We analyze different model experiments to investigate irrigation forcing of climate, all using the GCM ModelE2 (2∘ latitude × 2.5∘ longitude resolution), the latest version of the GISS atmosphere general circulation model, with 40 vertical layers in the atmosphere and updated physics . Irrigation water is added to the vegetated fraction of the grid cell at the top of the soil column, beneath the vegetation canopy. Irrigation rates are nominally for the year 2000, taken from a global gridded reconstruction (Fig. ). This reconstruction estimates irrigation demand based on combining maps of irrigated areas and crop types with crop-specific evapotranspiration scale factors, with a special allowance for maintaining a constant flood depth in paddy rice areas . Water for irrigation is initially withdrawn from rivers and lakes in the same grid cell. If irrigation demand is not satisfied by these surface sources, water is added under the assumption that it is taken from groundwater sources that are not represented in the model (i.e., “fossil” groundwater). The irrigation rate is kept constant over the course of the day and applied for every sub-daily time step. Irrigation water will either infiltrate the soil column or run off to the streams in the grid cell. The total amount of irrigation water averaged 0.019 mmday-1 (6.8 mmyear-1) globally (3500 km3year-1total), with a mean of 0.46 mmday-1 (168 mmyear-1) over irrigated land grid cells (defined as those for which the average irrigation amount was at least 0.1 mmday-1). Globally some 42 % of the applied irrigation water (1500 km3year-1) was modeled as coming from groundwater, a proportion similar to that found in an independent modeling study of the fraction of consumptive irrigation water use deriving from groundwater . Additional details and discussion of the irrigation scheme are in and . As opposed to “irrigation” (irrig) runs, in “control” (ctrl) runs no irrigation water is applied. Forcings such as greenhouse gas concentrations were the same for all the experiments, and based on values from around the year 2000 .

Irrigation and control simulations were carried out with two different ocean configurations. The first involves forcing the atmosphere model with prescribed, annually repeating monthly sea surface temperatures (SSTs) and sea ice. The SSTs and sea ice are based on average 1996–2004 data from the Hadley Center analysis . We refer to this as the atmosphere-only, fixed-SST, or A configuration. In the second configuration (referred to as “q-flux” mode, interactive-SST, interactive-(surface)-ocean, or O configuration), the ocean is represented as a 65 m deep mixed layer, with a prescribed distributed heat source to represent the effects of horizontal and vertical ocean mixing and advection.

The four simulations – irrig-A, ctrl-A, irrig-O, ctrl-O – were run 60 years each. The q-flux mode takes approximately 10 years to reach equilibrium under constant forcings, so we analyzed only the last 50 years of each simulation, which represent approximately steady-state conditions that show internal system variability under the different model configurations (i.e., with fixed SST or q-flux ocean, and with or without irrigation). Figure  illustrates the approach to equilibrium of the simulations. The A runs stayed at essentially the same temperature (with internal year-to-year variability) from the first year, but had 0.4 Wm-2 more radiation entering Earth than leaving. This was because the observation-based fixed SST was cooler than needed for radiative equilibrium with the imposed greenhouse gas concentration , although surface temperature and other climate variables did remain at steady state within the fixed-SST model configuration. In the O runs, the radiative imbalance largely resolved itself within a few years by SSTs and surface air temperatures warming around 0.3 K, and a difference of 0.1 K between the irrig and ctrl runs in the equilibrium mean temperature was evident (Fig. ).

For climate variables of interest, we considered irrig–ctrl differences in the monthly fields for both the A and O configurations. The irrig–ctrl difference field for the A set of experiments is referred to as ΔA, and the irrig–ctrl difference field for the O set of experiments is referred to as ΔO. The impact of interactive SSTs on the equilibrium irrig–ctrl difference was then obtained as ΔΔ≡ΔO-ΔA.

The significance of differences ΔA, ΔO, and ΔΔ, either at individual grid points or spatially averaged, was estimated using a Student's t test on their time series over the 50-year analysis period, with the degrees of freedom adjusted based on the lag-1 autocorrelation of the time series. This adjustment is based on the notion of effective sample size in time series analysis, taking as the null hypothesis that the difference time series is red noise with zero mean .

As metrics of overall irrigation and ocean configuration impacts, we looked at the mean and the root mean square (rms) of ΔA, ΔO, and ΔΔ aggregated globally over irrigated areas (which we defined as grid cells and months where the applied irrigation was over 0.1 mmday-1), non-irrigated land areas, and ocean areas. We considered annual means of these quantities as well as seasonal means. For seasonal means, aggregation was performed only over the zone of the Northern Hemisphere where the vast majority of the irrigation takes place (8–46∘ N, 92 % of global irrigation), to preserve consistent seasonality.

We focus on model climate variables that directly quantify conditions and moisture status at Earth's surface (surface air temperature, SST (only over ocean), precipitation, soil moisture (only over land), cloud fraction), terms in the surface energy balance that are affected by irrigation (latent and sensible heat fluxes), and circulation-related quantities (sea-level pressure, geopotential height, and meridional jet stream velocity fields) that can provide insight into how irrigation effects on surface energy and water balance could propagate to impact climate in distant regions. (The jet stream velocity is defined to be that at a pressure of around 250 mb, in the upper troposphere.)

The irrigation-induced surface air cooling, though still concentrated over irrigated land areas, spread over ocean areas in the interactive-SST simulation, consistent with our expectation that fixing SST would tend to stabilize air temperatures over the ocean. Global-mean above-ocean surface air temperatures and sea surface temperatures both decreased by 0.08 K (Table ). In the fixed-SST irrigation simulation, precipitation slightly decreased over the irrigated areas and increased elsewhere. Compared to the fixed-SST irrigation simulation, the cooling over the oceans slightly reduced evaporation and precipitation in the interactive-SST simulation. Interactive SST did not significantly modify the global-mean enhancements in soil moisture and cloudiness due to irrigation (Table ). Irrigation-induced latent and sensible surface heat fluxes were both slightly diminished in the interactive-SST simulation, consistent with the cooler surface temperatures and reduced precipitation (Table ). The mean atmospheric pressure responded inversely to the temperature change, with higher pressure in the cooled irrigated areas (consistent with the reduced precipitation there). The mean 300 mb height decreased significantly more in the interactive-SST simulation even in the irrigated areas, showing that, compared to fixed SST, interactive SST spreads the cooling due to irrigation throughout the atmospheric column (Table ). The meridional jet stream velocity was slightly higher in the runs with irrigation, although the effect of interactive SST (ΔΔ) was only significant over irrigated areas (Table ).

Table  shows changes by season (averaged over 8–46∘ N) for surface air temperature. Over land, the cooling is greatest in the summer and fall, when the largest amount of irrigation water is applied, and the mean amount is not significantly affected by whether SST is interactive. Over the ocean, cooling is more uniform across seasons, and is much greater in the interactive SST simulation (Table ).

The global or Northern Hemisphere mean impacts just shown conceal much spatial variability in the response to irrigation. The rms of the spatial field of irrigation response for the same climate variables (Table ) shows that interactive SST tends to increase this spatial variability over the ocean and non-irrigated land, even for variables such as over-ocean cloud cover and jet stream velocity for which the mean response is not significantly affected, implying that interactive SST on the whole enhances non-local irrigation impacts on climate. One exception is that interactive SST decreases the spatial variability in latent and sensible heat irrigation responses over the ocean (Table ), presumably because the interactive SST adjusts to changes in air temperature in a way that reduces the equilibrium change in surface fluxes.

We show illustrative maps of the seasonal-mean irrigation response with and without interactive SST (ΔA, ΔO). Under fixed SST, irrigation-induced changes in surface air temperature (Fig. ) are primarily local to major irrigation regions such as India, China, and the United States (USA), and effects in the ocean tend to be small, except in the North Pacific. Under interactive SST, irrigation-induced regional changes tend to have larger amplitude (∼ 0.8 compared to ∼ 0.4 K; Table ) and are also found in the tropical and southern oceans. Under fixed SST in boreal winter, the middle and high northern latitudes show a stationary wave pattern of alternating warm and cool anomalies due to irrigation (which during that season is concentrated in the Indian subcontinent). Under interactive SST, these boreal winter anomalies shift locations somewhat (for example, the cooling centered in the eastern USA under fixed SST is attenuated) and persist to a greater extent during the other seasons, and analogous wave patterns in the Southern Ocean and Antarctica are considerably stronger than under fixed SST. Under interactive SST, surface air temperature anomalies outside irrigated areas tend to be closely associated with SST anomalies of the same sign (Fig. ), which provide a mechanism for the surface air temperature anomalies to persist across seasons and supports the role of air–sea interactions in driving the divergence in irrigation responses between the fixed-SST and interactive-SST simulations.

Under fixed SST, a reduction in 300 mb height (corresponding to cooling of the atmospheric column; Fig. ) is seen primarily around irrigation regions in the northern midlatitudes, while under interactive SST the reduction in the northern midlatitudes is more zonally uniform, and there is also a stronger stationary wave pattern in the Southern Hemisphere roughly corresponding to locations of surface air temperature changes there. Particularly in boreal winter, the stationary wave pattern seen for surface air temperature is also found in the upper atmosphere, with phases shifted between the interactive and fixed SST simulations (Fig. ). The meridional jet stream velocity (ujet) changes correspondingly, consistent with geostrophic adjustment of the atmospheric circulation: ujet tended to increase on the north side, and decrease on the south side, of areas where 300 mb geopotential height rose, and vice versa where geopotential height dropped (Fig. ). The changes seen in 300 mb height are of the order of 20 m or ∼ 0.2 %, while the changes seen in meridional jet stream velocity are of the order of 2 ms-1 or ∼ 10 % (cf. Tables  and ).

Precipitation impacts (Fig. ) are strongest over the tropics and subtropics and appear to reflect, for example, a northward shift due to irrigation in the Intertropical Convergence Zone (ITCZ) in and south of India in boreal winter and a southward shift in boreal summer, with the zonal-mean effect of irrigation under interactive SST being to decrease tropical precipitation north of the Equator and increase it south of the Equator. The summer monsoon precipitation over India is reduced under both fixed and interactive SST, but with interactive SST, impacts of irrigation on summer precipitation appear to also be more widespread across southeast Asia (Fig. ). Latent heat impacts (Fig. ) reflect both increased evapotranspiration where there is irrigation and the impacts of nonlocal changes in temperature and precipitation, e.g., less evaporation over western Australia in austral summer associated with reduced precipitation there due to irrigation under interactive SSTs.

One measure of the increased modeled global impacts of irrigation under interactive SST is provided by the fraction of the global area with significant (p<0.05) change in the annual mean of each climate variable due to irrigation with fixed SST (ΔA) versus interactive SST (ΔO). This area fraction increases substantially with interactive SST for most of the variables discussed here, for example more than doubling (from 21 to 46 %) for surface air temperature and almost doubling (from 15 to 27 %) for precipitation.