2. Precipitation elasticity of streamflow.

In order to quantify the response of discharge to the precipitation change, we compute the coefficient of elasticity (ε_p, equation (4)) which is the relative change in mean annual streamflow divided by the relative change in mean precipitation (Schaake 1990, Sankarasubramanian et al 2001). The streamflow elasticity can vary usually between 1.0 and 4.0, i.e. a 10% change in mean annual precipitation results in a 10–40% change in mean annual runoff. The precipitation elasticity of streamflow is strongly correlated to the runoff coefficient, the mean annual precipitation and the streamflow according to Chiew (2006). Thus, river basins with low (high) runoff coefficient will be more (less) sensitive to precipitation change.

$$\begin{eqnarray} &&{\varepsilon }_{\mathrm{p}}=\mathrm{median}\left [\frac{\left ({Q}_{y}-\bar {Q}\right )}{({P}_{y}-\bar {P})}\cdot \frac{\bar {P}}{\bar {Q}}\right ] \end{eqnarray} \tag{ 4 }$$

P_y and Q_y (both in kg m⁻² s⁻¹) are the annual precipitation and discharge respectively, averaged over the year y of the time period.

The potential evaporation elasticity of streamflow is computed in the same way according to equation (4).

2.3.3. First and last deciles of streamflow.

The deciles are used to distinguish the monthly low flow (i.e. first decile) and monthly high flow (i.e. last decile) for a given sub-basin. The first (last) decile of streamflow is the value not exceeded by the lowest (highest) 10% of all streamflow values. They are calculated for each sub-basin, each year of future climate forcing and scenario, for present and future time horizons. For each future simulation, the significance of the median change of the first and last deciles is assessed using the Wilcoxon signed-rank test at the 95% level.

The ability of ORCHIDEE to simulate the streamflow variations over the end of the 20th century was thoroughly tested over the Amazonian sub-basins by Guimberteau et al (2012). The streamflow simulation has been improved using the new ORE HYBAM precipitation forcing which better takes into account the precipitation regimes. They are very different according to the regions of the Amazon River basin due to its huge size and its extension in both hemispheres. We can distinguish different seasons of precipitation according to the regions (Figueroa and Nobre 1990, Nobre et al 1991, Ronchail et al 2002, Espinoza et al 2009a). From southeastern to southwestern regions, the precipitation season occurs in DJF (December, January and February) and the dry season in JJA (June, July and August). The precipitation seasonality in western regions is similar but less pronounced. The rainiest and driest seasons in northwestern regions are slightly more marked; they occur in MAM (March, April and May) and DJF respectively. The precipitation regime slightly differs in the northernmost and the northeastern regions where the rainiest seasons occur in MJJ (May, June and July) and MAM, respectively.

The hydrological signatures describing the link between precipitation and runoff (runoff coefficient and precipitation elasticity) in present time are briefly analyzed in sections 3.1 and 3.2, then used to assess future streamflow variations in a climate change perspective (section 4). In the latter section and section 5, we focus on the future change of the annual extreme values of streamflow using the hydrological signature of streamflow deciles.

3.1. Annual runoff coefficient

3.2. Precipitation elasticity of streamflow

We assess that low flow (high flow) is strongly related to the dry season (rainy season). This simple hypothesis is limited by the complex relation between precipitation and streamflow regimes within the Amazon River basin. Floodplains extension can modify the relationship between precipitation amount and extreme stream flows. However, for a small sub-basin such as the Mamore at Guaraja-Mirim, the lag between precipitation and runoff is not significant at monthly timescale. In contrast, it can be significant (about two months) over larger sub-basins including floodplains areas (Óbidos, Manacapuru, Fazenda Vista Alegre). After a brief overview of the climate change projected over the basin for the 2046–2065 time horizon under the SRESA1B scenario (section 4.1), the regional changes in extreme flows related to the seasonal precipitation changes (see section 3) are presented for three large coherent regions of the basin (see color code in table 1) in section 4.2: southeastern/southern region (SE–S, section 4.2.1), western region (W, section 4.2.2) and northern region including northwestern/northern/northeastern regions (NW–N–NE, section 4.2.3). A separate section is dedicated to Óbidos results (section 4.2.4), representing the hydrological behavior of the Amazon River basin.

4.1. Overview of projected climate change

Annual ensemble mean temperature increases by 2.04 °C (1.49 °C–2.63 °C according to the different simulations) on average over the basin. Higher annual increase occurs in southern sub-basins with the highest ensemble increase over the Mamore River basin (2.15 °C). The temperature increase leads to an enhanced ensemble mean ET by 7.0% on average over the basin. A slight precipitation increase by 1.1% is simulated over the basin. The precipitation change is spatially contrasted. At the annual timescale, few GCMs (generally 3 out of 8) simulate a precipitation increase (i.e. most GCMs project a precipitation decrease) in eastern (Xingu and Tapajos Rivers), northernmost (Branco River) and, to a lesser proportion, southern (Mamore and Beni Rivers) regions (figure 2(a)). In contrast, wetter conditions are found by a majority of GCMs (generally 7 out of 8) in western (Amazonas River at Tamshiyacu outlet, Purus and Jurua Rivers) and northwestern regions (Japura and Negro Rivers).

Zoom In Zoom Out Reset image size

4.2. Regional response of streamflow to precipitation change

4.2.1. Southeastern/southern region.

The annual precipitation decrease predicted by most models, associated to an ET increase (not shown), leads to a runoff coefficient decrease by more than 5% at the six stations of the region (figure 3). The persistent south-easterly precipitation decrease from March to November predicted by most GCMs (figures 2(c)–(e)) particularly affects streamflow at Altamira where a near 10% decrease in runoff coefficient occurs (figure 3). A south-easterly decrease in dry-season precipitation (figure 2(d)) leads to more pronounced low flows at all the stations of the region (figure 4(b)). The signal is rather robust in the region because at each station, 75% of the ensemble low-flow differences between future and present are negative. The effect is significant over the Madeira (Porto Velho and Fazenda Vista Alegre stations) and the Xingu (Altamira station) rivers where JJA precipitation decreases by 9% and 22% respectively and is associated with a 5% ET increase due to temperature increase (about 2.5 °C) over these basins (not shown). Up to 7 out of 8 simulations give a significant average decrease in the median low flow by about 30% (50%) at Porto Velho (Altamira). No change in high flow is simulated at southerly stations except at Altamira where 6 future simulations out of 8 give a significant decrease by 18% in average (figure 4(a)). Southern sub-basins tend to become more responsive to precipitation. Indeed, precipitation elasticity of streamflow increases by more than 15% at Guajara-Mirim, Altamira and Itaituba (figure 5). Potential evaporation elasticity of streamflow also increases in these stations and in Porto Velho (between 5 and 19%), except at Altamira where it decreases by 66% (not shown) suggesting that the Xingu River basin becomes more (less) responsive to precipitation (evaporation) change.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

4.2.2. Western region.

Most GCMs project an annual precipitation increase over all the western part of the Amazon River basin (figure 2(a)). However, annual runoff coefficients are little affected at Tamshiyacu and Sao Paulo de Olivença (less than 2% decrease), whereas they decrease by more than 5% at Gaviao and Labrea (figure 3), where the lower proportion in swamps and floodplains does not compensate for the change in precipitation contribution to runoff. The wetter projected conditions persist seasonally during the wet season (DJF), in particular in the northwesternmost part of the basin and along the eastern slopes of the Andes, where all the models simulate a precipitation increase (figure 2(b)). High-flow variation is not significant in the four westerly stations. At Tamshiyacu (upper Solimões), all the simulations show a median high-flow increase ranging from 5 to 25% (not shown) but only 3 simulations are significant (figure 4(a)). No significant low-flow decrease is simulated over the western part of the basin (figure 4(b)). Little change in precipitation elasticity of streamflow (less than 5% except at Labrea, figure 5) occurs at the westerly sub-basins, which include flooded regions that regulate the surplus of precipitation. These basins become more responsive to evaporation change where potential evaporation elasticity of streamflow increases up to 56% at Tamshiyacu (not shown).

4.2.3. Northern region.

Most GCMs simulate an annual precipitation decrease in the north of the river mouth region and in the northernmost part of the basin (figure 2(a)). The tri-monthly decreases are projected with good confidence in the northeastern region except in MAM (figures 2(b)–(e)). Northerly streamflow decrease is going to be more severe from west (Acanaui) to east (Caracarai and Sao Francisco). Runoff coefficient highly decreases at Sao Francisco and Caracarai (10.5 and 13% decrease respectively) while low change (less than 2%) occurs at the most westerly stations (Acanaui and Serrinha) (figure 3). The same kind of severity is found regarding northerly low flows. They significantly decrease in at least six simulations out of eight at each station (figure 4(b)). Ensemble median low flows significantly decrease by 20% in the Japura (Acanaui station) and the Negro (Serrinha station) rivers, and it is much higher (55%) in the Branco River (Caracarai station) (figure 4(b)). The signal is rather robust for Serrinha and Caracarai, as 75% of the ensemble low-flow differences between future and present are negative. The discharge decreases at Caracarai during the high flow and recession periods. Overall, the seasonal variations of the ensemble are less contrasted than in the ORCH4 simulation (figure 6(a)). No significant change in ensemble high flow is simulated at Acanaui and Serrinha stations (figure 4(a)). Generally, the northern sub-basins at Acanaui and Serrinha are not responsive to future precipitation change (figure 5) but rather to evaporation; streamflow elasticity to potential evaporation increases by 22% in the Japura River basin (Acanaui station) (not shown). In contrast, the streamflow elasticity to precipitation (potential evaporation) increases (decreases) by 12% (22%) at Caracarai, suggesting that the Branco River basin becomes more (less) responsive to precipitation (evaporation) change.

Zoom In Zoom Out Reset image size

4.2.4. Main stem of the Amazon River.

The precipitation change is low on average over the basin, even though it is strongly spatially contrasted. At Óbidos station, the last gauged station before the mouth of the Amazon River, the runoff coefficient slightly decreases (about 4%) (figure 3) and no change in precipitation elasticity of streamflow is found (figure 5). The streamflow seasonal cycle only changes during the recession period when all the future simulations project more pronounced low flow than ORCH4 (figure 6(b)). Ensemble simulations provide a 10% median low-flow decrease (figure 4b) but only five future streamflow simulations out of eight give a significant decrease. No change in high flows is simulated at Óbidos station (figure 4(a)).

The annual increase in mean ensemble temperature over the Amazon River basin does not depend much of the SRES scenario at the middle of the 21st century. But differences between scenarios are pointed out at the end of the century (+2.0, +3.0 and +3.8 °C for SRESB1, SRESA1B and SRESA2 respectively). The spatial contrast between an eastern decrease and a western increase in ensemble precipitation occurring during the middle of the century under SRESA1B is maintained for the end of the century. However, rainfall increase is accentuated in the west while the precipitation decline is confined in the north-east. In contrast, the eastern rainfall decrease is enhanced under SRESB1 at the end of the century. Regarding SRESA2, the eastern decrease is much weaker than under SRESA1B at the middle of the century and increased rainfall dominates at the end of the century, being much stronger in the west and in the north.

Considering the A1B scenario, the mean ensemble discharge simulations for the end of the century present a low-flow attenuation in western rivers (Madeira, Purus and Jurua) and on the main stem, when compared to the middle of the century. The most striking are observed at Gaviao and Labrea, on the Purus and Jurua Rivers, where the deficit falls to −5%, when compared to the present, instead of −30% at the middle of the century. In contrast, northwestern high-flow increase (Amazonas and Negro rivers) is enhanced (+12% at Tamshiyacu instead of +7% at the middle of the century) in agreement with the western rainfall increase at the end of the century.

No change in low flow under SRESB1 is found when compared to SRESA1B results for both time horizons. Eastern high flow decreases are slightly attenuated under SRESB1 when compared to SRESA1B at the middle and the end of the century. Western high-flow increase under SRESB1 is lower than under SRESA1B at both time horizons. No change in high flow is observed at Óbidos.

The comparison of the results obtained under SRESA2 with those under other scenarios is biased by the fact that only 5 climatic simulations are available instead of 8. Nonetheless, the available data show that extreme discharge anomalies do not differ much from SRESA1B results in the middle of the century. By contrast, at the end of the century, and contrarily to the results obtained under B1 and A1B scenarios, the low flow increases consistently in a large western part of the basin (from +10 to +30%) and consequently on the main stem (+10% at Manacapuru, +5% at Óbidos). Elsewhere, the low-flow decrease is of the same order than for SRESA1B. Western and northern high flow increases substantially, in agreement with rainfall variation, leading to a 10% increase in Manacapuru and Óbidos. Interestingly for impact applications, intensification in extreme values (low and high flow) predominantly occurs in the Negro River at Serrinha (−20% and +20% respectively).

In conclusion, river discharge changes and differences between scenarios are less pronounced at the middle of the century and more accentuated at its end, especially between SRESA2 and the others; indeed, all year round, a large runoff increase is observed in a large part of the western Amazon River basin.

Future climate change should substantially modify the Amazon River basin hydrology and impact the water resources availability. In this study, we use the LSM ORCHIDEE to provide discharge projections, according to future climate forcings built from several GCM projections under the emission scenarios B1, A1B and A2. Various hydrological responses are found in the different sub-basins of the Amazon under SRESA1B scenario. Some of them are not significant since climate projections from the different GCMs diverge, as already assessed in previous studies (Bates et al 2008, Kay et al 2009, Blöschl and Montanari 2010, Nobrega et al 2011, Paiva et al 2011a), but in some sub-basins and for specific periods of the year, the projections are more reliable. In particular, low flows are projected to decrease severely in most stations, especially in the southern Madeira and Xingu Rivers and in the northern Branco River, and to a lesser extent in the northern Negro and Japura Rivers. By the middle of the 21st century, the decrease is projected to reach −50% (−55%) in the Xingu (Branco) Rivers. A low-flow decrease is projected in Óbidos, on the main stem of the Amazon River, which has already been experiencing an increasing number of droughts over the last twenty years. In contrast, the western and upper part of the Amazon are projected to undergo an annual precipitation increase so that high-flow discharge is expected to increase by 7% in the middle of the 21st century and even more by the end of the century. Therefore, more flooding events are expected in this region which is already affected by increasing extreme discharge. In the main stem of the Amazon River, however, there is no consistent signal of more flooding, neither of varying mean discharge, although previous studies projected a positive change (Nohara et al 2006) or a negative change (Milly et al 2005) of the Amazon runoff. By computing elasticity coefficients, we also show that the southern basins with low runoff coefficient are generally more responsive to precipitation change than the western basins with a high runoff coefficient. This result is in agreement with the findings by Chiew (2006) and Paiva et al (2011a).

Some uncertainties relative to the hydrological parameterizations of the LSM can be addressed. The present-time ET underestimation (Guimberteau et al 2012) and the infiltration parametrization are sources of uncertainties in future land water budget. Moreover, the routing module of ORCHIDEE does not represent the river dams and the backwater effect likely to affect the extreme stream flows in some tributaries of the Amazon (Meade et al 1991, Tomasella et al 2011). An advanced large-scale river routine module which represents backwater effect would represent more realistically the simulated water level in the rivers (Paiva et al 2011b, Yamazaki et al 2011, 2012). The modeling of vegetation growth with its interactive environment such as an interannual LAI variation and ET response to elevated CO₂, would lead to better realism using the STOMATE (Saclay-Toulouse-Orsay Model for the Analysis of Terrestrial Ecosystems, Viovy 1996) and LPJ (Lund-Postdam-Jena, Sitch et al 2003) modules. Other errors can result from the downscaling method. First, we assume that climate change varies only over large areas (i.e. as large as a GCM cell area) what leads to underestimate landscape heterogeneity effect due to the coarse resolution of the GCM. Thus, further simulations with better resolution than one degree would test if there is a dependence of simulated Amazonian hydrological processes on the spatial resolution of the forcing, as suggested by Verant et al (2004). Secondly, precipitation interannual variations in response to climate change could not be taken into account due to the simple downscaling method adopted in this study. Some bias-correction methods which conserve the changes of mean and standard deviation of the model simulation data (Watanabe et al 2012) could be adopted to preserve interannual variations. Uncertainties remain also in the emission scenario or the magnitude of rise in mean global temperature and the choice of the GCM. Indeed, a comparison between more AR4 GCMs and a new generation of GCMs (AR5) corroborates high consensus on drying during the dry season but highlights a lower confidence in wettening over the western part of the basin (Joetzjer et al 2013). Finally, considering land surface changes and deforestation in our LSM would probably affect the simulated mean and extreme flows over the Amazon River basin, since several studies reveal that land-use change accounts for at least 50% of the reconstructed global runoff trend over the last century (Piao et al 2007, Sterling et al 2012). The important deforestation in the Amazon River basin may in turn strongly alter the climate and eventually amplify the climate change (Malhi et al 2008). This topic, as well as improvement in the model parametrization, will be addressed in future work.

This work was financially supported by the GIS (Groupement d'Intérêt Scientifique) REGYNA (REGioNAlisation des précipitations et impacts hYdrologiques et agronomiques) and the EU-FP7 AMAZALERT (Raising the alert about critical feedbacks between climate and long-term land-use change in the Amazon) projects. Simulations with ORCHIDEE were performed using computational facilities of the Institut du Développement et des Ressources en Informatique Scientifique (IDRIS, CNRS, France).