The ISBA model (originally described in Noilhan and Planton, 1989) is a LSM developed by the research department of Météo-France (Centre National de Recherches Météorologiques, CNRM). It is embedded into the SURFEX modelling platform (Masson et al., 2013; Voldoire et al., 2017; Le Moigne et al., 2018) and can provide initial land surface conditions to various atmospheric models (e.g. ALADIN in Fischer et al., 2005) or be driven by atmospheric conditions in offline (i.e. standalone) mode. SURFEX integrates different models describing ocean and terrestrial surfaces. Over land, specific models are used to represent waterbodies, cities, and the soil–plant system. The latter is modelled by the ISBA LSM. The ISBA model can be coupled to the CTRIP model (Decharme et al., 2019; Munier and Decharme, 2022), which is specifically designed to represent water dynamics within rivers and aquifers. Only offline ISBA simulations are considered in this study.

An old irrigation scheme working at a local scale (Calvet et al., 2008) is available in the ISBA LSM. Major limitations of the old scheme are the lack of (1) spatialisation on a global scale, (2) representation of harvest, (3) diversity of irrigation types and irrigated vegetation types, and (4) interoperability with the multi-layer soil hydrology scheme. Key processes implemented in this scheme are briefly described below. The irrigation can be activated for ISBA versions able to simulate photosynthesis, interactive vegetation biomass, and leaf area index (LAI; ISBA-A-gs; Calvet et al., 1998; Gibelin et al., 2006). Sprinkler irrigation is represented by imposing an additional water flux forcing to the soil–plant system. Water is applied at a given time and over a certain period of time. A number of irrigation parameters need to be assigned, such as the irrigation amount, the irrigation interval, and the irrigation start and end times. A parsimonious approach is used in order to limit the number of parameters of the model. Table 1 lists the parameters and the values used by default in the model. Using these values allows the model to predict a realistic amount of irrigation water over irrigated corn in southern France (Bonnemort et al., 1996; Voirin-Morel, 2003; Calvet et al., 2008). Irrigation is triggered using thresholds of the simulated extractable soil moisture content when vegetation growth is limited by a soil moisture deficit. The plant water stress level is evaluated using a unitless soil wetness index along the root profile (SWIroot_zone). A SWIroot_zone value close to 1 corresponds to a well-watered soil, while a value close to 0 indicates extreme stress. In order to trigger irrigation, the SWIroot_zone value is compared to predefined SWI thresholds given as input parameters. These SWI thresholds are evolving during the irrigation season, and default values are fixed to 0.7 for the first irrigation, 0.55 for the second irrigation, 0.4 for the third irrigation, and 0.25 thereafter. The use of these values was validated by Bonnemort et al. (1996), Voirin-Morel (2003), and Calvet et al. (2008). The idea behind this approach is that irrigation does not completely refill the soil, especially at the end of the growing season. Mechanical harvest requires relatively dry conditions to avoid soil compaction. The crop is allowed to use rainwater together with the initial available water content of the soil. This irrigation strategy allows the optimisation of water withdrawal according to plant water extracting abilities at different crop growing stages. When a SWI threshold is reached, irrigation is triggered with a predefined quantity of water of 30 mm (by default), following Calvet et al. (2008).

The simulations and the evaluation of the new scheme are made over the state of Nebraska (United States of America, USA). This area presents a high density of irrigated fields (Fig. 2) and large, freely available observational datasets for evaluation. In this area, most irrigated fields consist of corn (Zhang et al., 2020). In particular, we focus on a region where the irrigation is prominent, i.e. the south of the state of Nebraska (100–97∘ W, 40.25–41.25∘ N, Fig. 2e). The objective of the model evaluation is to demonstrate that the model is able to reproduce irrigation activities and that the irrigation scheme improves vegetation modelling and the associated surface fluxes as compared to observations. It must be noticed that the new irrigation module represents the water demand for irrigation only, and irrigation is not explicitly limited by the lack of water resources. This has consequences on water conservation. Water used for irrigation is usually withdrawn from aquifers, rivers, or reservoirs. These compartments are not represented in ISBA, but a new module dedicated to dam/reservoirs is currently under development.

The SURFEX v8.1 version (Le Moigne et al., 2018) is used to do the simulations. Since this study focuses on irrigation, only the tile of natural and cultivated lands is simulated with ISBA, representing the evolution of soil (temperature and water profiles), vegetation (leaf-level and canopy-level photosynthesis, biomass, LAI, and carbon fluxes), surface hydrology (runoff and drainage), and snow conditions. To represent the global-scale diversity of continental natural surfaces, 20 different land surface types can be used in ECOCLIMAP-SG (see Fig. S1 and Table S2). The ISBA LSM simulations are made at a spatial resolution of 0.25∘ × 0.25∘ over a 40-year period from 1979 to 2018. The initial values of the soil moisture and soil temperature profiles are derived from a 20-year spin-up simulation by repeating year 1979. The same initial conditions are used for all the simulations, with and without crop phenology and irrigation modelling. All land surface types are grouped into 15 patches, including three irrigated ones, namely shrubs (orchards), C3 crops (typically wheat and rice), and C4 crops (corn). This study focuses on the results of these last two land surface types because there are hardly any irrigated orchards in Nebraska in the irrigation map described in Sect. 2.4.1. The dates of the irrigation season for corn are chosen in accordance with the literature (USDA and NASS, 2010) from May (emergence) to September (harvest), with a random picking of the day within those specific months. Three types of simulations are performed (Table 2), i.e. ISBA_ref, without irrigation or crop phenology (the benchmark), ISBA_pheno, with only crop phenology attributes (emergence and harvest dates), and the complete ISBA_pheno_irr simulation, with irrigation and crop phenology attributes. For the intercomparison of the simulations, we select areas where the irrigation fractional coverage is larger than 50 %, as determined from the irrigation map, in order to better assess the local effects of irrigation in offline simulations.

The reference ISBA_ref LAI simulations are compared with those from ISBA_pheno and ISBA_pheno_irr experiments and with the 0.01∘ × 0.01∘ LAI satellite observations over areas in Nebraska where the vegetation is considered to be C3 or C4 irrigated crops by ECOCLIMAP-SG. In addition to LAI, other variables are considered, including gross primary production (GPP) and land surface temperature (LST). In order to compare the time series simulations with observations, the Pearson correlation coefficient (r) and the root mean square difference (RMSD) scores are used. For water and carbon fluxes, they are calculated using daily values. 2r=∑i=1Nyi-y¯xi-x¯∑i=1Nyi-y¯2∑i=1Nxi-x¯2, where y and x stand for observations and model simulations, respectively, and 3y¯=1N∑i=1Nyi,x¯=1N∑i=1Nxi are observation and model simulation means, respectively. 4RMSD=1N∑i=1Nyi-xi2. N represents the number of observations interpolated or aggregated to the considered model grid cell, which is used in the calculation of the scores.

The significance of r, r differences, and RMSD differences is tested using Fisher's test, Fisher's z test, and a paired sample Student's t test, respectively. Significance levels of 0.01, 0.05, and 0.05 are used to reject the null hypothesis of the tests, respectively.

In Fig. 3, the mean yearly irrigation amount for C3 and C4 crops for the ISBA_pheno_irr experiment is compared to the values derived from the observations from the USGS. The simulated irrigation amount presents a large interannual variability, with a minimum of 60 mm in 1993 and a maximum close to 390 mm in 2002. It must be noted that 1993 is one of the wettest year recorded at the Lincoln weather station (https://lincolnweather.unl.edu/records/annual.asp, last access: 15 November 2022). The mean simulated value of the yearly irrigation water amount used for irrigation (271±75 mm yr-1) slightly overestimates the observed one (264±65 mm yr-1), with a difference of +2.7 %. This difference could be explained by the availability of the water resource, which is not explicitly accounted for by the model yet. The large observed irrigation amounts in 2000 and 2005, larger than 300 mm yr-1, are relatively well represented by the model. On the other hand, the observed small irrigation amount for the 2010 wet year is overestimated by about 110 mm yr-1. In situ precipitation observations over Grand Island indicate that year 2010 is wetter than 2005 and 2000 during the growing season, with 575, 508, and 277 mm from May to September, respectively. In 2010, the ERA5 precipitation bias from July to September triggers a cumulated precipitation gap of 103 mm (Fig. S18a). The model responds to this water deficit by triggering irrigation at the end of the growing season, especially in August (Fig. S18c). On the other hand, ERA5 is unbiased at the beginning of the growing period (May–June 2010).

Figure 4 illustrates the mean seasonal and interannual variability in LAI of irrigated crops (either C3 or C4) in the most densely irrigated part of Nebraska for areas with a fraction of irrigated crops larger than 50 % in Fig. 2e, from 1999 to 2018. Table 4 presents the peak LAI characteristics. While the satellite LAI observations present a peak at the end of July, the modelled LAI is plateauing in August (Fig. 4). Figure 2 in Boedhram et al. (2001) shows that the modelled LAI plateau in August at LAI values of about 3.7 m2 m-2 is realistic for irrigated corn. The satellite LAI observations are sensitive to both rainfed and irrigated vegetation. A comparison across all vegetation types is presented in Sect. 3.3. In all ISBA LAI simulations, the start of the growing season corresponds to a gradual increase in LAI from the initial value of LAI = 0.30 m2 m-2 imposed to the model in winter. The observed LAI presents a smaller, minimum LAI value of 0.15 m2 m-2, which starts increasing in April, and a value of 0.30 m2 m-2 is reached at the end of April. Then, plant growth continues at about the same low rate till the end of May. The LAI growth rate increases in June and LAI reaches a mean peak value of 4.9 (±0.8) m2 m-2 on 31 July (Table 4). The observed LAI then sharply decreases to reach its minimum value at about the end of September.

The ISBA_ref LAI simulations do not mirror the observed late growing season and rapid senescence. The ISBA_ref vs. observations comparison shows that, without crop phenology and without irrigation, the simulated LAI generally starts increasing in March. On average, a peak LAI value of 3.6 (±0.2) m2 m-2 is simulated by ISBA_ref on 2 July, before slowly decreasing until the end of December. The ISBA_ref growing season is much longer than observed. It starts 2 months before the observations and stops 3 months after the observations. The simulated LAI peaks 1 month before the observations. The simulated yearly LAI amplitude is 28 % smaller than observed.

The ISBA_pheno LAI simulation is much more consistent with the LAI observations. The growing season starts in mid-May, and the senescence ends at the end of September. However, the simulated peak LAI is still 30 % smaller than observed (LAI = 3.5 (±0.2) m2 m-2). The peak LAI is reached on 26 August, much later that the ISBA_ref peak LAI, and about 1 month after the observed peak. The sharp decrease in LAI in September results from harvests at random dates in September. Adding irrigation (ISBA_pheno_irr) does not change the general pattern of the LAI curve but increases the LAI amplitude, with a mean peak LAI value of 3.7 (±0.1) m2 m-2 on 28 August, which is larger (+8 %) than for ISBA_pheno but still below the observation (-24 %).

The interannual variability in simulated and observed LAI values is illustrated in Fig. 4b, from 2002 to 2008. The ISBA_ref LAI presents a systematic drop in summer, which is neither present in the observations nor simulated by the ISBA_pheno and ISBA_pheno_irr experiments. Without the regular seasonality imposed by crop phenology parameters, the model may simulate a regrowth of vegetation in autumn (e.g. in 2003) that is not present in the observations. The ISBA_pheno and ISBA_pheno_irr simulations are more consistent with the observed seasonality.

This section is focused on the impact of irrigation practices for the southern Nebraska zone (as defined in Fig. 2e), and all land surface types are considered for the comparison with observations at a spatial resolution of 0.25∘ × 0.25∘.

Figure 5a shows the seasonal mean LAI variations from 1999 to 2018. This figure is similar to Fig. 4a, but all land surface types are considered. Peak LAI characteristics are given in Table 4. They differ from the irrigated crop LAI peaks. While the observed LAI peaks at 3.8 (±1.5) m2 m-2 on 31 July, LAI peaks at 3.3 (±0.3) m2 m-2 on 1 July for ISBA_ref, 3.1 (±0.3) m2 m-2 on 16 July for ISBA_pheno, and 3.1 (±0.3) m2 m-2 on 16 July for ISBA_pheno_irr. Compared to irrigated crop simulations, the experiments with crop phenology (ISBA_pheno and ISBA_pheno_irr) present earlier peak LAI dates because they include rainfed crops and natural vegetation. Emergence dates are not imposed on rainfed crops and natural vegetation. This allows earlier leaf onset. The irrigated crop signature is visible in the second peak of the annual LAI cycle simulated by ISBA_pheno and ISBA_pheno_irr experiments at the end of August. More often than not (83 % and 88 % of the grid cells for r and RMSD, respectively), the LAI score differences between ISBA_pheno_irr and ISBA_ref shown in Fig. 6 correspond to an improvement in the LAI simulation with the representation of irrigation. A month-by-month analysis of the scores (Fig. 7) shows a significant improvement in r values in June and September. The r value can be increased by 30 %. Lower RMSD values are observed from April to November but more frequently in May and in October. In July, RMSD differences are not significant. In April, October, and November, the main cause of the reduction in RMSD values is the imposed minimum value of 0.3 m2 m-2 before the emergence (in May) and the harvest (in September). The ISBA_pheno correlation and RMSD differences with respect to ISBA_ref are nearly identical to those shown for ISBA_pheno_irr in Figs. 6–7 (not shown).

As the vegetation productivity is linked to LAI, the seasonality pattern of GPP (Fig. 5b) is comparable to the one of LAI (Fig. 5a), but the observed GPP peak (9.2±2.1 g [C] m-2 d-1) occurs in mid-July while the observed LAI peaks on 31 July. During the plant growth period, the smallest differences between all the simulations and the observations occur at about the same time as the observed GPP peak. For all simulations, a GPP plateau with a value of 9.0±1.8 g [C] m-2 d-1 is reached at the beginning of July and lasts until mid-July. Finally, the simulated GPP decreases in September with a delay of about 2 weeks with respect to the observations.

Before July, the ISBA_ref photosynthetic activity is well in advance, as compared to the observations, of about 20 d in May. This is consistent with the very large LAI values simulated by ISBA_ref in May, with about 2 m2 m-2, while the mean LAI observation hardly exceeds 0.5 m2 m-2. The simulated GPP maximum (9.7±2.0 g [C] m-2 d-1) is reached before the end of June. After a sharp decrease at the end of June, the ISBA_ref GPP decreases at a slower rate than the observations. From mid-September to the end of October, the simulated GPP is much larger than the observed GPP.

The ISBA_ref flaws are much less pronounced in ISBA_pheno and ISBA_pheno_irr experiments. In the latter simulations, the increase in the GPP occurs at about the same time as in the observations. The GPP values are systematically larger with irrigation in July and August than for other simulations. As for LAI, the GPP, r, and RMSD scores (Fig. 6c and d, respectively) are better for ISBA_pheno_irr than for ISBA_ref, with an improvement of 87 % and 81 % over the domain, respectively.

In order to evaluate the impact of irrigation on the land surface energy budget, Fig. 8 shows the land surface temperature at 12:00 local solar time simulated by the three model configurations and derived from the CGLS product. Overall, ISBA tends to overestimate the LST at noon, especially in April–May, by up to 7 ∘C (Fig. 8a). The bias is reduced during the summer.

Due to the difficulty of observing the differences between the simulations, Fig. 8b presents differences in ISBA_pheno and ISBA_pheno_irr vs. ISBA_ref. With crop phenology (with or without irrigation), the simulated LST is globally higher from April to June and from mid-September to November. The maximum difference with respect to ISBA_ref is +0.7±0.3 ∘C. It is observed for ISBA_pheno in September. During the summer (July and August), the new model versions tend to present lower LST values with temperature differences close to -0.2±0.1 ∘C in ISBA_pheno_irr. Moreover, from May to mid-September, the temperature in ISBA_pheno_irr is lower than in ISBA_pheno, and this difference can reach -0.9 ∘C locally in summer.

The crop phenology model is very simple, and adding more parameters related to phenology could be a way to further improve the model performance. Integrating satellite LAI observations in ISBA, using sequential data assimilation, is also an option (Mucia et al., 2020). The results of our numerical experiments over Nebraska show that considering crop phenology improves the consistency of the simulations with LAI and GPP observations. The corresponding correlation and RMSD scores are improved. The crop phenology parameters used to force emergence and harvest dates reduce the length of the growing season, delay spring growth, and avoid a regrowth in the autumn. It seems that irrigation only plays an additive role in improving the vegetation seasonal cycle as compared to the role of including crop phenology (Sect. 3.3). Both crop phenology and irrigation models have shortcomings, and their performance is limited by difficulties in simulating processes that are not directly related to irrigation.

First, the same emergence and harvest dates are imposed for all years, while in reality crop phenology may present an interannual variability related to climate conditions. This is particularly the case for Nebraska because the start of the growing season depends on the snowmelt and soil thawing dates. These processes are represented in ISBA and crop phenology parameters could be related to snow melting and soil thawing, but this would require extensive developments to be implemented on a global scale. Moreover, the representation of the cold season processes is not perfect in ISBA (Decharme et al., 2019), and the model tends to underestimate snow depth and the length of the snow season. This could explain biases in soil temperature and LST simulations before and after the irrigation time period. Figure 8 shows that LST values below the freezing level can be observed in April, and that their model counterparts are about 7 ∘C warmer. The earlier thawing in model simulations is reflected in the much earlier leaf onset in LAI simulations. Figure 5 shows that, while the observed LAI does not exceed 0.5 m2 m-2 at the end of April, the ISBA_ref LAI reaches the same value about 1 month earlier. The unrealistically early leaf onset is consistent with the warm model bias at the end of the cold season. This shows that improving the representation of the cold season by assimilating satellite-derived or in situ snow cover fraction observations could improve the simulation of the crop growing period in this area. Also, emergence and harvest dates could be derived from the LAI observation in order to better represent the interannual variation. However, the currently used random approach to establish the crop season provides robust results over the irrigated grid cells (Fig. 4).

Second, the irrigation itself is based on fixed parameter values such as the minimum period between two consecutive irrigations (1 week) and SWI levels triggering irrigation events. The simulations over the Hampton grid cell show that the first irrigation can start at quite low levels of the SWI (Fig. S7), even below the second irrigation threshold of 0.55 defined in Sect. 2.2. Suppressing the 1-week constraint of irrigation events improves the simulation of the peak LAI, which otherwise is rather poorly simulated (Fig. 4). However, this change triggers unrealistic large irrigation water amounts (not shown). A lack of irrigation water amount cannot explain the excessive soil water deficit. A possible explanation could be that the initial soil water storage value between the end of the cold season and the first irrigation event is withdrawn too quickly from the soil by the model. One could also challenge the quality of the ERA5 precipitation. ERA5 precipitation compares rather well with in situ observations, and the seasonal and interannual variability is fairly represented (Figs. S13 and S14). The comparison with in situ precipitation observations at the Grand Island station shows that ERA5 tends to markedly overestimate precipitation in April by 0.57 mm d-1, i.e. 27 % on average from 1985 to 2018 (Fig. S15). This is rather systematic, i.e. 8 years out of 10 (Fig. S16). In July, ERA5 precipitation can be much smaller than the observations, for example, in 1991 and 2007 (Fig. S17).

A possible limitation of using a global low-resolution reanalysis such as ERA5 is that changes to the local climatic conditions caused by irrigation may not be represented. Over Nebraska, Szilagyi and Franz (2020) show that the decadal increase in irrigated land tends to trigger a reduction in precipitation over the most densely irrigated areas of about -10 mm per decade. The largest precipitation suppression is observed in spring, in March, before the corn growing season, in relation to larger soil water content values. In our simulations, ISBA_pheno_irr presents larger soil moisture values than ISBA_ref in March (see Fig. S7), but this is mainly due to crop phenology. The ERA5 screen-level 2 m air temperature and relative humidity are analysed together with soil moisture by assimilating in situ observations from ground weather stations (Hersbach et al., 2020). In large irrigated areas where weather stations are present, the assimilation should be able to represent the soil moisture effect on these variables, even at coarse spatial resolution. A large-scale experiment involving ground and airborne measurements was recently performed in northeastern Spain to assess the impact of irrigation on atmospheric model simulations (Boone et al., 2021). In this context, high-resolution atmospheric data from the Application of Research at the Operational Mesoscale (AROME) numerical weather forecast model are available to drive the ISBA model. AROME is run operationally at 1.3 km over western Europe. Future developments will focus on the intercomparison of ISBA simulations at various spatial resolutions and under different conditions, as the choice of spatial resolution may affect the simulation of the smaller irrigation variabilities.

All evaporation terms (plant transpiration, soil evaporation, and interception) are simulated by ISBA. Under given environmental conditions, the simulated plant transpiration is not proportional to LAI. A simple canopy radiative transfer model is used to simulate the available photosynthetically active radiation (PAR) within the vegetation canopy. The response of GPP and transpiration to PAR and to LAI is controlled by this radiative transfer model. Photosynthesis and transpiration are calculated for three layers and summed to calculate canopy-level values. For large LAI values, the mean leaf-level GPP and transpiration simulations are reduced in relation to smaller vegetation transmittance to solar radiation. The impact of changes in LAI on mean leaf-level GPP and transpiration is large at intermediate LAI values ranging from 1 to 3 m2 m-2. It is much reduced for LAI values larger than 3 m2 m-2. An improved version of this radiative transfer model able to represent 10 canopy layers, and a more realistic response to solar zenith angle will be available in the next version of SURFEX (Delire et al., 2020). Over the Hampton irrigated area, total evapotranspiration of ISBA_ref and ISBA_pheno_irr is quite similar (Fig. S9). On the other hand, soil evaporation and plant transpiration differ. In the ISBA_pheno_irr simulation, transpiration is reduced in spring by more than 30 % compared to ISBA_ref. The lower transpiration is offset by larger soil evaporation values (Fig. S9c). As a result, total evapotranspiration does not change much. Also, the new crop phenology and irrigation module does not affect interception much (Fig. S10). The soil evaporation component could be overestimated because (1) the soil is too warm in relation to a poor representation of thawing or because (2) crop residues at the soil surface are not represented. Wortmann et al. (2012) show that, in this area, not harvesting crop residues tends to reduce soil evaporation and increase crop yield, limit water runoff, soil erosion, and contributes to maintaining soil fertility. Suyker and Verma (2009) show that increasing surface mulch dry mass from 50 to 150 g m-2 can decrease the non-growing season evapotranspiration by more than 20 %. The ISBA model includes a representation of litter in forests (Napoly et al., 2017) that will be generalised to low vegetation in the next version of SURFEX. Using this new capability could improve our simulations.

The use of independent evapotranspiration datasets is investigated in Sect. S3. In particular, in situ observations over an irrigated corn field (Suyker and Verma, 2009) are used. During the non-growing season, the ISBA_pheno_irr model overestimates evapotranspiration by 48 % (Table S4). The ISBA_pheno_irr evapotranspiration peak (in June) tends to happen too early (Table S5; Fig. S11). This bias in spring could be caused by too large soil evaporation values. Mean values of near-surface wind speed are particularly large over Nebraska, especially during wintertime and springtime (Chen, 2020). This feature could exacerbate the impact of a misrepresentation of soil evaporation.

In this study, sprinkler irrigation is considered. The model is able to represent other irrigation systems such as flooding irrigation but more developments are needed to limit the runoff to the irrigated plot, and this options needs to be validated. The newly implemented irrigation processes, along with the new ECOCLIMAP-SG vegetation description, let users choose which land surface type should be irrigated. Irrigation can be represented at various spatial scales, ranging from the field scale for agricultural studies to the global scale for climate studies. Model parameters can be specified using new datasets or local characteristics. For example, in this article we use a unique date for starting and ending the crop growing season with a random variability, but more accurate dates can be prescribed (varying spatially and from one vegetation type to another or using crop calendars). Moreover, the better spatial resolution of ECOCLIMAP-SG allows the use of high-resolution atmospheric forcing. This provides new opportunities for assessing the impact of irrigation on local climate and water resource conditions.

This study is mainly focused on a zone in the south of Nebraska where the irrigation density is relatively high (Fig. 2), and results could differ in other regions. Except for the fixed emergence and harvesting dates corresponding to regional crop phenology (from USDA and NASS, 2010), default values are used for all the other parameters (Sect. 2.4). Tests performed in southwestern France (Sect. S2) allow us to ensure that the model is able to work in contrasting climate conditions.

In this study, the ISBA simulations are neither coupled to the atmosphere nor to the CTRIP river routing system. Such coupled numerical experiments can be performed thanks to the SURFEX modelling platform. However, more developments are needed in order to ensure water conservation in the hydrological system. In particular, irrigation water amounts should be consistent with the available water resource in rivers, groundwater, and dams.