4.1. Seasonal Mean Geographical Patterns of Precipitation Changes
 In this section, projected changes at the end of the 21st century (2075–2099) compared to the present (1979–2003) are discussed. Figure 4 shows the seasonal mean precipitation changes between the present and the end of the 21st century for the 60 km mesh model ensembles and the 20 km mesh model. For the 60 km mesh model, results from the four different SST experiments are averaged. For each SST experiment, there are three realizations with different initial conditions (Table 1), so that a total of 12 members exist for 60 km mesh model data. The statistical significance is calculated by 3 (present) plus 12 (future) members. Each year is assumed independent. The degree of freedom is thus 373. The areas with significance greater than 95% are shaded in color. As another measure of robustness of the results, grid points where all four SST experiments (each is three member ensemble mean) agree in the sign of the change are marked with diagonal lines. For the 20 km mesh model, the differences with significance greater than 95% are shaded in color. The degree of freedom is 48. As will be shown later, the three member simulations with the 60 km mesh model yield mostly consistent responses, thus making more significant results in the 60 km mesh model than in the 20 km mesh model.
Figure 4. Seasonal mean precipitation changes (mm/d) between the present and the end of the 21st century for 60 and 20 km mesh atmospheric general circulation models. (a) December–January–February (DJF) 60 km, (b) DJF 20 km, (c) March–April–May (MAM) 60 km, (d) MAM 20 km, (e) June–July–August (JJA) 60 km, (f) JJA 20 km, (g) September–October–November (SON) 60 km, and (h) SON 20 km. For the 60 km model, areas statistically significant at 95% level are colored, and areas where all four different sea surface temperature experiments show consistent changes in sign are hatched. For the 20 km model, areas statistically significant at 95% level are colored. Contour interval is 1 mm/d. ENS denotes ensemble mean.
Download figure to PowerPoint
 In general, 20 km mesh model results are consistent with 60 km mesh model results. In austral summer (DJF), the both models show an increase in precipitation over all tropical and subtropical South America and a decrease over the Andes south of 20°S. In March–April–May (MAM), changes in precipitation are concentrated over the ITCZ and northwest Amazonia near the equator. Increased precipitation over Amazonia is largest in this season both for the 60 and 20 km mesh models. A decreased precipitation in southern Chile is much enhanced from DJF. In JJA, an area of increasing precipitation moves further to the north of the equator compared with the MAM case, and precipitation is projected to decrease over most of Amazon and southern South America. Southern Brazil may be a region of increased precipitation. In September–October–November (SON), an increase of precipitation over southeastern South America becomes more significant both in the 60 and 20 km mesh models, but changes over Amazon are not so systematic. The CMIP3 multimodel mean projections show that over the Amazon Basin, monsoon precipitation increases in DJF and decreases in JJA [IPCC, 2007], although model agreement is not high [Vera et al. 2006b]. This seasonal contrast over the Amazon Basin is also seen in our model projections. Most CMIP3 models project a wetter climate over the Parana River Basin [IPCC, 2007]. The 60 and 20 km mesh models show the increase in precipitation over the Parana River Basin in DJF and MAM, while the wetter region shifts northeast toward the upper Parana River Basin in JJA and SON. While the CMIP3 multimodel mean projections [IPCC, 2007] show a wide area of decreased precipitation over southern South America, the decrease over Chile and Patagonia is much concentrated windward side of the Andes in our model. This difference may be attributable to high-resolution nature of the current model such as sharp and high mountain ranges.
 In order to investigate possible reasons of regional changes in precipitation, atmospheric circulation changes are studied. As in other model studies, the Atlantic and Pacific subtropical anticyclones are intensified and a poleward shift of midlatitude westerlies are simulated in future climate both in the 20 and 60 km mesh models (not shown). Over the Amazon, westerly wind anomalies (i.e., weakened easterly winds) are simulated near the surface and in the lower troposphere from the equatorial Atlantic to the Amazon Basin (not shown). However, atmospheric moisture buildup due to rising temperature could result in larger moisture flux as found in the Indian monsoon region [Kitoh et al., 1997].
 Figure 5a shows the DJF mean vertically integrated moisture flux and its divergence for the 20 km mesh model present-day climate, while Figure 5b shows the difference between the present and the future. The vertically integrated zonal and meridional moisture flux is calculated at every time step, and then monthly mean value is stored. This vertically integrated moisture flux is dominated by the lower troposphere owing to abundance of moisture near the surface, and thus is representative of the lower tropospheric moisture fields. Large easterly moisture flux from the tropical Atlantic Ocean through ITCZ region into the Amazon is simulated in the present-day climate. This moisture flux hits the Andes and deviates to the south associated with the South American Low-Level Jet (SALLJ) with its maximum around 60°W, 20°S. This corresponds well to the observed analysis. As shown in Figure 5b, there is an increase in moisture flux convergence over Amazon. There are significant moisture flux changes into the Brazilian Plateau. Northerly moisture flux along the SALLJ significantly increased. Intensification of the SALLJ with the global warming is also reported by Soares and Marengo , who found an intensified and more frequent SALLJ in their HadRM3P RCM results under the IPCC A2 and B2 scenarios in the future, transporting more moisture from the Amazonia to the Parana River Basin to result in more frequent and intense heavy rainfall events [Marengo et al., 2009].
Figure 5. (a) December–January–February (DJF) mean vertically integrated moisture flux vectors and its divergence in millimeters per day by the 20 km model. (b) Same as Figure 5a but for the difference between the future and the present. (c) Same as Figure 5a but for June–July–August (JJA). (d) Same as Figure 5b but for JJA.
Download figure to PowerPoint
 The vertically integrated moisture flux and its divergence in JJA are shown in Figures 5c and 5d. In the present-day simulation, the southeasterly flux dominates in the tropical South America near the Atlantic coast, which transport moisture from south of the equator to north of the equator, with moisture flux divergence to the south and convergence to the north of the equator. This mainly contributes to the dry season in northeast Brazil and the southeastern Amazon. There are northerly moisture flux in Bolivia and Paraguay, with the SALLJ maximum around 60°W, 20°S, bringing about moisture into the Parana River Basin and southeastern South America. In the future, the difference pattern is almost similar to the present-day moisture flux and its convergence, thus enforcing the present-day characteristics. This suggests that the contribution from the moisture buildup dominates that from the circulation changes, thus enhancing the wet-dry contrast; that is, the wet region becomes wetter and the dry region becomes drier.
 Figures 6–8 show seasonal mean changes in evaporation, soil wetness at the uppermost layer of the model, and surface runoff between the present and the future for the 60 km mesh model ensembles and the 20 km mesh model. Owing to an overall temperature increase, the evaporation increases throughout a year almost all over South America (Figure 6). An exception is found in northeastern Brazil where evaporation is projected to decrease in the dry season (JJA and SON). This is associated with drier soil over that region (Figure 7). Drier soil is not restricted to northeastern Brazil, but is projected to occur over most of the continent except for northern and central Argentina. Soil dry-up is particularly large in the dry season. Even in the wet season (DJF), Amazon is expected to be much drier in the future at the uppermost layer of the soil. Eastern Brazil is expected to have wetter soil in DJF by the 60 km mesh model results. The soil wetness change in DJF by the 20 km mesh model over eastern Brazil is, however, not significant. Significant changes in the runoff by the 20 km mesh model are limited to some regions (Figure 8), but in those regions the 60 km mesh model also shows significant changes with the same sign with the 20 km mesh model. In SON and DJF, runoff in western Amazon will decrease while it increases in eastern and southern Brazil. In MAM, northwestern Amazonia experiences more runoff. Further analyses for regional hydrological changes are done in section 4.3.
4.2. Precipitation Extremes
 Global warming would result in not only changes in mean precipitation but also increases in the amplitude and frequency of extreme precipitation events. Changes in extremes are more important for our adaptation to climate changes. Here two extreme indices for precipitation are used to illustrate changes in precipitation extremes over South America, one for heavy precipitation and one for dryness. The extremes indices are calculated for each year on the annual basis. Statistical significance is calculated using these annual values. Statistical significance is only plotted for the 60 km mesh model. Significance level for the 20 km mesh model is actually very low, and barely surpasses the 95% level over the South America land area on the grid point basis.
 Figure 9 shows the changes in maximum 5 day precipitation total (RX5D) for the 60 and 20 km mesh models. Almost all over South America except for Chile and some spotty areas, RX5D is projected to increase in the future warming world both in 20 and 60 km mesh models. Large RX5D increase is found over Amazonia, southern Brazil and northern Argentina. This is consistent with Kamiguchi et al. , but is different from Marengo et al.  where increased intensity of precipitation is found over southeastern South America and western Amazonia while a slight decrease or no change is projected in northeast Brazil and eastern Amazonia. The reason for the difference is not clear.
Figure 9. Changes in maximum 5 day precipitation total (mm) between the present and the end of the 21st century for the (a) 60 km (b) 20 km mesh atmospheric general circulation models. For the 60 km model, areas where all 4 different sea surface temperature experiments show consistent changes in sign are hatched. For the 20 km model, statistical significance level is not shown. Zero lines are contoured.
Download figure to PowerPoint
 It is noted that the 20 km mesh model projects larger increase in RX5D than the 60 km mesh model. However, this is only true in grid point scale. When averaged over the whole land area of Figure 9, RX5D increases from 99.0 to 112.9 mm/d in the 60 km model, while it is from 101.8 to 116.1 mm/d. Therefore the area average change itself is almost similar to each other (13.9 versus 14.3 mm/d), but the higher-resolution model shows the larger increase in concentrated areas. This implies the resolution dependency on extremely large precipitation and also the difficulty in quantitative assessment of those changes at the local climate level.
 Figure 10 shows the changes in maximum number of consecutive dry days (CDD) for the 60 and 20 km mesh models. Here “dry day” is defined as a day with precipitation less than 1 mm/d. Large CDD change is projected over the Brazilian Plateau, and northern Chile and Altiplano. Kamiguchi et al.  analyzed future increases in CDD over the Brazilian Plateau using a former experiment data set with the 20 km mesh model. In the present-day simulation, some intermittent rain occurs in dry season (JJA). This intermittent rain in the dry season is verified by station data. In the present-day simulation, when the rain takes place in the dry season, equatorial easterly low-level winds from ITCZ brings about scattered clouds to the Brazilian Plateau. In the future climate, a weak Walker circulation associated with the El Niño–like SST changes and a weak Atlantic anticyclone contribute to the weakening of the equatorial easterly wind and the suppression of rainfall over land, and thus, this weak rainfall in the dry season vanishes, and thus longer CDD occurs.
4.3. Regional Average: Precipitation and Seasonal Cycle
 We have selected five domains for the regional analyses (Figure 11). Rather than defining the small local area, we keep a broad region as much as possible (most have 15 degree longitude by 15 degree latitude area). So some regions such as Patagonia include both the upstream and downstream sides of the westerlies across the Andes mountains with different climate regimes. This caveat can be remedied by looking into geographical figures in other sections.
Figure 11. Definition of target regions for regional analyses of land-only precipitation. Amazon (AMZ), 72.5°W–57.5°W, 10°S–5°N; Nordeste (NOR), 50°W–35°W, 15°S–0°S; Parana (PAR), 63°W–48°W, 40°S–25°S; Andes (AND), 75°W–67°W, 40°S–15°S; Patagonia (PAT), 75°W–60°W, 55°S–40°S. The locations of the six gauging stations used in section 4.4 are also shown.
Download figure to PowerPoint
 Figures 12, 14, 16, 18, and 20 show the DJF and JJA mean land-only precipitation averaged for these domains. In each figure, three observations (CMAP data, CRU 2.1 data and GPCP 1.0 degree data) are compared with the present-day simulations by the 20 and 60 km mesh models. For the future, the 20 and 60 km mesh model projections are plotted for the CMIP3 SST experiment, and 3 member 60 km mesh model data are plotted for three different SSTs used. A horizontal line denotes the ensemble mean of three individual SST experiments. Note that vertical scales are not the same for each figure. Generally speaking, the domain averaged precipitation of the three member ensemble simulations with different initial conditions by the 60 km mesh model does not show much difference as three crosses and a circle have almost same values. Second, the ensemble mean of three individual future SST experiments is close to the CMIP3 SST experiment results.
 Figures 13, 15, 17, 19, and 21 show the seasonal cycle of the monthly mean precipitation, evaporation and runoff averaged over these five selected domains. For the difference between the future and the present, both the 20 and 60 m mesh model results are shown. When the change is statistically significance at the 95% level in each month, solid circles and open circles are marked for the 20 and 60 km mesh models, respectively.
 In the Amazonia region (Figure 12), the models overestimate the observed DJF and JJA precipitation. It is noted that the 20 km mesh model more matches the observations than the 60 km mesh model. In the future, the DJF and JJA precipitation is projected to increase. The results are similar between the 20 and 60 km mesh models. Here in Amazon, scatter among four 60 km mesh realizations with different SSTs is largest compared with other four domains. The 60 km mesh model with the MRI SST shows the little change. The annual mean precipitation change shows an increase in northern Amazon, and a decrease in southern Amazon. As pointed out by Li et al. , the SST distribution, such as El Niño–like or La Niña–like, affects atmospheric circulation and thus precipitation changes. Figure 13 shows that the future precipitation is projected to increase in the wet season and changes are smaller in the relatively dry season (July–September), thus increasing the seasonality in precipitation. There is a larger increase in April–May when there are precipitation maxima in the present. In this region, evaporation takes the maximum in the relatively dry season of August–October. In the future, evaporation will increase throughout the year owing to surface temperature increase. In Amazonia, precipitation exceeds evaporation in every month, and thus there is considerable runoff with its maximum in April–May and minimum in September–October. Future climate change will exaggerate this seasonal cycle in runoff. An increase in runoff is projected in this region during a first half-year (March–July), and a decrease during September–November.
Figure 12. Amazon (AMZ, 72.5–57.5°W, 10°S–5°N) region average land-only precipitation during (a) December–January–February (DJF) and (b) June–July–August (JJA). Units are in millimeters per day. The left sides of Figures 12a and 12b show the present-day climate, and the right sides show future climate. Observational data are CMAP (2.5 degree, 1979–2003, 25 years), CRU 2.1 (0.5 degree, 1979–2002, 24 years), and GPCP 1DD (1.0 degree, 1997–2008, 12 years). Red diamonds show the 20 km model simulations. Black crosses show the 60 km model simulations. Green circles show the ensemble average of three 60 km model simulations with different initial condition. Horizontal lines at the top right show the average of all 9 ensemble simulations by the 60 km model forced with sea surface temperatures (SSTs) of CSIRO, MRI, and MIROC atmosphere-ocean general circulation models.
Download figure to PowerPoint
Figure 13. (top) Seasonal cycle in (a) precipitation, (b) evaporation, and (c) runoff averaged for Amazonia (AMZ, 72.5°W–57.5°W, 10°S–5°N). Units are in millimeters per day. The horizontal axis shows month. Only the land grid points are used. Dashed lines show the present-day 20 km model. Solid lines show the future 20 km model. (bottom) Future minus present day for the 20 km (thick solid lines) and 60 km models (thin solid lines). Note the different ranges between Figures 13a, 13b, and 13c. Significance at the 95% level in each month is denoted by solid and open circles at the upper or lower end of the box for the 20 and the 60 km models, respectively.
Download figure to PowerPoint
 In the Nordeste region (Figures 14 and 15), the models overestimated the precipitation in DJF, but well reproduced the JJA precipitation. In the future, the DJF precipitation will increase while the JJA precipitation will not change much. The results are similar between the 20 and 60 km mesh models. The model projects an increase in precipitation during the pre–wet season (January–February). Evaporation will decrease in a dry season (August–October) even with higher temperature due to decreased precipitation and drier soil conditions. The runoff will increase during the wet season.
 In the Parana region (Figures 16 and 17), the models overestimated the precipitation in DJF and JJA compared to CMAP and GPCP, but in close agreement with the CRU data. In the future, the DJF precipitation will increase while the JJA precipitation will not change much. The results are similar between the 20 and 60 km mesh models. Here, all the precipitation, evaporation and runoff will increase in the future almost throughout the year, except for winter precipitation that does not show significant changes. An increase in precipitation and runoff during March–April and October–December is notable, thus contributing a prolonged wet season.
 In the Andes region (Figures 18 and 19), the models overestimated the precipitation in DJF and JJA. In the future, the DJF precipitation will not change while the JJA precipitation will slightly decrease. The results are similar between the 20 and 60 km mesh models. The seasonal cycle in the present-day climate shows that there is a double peak in precipitation, one in January–February and the other in June–July. The model projects a notable decrease in this second peak (a decrease in May–June is significant). As the evaporation has a single peak in summer season (December–January–February), the runoff has a double peak of comparable magnitude. A winter runoff peak will decrease in future. As the change in evaporation is small, circulation changes mostly explain the precipitation changes.
 In Patagonia (Figures 20 and 21), the models overestimated the precipitation in DJF and JJA. The MM5 simulations [Solman et al., 2008] also overestimate the austral winter maximum over this region. It is argued that this difference may be partly plausible due to insufficient observation network in this mountainous area. It is also noted that a steep Andes makes different climate regimes between the western side and the eastern side of the Andes, making these broad area averages obscure. This is also the case for the future projections. There is a clear annual cycle in precipitation with a maximum in winter. In this region, the 20 and 60 km mesh models contradict each other in precipitation changes in future in austral winter. The 20 km mesh model projects large increase in winter precipitation west of the Andes (Figure 4f), while the 60 km mesh model shows an overall decrease in this region (Figure 4e). The 60 km mesh model projects a decrease in precipitation during February–June, while the 20 km mesh model shows a decrease only in February–March. The MAM mean precipitation changes are mostly limited to the west of the Andes in the 20 km mesh model (Figure 4d), while the 60 km mesh model projects a decrease for the entire area (Figure 4c). This clearly illustrates the effect of horizontal resolution to the future climate projections.
 We select 6 rivers to assess the future projection of climatological mean streamflow. Table 2 lists the characteristic features of the 6 rivers including gauging stations with basin areas.
Table 2. Selected Rivers and the Gauging Stations With Areas and Reproducibility of Annual Streamflow in the 20 km Mesh Present-Day Climate Simulation
|River||Total Area (km2)||Station||Location (Latitude, Longitude)||Subarea (km2)||Annual Streamflow Ratio (Simulated/Observed)|
|Orinoco||945,000||Puente Angostura||8.15°N, 63.6°W||836,000||1.03|
|Sao Francisco||630,000||Juazeiro||9.42°S, 40.52°W||510,800||0.87|
|Negro||116,000||Primera Angostura||40.43°S, 63.67°W||95,000||0.17|
 The annual streamflow is well reproduced in the 6 rivers except for the Parana and Negro Rivers (see Table 2). Both the river basins receive precipitation similar between the present-day climate simulation and in the observation (figure not shown). Evaporation, however, may be more overestimated than precipitation; therefore, streamflow as the residual is underestimated in the Parana and Negro Rivers.
 Figure 22b shows the percentage change in climatological annual streamflow between the present-day and future climates in the 20 km mesh simulations. The percentage change ratio in annual streamflow in the future climate projection shows some regional characteristics. It is projected that annual streamflow increases in the major stream region and some part of northwest Amazonia. There will be large decrease in annual streamflow in the southwestern Amazonia (75°W–60°W, 15°S–5°S). Increasing annual streamflow is projected in the central western part of the Parana River basin. Annual streamflow will decrease over southern Argentina including the Negro River basin. Indistinctive annual streamflow change at each gauging station is projected in the Orinoco, Tocantins, and Sao Francisco River; however some areas of these basins show statistically significant changes: increase in the upper basin areas of the Tocantins and Sao Francisco River, and decrease in the middle basin area of the Orinoco River. An overall spatial pattern of projected change in annual streamflow by the 20 km mesh model is similar to the CMIP3 multimodel analysis by Nohara et al. , except for the western Amazonia where the 20 km mesh model shows a distinct difference south of 5°S and north of that latitude, while the CMIP3 multimodel analysis shows general increase in annual streamflow there.
Figure 22. Percentage changes in streamflow between the present and the end of the 21st century for the (a) 60 and (b) 20 km mesh atmospheric general circulation models. For the 60 km model, areas where all 4 different sea surface temperature experiments show consistent changes in sign are hatched. Zero lines are contoured.
Download figure to PowerPoint
 Most of the 60 km mesh model experiments agree on the change direction of the 20 km mesh model experiment in the future climate as shown in Figure 22. Poor agreement areas overlap the areas with no statistical significance or transition zone from one sign to another in Figure 22a. The contrast of annual streamflow change between southwest and northeast Amazonia is robust. So are increasing annual streamflow in the central western part of the Parana River basin and decreasing one in the Negro River basin. Nakaegawa and Vergara  projected annual streamflow in the Magdalena River, Colombia, bordered by the west of the Orinoco River using the 20 km mesh model without quantitative robustness, and Figure 22b can add information that the signs of the changes in the basin are robust. Milly et al.  presented similar results but with 24 CMIP3 multimodel analysis. An overall spatial pattern of high intermodel agreements on the change direction in the multimodel analysis is similar to that of Figure 22 except for the northeastern South America and small-scale features appearing in only our results. It is plausible that the horizontal resolution of the CMIP3 models is too coarse to resolve complex terrain in their simulation of streamflow and/or the multimodel ensemble mean process has diluted regional-scale characteristics in annual streamflow distribution. The high-spatial-resolution results presented here are crucial, since countermeasures to climatic hydrological changes highly depend on locations.
 Figure 23 shows the seasonal hydrographs for the six rivers in the present day and the future climates and the percent change in monthly streamflow. Seasonal hydrograph for each river is very well reproduced. Seasonal peak streamflow occurs in March–April in the Tocantins River, in May–June in the Amazon River, and in July–August in the Orinoco River. The seasonal variations of hydrographs for both the Parana and Negro Rivers are not so well captured, although this reproducibility is better than the reproducibility of annual streamflow. The amplitudes of hydrographs in the Tocantins, Sao Francisco and Parana Rivers are larger in the present-day climate simulation than in the observations. Juazeiro in the Sao Francisco River is located just after the Sobradinho Reservoir, the capacity of which corresponds to 2.5 year annual streamflow. A large reservoir, the Tucuruí Reservoir exists in the Tocantins River as well. The low (high) flow is controlled to increase (decrease) in the real world and therefore the amplitude of hydrograph is larger in the present-day climate simulation in Figures 23c and 23d and the peak of hydrograph is shifted. The Parana River has two massive reservoirs, Yacyretá and Itaipu Reservoirs, which reduce the amplitude of hydrograph artificially as well. Since the model simulation does not include any effects of artificial intervention by human such as dams, these differences are reasonable. In addition, the overestimation of precipitation or runoff in the present-day climate simulation may be a possible reason for these overestimations.
Figure 23. Climatological monthly streamflow (m3/s) for the six rivers in the present-day (black) and future climates (red) and the percentage change in climatological monthly streamflow (blue). Long-term mean values (dotted line) with the range of interannual standard deviations (shadings) of streamflow at the gauging stations listed in Table 2 by the Global River Discharge Center are shown together. A blue circle denotes a statistically significant change at the 95% level. A black line with square shows consistent changes in sign for the four different sea surface temperature experiments.
Download figure to PowerPoint
 Climatological monthly streamflow at Puente Angostura in the Orinoco River for July and August is projected to significantly increase, although the annual one is not so. The amplitude of the seasonal hydrograph is projected to increase at Obidos in the Amazon River, since the future streamflow increases in high-flow season and decreases in low-flow season, implying more floods in wet season and droughts in dry season. The peak month of the high-flow season is also seen to delay and high-flow season becomes longer. Increasing climatological monthly streamflow at Posadas in the Parana River for November and December is projected. No significant change in climatological monthly streamflow is projected in the Tocantins and Sao Francisco Rivers where no significant change in annual streamflow is done as well. These features for each river resemble those of runoff depicted in Figures 13, 15, 17, and 19 except for Figure 21 where the difference in runoff change between the 20 and 60 km mesh AGCM experiments is large.
 The robustness indicated in Figure 23 is consistent with the statistical significance. Statistically significant change in climatological monthly streamflow for the 20 km mesh AGCM experiment mostly corresponds to consistent changes in sign for all the four different SST experiments not vice versa. Therefore, the statistically significant changes for the 20 km mesh AGCM experiment are robust as well. The changes in climatological monthly streamflow in the Tocantins and Sao Francisco Rivers are not statistically significant, but the increase in high-flow season seems robust. Such information may be useful for impact assessment and decision making.