The historical climate data were extracted from the AGRI4CAST observational dataset, specifically from the grid cell nearest to the respective geographical coordinates of each study site. The data were utilized to train the Long Ashton Research Station Weather Generator (LARS-WG) 7.0 (Semenov et al., 2002). Once trained, LARS-WG was used to generate 100 years of synthetic local-scale daily meteorological data (T_min, T_max, rainfall, solar radiation) for both the baseline period (1981–2010) and future climate scenarios. Specifically, three future time periods were considered: near future (2041–2060), mid-future (2061–2080), and far future (2081–2100). These time periods were considered along with three emission scenarios according to Shared Socio-economic Pathways (SSPs) (SSP1-2.6, SSP2-4.5, and SSP5-8.5), and three General Circulation Models (GCMs) (ACCESS-ESM1-5, HadGEM3-GC31-LL, and MRI-ESM2-0).

The SSP-based scenarios were used in the Coupled Model Intercomparison Project Phase 6 (CMIP6) which is the most recent set of climate model experiments. SSP1-2.6 (sustainable path), SSP2-4.5 (intermediate path), and SSP5-8.5 (fossil fuel-dominated development path) were chosen out of the five SSP families because they belong to the high priority scenarios according to the Intergovernmental Panel on Climate Change Sixth Assessment Report (IPCC AR6). Similarly, the three GCMs used were available from CMIP6 used in the recent IPCC AR6 and were selected because they are reliable to use with different emission scenarios and they are available GCMs in LARS-WG.

Synthetic time series produced with a weather generator does not represent actual weather but rather the potential realization of weather conditions within a given climate scenario. As such, the choice of 100 years was made to minimize the uncertainty in the stochastic generation of data by the weather generator, as it is recommended to use at least 60 years (Lawless and Semenov, 2005). The 100-year time series for baseline and future scenarios were then used as weather inputs for the crop model.

For all locations, the reference evapotranspiration (ET_o) was calculated on a daily basis using the Hargreaves equation (Equation 2), as described in “FAO Irrigation and Drainage Paper No. 56” (Allen et al., 1998), and then directly imported into the model. This method was used instead of the standard Penman-Monteith equation due to data limitations, and also because it is a recognized and accepted approach for estimating ET_o. The equation is as follows:

where H_A and H_E are standard values: H_A = 0.0023 and H_E = 0.5, R_e is the water equivalent of the radiation measured on the ground or terrestrial radiation (mm d⁻¹), T is the mean temperature ((T_min + T_max)/2°C) and ΔT is the difference between maximum and minimum temperatures.

The average CO₂ concentration (366 ppm) used for the baseline period simulations was obtained from Meinshausen et al. (2017). Instead, the projections were based on SSPs in which the average CO₂ concentrations were obtained from Meinshausen et al. (2020). The summary of the meteorological variables is presented in Supplementary Table 1.

Data from soil analyses at the respective study sites (Supplementary Table 2) were used as inputs to simulate the soil water movement and retention of water in the ERZ of soybean (Raes et al., 2022). The simulations were based on fine soil texture thus, neglecting the effects of the soil skeleton on the retention capacity and infiltration. Moreover, the groundwater table was not considered. This was based on the assumption that due to the depth of the groundwater table, it would have had negligible effects and that capillary rise would not have contributed moisture to the root zone.

The indicative values of crop parameters for soybean by Raes et al. (2022) and the agronomic data from the field experiments in the study sites were used to parametrize and calibrate the crop model (Supplementary Table 3). In the model, soybean was considered as a fruit or grain-producing crop with a yield formation period that started at flowering, coinciding with the progressive increase of the harvest index (HI). In the model, crop characteristics can be distinguished as either conservative or non-conservative. Thus, the cultivar-specific and non-conservative crop parameters were adjusted since they vary with the selected cultivar and may also be affected by field management, conditions in the soil profile, and climate. The parametrization and calibration were performed according to the user guide and procedure provided by Raes et al. (2022) and Vanuytrecht et al. (2014). Performance evaluation indicated that the AquaCrop model successfully simulated soybean grain yield in the study sites.

The simulations were conducted under both rainfed and irrigated conditions. Various levels of RAW depletion within the soybean ERZ were considered as thresholds to start the irrigation. These RAW depletion thresholds ranged from 25%, 50%, 75% to 100%. In this study, these are referred to as irrigation strategies and were evaluated in terms of change in yield and WF in relation to climate change. The maximum effective rooting depth in Castelfranco and Cesa was 1.0 m, whereas it was 0.70 m in Ljubljana. It is important to note that the depletion considered in the simulations was the RAW in the root zone and not the total available water between field capacity and the permanent wilting point. An irrigation file was created for each irrigation strategy by specifying each threshold in the root zone. That means that if the threshold was reached, irrigation was added to the soil profile to keep the moisture in the ERZ above the threshold. Thus, the total irrigation added to the soil profile throughout the growing period constitutes the net irrigation requirements (NIR) of soybean (Raes et al., 2022). On the other hand, no irrigation file was created for the rainfed conditions since it represented the default in AquaCrop.

The AquaCrop model assumes that the crop grows under optimal conditions, in that crop water demand is fully satisfied. It simulates daily ET_c (mm) using the two-step approach by multiplying the crop coefficient (K_c) by ET_o (Equation 3). The dual K_c (Equation 4) consists of a basal crop coefficient (K_cb) and an evaporation coefficient (K_e) (Allen et al., 1998; Pereira et al., 2015).

where K_cb and K_e are related to the portion of crop transpiration (T_c) and soil evaporation (E_s) over ET_o, respectively.

In this study, the irrigation and rainfall were assumed to be the only sources of blue and green WFs, respectively. As mentioned previously, the contribution of blue water from capillary rise was not taken into consideration due to negligible effects from the groundwater table. By tracking the proportion of daily rainfall and irrigation added to the soil profile under rainfed and irrigated conditions, the daily ET_c of soybean was divided into ET_c-blue and ET_c-green as described in the following section.

The sum of the daily ET_c for the entire crop growth period is equivalent to CWR (Pereira and Alves, 2013). Thus, the CWRs under rainfed (CWR_rainfed) and irrigated (CWR_irrigated) conditions were calculated for the crop growth period according to Equation 6 and Equation 7 as shown below. The CWR was based on the cumulative values of the corresponding daily ET_c under rainfed (ET_c-rainfed) and irrigated (ET_c-irrigated) conditions, respectively, as simulated by AquaCrop.

where gp is the duration of the soybean growing period (days) and d=1 is the day of sowing.

Under rainfed conditions, the amount of water that undergoes evapotranspiration is derived solely from the rainfall. Therefore, the green water (CWR_green) is equivalent to CWR_rainfed (Equation 8). On the other hand, the evapotranspiration under irrigated conditions is derived from both rainfall and irrigation. As such, the CWR_irrigated is equal to the sum of blue water (CWR_blue) and CWR_green (Equation 9). Therefore, CWR_blue can be derived from the difference between CWR_irrigated and CWR_rainfed (Equation 10) (Ventrella et al., 2017).

The only source of water under rainfed conditions is rainfall. Thus, the consumptive WFs throughout the growing period (WF_rainfed) are equivalent to WF_green (Equation 11). On the other hand, under irrigated conditions, the source of water is derived from both rainfall and irrigation. Thus, the consumptive WFs under irrigated conditions (WF_irrigated) are equivalent to the sum of WF_blue and WF_green (Equation 12).

The estimated WF_rainfed and green water for soybean production across the study sites in the baseline period are presented in Figure 2 and Table 1. The WF_rainfed varied significantly across locations (p < 0.01) and was the highest in Cesa (1984 m³ ton⁻¹), followed by Castelfranco (1570 m³ ton⁻¹) and Ljubljana (1178 m³ ton⁻¹), respectively.

The green water varied significantly across locations with an average of 315 mm in Castelfranco, 209 mm in Cesa, and 481 mm in Ljubljana. The ANOVA showed that the rainfall varied significantly across locations (p < 0.01). Thus, Ljubljana having the highest seasonal cumulative rainfall registered the highest green water. Similarly, Cesa had the lowest rainfall and, accordingly registered the lowest green water (Supplementary Table 1). The highest estimated grain yield under rainfed conditions was observed in Ljubljana, followed by Castelfranco, and finally Cesa (Supplementary Table 4). Thus, Ljubljana, with the highest green water and yield, registered the lowest WF_rainfed.

The estimated WF_irrigated, blue WF, and blue water of the respective study sites in the baseline period are presented in Figure 3 and Tables 1–3. Under irrigated conditions, the WF_irrigated was equivalent to the sum of green and blue WFs. The results showed that soybean production relied primarily on green WF, contributing to about 68.3 to 100% of the consumptive WFs, whereas the blue WF contribution ranged from 0 to 31.7%. The WF_irrigated was only significantly different across locations (p < 0.01) and was not significant across irrigation strategies (p = 0.08). The WF_irrigated ranged from 1922 to 1976 m³ ton⁻¹ in Castelfranco, from 2535 to 2694 m³ ton⁻¹ in Cesa, and from 1295 to 1339 m³ ton⁻¹ in Ljubljana.

The blue WF was significantly different across both irrigation strategies and locations (p < 0.01). In all locations, the 25% RAW depletion registered the highest blue WF with an average of 406, 710, and 161 m³ ton⁻¹ in Castelfranco, Cesa, and Ljubljana, respectively. Conversely, the lowest blue WF was observed at 100% RAW depletion with an average of 352, 552, and 118 m³ ton⁻¹ in Castelfranco, Cesa, and Ljubljana, respectively (Table 2). Considering the blue WF reduction, the results showed that up to 50% RAW depletion can be implemented in Castelfranco and Cesa without a significant reduction in grain yield. In Ljubljana, the RAW depletion can be increased up to 75%.

Generally, all locations showed an increasing pattern of blue water as the allowable RAW depletion decreased, although this was not as apparent in Ljubljana (Table 3). The ANOVA revealed significant variation in blue water across different irrigation strategies and locations (p < 0.01). In Castelfranco, the blue water ranged from 165 to 214 mm whereas, from 141 to 226 mm in Cesa, and from 66 to 101 mm in Ljubljana.

The WF_rainfed and green water for soybean production in the study sites across the SSPs and future time periods are presented in Figure 2 and Table 1. In Castelfranco, the WF_rainfed ranged from 1398 to 1604 m³ ton⁻¹, whereas it ranged from 1738 to 1914 m³ ton⁻¹ in Cesa and from 1130 to 1378 m³ ton⁻¹ in Ljubljana. In Castelfranco and Cesa, climate change reduced the WF_rainfed as indicated by the lower values in SSPs compared to the baseline. The exception was for SSP1-2.6 in the far future in Castelfranco. In contrast, WF_rainfed remained relatively stable in Ljubljana, although an increase was observed in the far future under SSP1-2.6 and SSP5-8.5.

Ljubljana had the highest green water, ranging from 347 to 438 mm followed by Castelfranco (243–293 mm) and Cesa (149–196 mm), respectively. Across the SSPs, the same pattern was observed in all locations which coincided with the trend in rainfall. The green water was relatively stable under SSP1-2.6 and SSP2-4.5 across time periods but was shown to decrease under SSP5-8.5.

The lowest WF_rainfed was observed in Ljubljana due to high green water, yield, and rainfall. Instead, Cesa had a higher WF_rainfed compared to the remaining two sites due to low green water, yield, and rainfall (Supplementary Tables 1, 4).

The WF_irrigated of each study site across the SSPs and time periods is presented in Figure 3 and Table 1. In Castelfranco, the WF_irrigated ranged from 1739 to 2039 m³ ton⁻¹, from 2300 to 2653 m³ ton⁻¹ in Cesa, and from 1309 to 1697 m³ ton⁻¹ in Ljubljana. The green WF constituted a significant portion of the consumptive WFs for soybean production, and ranged from 69.9 to 87.8% in Castelfranco, 59.6 to 94.5% in Cesa, and 70.9 to 100% in Ljubljana. In contrast, the contribution of blue WF accounted for approximately 12.2 to 30.1% of the consumptive WFs in Castelfranco, 5.5 to 40.4% in Cesa, and 0 to 29.1% in Ljubljana.

In Castelfranco, climate change reduced the WF_irrigated as indicated by lower values in SSPs compared to the baseline. The exception was noted for the far future under SSP1-2.6 and SSP5-8.5. Climate change also reduced the WF_irrigated in Cesa, but caused an increase in Ljubljana. All locations followed the same dynamics of WF_irrigated among the SSPs across time periods. Under SSP5-8.5, the WF_irrigated is expected to increase from the near to far future. Conversely, a reduction in WF_irrigated is expected under SSP2-4.5. Instead in Ljubljana, a slight increase was observed from the near to mid-future. Under SSP1-2.6, a decrease and increase in WF_irrigated is expected from the near to mid-future and from the mid to far future, respectively.

The WF_irrigated across irrigation strategies did not vary significantly in Castelfranco and Ljubljana. However, in Cesa, the WF_irrigated at 100% RAW depletion was significantly lower than the other irrigation strategies. In all locations, the same trend of increasing WF_irrigated was observed as the allowable RAW depletion decreased. The highest WF_irrigated was observed at 25% RAW depletion, and ranged from 1779 to 2039 m³ ton⁻¹ in Castelfranco, from 2412 to 2653 m³ ton⁻¹ in Cesa, and from 1347 to 1697 m³ ton⁻¹ in Ljubljana. Conversely, the lowest WF_irrigated was observed at 100% RAW depletion, ranging from 1739 to 1998 m³ ton⁻¹ in Castelfranco, from 2300 to 2539 m³ ton⁻¹ in Cesa, and from 1309 to 1672 m³ ton⁻¹ in Ljubljana.

The blue WF of each study site across the SSPs and time periods is presented in Table 2. The highest blue WF (561–784 m³ ton⁻¹) was observed in Cesa, followed by Castelfranco (341–447 m³ ton⁻¹), and Ljubljana (171–319 m³ ton⁻¹), respectively. The dynamics of blue WF among the SSPs across time periods followed the same pattern as WF_irrigated. In all locations, the blue WF is expected to increase from the near to far future under SSP5-8.5. On the other hand, it is expected to decrease from mid to far future under SSP2-4.5. Using the same emission scenario, blue WF is expected to increase from the near to mid-future, except in Cesa. Finally, under SSP1-2.6, the blue WF is expected to decrease from the near to mid future and increase from the mid to far future.

In all locations, the highest blue WF among irrigation strategies was observed at 25% RAW depletion, and ranged from 381 to 447 m³ ton⁻¹ in Castelfranco, from 674 to 784 m³ ton⁻¹ in Cesa, and from 213 to 319 m³ ton⁻¹ in Ljubljana. Consequently, the lowest average blue WF was observed at 100% RAW depletion, and ranged from 341 to 429 m³ ton⁻¹ in Castelfranco, 561 to 739 m³ ton⁻¹ in Cesa, and 171 to 295 m³ ton⁻¹ in Ljubljana.

The blue water varied significantly across locations, SSPs, and irrigation strategies (Figure 3, Table 3). In Castelfranco, the blue water ranged from 175 to 244 mm, from 147 to 289 mm in Cesa, and from 97 to 180 mm in Ljubljana. In all locations, the blue water increased under climate change scenarios. The increases were prominent under SSP5-8.5 in the mid or far future, although some variability did exist.

The dynamics of blue water across irrigation strategies followed the same pattern as WF_irrigated. In all locations, the highest average blue water was observed at 25% RAW depletion. Values ranged from 225 to 244 mm in Castelfranco, from 270 to 289 mm in Cesa, and from 135 to 180 mm in Ljubljana. Consequently, the lowest average blue water was observed at 100% RAW depletion, and ranged from 175 to 195 mm, 147 to 162 mm and 97 mm to 144 mm in Castelfranco, Cesa and Ljubljana, respectively.

The relative changes in yield and consumptive WFs were both positive indicating that irrigation increased the soybean yield with a proportionate increase in consumptive WFs depending on allowable depletion of RAW. There were noticeable differences in yield increases between 100% RAW depletion and other irrigation strategies, particularly evident for Cesa. The yield increases in Castelfranco (121.3–148.8%) and Ljubljana (36.6–50.1%) across irrigation strategies were relatively stable compared to Cesa (137.8–239.9%). In contrast, the change in consumptive WFs across all irrigation strategies remained relatively uniform in each location. Values ranged from 22.4 to 25.9% in Castelfranco, 27.8 to 35.8% in Cesa, and 9.9 to 13.7% in Ljubljana.

The changes in yield and consumptive WFs of the various irrigation strategies relative to rainfed conditions are presented in Figure 4. In all locations, the magnitude of change was determined by the RAW depletion. The lowest changes were evident at 100% RAW depletion, and the highest between 25% and 50% RAW depletion with negligible differences. The relative yield change in Castelfranco (140.2–236.9%) was comparable with that of Cesa (141.5–329.4%), whereas it was substantially lower in Ljubljana (42.4–109.4%). Across the irrigation strategies, the trend was the same in Castelfranco and Ljubljana with minimal differences. Notably, in Cesa, the change in yield was prominent at 100% RAW depletion and was significantly different from the three other irrigation strategies. Consequently, the increased yield across irrigation strategies coincided with a proportional increase in consumptive WFs. These values ranged from 23.9 to 29.0%, 30.7 to 43.6% and 14.6 to 23.1% in Castelfranco, Cesa and Ljubljana, respectively. In all locations, the increase in consumptive WFs was relatively uniform across strategies and followed the same trend as the relative yield change.

The change in yield and consumptive WFs of the various SSPs relative to the baseline is presented in Figure 4. Under rainfed conditions, climate change reduced soybean grain yield by -23.7 to -0.28% in Castelfranco, -24.3 to -0.63% in Cesa, and -37.1 to -9.4% in Ljubljana. Under irrigated conditions, the relative change in yield in Castelfranco and Cesa across the SSPs was generally positive, except for the far future under SSP5-8.5 in Cesa. The change in yield was negligent to negative in Ljubljana regardless of the SSPs and time periods. The change in yield under irrigated conditions ranged from -2.4 to 12.2% in Castelfranco, -15.7 to 9.2% in Cesa, and -14.9 to 0.32% in Ljubljana. The magnitude of change was dependent on the SSPs, which were the highest and lowest under SSP5-8.5 and SSP2-4.5, respectively. A notable yield reduction was observed in the far future under SSP5-8.5 in both Cesa (-15.7 to -4.4%) and Ljubljana (-14.9 to -12.1%). Regardless of whether the cultivations were irrigated or rainfed, climate change reduced the WF of Castelfranco and Cesa. The exception was for the far future under SSP1-2.6 and SSP5-8.5 in Castelfranco, where a minimal increase was observed. Conversely, consumptive WFs generally increased in Ljubljana under both rainfed and irrigated conditions. The relative change in consumptive WFs ranged from -11.0 to 4.0% in Castelfranco, -12.4 to 0.16% in Cesa, and -4.1 to 29.1% in Ljubljana. In all locations, the change in consumptive WFs among SSPs was relatively uniform in each time period and followed the opposite trend to the relative change in yield.

The simulations showed variable impacts of climate change on WF_rainfed across locations. In Castelfranco and Cesa, climate change generally reduced the WF_rainfed as indicated by lower values across SSPs compared to the baseline. Furthermore, in Castelfranco, the slight increase in WF_rainfed in the far future under SSP1-2.6 compared to the baseline may be attributed to a reduction in yield. The yield reduction was a consequence of decreased CO₂ fertilization effects due to the combination of increased temperature and the absence of irrigation water. Consequently, in Ljubljana, climate change increased the WF_rainfed in all time periods. The exceptions were for SSP2-4.5 and for SSP1-2.6 in the mid-future in which a slight decrease was observed. This signified that even the less detrimental increase in temperature in these two emissions scenarios, coupled with elevated CO₂ levels, could facilitate a change in WF. The dynamics of WF_rainfed across locations may also be explained by the combined effects of reduced grain yield and CWR (Supplementary Tables 4, 5), resulting from a shortened growing period. This shorter cycle was due to the faster accumulation of GDD as a consequence of increased temperature, which in turn, reduced the time available for biomass accumulation and yield formation.

Across locations, the WF_rainfed dynamics of each SSP varied across time periods. However, in general, WF_rainfed increased from the mid to far future under SSP1-2.6 and SSP5-8.5 and decreased under SSP2-4.5 from the near to far future. Wang et al. (2023) mentioned that rainfall and temperature are two of the main climatic factors influencing WF in crop production. Thus, the increased temperature and reduced rainfall under SSP5-8.5 (Supplementary Table 1) negatively affected soybean performance by increasing WF. The overall results followed the opposite trend of grain yield, indicating that the soybean WF dynamics in the future will be primarily associated with future changes in crop yield (Arunrat et al., 2022; Mekonnen and Hoekstra, 2011).

Rainfall is the only source of green water under rainfed conditions. The green water was highest in Ljubljana and lowest in Cesa. Thus, these results were expected given the abundant and scarce rainfall in Ljubljana and Cesa, respectively (Supplementary Table 1). In all locations, the green water was higher in the baseline period compared to the SSPs, indicating that climate change reduced green water. This may be attributed to the reduced rainfall and shortened growing cycle due to the accelerated accumulation of GDD resulting from increased temperature.

Irrigating soybean would require additional WFs, thus the WF_irrigated was foreseen to be higher than WF_rainfed due to the addition of blue WF from irrigation. The impact of climate change on WF_irrigated varied across locations. In Castelfranco and Cesa, irrigation acted as a buffer against the negative effects of climate change and maximized CO₂ fertilization effects that ultimately resulted in a slight yield increase. This was further supported by our results in rainfed conditions, indicating that despite increased CO₂ levels, low yield under SSPs compared to the baseline was observed due to lack of irrigation (Supplementary Tables 1, 4). This result confirmed previous work showing that CO₂ fertilizing effects can only be realized if soybean is irrigated (Ainsworth and Long, 2021). The CWR decreased both under rainfed and irrigated conditions (Supplementary Table 5), which confirmed that temperature was the main driving factor for the change. Thus, the reduction in WF_irrigated in Castelfranco and Cesa under climate change scenarios was attributed to a slight increase in grain yield due to irrigation and CO₂ fertilization, as well as a decrease in CWR. The reduction in CWR resulted from a shortened growing cycle, driven by faster GDD accumulation caused by both rising temperatures and the CO₂ fertilization effect, which also reduces crop transpiration (Raes, 2023). Consequently, no increase in grain yield under climate change scenarios was observed in Ljubljana due to a reduced buffering effect of irrigation as a result of the abundant projected rainfall in the area. However, reduced CWR was still observed due to increased temperature (Supplementary Tables 1, 5). On one hand, when rainfall is abundant, the benefits of irrigation are reduced and the negative impact of climate change intensifies leading to an increased WF. On the other hand, when rainfall is reduced, the benefits of irrigation are maximized by increasing grain yield which is further enhanced by CO₂ fertilization effects, ultimately leading to a decreased WF. Nevertheless, the overall results conformed to the expectation that both temperature and rainfall will facilitate future changes in the WF of crops (Wang et al., 2023).

Considering the dynamics of WF_irrigated among SSPs across time periods, as was expected, an increase was confirmed in all locations under the most pessimistic scenario (SSP5-8.5). In addition, the reduced CO₂ concentration from the mid to far future under SSP1-2.6 also led to an increased WF. Instead, following the intermediate path (SSP2-4.5), the WF_irrigated decreased in all locations from the near to far future. These results further confirmed the influence of CO₂ fertilization and increased temperature on WF, as discussed previously.

In all locations, the blue water was higher across SSPs compared to the baseline, indicating negative impacts of climate change and confirming the expectation that climate change will increase both the evaporative demand and blue WF of crops in the future (Arunrat et al., 2022; Toreti et al., 2022; Wang et al., 2023). The results corroborated the NIR results, where the same pattern was observed across locations (Supplementary Table 6). In our simulations, given that irrigation was the only source of blue water, these results were anticipated. Thus, blue water is projected to increase in the future. This signifies that more irrigation volume will be needed to produce soybean with values exceeding those in Cesa, compared to Castelfranco and Ljubljana, due to reduced rainfall.

Considering irrigation strategies, a general pattern of decreased blue WF was observed as the percentage of RAW depletion to start irrigation increased. The result was a consequent reduction in grain yield. This is explained by the direct relationship of blue WF with blue water and NIR. Our results highlighted that trade-offs like yield reduction need to be considered when implementing irrigation strategies to reduce WFs of soybean production. Nevertheless, the results indicated that in all locations, 50% RAW depletion in the ERZ should be maintained to reduce the blue WF without a significant reduction in soybean grain yield.