Introduction

Improving agricultural production is a pressing global need to meet the rapidly increasing demands of a growing population1. By some projections, food production must rise by 50–110% to meet these demands by the year 20502,3,4. However, feeding societies has become increasingly challenging due to the rising frequency of climate extremes on a warming planet5. In this regard, global rainfed breadbaskets, which contribute ~60% of the world’s food production6, are expected to play a critical role in ensuring food security7,8. Brazil’s rainfed breadbaskets, in particular, supply ~34% of the world’s soybeans, and contribute around 20% of the country’s gross domestic product9. Driven by rising global food demand, Brazil has seen a more than fourfold increase in grain production since 198010. This expansion has largely been achieved through extensive deforestation in the Cerrado region (south of the Amazon basin) and along the southeastern border of the Amazonian rainforest11,12, as well as through the adoption of soybean–maize double cropping systems during the wet season13. Consequently, Brazilian soybean agriculture has become essential for global food supply. Nevertheless, the productivity of the region is highly vulnerable to climate change, with 26–34% of its yield variability attributed to climate and weather fluctuations14,15. Stable and predictable rainfall is particularly crucial for planting and growth16,17,18. Recent trends, including shorter wet seasons19 and increased occurrences of dry and hot weather events20, are expected to threaten soybean productivity in Brazil, especially in the water-limited regions in the south and east.

Rainfall variability in Brazil, as in any other region of the world, is ultimately determined by changes in (a) local and upwind evaporation from land and oceans, (b) atmospheric dynamics (or circulation) that transport the evaporated moisture downwind, and (c) the tendency of the atmosphere to rain out (often referred to as ‘precipitation efficiency’) from source to sink (which is largely controlled by thermodynamic stability). Geographically, the atmospheric moisture resulting in rainfall over a region may come from evaporation from the region itself (precipitation recycling) or from upwind areas (both land and/or ocean), after being transported into the region by winds21. Rainfall changes over Brazilian breadbaskets are influenced by variability in each of the factors described above. During Brazil’s wet season from October to April, the heating of the land (and the subsequent lower troposphere warming) reduces air pressure. This pressure gradient draws moisture from the Atlantic Ocean and Amazon forest towards Brazilian croplands, contributing to rainfall in the region22. This process is largely governed by the South American Monsoon System (SAMS), which is driven by trade winds and the temperature difference between the Tropical Atlantic Ocean and South America23,24. Trade winds transport moisture from the Tropical Atlantic Ocean to the northern parts of South America, where it is redirected southwards by the Andes to form the South American Low-Level Jet (SALLJ)25,26. The SALLJ and other atmospheric processes further influence the South American Convergence Zone (SACZ), a region of intense rainfall stretching across central-eastern Amazon and southeast Brazil27. In addition, large-scale climate phenomena, such as the El Niño Southern Oscillation (ENSO), Atlantic Multi-decadal Oscillation (AMO), and Madden-Julian Oscillation (MJO), are known to modulate the transport of moisture from upwind regions into the Brazilian breadbaskets28,29.

Long-term changes in precipitation in the region have not only been attributed to climate change24 but also to the ongoing Amazon deforestation over the past decades30. A major terrestrial source of moisture that regulates rainfall across South America is the Amazon rainforest31,32,33. Evaporation from the Amazon contributes significantly to rainfall in southern South America, including Brazil’s southernmost soybean croplands25. During drought periods, forests mitigate precipitation deficits by accessing soil moisture from deeper soil layers34,35,36,37. Therefore, disturbances in the Amazon are expected to affect land–atmosphere fluxes, transported moisture, and finally rainfall over downwind regions. Deforestation, especially in the humid tropics, is expected to reduce evaporation, subsequently decreasing atmospheric moisture and downwind precipitation38, assuming that the prevailing large-scale circulation and stability remain constant. Thus, the ongoing deforestation driven by agricultural expansion is (in principle) expected to adversely impact the soybean growth in these croplands by potentially reducing rainfall volumes, delaying the onset of the rainy season, and enhancing air temperatures20,39. On the other hand, deforestation may increase moisture transport due to reduced surface roughness and increased low-level wind, potentially compensating for the impact of evaporation losses on downwind precipitation; these influences on circulation have rarely been studied40,41.

Several studies have examined the importance of wet-season rainfall variability in influencing soybean productivity in Brazil14,15,42,43,44. However, the local and upwind drivers of rainfall over the region remain relatively unknown and have typically been studied using climate model experiments45,46,47,48. Specifically, the role of Amazonian deforestation in modulating atmospheric moisture transport to Brazilian soybean breadbaskets has not been previously quantified, and certainly based on observational data. Here, we study the variability of wet-season rainfall over the key soybean breadbaskets in Brazil from the perspective of atmospheric moisture transport. We hypothesise that deforestation has influenced moisture transport, which in turn has affected rainfall and potentially soybean productivity in water-limited agricultural regions. To test this hypothesis, we disentangle the relative contributions of local and upwind oceanic and land-based moisture sources to rainfall over two major Brazilian soybean breadbaskets in the Amazon–Cerrado and Southern Brazil regions. First, we employ a moisture tracking framework based on a Lagrangian model constrained by satellite observations (see Methods) to identify the major evaporative moisture source regions. Subsequently, we attribute long-term changes in precipitation to changes in evaporation and other factors (represented by the parameter β), including changes in atmospheric circulation and precipitation efficiency throughout the entire atmospheric trajectory (see Methods). Our findings offer new insights into the understanding of long-term precipitation variability in Brazilian breadbaskets, particularly under pressures from climate change and Amazonian deforestation.

Results

Brazilian soybean breadbaskets and their contrasting dependence on rainfall

We identified two main soybean regions in Brazil based on the rainfed area fraction greater than 95% and harvested area greater than 20,000 ha in 2000 (for details, see Methods). The resulting regions encompass large parts of the southern Amazon rainforest and northern Cerrado region (hereafter ‘ACR’) and the southern Brazilian region, including the Paraná and Rio Grande (hereafter ‘SBR’). The soybean growing season is defined from October to April of the following year for both regions, based on data from the global crop calendar49, which is in line with the onset and termination of the wet/rainy season in most regions of Brazil42. These two regions contribute ~31% and ~50% of the soybean production in Brazil, respectively (pie chart in Fig. 1a). Together, they account for more than 25% of the world’s soybean production9. Note that SBR is a traditional agricultural region in Brazil50, while the expansion of cropland area in ACR has been more recent, and mainly at the expense of forests and shrublands20, which highlights a trade-off between ecological protection for tropical forests and agricultural production51. ACR receives annually ~1500 mm of rainfall, and has a mean temperature of 25.2 °C, making it a preferentially energy-limited region, with an aridity index, defined as precipitation over potential evaporation52, of 1.61, in which water availability is sufficient to meet the atmospheric demand for evaporation. On the other hand, SBR is more water limited (aridity index is ~1), with an average temperature of 21.8 °C, and on average receives less than 1000 mm of rainfall during the growing season.

Fig. 1: Brazilian soybean breadbaskets, agricultural dependency on rainfall, and growing season rainfall variability.
figure 1

a Soybean production (tons) in Brazil based on the year 2000, and calculated as yield in tons per hectare69 multiplied by harvest area in hectares49. The two contours in the figure, green and blue, enclose the two cropland regions studied: the southern Amazon and Cerrado region (ACR) and the southern Brazilian region (SBR). The pie chart in the lower right corner shows the relative contribution of each of them to the total soybean production of Brazil. The background map shows the elevation of South America. b Time series of rainfall over the growing period (from October to April) for both ACR (green) and SBR (blue) from 1982 to 2018. The growing-season precipitation was calculated as an area-weighted average using the Multi-Source Weighted-Ensemble Precipitation (MSWEP65). In the upper left corner, we show the long-term trend of rainfall (lines), calculated as the slope of a linear regression (S), and the p value associated with it. c Relationship between soybean yield and growing season rainfall over the rainfed breadbaskets. In the upper left corner, we show the Pearson’s correlation between the growing-season rainfall and relative soybean yield anomaly (Yp) that is calculated by removing the long-term trend (for details, see Methods).

Full size image

The rainfed nature and vulnerability to dry periods are reflected in the positive correlation between soybean yield and growing-season rainfall in SBR (r = 0.41, p = 0.01, Fig. 1c). However, the slightly negative correlation between soybean yield and rainfall in ACR (r = −0.25, p = 0.164) is the result of the region’s higher water availability during the growth period, which renders solar radiation and temperature as primary controls on photosynthesis (and thus agriculture productivity) in ACR. Since radiation and temperature tend to be higher during periods with lower cloud cover and rainfall, the correlation between rainfall and soybean yield becomes negative53,54. Growing-season rainfall does not show a significant trend over the period 1982–2018 in either region, despite potential influences of deforestation (Fig. 1b). This analysis is further corroborated by other observation-based precipitation products (Supplementary Figs. S1 and S2).

Rainfall origins and their variability for the studied soybean breadbaskets

To understand the long-term variability of rainfall over the past decades over the two studied soybean production regions, we calculate their year-to-year precipitation sources using the Lagrangian moisture tracking methodology (described in Methods). These precipitation sources indicate the regions from which rainfall in the breadbaskets originated through evaporation, or in other words, where the surface evaporation that contributes to precipitation (hereafter referred to as E2P) in the agricultural regions happens. At the climatological scale, our analysis shows that growing-season rainfall over the two soybean croplands is mostly of terrestrial origin (Fig. 2a, c). A large portion of the terrestrial contribution comes from local recycling, i.e. evaporation from the soybean croplands themselves. Moisture sources extend mainly northeastward for ACR and north-northwestward for SBR, as expected from their preferential wind regimes during this time of the year (see Supplementary Fig. S3). On average, terrestrial evaporation contributes ~80% of the rainfall over ACR and SBR during the growing season (Supplementary Fig. S4). The oceanic contribution is mainly from the tropical Atlantic for ACR (Fig. 2a) and from the South Atlantic for SBR (Fig. 2c). Above all, the high continental contributions to growing season precipitation highlight the potential influence of land cover and land-use changes on precipitation over both breadbaskets55.

Fig. 2: Rainfall origins and their variability during 1982–2018.
figure 2

a, c Climatological evaporation contribution to precipitation (E2P) in ACR and SBR, i.e. the average amount of surface evaporation that ends up as precipitation in these two studied regions. b, d Long-term change in E2P in ACR and SBR, calculated by multiplying the linear trends (mm per year) by 37 years, i.e. the length of the period considered in the study. Significant trends (p < 0.05) are highlighted with black hatches.

Full size image

When we examine the long-term changes in E2P in the agricultural regions (Fig. 2b, d), a very evident spatial pattern appears, which partly explains why in recent decades there has been no significant change in total precipitation during the growing season (as shown in Fig. 1b). The local moisture contribution to rainfall significantly decreased, but this decline was offset by a widespread increase in the contribution from remote sources (including Amazon rainforest and Atlantic Ocean, see Fig. 2b, d). For both regions, negative trends in E2P predominantly occur within the cropland regions themselves and their adjacent areas (p < 0.05), while the respective upwind terrestrial source regions show strong positive trends. Although to a lesser extent, (significantly) positive trends in E2P extend into the Atlantic for both ACR and SBR, indicating an increase in oceanic moisture contribution in line with previous studies56.

Unravelling the factors behind the changes in E2P

As explained in Methods and Introduction, we decompose the trend in E2P into two components: one corresponding to the change in evaporation, and the other encompassing all other factors affecting precipitation (represented by β, mainly including atmospheric circulation and precipitation efficiency), representing how evaporated moisture is retained in the atmosphere and transported from source to sink regions. For ACR—located just downwind of the Amazon deforestation belt (see Supplementary Fig. S3)—the impact of evaporation declines in the Brazilian arc of deforestation is a decline in E2P (Fig. 3a). For SBR, these declines also cause a reduction in E2P (albeit of lower magnitude), which is offset by increases in evaporation from nearby oceans (Fig. 3c). Nonetheless, the spatial dipole in the E2P trend in Fig. 2b, d, with increases in the north and decreases to the south, appears to be primarily driven by changes in atmospheric conditions (β), as shown by the greater contribution to long-term moisture supply changes in Fig. 3b, d. For both ACR and SBR, long-term β changes have led to a greater proportion of regional precipitation originating from the north of each region, most especially over land and, to a greater extent, over the Atlantic Ocean (Fig. 3b, d). Negative trends resulting from changes in atmospheric conditions (i.e. β) are found in the soybean breadbaskets themselves, as well as in adjacent regions, and they are potentially associated with stronger winds from northern regions (see also Discussion).

Fig. 3: Attribution of the change in evaporation contribution to precipitation (E2P) during the past 37 years.
figure 3

a, c Contribution of evaporation to the change in E2P for ACR and SBR. b, d Contribution of atmospheric conditions (β) to the change in E2P, mainly related to long-term changes in atmospheric circulation and precipitation efficiency. Red contours indicate the soybean breadbaskets, i.e. ACR and SBR.

Full size image

Contribution of specific sources to rainfall changes in soybean regions

To better understand the physical mechanisms behind the trends in E2P, we separately analyse different moisture source regions and scrutinise the two distinct components of the E2P trend for each of them (Fig. 4). First, given that the region in which the E2P trend associated with evaporation declines coincides with part of the Brazilian arc of deforestation (Fig. 3a, c), the trends in E2P from that particular source region are separated from those in the rest of the South American continent (see Fig. 4c). During the past 37 years, the annual contribution of moisture from the deforestation belt to the precipitation in ACR has significantly decreased by ~120 mm (black diamond in Fig. 4a), which is ~8% of the region’s mean precipitation (~1500 mm). Over two-thirds of this E2P decrease (~80 mm) was caused by a decline in evaporation in the deforestation region over the study period (blue bar in Fig. 4a). For SBR, the decrease in evaporation in the deforestation region also led to a rainfall decline of 20 mm, but this influence was overcompensated by an increase in moisture contribution due to changes in atmosphere conditions (β) changes (~60 mm, see red bar in Fig. 4), leading to a net E2P increase (black diamond in Fig. 4b). The link between deforestation and the positive trend in moisture contribution arising from atmospheric circulation changes is discussed in the following section. Furthermore, for both ACR and SBR, the contribution of the E2P trend associated with the other terrestrial source regions (outside the deforestation belt) is primarily a result of β changes (Fig. 4a, b).

Fig. 4: Influence of different sources on the changes in evaporation contribution to precipitation (E2P).
figure 4

a For the ACR. b For the SBR. From top to bottom, the panels show the changes in moisture contribution from 1982 to 2018 from the following sources: (1) the arc of deforestation (‘Deforestation’) and the rest of the South American continent (‘Other land’), (2) the soybean croplands (ACR or SBR, ‘local land’) and the terrestrial areas outside these regions (‘Upwind land’), and (3) the oceanic source region. The delineation of the different sources can be found in (c). In a, b the blue colour indicates the contribution of the evaporation, and the red colour indicates the contribution of the atmospheric conditions (β, including atmospheric circulation and precipitation efficiency) to the E2P changes. The brown colour denotes uncertainties in the decomposition of the trend (‘Residual’). The diamonds indicate the total change, i.e. the sum of the two components (evaporation and β) and the residual term.

Full size image

Moreover, we divided the terrestrial moisture sources into breadbasket recycling itself, upwind land regions, and oceanic sources (Fig. 4c). The trend in local recycling observed in Fig. 2 is negative for both regions, and it is dominated by β changes rather than by the mild decline in evaporation (Fig. 4). In fact, the decline in moisture recycling in ACR (from Fig. 2b) is associated with an increase in the moisture contribution from this region to SBR precipitation (from Fig. 2d). The potential relationship between this decline in local recycling and the strengthening of the winds coming from the north is discussed in the next section. Finally, as expected from earlier results in Fig. 2, upwind terrestrial sources increased their moisture contribution to both ACR and SBR, with the moisture contributed by the ocean also increasing. In both cases, the positive trend is primarily due to β changes (Fig. 4).

Changes in atmospheric conditions under deforestation

Since 1982, widespread tree cover loss has occurred across South America, with the most substantial losses (exceeding 20%) concentrated in the arc of deforestation (ARC, black contour in Fig. 5a). Located between southern Amazonia and the northern Cerrado, the ARC has been identified as the world’s primary deforestation hotspot57. Consequently, South America has experienced significant evaporation declines during the study period (Fig. 5b). The most substantial decrease occurred within the ARC, where evaporation significantly decreased by over 300 mm during the past 37 years, primarily due to reduced tree transpiration and interception loss (Supplementary Fig. 5). Our study shows that the reduction in evaporation over the arc of deforestation had a detrimental effect on rainfall for downwind soybean breadbaskets (Fig. 3), especially for ACR, which is closer to the Amazon belt of deforestation (Fig. 4a, c). However, this negative impact is masked by other factors that govern how moisture is transformed into downwind moisture transport and ultimately rainfall (Fig. 3b, d).

Fig. 5: Deforestation, evaporation, and atmospheric conditions’ changes during the period of 1982–2018.
figure 5

a Change in evaporation, calculated by multiplying the linear trend by the period length (37 years). b Decline of tree cover during the study period. c, d show the long-term change in integrated water vapour transport (IVT) and convective available potential energy (CAPE), respectively. The arc of deforestation around the Amazon rainforest is shown as the black contour. The breadbasket ACR and SBR are shown as the green and blue contour, respectively. Significant trends (p < 0.05) are highlighted with black hatches.

Full size image

Changes in integrated water vapour transport (IVT) during 1982–2018 (Fig. 5c) corroborate our results, emphasising the predominant role of large-scale atmospheric circulation in precipitation patterns. During the study period, IVT has significantly increased in central and northern South America, and specifically over the tropical Atlantic (Fig. 5c), both crucial moisture sources for ACR and SBR (Fig. 2). This pattern reflects an intensification of the northeasterly moisture flux across the Amazon58, which is associated with a larger continental penetration of the South Atlantic subtropical high pressure system toward South America, as shown by the enhanced anticyclonic wind pattern and reduced wind speed observed in recent decades over Southeast Brazil (Supplementary Fig. 6). Due to the geographical location of ACR and SBR and their prevailing wind direction, this changing pattern finally translates into an increase in water vapour advection from the north and northeast, in line with our moisture tracking results (Fig. 2b, d) and our attribution of mechanisms (Fig. 3b, d). Locally, however, increased wind speeds enhance moisture divergence and are therefore consistent with a reduction in the local moisture contribution to rainfall (Fig. 2b, d). In addition, trends in convective available potential energy (CAPE, Fig. 5d) suggest that increased stability upwind of ACR and SBR, especially south of the deforestation arc, may have favoured this process; more stability implies less precipitation and thus greater outflow at upwind sources. In other words, excess moisture is further transported to the downwind breadbasket regions, ultimately contributing positively to precipitation there.

Discussion

The importance of growing-season rainfall variability in securing soybean production in Brazil has long been recognised. For this reason, understanding the physical mechanisms that govern precipitation variability is fundamental. However, precipitation is only expected to be crucial to rainfed agriculture in regions that are water-limited59, while in energy-limited regions, yields are expected to be less sensitive to rainfall variability53. In this study, we focused on the two major rainfed soybean breadbaskets in Brazil, one of them being mostly energy-limited (ACR), and the other primarily water-limited (SBR). Despite climate change and deforestation, growing-season rainfall has not changed significantly during the period 1982–2018 in either of these two regions (Fig. 1b). Our analysis, based on Lagrangian moisture tracking simulations guided by observations, shows that the lack of trend in precipitation is due to different and competing factors, namely changes in evaporation, atmospheric dynamics and precipitation efficiency. Understanding how these factors compensate each other is crucial to understanding the climate resilience of these agricultural regions and anticipating potential future changes.

We demonstrate that the reduction in evaporation over the arc of deforestation has had a detrimental effect on rainfall, consistent with previous studies31,33, especially for ACR, which is closer to the Amazon belt of deforestation. However, changes in atmospheric (thermo)dynamics have led to increased moisture contribution from remote sources to both ACR and SBR (Fig. 3), mainly from northern terrestrial sources and the Atlantic Ocean. This increase has compensated for the decline in E2P associated with the decrease in evaporation following deforestation (Fig. 4). Importantly, trends in IVT and CAPE (Fig. 5c), as well as in low-level wind speeds (Supplementary Fig. 6), reinforce our moisture tracking-based results (Fig. 3) and confirm that the increase in E2P from remote sources is mainly driven by changes in atmospheric (thermo)dynamics, likely related to deforestation. Previous studies had already reported this compensatory effect between the loss of evaporation over the deforested area and the reinforcement of moisture transport from upwind regions40,41,46. They showed that the increase in wind speed relates to a reduction in surface roughness after deforestation, which can affect even synoptic-scale atmospheric circulation45,60,61. While these studies were largely model-based and/or considered idealistic or future deforestation scenarios, our study, directly constrained by observations, supports the hypothesis that this compensatory effect simulated by Earth System Models has already been occurring in recent decades.

However, several methodological limitations are acknowledged here. While our bias-correction approach62 improves the quantification of moisture exchanges along trajectories by incorporating observational datasets63,64,65, the moisture transport pathway patterns remain governed by ERA-Interim wind fields66. This dependence introduces inherent uncertainties, particularly in the representation of atmospheric circulation. Our moisture tracking framework could potentially be enhanced using newer reanalysis products like ERA567, which offers improved spatial resolution and more advanced data assimilation techniques. Additionally, the parameter β in our decomposition method encompasses both thermodynamic and dynamic (circulation) components without explicitly separating their relative contributions, which represents another limitation of our methodology. Beyond these methodological considerations, we also acknowledge that internal climate variability may contribute to the observed long-term trends in atmospheric circulation and thus moisture transport across South America28. Therefore, the compensatory effect in the historical periods may be influenced by the phase of these oscillations and could potentially change in the coming decades. As climate change progresses and deforestation continues, the balance between these competing mechanisms may shift, potentially altering precipitation patterns in these critical agricultural regions.

Despite these methodological and interpretative limitations, the consistency between our moisture tracking results (Fig. 3) and atmospheric circulation changes (Fig. 5) provides confidence in the robustness of our main findings regarding the compensatory effects between evaporation declines and enhanced moisture transport following deforestation. Overall, this study unravelled the contribution of evaporation and atmospheric (thermo)dynamics to long-term rainfall changes in Brazil, particularly from the perspective of atmospheric moisture transport. Our results showed that the negative impacts of Amazonian deforestation on evaporation lead to declines in downwind rainfall, but these are offset by enhanced moisture transport from remote regions, including Amazonia itself and the Atlantic Ocean. Thus, we highlight that compensatory interactions between evaporation and atmospheric (thermo)dynamics are shaping Brazilian rainfall trends under deforestation and climate change, which may inform the development of forest management strategies and agricultural practices in the region.

Methods

Definition of the studied soybean regions in Brazil

The main rainfed soybean breadbaskets are identified according to the following criteria. Firstly, the rainfed area fraction relative to the total harvest area (the sum of irrigated and rainfed area) must be greater than 95% and the total harvest area must exceed 20,000 hectares in each of the 0.5∘ pixels within the Brazilian territory. The results at 0.5 degrees are subsequently smoothed to 1.0∘ to ensure compatibility with the grid used by the moisture tracking model (next section). The resulting boundaries encompass large parts of central-western Brazil (i.e. Mato Grosso and Goiás, located in the Amazon and Cerrado region, hereafter ‘ACR’) and southern Brazil (i.e. Paraná, Rio Grande region, hereafter ‘SBR’). The soybean production of these two major breadbaskets comprised over 80% of Brazil’s total soybean production in 2000. To represent the harvesting areas, a gridded data set on the crop area around the year 2000 (MIRCA2000,68) is used. The main soybean growing season in Brazil is estimated from a global crop calendar dataset49, considering that soybean in the selected regions is usually planted in October of the previous year and harvested in April of the current year. For simplicity, we refer to a soybean growth year as the harvest year, meaning that the yield in the year 2001 corresponds to the growing season from October 2000 to April 2001. The interannual variability of soybean yield in these agricultural regions is calculated from a historical yield dataset that combines agricultural census statistics and satellite remote sensing (GDHY,69). In addition, the relative yield anomaly (%) is calculated by removing the long-term trends of soybean yields (ton per hectare) using a locally weighted scatter plot smoothing (LOWESS) method70, because this method can account for possible nonlinearity in time, and is therefore more suitable for non-stationary time series than other approaches. To characterise the hydroclimatic conditions of these regions, we calculated the aridity index, defined as the ratio of precipitation to potential evaporation at the annual scale. When defined this way, a higher aridity index indicates a lower importance of water as a limiting factor for the ecosystem52. In addition, the arc of deforestation of the Amazon rainforest is identified when the loss of tree cover71 over the 0.5∘ grid cells exceeds 10% during the past 38 soybean growth years (1982–2018). Afterwards, it is smoothed to the 1.0∘ resolution to be compatible with the FLEXPART analysis.

Moisture source identification

We use a novel moisture tracking framework to determine the spatio-temporal origins of moisture feeding precipitation in the two soybean breadbaskets considered. The framework is based on the Lagrangian particle dispersion model FLEXPART v10.472, used to track air parcels in space and time. FLEXPART was forced with ERA-Interim reanalysis data66 every 3 h at 1.0∘ resolution. Our simulations, of a 40-year period (1979–2019) duration, were performed by tracking 3 million air parcels homogeneously filling the entire globe. The trajectories of air parcels residing in each rainfed soybean breadbasket were then evaluated to determine their moisture sources, i.e. the regions in which air parcels arriving at each breadbasket were moistened by evaporation. Specifically, our methodology follows the process-based trajectory analysis approach introduced by Sodemann et al.73 and applied by Cheng and Lu74, which allows for quantification of individual evaporation and precipitation events along trajectories, rather than diagnosing net moisture flux changes. The detailed criteria for detecting precipitation and evaporation events are described in ref. 62. Therefore, air parcels over the cropland were identified for each day between sowing (October) and harvest (April), and subsequently tracked backwards in time for 15 days. The resulting backward trajectories were evaluated to establish the moisture source field (E2P), i.e. the amount of evaporation fuelling precipitation in the soybean regions.

Moreover, in order to ensure the reliability of the results, we account for en route moisture losses due to rainfall and apply a bias-correction technique by constraining the output of our model with observed precipitation from MSWEP65 and evaporation from GLEAM63 and OAFlux64. The bias-correction approaches typically provide more realistic estimates of moisture contributions while maintaining the spatial patterns of moisture source regions62. In addition, while ERA-Interim employs a fixed land cover map and static LAI climatology as boundary conditions, the atmospheric effects of deforestation are still partly reflected through the assimilation of meteorological observations, which have been influenced by historical vegetation changes66,75. Furthermore, the use of GLEAM for bias correction enhances the sensitivity to land use modifications, as GLEAM explicitly accounts for these changes in its evaporation estimates based on satellite data63, which is particularly valuable for studying the effects of deforestation.

Linear decomposition method for disentangling E vs. circulation and stability impacts in E2P

The evaporation contribution to precipitation (E2P) describes the gain in moisture via evaporation of the parcels that will eventually travel to the sink region (i.e. the two breadbasket regions) and precipitate. Therefore, we conceptualise here E2P as the product of the evaporation (E) and the parameter β, with the latter accounting for atmospheric circulation and precipitation efficiency throughout the entire atmospheric trajectory (and in fact any other factors that may affect the E2P/E ratio).

$$E2P=\beta \times E$$
(1)

Thus, the change over time of E2P can also be evaluated as a function of the time variation of E and β by simply taking the derivatives with respect to time in Eq. (1):

$$\frac{\partial E2P}{\partial t}=\frac{\partial E}{\partial t}\times \beta +\frac{\partial \beta }{\partial t}\times E$$
(2)

Taking time averages, Eq. (2) can be rewritten as follows:

$${S}_{E2P}={S}_{E}\times \bar{\beta}+{S}_{\beta}\times \bar{E}+\delta$$
(3)

where SE2P, SE, and Sβ are the slope of the linear regression (trend) of E2P, E, and β, respectively, with bars denoting climatological averages and δ being the residual of the approximation. Note that, when transitioning from Eqs. (2 and 3), We have approximated the average of the derivative as the slope of a linear regression and the average of the product as the product of averages, which has proven to be a good approximation in this case due to the low residual values found in Supplementary Fig. S7.

Finally, note that in principle, Eq. (1) is not an approximation; E2P is equal to a fraction of E by definition. The approximation relies on the assumption that β and E are independent variables, which is what finally allows linearising the trend of E2P Eq. (3). Although for short time scales this approximation would not be valid in many cases, for the climatological timescale of this study, we consider it a reasonable assumption.