This study uses the atmosphere and land surface components of the IPSL-CM, which has been a regular participant in CMIP, including CMIP6 . The simulations are run with prescribed sea surface temperature and sea ice content from the Atmospheric Model Intercomparison Project (AMIP) dataset.

The Iberian Peninsula was chosen as the study region since it is a hotspot for irrigation in Europe, particularly in the Ebro Valley (north-eastern Spain, see Fig. ). In addition, it is clearly bounded by the Atlantic Ocean, the Mediterranean Sea, and the Pyrenees mountain range in the north, which ensures that all watersheds are fully included in the simulation domain, making the regional study of the continental water cycle rather straightforward. Moreover, although it was not identified by as a region of strong coupling, probably owing to its small size relative to the resolution of the model used and to other identified hotspots, its semiarid climate should make it very sensitive to the influence of land-atmosphere coupling processes. Therefore, modifications of the SM heterogeneity patterns induced by irrigation are expected to have a clear influence on the local and regional climates.

To study this region, the LAM simulation domain is a hexagon (centred at 40.4° N, -3.7° E) with a radius of 1500 km (Fig. ). The radius is composed of 60 grid cells, so the diameter of each cell is 25 km. The simulations are run for 13 years from 2010 to 2022, which enables capturing some interannual variability of the current climate, making the averages less sensitive to anomalies or biases from any single year or extreme event. Before this 13 year simulation, 3 years were run as a spin-up to allow for the hydrological and vegetation variables to reach a satisfactory equilibrium (see Appendix  for more details). Two simulations are run with the same setup except for the inclusion of irrigation. They are referred to as irr (with irrigation activated in ORCHIDEE) and no_irr (without irrigation). Considering the focus on land-atmosphere coupling, the analysis is restricted to the Iberian Peninsula, thus excluding ocean grid cells and the land grid cells north of the Pyrenees, which flow to France. To prevent mismatches between the simulations and the evaluation datasets because of different coastlines, the analysis is further reduced to grid cells where the continental fraction is greater than 95 % .

The simulations are evaluated against the monthly mean values of the reference gridded products listed in Table . Precipitation data from the Global Precipitation Climatology Centre (GPCC) Full Data Monthly Product Version 2020 is used. This reanalysis product provides precipitation data until 2019 over land on a 0.25° × 0.25° grid using in situ rain gauges. For ET, the Global Land Evaporation Amsterdam Model (GLEAM) dataset is used, in its fourth version . This product computes ET using a large set of input variables obtained from reanalyses as well as in situ and satellite observations. Monthly values at 0.25° resolution are used, initially given in mm per month but converted to mmd-1. GLEAM4 is available until 2022 but since precipitation and ET are evaluated jointly, it is only used over the availability period of GPCC data (2010–2019). The evaluation of surface soil moisture uses the European Space Agency Climate Change Initiative ESA CCI, COMBINED product, which merges multiple active and passive satellite remote sensing instruments in the microwave domain to provide daily SSM (in m³ m⁻³) at 0.25° resolution. This product contains low-quality flags (e.g. in the presence of snow, dense vegetation, or radio-frequency interference in the measurements) and an estimated uncertainty of the measurement. A screening of the data for the Iberian Peninsula over the period 2010–2022 was conducted, following the procedure described in . Data records with quality flags other than zero were excluded (representing 19 % of daily data records), as well as those with an uncertainty larger than 0.6 m³ m⁻³ (representing only 0.1 % of the remaining data records). The remaining daily values were then averaged to monthly time steps. Direct comparison of the CCI SSM product with the ORCHIDEE model is limited by the inherent differences in soil representation between ORCHIDEE and the LSM used to scale the satellite measurements (GLDAS-Noah), as analysed in . Therefore, a spatio-temporal normalization was applied following to compare statistically normalized values rather than absolute values: 1SSMnorm=SSM-SSM‾σSSM, where SSMnorm is the normalized SSM, and SSM‾ and σSSM are the mean and standard deviation, respectively, of the available SSM values over the Iberian Peninsula for all the considered monthly time steps. Similarly to ET, although the product is available until 2022, the evaluation was restricted to 2010–2019 to match the evaluation of precipitation, and the normalization is conducted over this period.

Irrigation is evaluated in the Ebro Valley region using a high-resolution remote sensing product from the European Space Agency (ESA) Irrigation+ project . This product estimates irrigation with a soil moisture based approach using satellite measurements from Sentinel-1, and provides data for three intensely irrigated areas: the Ebro Basin in Spain, the Po Valley in Italy and the Murray-Darling Basin in Australia. From 2016 to 2020 in the Ebro Basin, the median values of the RMSE, Pearson correlation coefficient r and bias are 12.4 mm over 14 d, 0.66, and −4.62 mm over 14 d, respectively. Weekly values are aggregated to a monthly average in mmd-1. The simulated river discharge is evaluated against monthly observation data from discharge stations of the Global Runoff Data Center GRDC, https://grdc.bafg.de, last access: 14 May 2026,. Stations were positioned on the MERIT DEM grid with tools presented in , which use the GPS position of the stations as well as the upstream catchment area to find the most appropriate grid cell for comparison with the observations. The 18 selected stations with available data over the simulation period and an adequate position on the DEM grid are described in Table  and shown in Fig. . Most stations have available data from January 2010 to September 2017, and river discharge was therefore evaluated using the first eight years of simulation (2010–2017).

To assess the influence of irrigation on the model, the two simulations (irr and no_irr) are compared and a statistical significance test is used to filter out differences that may be the result of natural variability only. For the maps of changes induced by irrigation in Sect. 3.4, a Student t test is used to assess for each grid cell whether the mean difference between the two simulations (irr - no_irr) significantly differs from 0, with a p value of 0.05 as the limit to reject the null hypothesis. Grid cells with nonsignificant changes are partly hidden with hatches.

The computed irrigation demand (Fig. c) is highly dependent on the irrigated fraction (Fig. a) and much greater than the applied irrigation (Fig. d). This shows that irrigation is often constrained by water availability, with clear regional differences. Indeed, irrigation is much greater in the northern regions (Ebro and Douro river basins) than in southern regions (Guadiana and Guadalquivir basins) even though these regions have similar levels of irrigation demand (Fig. c, d). As shown in Fig. b, southern regions (Guadalquivir Basin, upper Guadiana Basin) are more dependent on groundwater equipment for irrigation water withdrawals than the Ebro Basin, where withdrawals are taken mainly from surface water (overland and river reservoirs in the model). Considering that the groundwater reservoir is much more depleted in the presence of irrigation than the river reservoir is (Fig. e, f), it is not surprising that the irrigation requirement cannot be met in these regions as much as it is in the north. This depletion can be explained by the fact that the groundwater reservoir can only be filled with drainage in the grid cell, whereas the river reservoir can be fed from upstream grid cells. It is also important to note that ORCHIDEE does not model deep groundwater storage, which is an important source of water for irrigation in southeastern Spain .

In the Ebro Basin, simulated irrigation is evaluated using the irrigation remote sensing product from with good overall performance, particularly in summer (Fig. a). In winter, the model simulates almost no irrigation since it requires a minimum LAI to be activated, whereas the product shows irrigation all year long, which can be explained by the presence of winter crops not represented in the model. A delay of the summer peak in the model compared with the product is noticeable, but the model might not be very far from actual irrigation since this product was found to be slightly ahead of actual irrigation based on benchmark volumes in some districts of the Ebro Valley Fig. 5 in. Spatially, the remote sensing product shows greater irrigation on the hillslopes than in the thalwegs, whereas the model simulates the opposite, with more intense irrigation next to the large rivers (Ebro, Segre, Cinca). The resulting bias pattern (Fig. b) can be explained by the fact that in the model, water is mainly withdrawn from the river reservoir in this region, which is much greater in grid cells holding a large river than in upper areas of the valley. In reality, infrastructures such as the Canal d'Urgell in the Segre basin , enable gravity irrigation of hillslopes by diverting water from large rivers to neighbouring croplands. Including a representation of water adduction in the irrigation scheme by enabling withdrawal from adjacent grid cells could be a way to improve this bias. Overall, the spatial biases of the simulated irrigation offset each other relatively well. Averaged over the subdomain where the satellite product provides values, the simulated irrigation is 0.20 mmd-1 while the product estimates it at 0.23 mmd-1.

Figure  shows the average seasonal cycle of river discharge for the five largest rivers of the Peninsula (Ebro, Douro, Tagus, Guadiana, and Guadalquivir), for the two simulations and the GRDC observation data. The station with the greatest river discharge was selected for each river, to reflect integrated impacts of irrigation over the basin. A similar figure with all eighteen stations is presented in (Fig. ). In most cases, the model shows a slight delay and a large overestimation of river discharge compared with observations, particularly in winter and spring. Snow melt does not seem to play a large role in the delay since it occurs rather concurrently with snowfall (Fig. a). These errors are most likely related to the precipitation biases of the simulations, described in the next section, and to the lack of river dams in the model, which have a strong impact on actual discharge, given their high density in the Iberian Peninsula Fig. ,. In particular, the model overestimates the winter and spring discharge, a period when water is stored in dam reservoirs, which reduces the actual river flow. The presence of dams also leads to unnatural seasonal cycles in the observations if river discharge is artificially increased in summer by the release of stored water (Fig. e).

The simulation of irrigation used cannot improve either of these aspects since it does not include a specific reservoir to store water. Its impact becomes noticeable in spring, with water withdrawals resulting in lower discharge during summer and autumn, generally leading to a much better match with observations (Fig. ). As shown in Table , for all fifteen stations where the mean bias is positive in the no_irr simulation, this bias is reduced in the presence of irrigation (in three cases it becomes negative but the absolute value is reduced). However, for the three stations where the no_irr bias is negative (no. 10, 13, 17), it is worsened in the irr simulation. On average, the irr simulation exhibits clear improvements of the mean bias (-41.8 %) and root mean square error (RMSE, -7.12 %). The Pearson correlation coefficient is 0.56 on average in no_irr and is slightly improved (+0.02), with mostly small changes except for stations 6 and 8 (+0.09 and +0.08). Improvements are also observed for Nash-Sutcliffe efficiency (NSE, +0.67) and Kling-Gupta efficiency (KGE, +0.2), mostly as a consequence of improvements in the mean bias. However, only eight out of eighteen stations have a positive value of NSE and KGE values in the no_irr simulation (not shown), limiting the relevance of this average increase. In particular, the average NSE value is strongly influenced by a few stations (no. 5, 8, 15) with initial NSE values below -10.

Overall, the performance is clearly improved for 12 stations but partly degraded for 6 stations, although three of them (no. 10, 13, 17) have a very small average discharge. This may explain why small changes in the model can lead to large changes in performance and limits their relevance compared with larger stations. If only stations with an average annual discharge greater than 10 m³ s⁻¹ are considered, irrigation improves performance in nine out of twelve cases, and degrades it for three stations. One station (no. 9) exhibits an unexpected increase in the spring discharge peak which originates mostly from a single year (2011, see Fig. ) and worsens an already-existing 250 % bias in this season. The other two (no. 4 and 5) are close to the Pyrenees Mountains and present very strong biases in both simulations (300 % overestimation in spring, unexpected double peak in March and June, Fig. ). These discrepancies are likely related to biases in precipitation (discussed hereafter) and were not positively impacted by irrigation, apart from the mean bias. It must also be mentioned here that in the Pyrenees, snowfall represents up to 35 % of annual precipitation in the simulation (Fig. b) and largely contributes to river discharge in these two stations via snow melt.

The simulated precipitation, ET, and normalized SSM are evaluated from 2010 to 2019 using the GPCC, GLEAM4 and CCI products, respectively (Fig. ). On average over the domain, the two simulations present very similar seasonal cycles of precipitation. The model is in good agreement with GPCC until June (Fig. a), but presents a strong underestimation of precipitation in summer, followed by a delayed and overestimated peak in autumn. This peak is mostly the consequence of precipitation in mountainous regions such as the Pyrenees and the northern coast (see Fig.  for seasonal comparisons to GPCC) and is dominated by rainfall rather than snowfall (Fig. a). The delay in precipitation likely contributes to the biases of river discharge winter peaks visible in Fig. . This seasonal cycle is largely representative of the whole peninsula, although some spatial disparities persist. The two simulations exhibit very similar spatial patterns of annual mean precipitation, with a strong overestimation in elevated areas (Fig. b, c), which is a known bias of climate models . This is partly compensated by smaller underestimates of precipitation over large neighbouring areas, as seen in the Ebro Valley.

Both simulations match the GLEAM4 ET product well from January to May but underestimate ET for the rest of the year, particularly in summer (Fig. d). As expected, ET increases when irrigation is accounted for, particularly from May to September, which is the period where vegetation is the most developed and irrigation is the greatest. This partially alleviates the dry bias, but ET remains underestimated, even in the irr simulation. No similar patterns of biases between ET and incoming radiative fluxes were identified, and the remaining ET bias can be related to the underestimation of precipitation in the southwestern part of the Peninsula and in plains, such as the northern Ebro Valley (Fig. b, e). The biases in ET and precipitation are not consistent in the north-west of the Peninsula, which is explained by seasonal differences. Indeed, the overestimation of precipitation mostly occurs in winter and spring and does not strongly affect ET since soils are already wet in this season, whereas the underestimation of ET is dominated by summer (see appendix Figs.  and ). Along large rivers, the ET underestimation almost disappears in the irr simulation (Fig. f). In contrast, the ET bias remains significant in lightly irrigated grid cells such as hillslopes, which is consistent with the limits of simulated irrigation described in Sect. 3.1 (Fig. ).

The seasonal cycle of the normalized SSM matches the CCI product in both simulations (Fig. g). It is very similar in the two simulations, although in absolute value, SSM is slightly higher in summer in irr, and annual increases up to 11 % are visible in the most irrigated grid cells (Fig. a–b). Compared to the increase in ET, this indicates that the water added by irrigation is mostly evaporated, as will be further illustrated in Sect.  and Fig. . This confirms that the transition regime described by is dominant in the region. Spatial biases in normalized SSM compared with the CCI product are very similar in the two simulations and match the precipitation underestimation in the Ebro valley, Guadalquivir river mouth, and overestimation in the Pyrenees, but show a contrasting pattern in the West compared to precipitation and ET, due to a positive bias in summer (Fig. ).

Here, the impacts of irrigation on atmospheric variables are studied in summer (JJA) since it is the season with the highest levels of simulated irrigation, with seasonal mean values of up to 1.5 mmd-1 in the most intensely irrigated grid cells (Fig. a). In the presence of irrigation, the simulated latent heat flux (LE) increases across the entire Iberian Peninsula, by up to 50 Wm-2 in the Ebro Valley. As expected from the surface energy partitioning, this is compensated by a decrease in the sensible heat flux (H), which is almost equivalent in irrigated areas and leads to large increases in the evaporative fraction (EF=LELE+H) shown in Fig. b. When the sensible heat flux decreases, less energy is transmitted from the surface to the air, leading to a decrease in the 2 m air temperature which spatially matches the increase in EF. The order of magnitude remains low over most of the Peninsula, with the most important changes reaching -0.35 K in the Ebro Valley (Fig. c). The decreases in sensible heat flux and temperature also lead to a more stable boundary layer over most of the peninsula, but mostly in intensely irrigated areas where it is lowered by 100 m (Fig. d). Moreover, the presence of irrigation results in a moister lower atmosphere, with an average specific humidity over the Peninsula increasing by 2.8×10-4 kgkg-1 in summer (+3.4 %) and maximal local increases in the Ebro Valley of 1×10-3 kgkg-1 (+10 %). Since air temperature changes in the atmospheric column are rather small, the lowering of the lifting condensation level (LCL) reflects this atmospheric moistening very well. It is most marked in the Ebro Valley, where the LCL is lowered by 250 m (-13 %) in the most intensely irrigated grid cells, and remains significant even in areas where irrigation is low (Fig. e).

The lowering of the ABL and LCL theoretically favour opposite effects on precipitation. On the one hand, a lower and more stable ABL inhibits vertical mixing and convection, reducing the likelihood of cloud formation and deep convection initiation. On the other hand, if the LCL is lower, air parcels do not need to be lifted as high to condense, which increases the likelihood of cloud formation. Over the most intensely irrigated areas, ABL stabilization seems to dominate and inhibit convective development since no significant change in precipitation is observed. However, mountainous areas surrounding the Ebro Valley show significant increases in precipitation (Fig. f). This can be understood because ABL stabilization remains weak in these zones whereas humidity can still be increased if moisture is advected (Fig. d, e). In particular, the dominant wind patterns in the Ebro Valley (Fig. a) indicate that the additional atmospheric moisture from irrigated areas is driven towards the valley slopes, which is consistent with the increases in moisture convergence (Fig. h) and precipitation over the Pyrenees. The competing interactions of ABL stabilization and atmospheric moistening are reflected by the increases in convective available potential energy (CAPE) which are most important in elevated areas around the valley (Fig. g), where increases in precipitation are significant.

In other seasons (see Appendix  for equivalent figures), almost no significant impacts are visible in winter (DJF) since irrigation is very low and no clear pattern emerges in other variables. In spring (MAM) and autumn (SON), the impacts resemble those of JJA with lower values for all variables and less statistical significance over the domain, and a greater amplitude in SON than MAM. Precipitation increases in the Pyrenees are also present in these seasons, but the area where changes are significant is reduced, and no clear signal is found in the southern rim of the Ebro Valley. The only difference in spatial pattern is found for CAPE in the spring, with increases at the centre of the Peninsula rather than the mountainous areas, although the amplitude of the change remains almost ten times lower than in summer.

On average over the continental domain, the monthly change in ET in the presence of irrigation is well correlated with the amount of water added by irrigation and even exceeds it, particularly in summer months (the orange JJA data points in Fig. a are all on or above the 1:1 line). In the simulation, ET is constrained by available water, and almost all the water added by irrigation is evaporated or transpired, meaning that this additional increase in ET comes from an additional input of water into the soil. This is explained by a systematic increase in precipitation over the domain (all the data points are on or above the x axis on Fig. b). This increase is also roughly proportional to the amount of applied irrigation, although the correlation is weaker than that for the increase in ET, and its values remain lower than the amount of water added by irrigation. Therefore, it appears that irrigation contributes to an increase in atmospheric moisture, and that a part of this moisture is recycled as continental precipitation, which can then be reevaporated.

To look further into this recycling, three subdomains were defined, namely, low, medium and high irrigation areas, on the basis of the mean irrigation thresholds given in Table . In the map of simulated irrigation (Fig. d), the low irrigation domain corresponds to the first colour bin (yellow), the medium irrigation domain to the second bin (light green), and the high irrigation domain to the eight other bins. The three subdomains are also shown distinctly in Fig. e. On average, the increase in ET is slightly superior to irrigation for each subdomain (Fig. ). However, the increase in precipitation is more than twice as large for the low irrigation subdomain than for the medium and high irrigation subdomains. Since irrigated areas are mostly in plains and valleys, this result is consistent with the increase in precipitation already described over mountainous areas in summer (Fig. f). It points towards a nonlocal moisture recycling, with atmospheric moisture transfer from intensely irrigated areas to neighbouring lightly irrigated areas, meaning that a significant part of the additional rainfall does not occur on irrigated crops. Over the entire Iberian Peninsula, the increase in precipitation represents 25 % of the irrigated volume, whereas the increase in ET amounts to 112 % of irrigation.