1. Introduction

Alteration of soil moisture by irrigation can influence local climate by an amount comparable to or, in some cases, in excess of the effects of other climate forcings such as greenhouse gas changes (Bonfils and Lobell 2007; Kueppers et al. 2007). Improved quantification of the effects of irrigation will be needed for projections of climate change that inform adaptation efforts, particularly in developing regions where irrigation expansion continues at a much faster rate than in developed countries. However, most studies to date have been limited to irrigated regions in the United States where datasets on both land use and climate are relatively rich (Kueppers et al. 2008; Mahmood et al. 2006). Recent modeling studies have also examined effects in India and Thailand but, in general, the impacts of irrigation on temperature in these regions are not well understood (de Rosnay et al. 2003; Douglas et al. 2006; Haddeland et al. 2006).

The response of climate to irrigation can be expected to vary by region for two main reasons. First, the extent of irrigation differs markedly, with nearly 100% of the land area irrigated in some regions, such as northwest India or the Central Valley of California, but substantially less intensive irrigation in other regions. As a result, the relative change in land surface properties will vary by region. Second, the climate regime of some regions may be more intimately tied to land surface processes than in others (e.g., Koster et al. 2004). For these reasons, it is unclear how transferable the results of well-studied regions, such as California (Bonfils and Lobell 2007; Kueppers et al. 2007), are to regions with fewer data and modeling efforts.

Previous global modeling studies have evaluated the impact of irrigation in multiple regions, but with simplifications that prevent detailed regional comparisons. For example, Boucher et al. (2004) prescribed a surface flux of evapotranspiration (ET) based on spatial maps of estimated irrigation ET, but this flux was seasonally constant and only annual average temperature changes were reported. Lobell et al. (2006a,b) performed a sensitivity analysis with irrigation applied in all agricultural areas regardless of their actual irrigation use.

The goal of this study was to evaluate regional differences in the response of near-surface (2 m) air temperature to irrigation, using a general circulation model (GCM) with a more realistic treatment of irrigation, in order to understand both the magnitude and causes of these differences. A secondary goal was to compare simulated temperatures with observations to test whether inclusion of irrigation in a GCM reduces or aggravates model bias on a region-by-region basis.

2. Model and methods

The study consisted of simulations for the period from 1978 to 1999 performed with version 3.3 of the Community Atmosphere Model (CAM3.3) (Collins et al. 2006). We used the finite volume scheme for atmospheric transport, and monthly sea surface temperatures and sea ice extents were prescribed according to values supplied by the Atmospheric Model Intercomparison Project (AMIP) (Gates et al. 1999). Our model experiments were performed at three different horizontal resolutions with latitude × longitude mesh sizes of 1.88° × 2.50°, 0.94° × 1.25°, and 0.47° × 0.63° to evaluate the dependence of results on the model spatial resolution. For brevity, we present only the results from the coarser model (1.88° × 2.50°) since results were qualitatively identical for all resolutions. Results are highlighted for eight major irrigated regions, as outlined in Fig. 1: California (CA), Nebraska (NE), Mississippi (MS), Spain (SP), Turkey (TU), Aral Sea Basin (AB), Indo-Gangetic Plains (IGP), and northeast China (CH).

In CAM3.3, land–atmosphere interactions are formulated according to version 3 of the Community Land Model (CLM3) (Oleson et al. 2004). For each land grid cell, bare-soil and vegetated fractions are specified, where the latter may include up to 4 of a total of 16 plant functional types (PFTs) that comprise a variety of trees, grasses, shrubs, and agricultural crops. In the first experiment (CONTROL), CAM3.3 was run in its default configuration, which utilizes a single crop PFT that is not irrigated. In the second experiment (IRRIG), we added a second crop PFT to represent irrigated agriculture. To ensure that the amount of irrigated land was as realistic as possible, the fraction of the irrigated crop PFT was set equal to an estimate of the area equipped for irrigation from the 5′ × 5′ resolution map distributed by the Food and Agriculture Organization (FAO) (Siebert et al. 2005; Fig. 1a). This fraction was then subtracted from the original crop PFT to ensure the same total agricultural area as in CONTROL. In cases where the FAO fraction exceeded the original crop PFT fraction in CLM, the irrigated crop PFT was restricted to the latter value and the original crop PFT was set to zero. This occurred in less than 3% of the grid cells with an original crop PFT greater than 5%.

We modified CLM to include a representation of irrigation for the fraction of the grid cell containing the irrigated crop PFT, which importantly has a separate soil column from the other PFTs within the grid cell, ensuring that soil moisture changes are confined to the irrigated crop PFT. Soil moisture in the top 30 cm of the irrigated column was modified at each time step (every 30 min) by adding water when soil liquid water fell below a critical value M*, with the quantity of added water set equal to the deficit below this critical value. No water was added for time steps when moisture was above M*. A single value of M* was used for all grid cells and all months and was set at 40% of soil saturation based on examination of moisture in the CONTROL simulation (Fig. 2). The average monthly moisture in the CONTROL simulation fell below 40% in the summer months for most irrigated regions but exceeded this value for the remainder of the year. To complement this simulation, we also conducted simulations with the threshold M* changed to 30%, 50%, and 90%.

3. Results and discussion

As expected, soil moisture in IRRIG was higher than in CONTROL for all months that had average values below 40% in CONTROL (Fig. 2a). Significant changes were also evident for other months because many individual time steps in these months were below 40% in CONTROL despite average moisture above this value. Soil moisture was essentially unchanged for summer months in India and China, which had relatively high levels of average moisture (>50% of saturation) in CONTROL. Changing the threshold from 40% to 30% resulted in relatively lower soil moisture in most months (Fig. 2a) but still above that of the CONTROL simulation, even in months with average moisture well above 30%.

The total amount of irrigation water applied was computed for each month by summing applied water for all grid cells without soil ice. The use of a simple threshold based on soil liquid water meant that irrigation was often applied when water was present in the soil, but frozen. Obviously, this is neither a time when actual irrigation would take place, nor a time when irrigation would be expected to affect climate (although water added in frozen soils could affect the subsequent spring thaw); therefore, we considered only water added to ice-free soil. The annual applied water averaged 5.9 × 10³ km³ for the simulation with a threshold of 40%, and 4.9 × 10³ km³ in the simulation with a threshold of 30%. These are comparable with an estimate of actual total annual withdrawals for irrigation of 2.5 × 10³ km³ by Döll and Siebert (2002). That is, our simulation applied an amount of water that was the right order of magnitude but roughly twice as much as withdrawn for irrigation in reality, which in turn is significantly higher than the amount of water actually applied because of losses before the water reaches the field.

Reductions in average monthly temperatures for IRRIG relative to CONTROL ranged from 0° to 10°C, depending on the month and region (Figs. 1b,c and 2b). Most regions were characterized by reductions of 5°C or more in the dry season [June–August (JJA) in most regions, December–February (DJF) in IGP and CH]. Very little change was observed for the wet season, even though irrigation was applied year-round. Reductions in CH were less than 3°C in all months as a result of the relatively high soil moisture in the CONTROL simulation for this region.

As discussed above, the sensitivity of the results to the prescribed value of M* was tested by repeating the experiment with values of 30% and 50% and then with an extreme value of 90%. The simulated air temperatures were nearly identical for the four cases (including the 90% case), which reflects the fact that latent heat (LH) fluxes in the model are extremely insensitive to moisture increases beyond 30% of saturation. Indeed, the equations in CLM that govern soil evaporation, which are common to many land surface models, saturate between 20% and 30% of soil saturation, depending on soil texture (Oleson et al. 2004). Canopy transpiration, which on average contributed 32% of total simulated LH flux in crop regions, is similarly insensitive to moisture levels above 30%. This implies that parameterization of the frequency of irrigation, which determines the length of time spent at very low moisture, is more important than the amount of water applied and that future implementations of irrigation in climate models should consider multiple timings, such as biweekly additions of water that mimic actual irrigation practices.

The lack of temperature sensitivity to further moisture increases in our experiments suggest that the model is providing an upper limit to the cooling effect of irrigation, limited only by the prescribed areal extent of irrigation in each region. However, comparison with observed temperatures, as represented in the Climatic Research Unit (CRU) TS2.1 datasets (Mitchell and Jones 2005), revealed substantial reductions in bias for IRRIG relative to CONTROL (Fig. 2b). In CA, NE, and TU, for example, a warm bias in the CONTROL simulation was significantly reduced in IRRIG.

Model bias was also reduced in IRRIG relative to CONTROL in MS, although a positive bias remained. In SP and AB, the positive bias in CONTROL was replaced by a negative bias in IRRIG, albeit one with a smaller magnitude. In IGP, the CONTROL simulation had a cold bias that was exacerbated by irrigation, indicating that model biases in this region are not likely associated with a lack of irrigation. Finally, biases in CH were relatively small throughout the year and unaffected by irrigation.

Given that moisture levels were held artificially constant in the irrigation experiment, one might have expected the cooling effect of irrigation to be overestimated. The small bias in some regions suggests that either the overestimate is small (i.e., the model estimates an appropriate magnitude of cooling) or the overestimate is balanced by other errors in the model. For example, several authors have discussed a widespread dry bias in CLM (Bonan and Levis 2006; Oleson et al. 2008). In the current study, for example, strong warm biases persist in IRRIG for nonirrigated portions of the United States (not shown), supporting the notion that lack of irrigation is not the sole explanation for biases in CONTROL. Moreover, global ET increased by 5.9 × 10³ km³ in IRRIG (with M* = 30%) compared to that of CONTROL, while estimates of actual ET from irrigation are less than half this value (Gordon et al. 2005), suggesting that the cooling effect of irrigation is, indeed, overestimated on a global basis. One reason for the overestimate of ET, in addition to the large amounts of total added water, is likely that irrigation water was artificially added rather than routed from local surface waters. Shibuo et al. (2007), for example, modeled irrigation in the AB region by removing the applied irrigation water from the regional inland water systems, and estimated a net change in water export from the AB region (ET − precipitation) of 40 km³ yr⁻¹. Our estimates of net change are roughly three times this value over a smaller domain, indicating that accounting for local water withdrawals is needed for accurate ET simulations.

As discussed in previous work (Boucher et al. 2004; Kueppers et al. 2007; Lobell et al. 2006b), the primary mechanism for cooling upon irrigation is an increase in the fraction of incident energy partitioned into LH flux rather than sensible heat (SH). Comparison of simulated changes in LH with daily maximum temperatures (T_max) revealed that much of the variation in temperature response between regions was indeed driven by differences in LH (not shown). There were notable exceptions to this trend, however, such as in IGP where DJF temperature reductions were similar to JJA cooling in NE or CA but where LH increases were much smaller. The presence of another mechanism of cooling is therefore indicated.

Feedbacks from clouds can also contribute to cooling, whereby greater moisture fluxes associated with LH lead to increased cloudiness and reduced net surface radiation (Lobell et al. 2006b). Net surface changes in LH, SH, and shortwave radiation (S) for IRRIG − CONTROL are shown in Fig. 3, revealing large reductions in S for many regions associated with cloud increases. In IGP, the effect of this feedback far exceeds that of the initial change in LH from irrigation in DJF. We also note that the simulated change in LH is similar to values of 10 W m⁻² estimated for the same region and season by an independent terrestrial water balance model (Douglas et al. 2006).

Representation of cloud processes is among the most uncertain aspects of climate models: Therefore, any cooling from cloud feedbacks should, in the least, be simulated by multiple climate models in order to be considered a robust result. We therefore consider our results suggestive but in no way conclusive. If cloud feedbacks are in fact important in IGP, then the rapid expansion of irrigation in the latter part of the twentieth century should have been associated with increased cloudiness in the winter. Several studies have noted large reductions in S in India since 1970 (Padma Kumari et al. 2007; Ramanathan et al. 2005), but have attributed this mainly to increased emissions of sulfate aerosols and black carbon. Padma Kumari et al. (2007) note that reductions in S have been largest in the northern part of India, where irrigation is more dominant, and that increased cloudiness may have played some role in this trend. In general, however, the relative influence of haziness versus cloudiness on surface radiation trends remains unclear.

4. Conclusions

Simulations with CAM3.3 indicated potentially large differences in the response of temperature to irrigation in different regions, with some areas cooling by 10°C and others exhibiting little change. These differences were largely explained by the change in latent heat flux that accompanies irrigation, which in turn depended on both the area of land irrigated and the simulated moisture under rain-fed conditions (i.e., in CONTROL). The results of this study were therefore dependent on biases in simulated soil moisture in CAM3.3. For example, the model indicated very little effect of irrigation in NE China because of high soil moisture levels in the CONTROL simulation relative to other irrigated regions.

The results were also dependent on the treatment of clouds in CAM3.3, as cloud feedbacks provided an important cooling mechanism in the Indo-Gangetic Plains and several other regions. To better understand the interaction of irrigation with climate, repetition of these experiments with multiple climate models would therefore be useful, as would evaluation of historical cloud cover trends in areas such as northwest India.

Inclusion of irrigation was found to reduce CAM3.3 temperature biases in several regions (Fig. 2b). While this result suggests that representation of irrigation can contribute to improved simulation of temperature climatologies, it does not alone justify adding irrigation modules in GCMs. In particular, the added irrigation may mainly be compensating for a dry bias in CLM. The ultimate value of including irrigation in climate models will depend on the specific goals of the GCM. For example, improving model biases may have relatively small effects on the simulated changes in climate for greenhouse gas forcing (Tebaldi and Knutti 2007). Simulations with an earlier version of CAM showed that warming for a doubling of atmospheric carbon dioxide was only slightly smaller for a model that included irrigation because the model did not contain feedbacks from soil drying that were simulated without irrigation (Lobell et al. 2006a). Thus, irrigation may strongly influence temperatures but, if the irrigated area does not change with time, then GCM simulations of temperature change from other forcings may still be reliable.

Finally, the results presented here indicate that future studies should consider a more realistic treatment of the temporal nature of irrigation, with water added on a weekly or semiweekly rather than subdaily basis, and only during the growing season months. In the latter case, soil water never falls below values at which evaporation is sensitive to moisture changes (20%–30% of saturation), and more irrigation water is added globally to the model than has been observed. In addition, rather than artificially adding water as in this study, the conservation of water by removing irrigation water from rivers or reservoirs (e.g., de Rosnay et al. 2003; Haddeland et al. 2006) and groundwater (e.g., Kollet and Maxwell 2008) may be warranted, particularly in cases where sea surface temperatures are not prescribed but are instead influenced by freshwater inflows.