Falkenmark and Rockström (2004) initiated a paradigm shift in water resource management by proposing to change the traditional notion of the freshwater source from rivers, reservoirs, and aquifers to precipitation. In their description of water resources, precipitation is partitioned at the soil surface into blue water (as surface or groundwater) and green water (as soil moisture regenerated by precipitation), typically absent in water management considerations (Falkenmark and Rockström, 2004, 2006). Traditional blue-water management considers ET as a flow of water lost to the atmosphere (Oki and Kanae, 2006), while green-water resource management calls for a focus on vapour supply to the atmosphere for precipitation recycling (Falkenmark and Rockström, 2004, 2006; Ellison et al., 2012). On the land, such vapour supply is represented by green water consumed mainly by ET but also by some unconsumed water that is returned to the atmosphere through evaporation from soil moisture, snow, or ice sublimation, and evaporation of water intercepted by human made or natural landscapes.

Blue and green water are distinguished by their physical state as well as the processes, frequency, and factors of influence that govern their consumptive uses (Table 1). Consumptive uses are different from water withdrawals in that withdrawals can be returned to the blue water cycle, whereas consumptive uses cannot (Rockström et al., 2010). Blue-water consumptive uses include some fraction of drinking water, evaporative losses through cropland irrigation or hydropower, and incorporation of water into products. Green-water consumptive uses exclusively occur through ET with a distinction between productive and unproductive vapour flows characterized respectively by transpiration and direct evaporation of soil moisture (Falkenmark and Rockström, 2006).

Aquatic ecosystems rely exclusively on blue water and may require inflows of surface water or groundwater to ensure proper function (e.g. wetlands, fisheries). Blue water is shared between humans and ecosystems such that a consumption activity will require blue-water trade-offs between users. Green-water resources are exclusively consumed through ET processes and, as such, are consumed only once by terrestrial ecosystems, in the case of productive green-water consumption, before returning to the atmosphere (Rockström and Gordon, 2001). Evaporation of soil moisture or water intercepted by canopies also regenerates precipitation unproductively, meaning the process does not support any additional human or ecosystem activity, although the latent heat uptake from evaporation can be important to regional water balances (Biggs et al., 2008). Water vapour flows resulting from blue-water consumption (Karlberg et al., 2009) illustrate the special case of ET from irrigation, a blue-water resource.

The above representation of the water source as precipitation and its partitioning into blue- and green-water resources brings new considerations to water resource management: the importance of vapour flows and precipitation recycling, as well as potential blue- and green-water trade-offs with land management discussed below.

By emphasizing the importance of green-water resources, Falkenmark and Rockström (2004) highlighted the need to regenerate rainfall through ET to ensure a continuation in the supply of precipitation, the water source in the above paradigm. Precipitation recycling occurs at different scales when considering local (e.g. same watershed), regional, or continental effects since regional terrestrial recycling is generally larger than local recycling (40 and 13 % respectively; Ellison et al., 2012) and increases as a function of area (van der Ent et al., 2010). It is estimated that about 65 % of global terrestrial precipitation is sourced from terrestrial water vapour flows to the atmosphere (Karlberg et al., 2009) (Table S1 in the Supplement), with the remaining 35 % sourced by evaporation from oceans and other water surfaces (liquid, ice, and snow) (Oki and Kanae, 2006). Green-water resources therefore recharge “atmospheric watersheds” or “precipitationsheds” which connect soil moisture evapotranspired from one source region to a sink region, further downwind (Keys et al., 2012). As ET is typically water-limited in arid, semi-arid, and highly seasonal environments, these regions are particularly reliant on green-water resources to regenerate precipitation (Falkenmark and Rockström, 2004, 2006; Rockström et al., 2010).

Key terrestrial ecosystem services are maintained through the consumption of green-water resources, including agro-ecosystems, with vapour flows ensuring 90 % of global human needs (Rockström and Gordon, 2001). Global terrestrial ecosystems were estimated to supply 42 900–45 646 km3 yr-1 of water vapour through transpiration, while 14 682–15 478 km3 yr-1 of water vapour is estimated to be supplied by rain-fed agroecosystems (cropland and pasture) (Gerten et al., 2005; Rost et al., 2008).

Since green- and blue-water resources emerge through the partitioning of precipitation at the soil interface, there are green- and blue-water trade-offs with every land use decision (Karlberg et al., 2009). Several models have attempted to provide annual average ET estimates for major biomes of the world (Table S3), which highlight partitioning of precipitation and potential trade-offs based on vegetation and climate. These trade-offs depend on the ecohydrological relationship between the vegetation and the water cycle, which can be explained through environmental and physiological controls on ET. Environmental controls are illustrated by the relationship linking energy, climate, vegetation, and land use. In 299 basins, Zeng et al. (2012) found strong correlations between ET and mean annual temperature (R=0.68), annual precipitation (R=0.87), and NDVI (R=0.70). Analysis of 21 tropical eddy covariance sites showed a strong correlation of latent heat flux with net radiation (R2=0.72) but weak correlations with vapour pressure deficit (R2=0.14) and NDVI (R2=0.09) (Fisher et al., 2009). These relationships show the dependence of ET on green-water supply and environmental demand (sensu Christoffersen et al., 2014).

Changing the landscape can affect both environmental and physiological controls on ET with consequences on green- and blue-water trade-offs. There being shallower root systems of crops or pasture compared to woodlands or tropical forest is one important morphological condition that explains differences in precipitation partitioning (Table S3) in addition to the amount of water supplied by precipitation or access to deep groundwater (blue water) reserves (Matyas and Sun, 2014). Declines in ET from the landscape due to changes in climate or land use can lead to an increase in blue water downstream due to runoff (Karlberg et al., 2009) and consequently a reduction in moisture recycling to regenerate precipitation locally (Savenije, 1995, 1996). However, the scale of these processes is an important consideration since reduced precipitation recycling at regional scales can also decrease blue water by lowering rainfall inputs into rivers (Ellison et al., 2012).

Research linking global agricultural practices to changes in vapour flows has emphasized the importance of deforestation and irrigation expansion in some regions with expected consequences on precipitationsheds (Keys et al., 2012). Deforestation was estimated to have reduced transpiration by over 100 mm yr-1 in regions of intensive land use and land cover change, with a 7.4 % decrease in transpiration simulated globally for the 1961–1990 period (Gerten et al., 2005). Another study concludes that agricultural expansion through deforestation led to a global water vapour loss of 3000 km3 yr-1, while irrigation expansion has increased flows by 2600 km3 yr-1, suggesting only a small net loss due to land use and cover change (Gordon et al., 2005). In parallel, Gerten et al. (2005) also predict a 2.2 % increase in runoff, while Rost et al. (2008) estimate a 5 % increase in river discharge accompanying a 2.8 % decrease in ET from land use change and a 1.9 % increase in ET from irrigation. Such trade-offs would tip the balance toward greater blue-water yields. However, the question of scale needs to be addressed as precipitation recycling might counteract this effect and so, just like global changes in ET, global changes in runoff might not represent regional effects. Reductions in ET can lessen precipitation hundreds or thousands of kilometres away, thus also impacting river discharge (Coe et al., 2009, 2011; Ellison et al., 2012; Spracklen et al., 2012; Stickler et al., 2013; Spracklen and Garcia-Carrera, 2015).

As the largest consumer of water resources, agriculture has been the focal point of early research on green and blue water. An additional 1700 and 1550 km3 yr-1 of water consumptive use is expected for increases in food production and carbon sequestration projected for 2050 respectively, compared to the current total blue-water consumptive use of 2600 km3 yr-1 (Rockström et al., 2014). This combined 5800 km3 yr-1 approaches the upper limit of the estimated planetary boundary of 4000–6000 km3 yr-1 (Rockström et al., 2014); given that blue-water resources are already stressed in many regions of the world, there seems to be limited opportunity to feed the world solely through irrigation expansion.

Green-water consumption in rain-fed agriculture represents about 75 % of total cropland consumptive use (green and blue water), which is 4–5 times greater than blue-water consumptive use in irrigation according to seven global models (Hoff et al., 2010). Estimates of global cropland ET were predicted to be between 3272 and 7200 km3 yr-1 based on models considered, while pasture ET exceeds 4000 km3 yr-1. Rain-fed agriculture is often key to securing the livelihoods of those living in poverty, especially in drylands or savannah regions where crop water requirements typically exceed precipitation (Rockström et al., 2009). Such regions are not necessarily considered water-scarce; rather it is the intensity and timing of precipitation throughout the year and its concentration in short wet seasons which present challenges for land and water resources management. As such, these regions, especially sub-Saharan Africa, have been the focus of research on upgrading rain-fed agriculture: under improved management, the current 10–30 % use of green water could increase to 50 % with significant increases in yields (Falkenmark and Rockström, 2006). Such a strategy aims to reduce evaporation (unproductive green water) and increase transpiration (productive green water) through a so-called “vapour shift” (Rockström, 2003; Rockström et al., 2007).

Improvements in water productivity, or the amount of crops produced per unit input of water consumed (Cai et al., 2011), could decrease the water requirements for food production in 2050 by almost 2850 km3 yr-1, broken down into 725 km3 yr-1 of blue water (i.e. irrigation) and 2125 km3 yr-1 secured through green-water resources (i.e. rain-fed agriculture) (Rockström et al., 2010). Strategies for increasing food production and reaching self-sufficiency differ based on the green- and blue- water resource potential of each country (Rockström et al., 2009). An increase in the use of blue-water resources for irrigation remains a viable option as long as it does not promote further water scarcity and does not impose damages to aquatic ecosystems or land subsidence. The expansion of rain-fed agriculture into native terrestrial ecosystems remains another option for countries that could significantly expand green-water resources for food production, although not without affecting biodiversity and precipitation recycling while increasing local runoff. Virtual water imports, or the import of water virtually via agricultural products from national and international trade, are the final option for countries that are already chronically blue- and green-water-short (e.g. Jordan, Israel, Pakistan, Iraq) (Rockström et al., 2009), and a strategy which has been under scrutiny in order to qualify water savings from trade (Dalin et al., 2012; Hoekstra and Mekonnen, 2012).

An important portion of Brazil's economic future is focused on the continuous increase of production for export of soybean and maize feed for cattle, and beef (MAPA, 2013), which itself relies on south-eastern Amazonia's strongly seasonal agricultural frontier, with semi-arid conditions during extended periods (Fig. 1). This frontier is located in the headwaters of main tributaries of the Amazon River, in which fisheries, navigation, and hydroelectric projects are important downstream blue-water users (Castello and Macedo, 2016). For example, the 176 000 km2 upper Xingu River Basin of Mato Grosso contains over 22 000 springs feeding the 510 000 km2 Xingu Basin (Fig. 1) (Velasquez and Bernasconi, 2010; Macedo et al., 2013) and may soon be home to the future Belo Monte dam, which will require significant amounts of blue water to operate (Stickler et al., 2013). For such a major production centre for commodities, increases in agricultural production will need to consider green- and blue-water trade-offs from possible production pathways such as expansion into natural ecosystems, expansion into pastureland, or intensification into current land, along with additional irrigation as insurance for dry spells and drought years. The additional water vapour supply from irrigation as well as other upstream water bodies (e.g. small farm dams) represents an important planning consideration for the regional water cycle. While much of previous research has focused on regional temperatures and greenhouse gases (Oliveira et al., 2013), precipitation recycling (Stickler et al., 2013; Bagley et al., 2014), river discharge (Coe et al., 2011, 2009; Panday et al., 2015), and impacts to biodiversity (Chaplin-Kramer et al., 2015), detailed modelling studies on how potential increases in regional water vapour flows from irrigation may impact the water cycle in Amazonia are still lacking despite the current state of knowledge on atmospheric feedbacks from land use change. We explore these implications by focusing exclusively on green- and blue-water trade-offs in the region.

The Amazon Basin is abundant in both green and blue water (see Supplement), whose trade-offs result from environmental and biological controls of ET. Environmental controls follow a precipitation gradient that declines from north to south over the 0–11∘ S latitudinal band (Manaus to Sinop in Table 2, Fig. 1). In equatorial Amazonia (e.g. Manaus, Santarem), ET seasonality is primarily driven by radiation but is also driven by morning fog, especially in the wet season (Anber et al., 2015). The dry season occurs later in the calendar year (July–November), when increasing solar radiation coincides with limited cloud cover, favouring photosynthesis and increasing ET to more than 100 mm month-1 (Restrepo-Coupe et al., 2013; Christoffersen et al., 2014). In equatorial Amazonia, latent heat flux is correlated to net radiation (R2=0.53), suggesting that available energy is a strong control on ET within the latitudinal band (Restrepo-Coupe et al., 2013). With little or no soil moisture stress affecting the productivity of broadleaf evergreen forests, ET in equatorial Amazonia is only mildly seasonal as green-water stocks remain largely available for ecosystems to consume all year round.

In contrast, ET in southern Amazonia is strongly seasonal. Remote-sensing observations from MOD16 (2000–2009) for the Amazon–Cerrado transition forest showed a forest-wide ET of 65 mm month-1 for August periods, only 60 % of rainy-season ET values of 105–115 mm month-1 between January and April (Lathuillière et al., 2012). Future increases in regional temperatures could lead to an overall basin-wide increase in ET due to an increase in potential ET, limited, however, by regional differences in soil moisture availability, as well as groundwater reserves, which can be deeper than 20 m in south-eastern Amazonia (Pokhrel et al., 2014).

Biological controls on ET have been shown to occur in southern Amazonia's transition forest where high vapour pressure deficit in the dry season can trigger stomatal closure and allow forest ecosystems to conserve water in water-limited conditions (Costa et al., 2010). Access to green water by deeply rooted trees has been suggested as a drought resilience mechanism for forest ecosystems in the region, with roots accessing soil moisture over 8 m deep (Nepstad et al., 1994; Davidson et al., 2011). Deeply rooted trees help sustain ET over southern Amazonia's dry season (Coe et al., 2009, 2011; Lathuillière et al., 2012; Christoffersen et al., 2014; Biudes et al., 2015; Panday et al., 2015; Silvério et al., 2015; Vourlitis et al., 2015) and will likely become more important with increased air temperatures (Pokhrel et al., 2014). As such, land use change resulting in the replacement of forest by more shallow-rooted pasture grasses or cropland reduces the amount of accessible green water and vapour flows to the atmosphere.

Given these well-defined processes across the basin and the important role of seasonality in the southern portion, south-eastern Amazonia appears as a region that requires special attention. The region's ET processes are water-limited during an extended dry season. A rise in the local dry-season temperatures shows the importance of soil moisture and groundwater as important water sources for deeply rooted trees to ensure continuous water vapour flows to the atmosphere (Pokhrel et al., 2014). Its geographical importance, both as the home to Brazil's expanding agriculture and as the region upstream of the Amazon River, make it an environmentally and economically important region that is sensitive to future land use and climate changes. Green- and blue-water trade-offs will be inevitable considering the current and future land use changes which decrease green-water consumption of terrestrial ecosystems and can increase blue water through runoff.

Brazil's internal colonization driven by the Agrarian reform of the 1960s brought intensive agricultural activity to the Amazon and Cerrado regions. The 1980s and 1990s saw the expansion of settlements in the region, starting with cattle ranching, later followed by soybean production, both of which created economic activity that also required road building and ever-increasing mechanization of deforestation and agriculture (Fearnside, 2001, 2005; Morton et al., 2006; Rudel et al., 2009). The country's expansion of soybean production followed a south-to-north progression into the Cerrado and closer to the Amazon biome (Simon and Garagorry, 2005), today reaching south-eastern Amazonia and the state of Mato Grosso (Fig. 1). Soybean expansion has occurred either directly through a forest-to-cropland conversion or indirectly through a pasture transition (Macedo et al., 2012; Spera et al., 2014; Silvério et al., 2015), which has displaced pasture further north into the Amazon (Barona et al., 2010).

The state of Mato Grosso (Fig. 1) is a hotspot for this expanding agricultural frontier with more than a decade of documented deforestation activity (Macedo et al., 2012). In accordance with land use change practices, government policies, and private initiatives, Nepstad et al. (2014) identified three distinct phases guiding deforestation: agro-industrial expansion (pre-2004), frontier governance (2005–2008), and territorial performance (2009–present). Mato Grosso and Pará have shown the greatest rates of deforestation in Amazonia, with an accumulated 138 289 and 137 923 km2 respectively for the 1988–2014 period, and, together, contributed to 70 % of total deforestation in Brazil (INPE, 2015). The Brazilian federal Forest Code is the main federal legislation controlling deforestation in Brazil. The 1965 version of the law requires a land reserve in which 80 % of native vegetation must be retained on properties located in the Amazon biome, but this requirement on native vegetation retained drops to 50 % for the Amazon–Cerrado transition zone and 20 % for the Cerrado (Brannstrom et al., 2008; Soares-Filho et al., 2014). A new version of the Forest Code was signed in 2012, which retains the old reserve and provides new rules for illegal deforestation prior to 2008, while adding new incentives to reduce deforestation such as trade in land reserves between properties (Soares-Filho et al., 2014).

The drop in Mato Grosso's deforestation rates in the late 2000s coincided with a drop in exchange rate of the Brazilian real, which increased the opportunity costs of deforestation (Richards et al., 2012); restrictive access to credit for producers located in municipalities labelled as hotspots of deforestation; and the Soybean Moratorium (2006) and Cattle Agreement (2009), which sought to remove any suppliers from the soybean and meat supply chains that have produced on land previously cleared from forests (Macedo et al., 2012; Nepstad et al., 2014; Gibbs et al., 2015). Soybean and beef production, however, continued to grow with further internationalization of commodity markets as China imported an ever-increasing amount of soybean to meet its increasing national demand, mostly for producing animal protein (Lathuillière et al., 2014). Deforestation has been apparently on the rise since 2012 (INPE, 2015), which coincides with the implementation of the new Forest Code (Soares-Filho et al., 2014; Spracklen and Garcia-Carrera, 2015). Land use change activities have been recognized to have an effect on the local climate (Davidson et al., 2012) with emerging evidence of changes in regional and continental precipitation recycling, with south-eastern Amazonia playing an important role in the Amazon Basin (Table 3).

Differences in the energy balance have been observed on different landscapes across Amazonia (Table 2). Therefore, land use change from one landscape to another is expected to affect radiation partitioning, with noted impacts on the water cycle. Model simulations in south-eastern Amazonia have shown that changes in land cover affect surface albedo, while morphological (vegetation height, root depth, albedo) and physiological changes (C3 to C4 photosynthetic pathways) can affect the magnitude of sensible and latent heat fluxes (positively and negatively), with possible effects on surface temperature (Pongratz et al., 2006; Davidson et al., 2012; Bagley et al., 2014). Analysis of satellite information obtained for the Upper Xingu River Basin of Mato Grosso showed that forest-to-cropland and forest-to-pasture land use transitions in the 2000s decreased ET (32 and 24 %), increased sensible heat flux (6 and 9 %), and increased surface temperature up to 6.4 ∘C (Silvério et al., 2015). In the Amazon Basin, deforestation reduced ET by 5 %, increased sensible heat flux by 2 %, and decreased precipitation by 6 % in the dry season (Bagley et al., 2014), all of which were exacerbated in drought years (6, 4, and 6 % respectively). Morphological differences in the root infrastructure can make green water more accessible to maintain ET processes during the dry season (Nepstad et al., 1994; Lathuillière et al., 2012).

The above changes in surface energy balance affect the partitioning of precipitation into blue and green water as quantified by runoff and ET. Field studies in south-eastern Amazonia have shown that soybean watersheds can have water yields up to 4 times greater than forested watersheds (Hayhoe et al., 2011; Dias et al., 2015). Coe et al. (2009) simulated runoff in the Amazon Basin through a coupled land–atmosphere and climate change numerical model. Most tested river basins exhibited an increase in discharge for 2000 and 2050 when compared to potential natural vegetation, even in a restrictive deforestation governance scenario. The Tocantins and Madeira rivers (Fig. 1) saw discharges increase from 26 and 7 % in 2000 to 18 and 32 % for 2050 respectively (Coe et al., 2009). Similarly, discharges of the Xingu and Araguaia rivers have increased 6 % (1970 to 2000s) and 25 % (1970 to 1990) primarily due to deforestation and climate (Coe et al., 2011; Panday et al., 2015).

Local changes in land cover also change vapour supply to the atmosphere, which can reduce regional precipitation and, indirectly, river discharge in the basin (Coe et al., 2009, 2011; Ellison et al., 2012). The Amazon Basin is the source of moisture to a precipitationshed that provides subtropical rainfall as far south as the La Plata Basin through the South American Low-Level Jet (Marengo, 2006; Keys et al., 2012). Vegetative surfaces promote additional vapour inputs into air masses that result in precipitation in downwind areas (Spracklen et al., 2012), with evaporated water sources contributing to continental precipitation less than 2000 km away (van der Ent and Savenije, 2011). Results for the 2000–2009 period show a 10 % drop in the contributions of forests to total water vapour flows to the atmosphere (a decrease of 119 km3 from 593 km3 yr-1 to 474 km3 yr-1) due to a shift of green water use from natural ecosystems to agricultural production in the state of Mato Grosso (Lathuillière et al., 2012). In the same time period, the Upper Xingu Basin experienced a 35 km3 yr-1 ET drop due to land use change (Silvério et al., 2015). This reduction in vapour supply to the atmosphere can also affect river discharges and hydropower generation within the basin (Stickler et al., 2013).

Local land cover change compounds the effects of inter-annual variability effects of regional precipitation in the basin. While pluvial and drought years affected regional precipitation regimes, local deforestation impacts on precipitation were at least as important as regional effects (Bagley et al., 2014). Areas of deforestation showed up to a 20 % decrease in precipitation during the dry months of drought years (Bagley et al., 2014) with interconnected regions of precipitation source, or precipitationshed (e.g. central and southern Amazonia), to distant sinks (north-western Amazonia). On a local scale, precipitation in the Xingu River Basin was found to be sensitive to potential future deforestation both inside the confines of the basin and in the rest of the Amazon forest (Stickler et al., 2013). This means that land cover change in one region can greatly affect precipitation in addition to local recycling, such as in south and central Amazonia (Spracklen et al., 2012; Bagley et al., 2014). Drought years were found to increase recycled evaporation from 67 to 74 % in the dry months of south-eastern Amazonia (Bagley et al., 2014), indicating that atmospheric water demand can be met, in part, from regional sources.

A diminished vapour supply to the atmosphere as a result of land use change can affect regional precipitation patterns, which can in turn impact ecosystem processes and services in the region. Precipitation has been declining in Amazonia in recent years (Hilker et al., 2014). Analysis of 280 meteorological stations across the basin showed a decline in precipitation of 5.3±0.7 mm yr-1 for the 1996–2005 period, and an increase to 7.8±1.6 mm yr-1 in areas with denser tree cover (Brando et al., 2010). More recent analysis of satellite imagery confirmed a 17–30 % decline in precipitation over the greater portion of landscapes of the Amazon region for the 2000–2012 period, especially in eastern and south-eastern Amazonia, which showed a 25 % decline in precipitation for the period (Hilker et al., 2014).

In contrast, the Cerrado did not show any decline in precipitation between 2002 and 2010, but the biome did see an average increase in ET of 51±15 mm yr-1 (Oliveira et al., 2014). The above results are in line with a review of 26 studies linking deforestation to reductions in precipitation (Marengo, 2006) as well as 96 simulations (Spracklen and Garcia-Carrera, 2015). Deforestation is also known to increase the length of the dry season, particularly in southern Amazonia, with possible changes in the onset of the wet season (Costa and Pires, 2010). Since 1979, the end of the dry season in southern Amazonia has been delayed by 4.5±2.0 days per decade, with serious implications on the integrity of the tropical forest ecosystem should the dry season length continue to increase this Century (Fu et al., 2013).

Deforestation in a business-as-usual scenario could lead to declines in precipitation by 8±4 % in the Amazon Basin, which was greater than annual natural variability of 5 % (Spracklen and Garcia-Carrera, 2015). Southern and central Amazonia regions can be considered a hotspot for changes in recycled precipitation. This is due to water limitations on ET (Biudes et al., 2015; Vourlitis et al., 2015), high precipitation recycling ratios, and the extended dry season, especially in the south-south-eastern region. Declines in precipitation can be accompanied by reductions in vegetation greenness, which further impacts the availability of green-water resources and the local water balance (Hilker et al., 2014). Reduced precipitation diminishes the amount of green water available for terrestrial ecosystems, with a possible impact on net primary production. Results of a 5-year rainfall exclusion experiment (near CUE, Santarem, Fig. 1) showed an increase in tree mortality of 5.7 % yr-1 compared to 2.5 % yr-1 for the control plot, along with a decrease in aboveground live biomass by 25 % and an increasing difference in wood production between experimental and control plots by up to 58 % (Nepstad et al., 2002; Brando et al., 2008).