Influence of irrigation on land hydrological processes over California

In this study, a regional climate model (RCM) is employed to investigate the effect of irrigation on hydrology over California through implementing a “realistic irrigation” scheme. Our results indicate that the RCM with a realistic irrigation scheme commonly practiced in California can capture the soil moisture and evapotranspiration (ET) variation very well in comparison with the available in situ and remote sensing data. The RCM results show significant improvement in comparison with those outputs from the default run and the commonly used runs with fixed soil moisture at field capacity. Furthermore, the model reproduces the observed decreasing trends of the reference ET (i.e., ET0) from the California Irrigation Management Information System (CIMIS). The observed decreasing trend is most likely due to the decreasing trend of downward solar radiation shown by models and CIMIS observations. This issue is fundamental in projecting future irrigation water demand. The deep soil percolation rate changes depending on the irrigation method and irrigation duration. Finally, the model results show that precipitation change due to irrigation in California is relatively small in amount and mainly occurs along the midlatitudes in the western United States.


Introduction
Anthropogenic land cover and land use changes have dramatically altered the global land surface, especially with regard to the conversion of agriculture and pastures from natural lands [Kueppers and Snyder, 2012;Ramankutty et al., 2008]. This conversion modifies the surface dynamical and thermodynamic features, such as roughness, albedo, and vegetation type and coverage, and affects the exchanges of water and energy between land surface and atmosphere and between land surface, soils, and aquifers. Agricultural practice processes, such as planting, fertilizing, irrigating, and harvesting, may also impact soilland-atmospheric exchange processes in both biophysical and biogeochemical ways. Therefore, many numerical models, including global climate models (GCMs) and regional climate models (RCMs), have been employed to simulate these processes and the impacts of land conversions and land management practices on global/regional climate [e.g., Avila et al., 2012;de Noblet-Ducoudré et al., 2012] (also, see reviews of Mahmood et al. [2010] and Pielke et al. [2011]). However, few studies have been reported to investigate the impacts on hydrological processes using RCMs and/or GCMs, although irrigation may result in an increase in soil moisture, deep-layer percolation (or recharge), as well as the modification of surface runoff and river streamflow peaking, etc.
The impact of agricultural irrigation in California's Central Valley on the local/regional climate has been well reported based on numerical models, in situ observations, and remote sensing retrievals [Adoegoke et al., 2003;Bonfils and Duffy, 2007;Christy et al., 2006;Kanamaru and Kanamitsu, 2008;Kueppers and Snyder, 2012;Kueppers et al., 2007Kueppers et al., , 2008Lobell and Bonfils, 2008;Lobell et al., 2009;Ozdogan et al., 2010;Sacks et al., 2009;Sorooshian et al., 2011Sorooshian et al., , 2012, and many others]. The results from models unanimously indicate that irrigation results in an increase in surface evapotranspiration (ET) and a decrease in low-level air temperature, although there are differences in amounts from different models. In order to mimic irrigation, the model root zone soil moisture in most of the numerical models is fixed at field capacity or even saturation in each integration time step. For example, Kueppers and Snyder [2012] investigated the surface flux (sensible, latent heat flux) variations from diurnal to seasonal scales through fixing the soil root zone moisture at field capacity at each time step. They argued that the surface ET and air temperature changes are very small when setting soil moisture to any values larger than 25% of the field capacity. Lobell et al. [2009] set the root zone soil moisture at saturation at each time step when they conducted the same study; they also concluded SOROOSHIAN ET AL. that the irrigation-induced surface air temperature is insensitive to soil moisture when soil moisture is above 30% of saturation. At the very least, this model configuration [e.g., Kueppers and Snyder, 2012;Lobell et al., 2009] results in improper description in soil hydrological processes, such as overestimating soil moisture variation and deep soil-layer percolation, which will be one of the items addressed in this manuscript.
Other investigators have made efforts to reasonably represent irrigation processes in the numerical models or to configure the proper amounts of water in the model root zone layers in order to improve model simulations. For example, Kanamaru and Kanamitsu [2008] found that setting up the root zone soil moisture at 50% of field capacity can reduce the model cool biases in a regional climate model (regional spectral model in their study). Ozdogan et al. [2010] conducted similar studies in an off-line land surface model (the Noah land surface model in their study), and their results (e.g., ET) are comparable with in situ observations. Sorooshian et al. [2011Sorooshian et al. [ , 2012 implemented the recommended irrigation method suggested by Hanson et al. [2004], hereafter "realistic irrigation," into RCMs (MM5 and WRF). The RCM results (including ET, surface temperature, humidity, and wind) match well the in situ and remote sensing observations in different time scales (diurnal, annual, and interannual). However, irrigation-induced changes in the subsurface were not addressed.
It should be pointed out that similar to meteorologists' use of GCMs/RCMs to investigate the effects of irrigation on weather and climate, hydrologists have been using hydrological models to investigate the effects of agricultural irrigation on basin streamflow and water resources. For example, Haddeland et al. [2006] investigated the possible effect of irrigation in the Colorado River basin on streamflow, with the soil moisture in the VIC model set at fixed field capacity at each time step. Pokhrel et al. [2012] incorporated irrigation processes into a land surface model, in which a simplified TOPMODEL is coupled, to address water consumption and supply at the global scale. In their study, the horizontal resolution is 1°, which is relatively too coarse to represent the irrigation spatial features in the relatively narrow Central Valley and Imperial Valley of California. Additionally, the model soil moisture is fixed at 70% of field capacity at nonrice paddy irrigation regions without considering the differences of crop types to soil water stress. Specific to the Central Valley, 70% of field capacity still may be too high for most crops in comparison with field experiment results reported by Hanson et al. [2004]. Hanson et al. [2012] considered Central Valley irrigation water amounts based on water usage and supply in the commonly used groundwater flow model MODFLOW. The temporal and spatial scales used in their study were biweekly time scale and a~0.2 km spatial resolution. In MODFLOW, the surface processes, including energy and mass processes, are treated in a relatively simple manner in comparison with grid point and physical-based land surface models (LSMs).
The investigation of irrigation-induced soil hydrology (i.e., water below the soil surface) is still necessary, both in understanding the physical processes such as water redistribution and variability, etc., and in examining physical model performances (see reviews of Mahmood et al. [2010] and Pielke et al. [2011]). In Sorooshian et al. [2011], an irrigation method is incorporated into the RCM, and the model performance was examined against available climate data with good results. In the second study [Sorooshian et al., 2012], the effect of irrigation on ET was examined against remote sensing observations. In this third paper in the sequence, we focus on studying the effects of irrigation on land hydrological processes (e.g., surface runoff, groundwater recharge, and soil moisture) using RCM simulations, available observations, and off-line land surface model outputs from the North American Land Data Assimilation System (NLDAS) Phase 2.

Model Configuration and Data Collections
The mesoscale model National Center for Atmospheric Research (NCAR)/PENN STATE MM5, which has been used to study similar topics [e.g., Segal et al., 1998;Kueppers et al., 2008;Sorooshian et al., 2011Sorooshian et al., , 2012, is employed as the integration model. Since MM5 provides multiple choices of model physics schemes and model parameters, it is important to ensure that the selections are appropriate for the region of application. For this reason, we used sensitivity tests from the summer of 2007 to select the physics schemes and parameters for long-term runs at 18 km spatial resolution. surface runoff is based on a simple water balance model, the excess precipitation not infiltrated into the subsurface [Schaake et al., 1996]. The total evapotranspiration (ET) includes evaporation from the bare soil and canopy as well as transpiration from vegetation. The deep-layer percolation is defined as the Noah LSM [Chen and Dudhia, 2001].
For this study, three types of model runs are conducted to be forced by National Centers for Environmental Prediction reanalysis-I data. Run 1 is the "control" run (hereafter MM5C), which is the normal MM5 simulation run without any modifications. Run 2 is called the "field capacity" run (hereafter MM5F), where the MM5 root zone soil moisture is set to near-field capacity (i.e., 0.90% of field capacity in this study) at each time step. The conditions created for Run 2 are similar to those mentioned in some previous studies [e.g., Haddeland et al., 2006;Kueppers and Snyder, 2012;Kueppers et al., 2007Kueppers et al., , 2008Lobell et al., 2009]. Run 3 is called the "realistic" run (hereafter MM5R). For this case, and as was done in our previously reported studies, the soil moisture conditions are set up based on a number of factors, including the realistic irrigation of Hanson et al. [2004] and closely practiced by irrigators. In our MM5R run, irrigation water is applied when the following three conditions are satisfied: (1) the root zone's relative available soil water (SW) content is less than the maximum allowable soil water depletion (SW m ), where the magnitude of SW m depends on the crop type and growth periods as discussed in Hanson et al. [2004], (2) when downward solar radiation is less than 50 Wm À2 for the purpose of model numerical stability during the simulations and for avoiding water loss due to ET at high solar radiation levels, and (3) soil temperature is greater than 10°C for the long-term run to avoid irrigating if the soil is frozen. Irrigation ceases when soil moisture reaches field capacity. In this paper, we mainly discuss variations of modeled hydrological variables, such as evapotranspiration (ET), soil moisture, surface and subsurface runoff, and percolation (recharge). The simulation studies are conducted at different time scales, and the results are compared with available observations and off-line land surface model outputs. The referenced data include the following: 1. NLDAS data. In brief, the NLDAS database provides model-generated hydrometeorological information from a number of available models (Noah, VIC, the Sacramento Soil Moisture Accounting, and Mosaic). For this study, ET, precipitation, gridded-surface runoff, soil moisture, and deep-layer percolation data generated from the Noah LSM in NLDAS version 2 [Xia et al., 2012a] are used as reference data. The data are hourly and 0.125°horizontal resolution and are available since 1980 from the website (http://disc.sci. gsfc.nasa.gov/hydrology/data-holdings) (hereafter NLDAS). In NLDAS, precipitation is downscaled to hourly time step using daily, 0.25°spatial resolutions from the Climate Prediction Center (CPC) based on diurnal features of Stage II radar data. The other forcing data used in Noah-NLDAS are mostly downscaled from the North American Regional Reanalysis data set. The Noah LSM in the NLDAS version 2 has a few modifications to improve model performances [Xia et al., 2012a;Wei et al., 2012] [Tang et al., 2009]. The data are monthly at 0.05°spatial resolution and cover the whole U.S. from 2001 to 2008 (hereafter MODW). Another ET data set, generated by the Numerical Terradynamics Simulation Group at the University of Montana, is also based on MODIS data and is a 1 km sinusoidal projection which covers the years from 2000 to 2011 on a global basis (http://ntsg.umt.edu/project/mod16) [Mu et al., 2011] (hereafter MODM). 3. Advanced Microwave Scanning Radiometer-EOS (AMSR-E) soil moisture. Surface soil moisture data generated from AMSR-E observations at daily and 0.25°are used in this study. The data set is provided by Owe et al. [2008]. The retrieved method for this data set uses a forward modeling optimization procedure to solve a radiative transfer equation for both soil moisture and vegetation optical depth [Owe et al., 2008]. 4. In situ data. In addition to gridded data, the following in situ observations are also used in this study: [2010] and Johnson et al. [2002Johnson et al. [ , 2004Johnson et al. [ , 2005. Therefore, the data quality should be reliable for this study. We should note that a few additional in situ measurement programs are also ongoing, but their data had not been released for public use at the time of this publication.

Precipitation
Precipitation is the key component in hydrological processes, and several studies have focused on changes in its trends and patterns [e.g., Bell et al., 2004;Duffy et al., 2006;Gershunov et al., 2000;Hao et al., 2013;Leung and Qian, 2009;Damberg and AghaKouchak, 2014]. Figure 1 shows the precipitation based on MM5C simulations and NLDAS forcing for June-September (i.e., the main irrigation months) from 1981 to 2007. As seen, summer precipitation is relatively small (primarily 0-20 mm/month), as confirmed by both data sets. However, in MM5C, precipitation is larger in midlatitude and mountainous regions than that in NLDAS forcing. Figure 1 also indicates that the model overestimates precipitation over most areas, except for the lee slope (eastward slope), where the model underestimates precipitation, and in the Tulare Basin in the southern Central Valley, where model precipitation and NLDAS precipitation are in good agreement.
Over the San Joaquin and Sacramento basins, precipitation is overestimated by 10-30 mm annually. However, in comparison with NLDAS precipitation, the MM5 overestimates by 50-80 mm annually, even more so in mountainous regions. NLDAS precipitation is originally based on 0.25°daily precipitation from the CPC. Some studies indicated that CPC precipitation data have dry biases, especially over the mountains [e.g., Crow et al., 2009].  Figure 2) and NLDAS (line NLDAS-Y in Figure 2). In this figure, precipitation is averaged over the irrigated region in California from MM5 (without an irrigation scheme). In comparison with NLDAS forcing precipitation, MM5 overestimates the annual precipitation by 30-50% (e.g., in 1983, 1984, 1989, 1995-1998, 2000, 2003, and 2006). Note that the precipitation variation of the two data sets indicates that the variability is mostly in phase with amplitude differences. Although some other hydrological variables, such as soil moisture and ET, are highly relevant to precipitation, our previous tests indicate that soil memory due to irrigation can last a maximum of about 1 month at the regional scale over the Central Valley [Sorooshian et al., 2011]. This suggests that model precipitation biases in the cold season may not influence summer soil hydrological processes in the current model configuration. In Figure 2, MM5-W and NLDAS-W represent the accumulated annual precipitation time series averaged at irrigation grids over California from May to September. We note that the 5 month precipitation amounts are small (<60 mm for most of the years), averaging about 8.5 mm per month for MM5 and 4.8 mm per month for NLDAS forcing. The warm season precipitation feature also favors the analysis of modeled irrigation hydrological processes in the same seasons. In other words, because precipitation bias-induced soil moisture and ET variation can be minimized in the model system, we can address irrigation-induced hydrological processes with more confidence.
[2012] examined the ET variation from diurnal to interannual time scales and concluded that using a realistic irrigation method, model ET matches available remote sensing observations better than the control run, as well as the irrigation run when fixing the model root zone soil moisture at field capacity at each integration step. Here we extend the previous studies in the following aspects: (1) comparing model ET with the newly released NLDAS ET and in situ ET observations and (2) examining the ET long-term trend based on the model results and CIMIS reference ET.  [Tang et al., 2009]; MODM refers to the remote sensing ET data generated by the Numerical Terradynamics Simulation Group at the University of Montana [Mu et al., 2011]. Although there are some differences in magnitudes, the ET spatial patterns are very consistent, indicating higher amounts along the coast and Sierra Nevada Mountains and lower amounts over the Great Basins and Death Valley. There are also high amounts along the irrigated agricultural areas over the Central Valley, the Imperial Valley, and the southern Idaho Valley. It should be noted that in MODM, ET over urban, water, and barren areas is masked.
The NLDAS ET shown in Figure 3 is based on the 2012 NLDAS version 2 Noah LSM [Xia et al., 2012a[Xia et al., , 2012bWei et al., 2012]. This version includes crop root distribution, leaf area index seasonal variations, etc., which theoretically should improve ET estimates. However, the ET from the Noah LSM/NLDAS-2 still cannot capture the ET pattern over irrigation areas as observed by remote sensing data, especially in the Central Valley. On the other hand, RCM results with the realistic irrigation scheme can capture the spatial patterns similar to the ones captured by remote sensing data. Figure 3 also indicates that irrigation-induced ET varies in space and from one location to another, depending on soil features and weather conditions. In comparison with available in situ ET at the two sites (see Figure 1), RCM ET with the realistic irrigation scheme captures the peaks and the annual variation very well. MODIS ET (here MODW and MODM) and NLDAS-2/Noah, as well as MM5C, underestimate the ET relative to in situ observations. At Five Points, MM5R and MM5F are in very good agreement, while MM5F overestimates at the KREC site. Furthermore, both NLDAS Noah LSM outputs and MM5C ET indicate a phase shift because no water is added into soil, thereby resulting in drier soil during the dry seasons. It should be noted that the observed site at KREC was planted with peach trees starting in 1988, with the trees maturing in 1993. It is understandable that there is some difference between modeled ET and observed ET because the real water stress and water applications vary from year to year as the trees grow, while in the RCM irrigation mode the averaged monthly water stress of crops is used, which remains constant through the study years [Sorooshian et al., 2011].  When fixing Noah LSM root zone layers at saturation, the Noah LSM ET matches CIMIS ET0 very well (not shown for brevity). This result is inconsistent with some of the previous studies, such as Kueppers and Snyder [2012] and Lobell et al. [2009], who concluded that when the amount of soil water is larger than 25% of field capacity or larger than 30% of saturation, the ET or near-surface air temperature is insensitive to soil water content. This difference between our results and the examples shown comes mainly from differences in the land surface parameterizations that are integrated into the climate model. Each LSM has its own parameterization scheme in estimating land-atmosphere water exchange coefficient, soil thermal and hydrologic conductivity, etc.
3. Both the model ET and CIMIS ET indicate a clear, long-term linear decreasing trend at annual mean as well as the average for July and August; furthermore, the more irrigation water that is applied, the more apparent the decreasing trend. However, both MM5C and NLDAS ET did not show these trends. As shown in Figure 2, both NLDAS precipitation and MM5 precipitation at the irrigation grid did not show any clear trends, either in annual mean or in the warm season accumulation. To identify a possible reason for this, we checked the model incoming radiation differences between irrigation runs and the control run. The results indicate that irrigation causes a decrease in downward long-wave radiation and appears to be statistically significant during the high-irrigation season (i.e., July and August) (see Figure 5d). Specifically, MM5R indicates an approximate 3.6 Wm À2 decrease in the downward long-wave radiation in comparison with MM5C, while MM5F shows a decrease of 5.51 Wm À2 relative to MM5C. A linear trend analysis indicates that MM5 model solar radiation decreased by about 3.7 Wm À2 per decade (Figure 5c), while downward long-wave radiation decreased by about 1.0 Wm À2 per decade (Figure 5d). We then evaluated the precipitation variation at the CIMIS sites, which is fairly close to the values of the nearest NLDAS grid cell. No long-term linear trends were detected (not shown for brevity). We also investigated the 2 m soil temperature, downward solar radiation, and humidity in some CIMIS sites. We noticed that the downward solar radiation shows a clear decreasing trend (about 4.0 Wm À2 per decade from 1984 to 2007), which is consistent with the model results.
The annual 2 m temperature and mixing ratio time series for July and August from MM5 is also exhibited in Figures 5e and 5f. In that figure, a decreasing trend for the 2 m temperature is shown, supporting the notion Based on the above findings, we conclude that the irrigation-caused ET exhibits a decreasing trend, most likely due to decreasing solar radiation, downward long-wave radiation, and 2 m air temperature. Decreasing solar radiation and downward long-wave radiation may have caused a decrease in net radiation, which can impact the redistributions of net radiation between sensible and latent heat fluxes since the albedo and emissivity in the Noah LSM are irrigation independent.
Recently, Pu and Dickinson [2012] reported a similar phenomenon, namely, that ET has been experiencing a decreased trend due to vegetation feedback owing to increases in the greenhouse effect.
4. NLDAS ET is smaller than MM5C ET mainly due to the fact that MM5 overestimates precipitation, as indicated in the previous section. However, the variations of the two ET data sets are consistently in phase and out phase with the ET when irrigation is added at the interannual scale.

Soil Moisture
We first evaluated the remotely sensed soil moisture data and noticed that their quality is quite low, both in terms of resolution and missing data. Figure 6 displays an example of AMSR-E topsoil wetness mean for  Figure 6 also shows that the irrigation signature is reasonably captured where irrigation areas are large enough to be resolved in current AMSR-E 25 km resolution (e.g., over the southern San Joaquin Valley and the Tulare Basin). The soil wetness over the valleys displayed in Figure 6 is about 3%-20%. Based on the definition of soil wetness (i.e., the ratio of soil moisture to porosity), one can estimate that if soil is saturated, the wetness is 100%, and if soil moisture is at the field capacity level, the wetness should be about 60%-85%, depending on soil texture. The wetness over the Central Valley indicated in Figure 6 is far below the field capacity. Therefore, we can safely state that the model root zone soil moisture reaching 70% of field capacity overstates the irrigation impact on hydrological processes over the Central Valley. Over other areas where the irrigation areal extent is smaller than the AMSR-E (~25 km grid) resolution, AMSR-E cannot be used to identify the irrigation-induced soil wetness variation.
Currently, the availability of in situ soil moisture data is a limiting factor for validation and model-testing purposes. We obtained one site where the soil moisture was measured at an irrigated peach tree field (i.e., at the KREC in Parlier, California). Figure    Comparing the soil moisture of the MM5 control run with NLDAS, one can see that soil water content estimates from NLDAS/Noah LSM are slightly lower (À1.8 mm on average) than the MM5 control run (MM5C) in the top layer and much lower (À33.8 mm on average) in the bottom layer. This is because (1) MM5 generates more rainfall than NLDAS forcing during the same periods (see Figures 1 and 2) and (2) in NLDAS/Noah LSM, the upper threshold of soil moisture at which the vegetation reacts to a soil moisture deficit has been adjusted [Xia et al., 2012a;Wei et al., 2012], compared to the Noah LSM in MM5. However, soil water contents from MM5C and NLDAS/Noah have the same interannual variations. Figure 8 indicates that MM5C has basically reproduced the soil water content variations as shown in the NLDAS-2 Noah LSM output with observation/analysis forcings, demonstrating that our model outputs are reliable.

Surface Runoff
In this study, we focus mainly on grid cell hydrology at irrigation areas in California. It should be noted that water in California is highly regulated by dams and multilevel (local, state, and federal) water transportation projects across the state, which makes it difficult to simulate the horizontal flow of water at basin scale. From June to September (JJAS) each year, rainfall precipitates primarily over the Serra Nevada Mountains and the elevated Northern California regions, with very limited rainfall over the irrigation areas (see outputs of   Figures 1-3). Therefore, surface runoff occurs mainly over the mountains and the higherelevation areas in Northern California.
In JJAS, and based on MM5C, MM5R, MM5F, and NLDAS/Noah outputs, the surface-gridded runoff over the irrigation areas (Central Valley and Imperial Valley) is less than 1 mm per month (not shown). Figure 9 exhibits the accumulated "gridded" surface runoff for MM5C/R/F over irrigation areas in California during the water year. In comparison with amounts of ET and precipitation presented earlier, surface runoff is a relatively small component of the water cycle during the summer, especially in the irrigation season (JJAS).

Deep-Layer Percolation and Water Balance
In the following section, we discuss the deep-layer percolation component of the soil hydrology over the study area. The amounts of the deep-layer percolation are the outputs of MM5-coupled land-vegetation (crop) atmospheric model, the NLDAS off-line simulations, and the results from previously published studies.
In the control and irrigation runs, grid cell soil moisture variation at a specific time interval ideally can be written as follows: where m is soil water content, P is precipitation, I is irrigated water, E is ET, R is surface runoff, and D is deeplayer recharge during the time period. In equations (1) and (2), the subscripts c and i represent control run and irrigation run, respectively.
From Figures 1, 2, and 11, one can conclude that the irrigation-induced precipitation and surface runoff are relatively small components of the water cycle (see Figure 10). The long-term soil moisture changes are not significant in arid/semiarid areas because the soil moisture depletes rapidly through evaporation and reaches the wilting point in nonirrigation areas, while the water from the irrigation areas is continuously added to keep the soil water content at a certain level. Therefore, the total irrigation water can be approximated based on the differences of ET and recharge in the irrigation and control runs. Over the entire California irrigation areas, the annual percolation rate is about 18% based on MM5R and 28% based on MM5F, respectively. These estimates indicate that with a realistic irrigation scheme incorporated into MM5R, the percolation rate due to irrigation is substantially less than the commonly used irrigation at field capacity assumed in MM5F. The obvious question is which is closer to in situ and more trusted values.
Unfortunately, consistent and long-term field observations of the percolation rates are not available. However, the closest we can get to in situ observations are those estimates provided by Wichelns and Nelson [1989].  in 1986, 20% in 1987, and 18% in 1988. These values are very close to the model results from MM5R. The California State Water Resources Control Board (1992) reported the relationship of the irrigated water and deep drainage. The average percolation rates of 1989 and 1990 are 14% for subsurface drip irrigation (similar to our MM5R setup), 26% for low-energy precision application, and 30%-64% for furrow irrigation, which means that the percolation rate depends highly on the irrigation methods as well. Nevertheless, based on these limited reported observation estimates, we feel comfortable in suggesting that the MM5R-generated percolation rate is close to the field measurement for most of the experiments.  In our 2011 work [Sorooshian et al., 2011], we analyzed the changes in meteorological variables at the locations more than 10 km in distances from irrigation areas relative to observations. Our results showed that the absolute differences between irrigation and nonirrigation runs were much smaller than those between irrigation/nonirrigation and observations. We concluded that irrigation-induced climatic change in Central Valley, California, is very small at the nonirrigation areas a few tens of kilometers away from the irrigation regions.
In this study, we analyze the effect of irrigation on precipitation in warm seasons (June-September) for the period of 1981-2007 in California and Southern Idaho. Figure 10 shows the changes in precipitation frequencies at different model resolutions due to irrigation (irrigation areas are marked with gray). In Figure 10, the monthly accumulated precipitation differences between the irrigation and control runs (À0.3 to 0.3 mm) is masked. The results show that while adding more water in the irrigation areas may generate slightly more precipitation outside the irrigation region, precipitation occurs mainly in midlatitude regions, such as Oregon, Idaho, Wyoming, and South Dakota, with no significant rainfall differences detectable in the southwestern United States. We have also used precipitation differences at 3 mm per month as a threshold to examine the effects of irrigation on precipitation and found no rainfall differences between the irrigation runs and the control run. In other words, our model-based results show that increases in irrigation-induced precipitation could be in the range of 0.01-0.1 mm d À1 , which indicates that the precipitation change due to irrigation is less than 1% during the warm season.
Recently, there have been a number of reported studies presenting model simulation results regarding the fate and the extent of the influence of ET water from California's Central Valley irrigation region. Using version 3.5 of the NCAR Community Atmosphere Model, combined with version 3.5 of the Community Land Model and prescribing irrigation water based on the work of Wisser et al. [2008] in the Central Valley irrigation areas at the resolution of T85 (approximately 1.4°× 1.4°), Lo and Famiglietti [2013] reported that precipitation in the southwestern U.S. increases by 15% in the summer (June-August) and streamflow in the Colorado River increases by about 30% because of Central Valley irrigation. In another study, Wei et al. [2013] used the Modern-Era Retrospective Analysis for Research and Applications system and the quasiisentropic back-trajectory method Brubaker, 1999, 2007] to trace the water vapor for each precipitation event backward in time. The tracing was done along the isentropic surface starting from the grid box with precipitation backward in space and time until all of the original precipitation is attributed to ET. They found that California irrigation-induced precipitation occurs mainly from the Sierra Nevada Mountains eastward along the midlatitude and with less than 1% of total precipitation in most of the areas (see their Figures 6 and 10). Our results, which are close to those of Wei et al. [2013], are different (in both spatial distribution and amount) than those recently reported by Lo and Famiglietti [2013]. The differences could be due to a number of factors, such as model resolution, model physics, parameterization, land use/land cover representation, and boundary conditions. Figure 11 presents the total column water vapor flux average from the control run in JJAS. This circulation pattern favors the scenario that the evaporated water (from irrigation regions) was transported eastward and into midlatitudes due to wind patterns.

Transferability of the Proposed Irrigation Scheme
Based on the comments from reviewers, we have implemented the same irrigation scheme into the Weather Research and Forecast (WRF) model (version 3.3.1), and the results are similar. As an example, Figure 12 shows the simulated average 2 m temperature changes due to the irrigation at different times for June-August

Conclusions
The impacts of agricultural irrigation on hydrological processes over California have been investigated using the regional climate model (RCM), an off-line land surface model output, and available in situ observation and remote sensing data. Our results indicate that by implementing a realistic irrigation scheme, one can reasonably capture the irrigation-induced ET, soil moisture, and percolation variations and reproduce observed patterns and previous field reports [e.g., Wichelns and   Nelson, 1989]. Without consideration of the irrigation processes in physically based models, neither the MM5 control run nor the improved Noah LSM off-line run (e.g., NLDAS) can capture the irrigation-induced hydrological patterns. On the other extreme, to fix the model root zone soil moisture as field capacity or saturated at each time step, the model overestimates evapotranspiration and deep-layer percolation, which is inconsistent with available field measurement reports and remote sensing observations. The realistic irrigation method adapted in our study can be used to more accurately estimate the influence of irrigation water applications over California, either in the hydrological model or in the water resources management model.
Model results also indicate that irrigation-induced ET exhibits long-term linear decreasing trends. The RCM ET long-term linear trend is consistent with the variation of the CIMIS reference ET, which is estimated from the in situ field measurements. The reason for this decreasing trend is the observed downward short-wave radiation trend, which consequently impacts the net radiation and its redistributions between sensible heat fluxes and latent heat fluxes. The results show that the irrigation-induced precipitation variation outside of the irrigation area is very small (0.01-0.1 mm d À1 ) and occurs mainly along the midlatitude areas.
Our estimated values are consistent with reported magnitudes from more realistic observation-based studies Brubaker, 1999, 2007;Wei et al., 2013]. Our results also show significant inconsistency with the recently reported model simulation studies of Lo and Famiglietti [2013] in both spatial distribution and amount.
Although a model with a realistic irrigation scheme can basically capture the hydrological features observed, the model deficiencies are also apparent. Despite significant progress in the past decades, regional model simulations of precipitation have substantial biases that could affect irrigation water demand. For example, a wet bias in precipitation leads to less irrigation water demand than what is actually needed and vice versa.
In order to fully and realistically estimate the amounts of ET, soil moisture, and deep-layer percolation, the resolution of the modeling application must be at the field scale where the variations in crop types and their respective water stress differences can be captured. At the same time, more ground-based observations are needed for model validation and verification purposes. This will greatly improve model performance evaluations and allow for more effective ways of improving understanding of land hydrological processes in irrigated agricultural regions.