Drivers of enhanced evaporative demand in U.S. croplands: Determining relative contribution using constrained input scenarios

Altered evaporative demand is a global phenomenon observed over recent decades, however, such change has not been attributed explicitly to specific meteorological drivers, hampering consensus on what has caused such change. Here we investigate exactly how much individual drivers have contributed to long‐term grass‐reference evapotranspiration (ETo) change within conterminous United States (CONUS), with an emphasis on agricultural croplands. Using scenarios that constrain individual drivers i.e., air temperatures (T), relative humidity (RH), solar radiation (Rs), and wind speeds (U2) to their climatologies, we determined their relative contribution toward ETo change at monthly and annual scales. Annual ETo increased by 111 mm, or >2 standard deviations (SD) relative to the 1981–2000 baseline, accompanied by strong increase in Rs (2.7 SD), U2 (2.5 SD), T (1.1 SD), and decreased RH (2.3 SD) in regions that account for one‐third of calories produced in the U.S. Annual ETo increase was attributed primarily to T (relative contribution of 36%), followed by Rs (29%), U2 (18%), and RH (17%) with significant spatial and seasonal variability. During agriculturally critical summer months, Rs was the dominant driver with a 40%–50% relative contribution, and other three drivers were roughly equally important. These findings address demand‐side of agricultural water use and imply long‐term change in crop functions and performance, water security, and planning across aridity gradients.


| INTRODUC TI ON
Moisture availability for terrestrial vegetation is governed by a combination of water supply, water use, runoff, and soil characteristics.Among these elements, water use (evapotranspiration or ET) has been recognized to be the most complex to understand, estimate, and interpret.
Progress made since the 1940s has revealed that ET is a coupled process encapsulating surface processes and the overlying atmosphere (Penman, 1948).To effectively decompose ET process, accurate determination of atmospheric evaporative demand is integral and significant effort has been invested by researchers to investigate, conceptualize, parameterize, and standardize operational frameworks for its estimation.Evaporative demand is driven by both radiative energy and aerodynamic factors, and thus, is well represented by combination-based models.McMahon et al. (2016) compiled 38 combination-based ET models and showed that substantial diversity exists in model definitions, structure, inputs, and applications.These models developed as early as 1948 (Penman, 1948) to as recent as 2015 (Tegos et al., 2015;Valiantzas, 2015) allow estimation of potential, reference crop, open water, deep lake, and pan evaporation from near-surface standard meteorological data.For agricultural applications, evaporative demand is predominantly determined using the American Society of Civil Engineers-Environmental & Water Resources Institute (ASCE-EWRI) Standardized Penman-Monteith (ASCE-PM hereafter) reference ET model (Allen et al., 2005).The ASCE-PM model relies on near surface air temperatures, relative humidity, solar radiation, and wind speed to calculate daily or hourly reference ET for a short or tall reference surface.Changes in any of these fundamental weather parameters will translate into altered demand for water in agroecosystems, given absence of biological feedback.Within this framework, the biological feedback from the soil and plant system is encapsulated in a crop coefficient term for operational simplicity for practitioners.Thus, within this widely recognized framework intended for agricultural practitioners and managers, the demand side of the evaporation process is solely captured by the reference ET term.Altered reference ET can cause complex changes in agricultural and hydrological systems, which are typically dependent on the aridity at the site of interest.For instance, in humid regions, increased grass-reference evapotranspiration (ET o ) will decrease surface water resources due to greater evaporation from open water, greater soil evaporation and plant transpiration since no water availability constraints exist (Friedrich et al., 2018).Greater ET generally reduces soil moisture (Manning et al., 2018;Teuling et al., 2013), further penetrating other components of the hydrological cycle such as reduced streamflow (Vicente-Serrano et al., 2014), especially under drought conditions when the assumption of unconstrained water availability may not hold true.On the contrary, water-limited regions show an opposite behavior between ET o and ET as explained by complementary relationship (Bouchet, 1963).A higher ET o in water-limited regions will not proportionally translate into hydrological effects due to low water availability and small runoff production (Le Houérou, 1996), although surface water resources will generally decrease (Vicente-Serrano et al., 2019).Agricultural and environmental systems productivity in humid regions are usually limited by temperature and radiative energy under normal conditions, and thus, increased ET o is not expected to have negative effects on plant canopies.Rather, a moderate increase in certain constituent variables within the ET o framework such as temperature, vapor pressure deficit (VPD), and radiation may improve photosynthetic activity (Anthoni et al., 1999;Niu et al., 2008;White et al., 1999) with a corresponding increase in water use (Doorenbos & Pruitt, 1975), given unlimiting soil water content.Extremely high VPD and extreme (killing) degree days will negatively affect plant growth and development even in humid regions (Butler & Huybers, 2015;Grossiord et al., 2020;Kukal et al., 2023).In water-limited regions, ET o influence on environmental and agricultural systems is relatively more important and complex.A greater ET o in these regions may not result in a substantial increase in ET due to limited soil water availability (Manning et al., 2018).In fact, biological feedback of plants to soil moisture deficit and higher VPD by increasing stomatal resistance allows for conservation of soil moisture by suppression of transpiration (Ball et al., 1987;Konings et al., 2017;Miralles et al., 2019).Despite this strategy, a greater deficit between actual and potential ET (under unstressed conditions) than what can be tolerated by plants can result in biomass penalty (Singh, 1991;Zhang et al., 2004).All these projected impacts assume that precipitation is unaltered but could be much more complex when considering temporal precipitation variability.It is well established that ET o has been increasing largely across the globe during the last three to four decades (Vicente-Serrano et al., 2019) and that global agroecosystems are already subjected to these divergent impacts.
A primary question that arises is what factors have historically been responsible for altered evaporative demand.As noted, the available energy (radiative component) and drying power of the air (aerodynamic component) can both drive evaporative demand.Several studies have investigated these controls, concluding that aerodynamic controls were important in some regions (Ali et al., 2009;Liang et al., 2008;McVicar et al., 2012;Vautard et al., 2010;Wang et al., 2012) with other regions showing greater importance of radiative energy controls (Albano

Research Impact Statement
A third of U.S. caloric footprint area has undergone significantly increased evaporative demand driven by summer radiation, underscoring adaptation needs for agricultural water resources consumption. et al., 2022;Fan & Thomas, 2013;Linacre, 2004;Matsoukas et al., 2011;McCuen, 1974;Roderick & Farquhar, 2002;Saxton, 1975;Stanhill & Cohen, 2001).Relevant investigations demonstrate use of advantageous approaches but also have overlooked some important methodological aspects.First, because of regionally varying importance of controls in altering evaporative demand, it is critical that spatial heterogeneities are accounted for.A large proportion of the literature deals with ground-based stations, which can be subject to non-uniformity in site density, period of record, station relocation, sensor malfunction, and data gaps.These challenges have been partially overcome by the advent of gridded meteorological datasets based on reanalysis products.These datasets allow for investigating controls of evaporative demand at finer (a few km 2 ) scales, enabling visualization of regional behaviors, although this has been accomplished by only a handful of studies (Albano et al., 2022;Ficklin et al., 2015;Hobbins, 2016;Vicente-Serrano et al., 2019).Second, it is highly likely that various drivers of evaporative demand can emerge as more or less important during different times of the year.However, investigating seasonal evolution of drivers' roles has received none to limited emphasis in the literature.Third, a large proportion of studies globally have used Penman-like formulations (Doorenbos & Pruitt, 1975;Penman, 1963) to examine trends and sensitivity of evaporative demand (Piper, 1989;Saxton, 1975;Sheffield et al., 2012).The structure of these formulations is such that it can be analytically represented as two exclusive terms each corresponding to aerodynamic and radiative evaporation.However, the ASCE ET o formulation has been altered in structure and reparametrized to aid practitioners in effective implementation and cannot be viewed as two exclusive terms.For instance, air temperatures are usually relied on when computing both net radiation and VPD.Thus, inferring sensitivity of ASCE-PM ET o requires that air temperatures (T), relative humidity (RH), incoming shortwave radiation (R s ), and wind speed (U) are used as fundamental drivers to investigate their explicit signatures in ET o trends.Investigating the sensitivity of ASCE ET o becomes even more significant from an operational standpoint, as (1) T, R s , RH, and U are typically measured across public and private agricultural weather monitoring sites, and (2) ASCE-PM ET o is currently the most dominantly employed methodology for agricultural water use estimations in the United States (U.S.).Lastly, differing approaches have been used to infer how much T, R s , RH, and U have contributed to change in ET o .Within the body of relevant research, either empirical or analytical approaches have been employed.Among most of the empirically based studies, the focus is on observing the change in ET o by perturbing each of the drivers individually where the perturbation can be a fixed increment or based on historical variability in each driver (Gong et al., 2006;Irmak et al., 2006;Jerszurki & de Souza, 2019;Koudahe et al., 2018;Liu et al., 2010;Porter et al., 2012).Some researchers have quantified the sensitivity of ET o to interannual variability in drivers (Hobbins, 2016) by deriving analytical expressions using differentiation technique (Hupet & Vanclooster, 2001;Piper, 1989;Saxton, 1975).Others rely on variance-based approaches to quantify uncertainty in ET o and the factors responsible for it (DeJonge et al., 2015;Hobbins, 2016).Most of the literature relies on using day-to-day variability in ET o and its drivers to study sensitivity.While useful, this approach does not further our understanding of how "long-term change" in a driver or multiple drivers, if any, have resulted in ET o increase.
In this research, we address the question "how have ET o and its drivers (comparatively) changed during the last four decades, and how can ET o change be explicitly attributed to each of its drivers during the year".We do so by ensuring effective inference of these findings by agricultural

| Data retrieval and preprocessing
Daily records for daily minimum and maximum air temperatures, daily minimum and maximum relative humidity, surface downward shortwave radiation (R s ), and wind speed at 10 m (U 10 ) during 1981-2021 were obtained for the entire geographical extent of CONUS from gridMET dataset (Abatzoglou, 2013).The gridMET provides daily surface meteorological data at a spatial resolution of 1/24 degree (~4 km) for the United States and has been evaluated against ground-based stations for relevant use (Blankenau et al., 2020).The daily CONUS-wide rasters were converted to tabular data by averaging grid-level data within each county using spatial reducer tools in Google Earth Engine.Air temperature (T) data were available in units of Kelvins and were converted to degree Celsius, R s data were converted from original units of W/m 2 to MJ/ m 2 /day, and relative humidity (RH) and U 10 data were retained in original units of % and m/s, respectively.U 10 was converted to U 2 using the following relationship: For each county, time series of daily T, RH, R s , and U 2 were compiled, each consisting of 14,975 (days) datapoints for 1981-2021.

| Quantifying change in ET o and its drivers
Trends in annual mean and monthly mean ET o , T, RH, R s , and U 2 were assessed for each county for the 1981-2021 period following the nonparametric Theil-Sen slope estimator (Sen, 1968).A historic baseline was considered (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) to represent the relative magnitude of trends, and z scores were calculated for annual mean and monthly mean ET o , T, RH, R s , and U 2 based on this baseline.The resulting trends were reported as total change in the parameter (over 41 years) in units of standard deviations (SD) from the baseline.The statistical significance for each of the monotonic trends was tested using the Mann-Kendall trend test (Kendall, 1975;Mann, 1945).The trend analysis was accomplished using the zyp package (Bronaugh & Werner, 2019) in RStudio, which is an integrated development environment for working with R programming language.

| Sensitivity analysis to determine relative contribution of drivers to ET o change
To infer relative contribution of the trends in T, RH, R s , and U 2 toward change in ET o during 1981-2021, we used a practical and simple sensitivity analysis based on constraining the inputs.The analysis was conducted via the following consequent steps: a. We calculated county-level climatologies (41-year mean) for T, RH, R s , and U 2 for each Julian day (day 1 to day 366).e. Trends in annual and monthly ET o were analyzed using methods described in section 2.2, with the exception that absolute values (mm) of trends were used rather than z scores.These trends were quantified for both observed and four constrained time series.
f. Trends (mm/year) in each of the four constrained ET o series were subtracted from trend (mm/year) in observed ET o .These differences represent deviation from the observed ET o trend when either of the four ET o drivers was constrained to their climatologies.
g.The fractional relative contribution of each of the four ET o drivers was calculated as the ratio of difference terms calculated in (f) to sum of the four difference terms.This relative fractional contribution ranges from 0 to 1.The variable with the highest fractional contribution was the dominant driver of ET o trends.For example, relative contribution of T can be represented as follows: where  3) can be changed to any of the four drivers for which relative contribution has to be quantified. (

| Spatial distribution of calorific footprint and crop acreage
To present the trends and relative contributions in a manner that is tailored to crop footprint within the U.S., we calculated weighted mean values for CONUS that account for crop distribution heterogeneity.To achieve this, we used two datasets: (a) Total kilocalories produced (Figure S1a) raster (Cassidy et al., 2013)  4-019-4594-4#avail abili ty-of-data-mater ials.The kilocalories production dataset was resampled to county-level for consistency.

| How have ET o and its drivers changed at the annual scale?
The CONUS has undergone widespread increase in ET o during the last four decades, as shown at county-level as both absolute rate of change (mm/year) and number of SD relative to baseline (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) in Figure S2 and Figure 1a, respectively.Fifty-three percent of the counties had statistically significant increase in annual ET o during 1981-2021 that range between 0.8 SD (28 mm) to 5.2 SD (255 mm) relative to the 1981-2000 baseline.On average across counties that have undergone statistically significant (α = 0.05) change, a mean increase of 2.1 SD (111 mm) was observed relative to the 1981-2000 baseline.No significant negative trends were found in annual ET o , however, trends in parts of Northern Rockies and Plains, Upper Midwest, and the South were statistically nonsignificant (Figure 1a).
Overall, all drivers of ET o show increase in magnitude except RH, which decreased.Among all drivers, U 2 and R s showed the highest magnitude of significant change across most of the CONUS.U 2 and R s increased significantly for 89% and 87% of the counties, respectively.
Across these counties, U 2 increased by 2.6 SD and R s increased by 2.5 SD, on average.Change in U 2 ranged from 0.8-6.6.SD while R s change ranged from 0.8-4.2SD.Change in RH followed that of U 2 and R s , with an average decrease of 1.7 SD across counties with significant change.RH showed a decrease in 10% of counties in the West and Southwest U.S., whereas a small percentage of counties (2%) in the far Northeast showed increased RH.In areas where RH decreased, the range encountered was −1.1 to −4.3 SD.Lastly, the driver that showed the least magnitude of change was T which increased by 1.1 SD on average across 77% of CONUS counties which underwent significant change.No significant negative T trends were found, and increase ranged from 0.4 to 3.0 SD.Spatial distribution of trends in the four divers is quite heterogenous, with areas of significant change distinctly distributed in different regions (Figure 1b,e) and do not necessarily coincide with spatial distribution of ET o trends (Figure 1a).Representation of change in ET o and its drivers using a normalized metric, i.e., total change during 1981-2021 reported as SD based on the 1981-2000 baseline has not been accomplished in relevant literature.As shown here, such representation allows for understanding relative change in otherwise physically distinct quantities that are typically reported using absolute units (e.g., mm ).Such reporting has hindered effective interpretation of long-term change in these quantities in the past.shown for all counties as both absolute rate of change (mm/year) and number of SD relative to baseline (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) in Figure S3 and Figure 2, respectively.Spatial distribution of significant trends in monthly ET o is vastly different than what was found for annual time-step.On an average across counties with significant trends, ET o increased between 0.2 SD (February) and 1.7 SD (March), although county-specific increases as high as 3.9 SD were found during August.Except for February (small but significant decrease in ET o ), ET o increase remained within the 1.1 to 1.7 SD range through the year.However, the percentage of counties where ET o increase was recorded was higher in the summer and fall than spring.During December, as much as 64% of counties showed increasing ET o .It should be noted that although similar relative increase (in SD)

|
was found, increase in summer ET o will translate to greater depth units of water (mm) owing to seasonality of ET o .Similarly, 1 SD increase in each driver will have varying impacts toward changing ET o , so caution should be exercised for interpretation.
In a similar fashion, long-term change in monthly mean T, RH, and U 2 and total R s for all CONUS counties were calculated and presented in Figures S4-S7, respectively.T has increased across CONUS for all months of the year except February.Anywhere from 2% to 82% of counties showed statistically significant trends during the year.In general, T increased by 0.3 to 1.7 SD during the year, with greater change occurring during June until October.As observed for annual mean RH, both positive and negative change occurred in monthly mean RH.While negative change was more dominant (1%-30% of counties), these were mostly limited to the Northwest, West, and Southwest.Increase in RH was observed for parts of Northern Rockies and Plains, Upper Midwest, and the Southeast, especially during winter and spring.During the summer, CONUS-aggregated significant change in RH was negative (−1.2 to −1.9 SD).A strong tendency of solar brightening was found for most of the year with R s increase ranging from 0.4 to 2.1 SD.Both the intensity of increase as well as proportion of counties with significant change peaked in summer.A small proportion of counties also showed decreased R s (solar dimming) during the winter (December until February) and July.U 2 showed largely similar (1.2-1.7 SD) increasing trends throughout the year except December (0.7 SD).During all months, significant positive U 2 change for 10%-81% of the counties.The only decrease in U 2 was found in the Northwest and West during July.All drivers have a distinct seasonality; however, the results are presented so that the seasonality effect is eliminated for better inference of trends and their comparisons across months.
A considerable portion of area within CONUS does not show significant trends in ET o , T, RH, R s , and U 2 .To aggregate our findings in a representative manner for the CONUS, we only consider trends in drivers for counties corresponding to significant ET o increase (Table 1).In doing so, we found that R s showed the greatest trend (2.7 SD), closely followed by U 2 (2.5 SD) and RH (−2.3 SD), while T showed the least increase (1.1 SD) when averaged over regions with significant ET o change.More than 90% of the counties where ET o was significant had significant T, R s , and U 2 trends, but only 17% had significant decrease in RH.Changes in drivers maintained the same direction (positive for T, R s , and U 2 , negative for RH) throughout the year.When analyzed comparatively (SD change), the changes in drivers were mostly of the same order (1-2 SD).During period of greater relevance of change (summer), change in R s was the greatest followed by RH, U 2 , and T. Overall, in regions where ET o has risen during the study period, R s and T have seen the most and the least increase, respectively.U 2 and RH have increased and decreased, respectively in these regions with mostly similar magnitudes.While it may seem that drivers' importance in affecting long-term trends in ET o are consistent with their intensity of change, it may or may not be the case.Such assessment would entail a sensitivity analysis of ET o equation that not only accounts for driver-specific long-term trends but also how sensitive ET o model is to each driver and their interactions.The following section presents the results of county-specific assessments that will inform us on what is the relative contribution of each driver to long-term ET o change.

| CONUS-aggregated importance of ET o drivers
Relative contribution of each driver to the long-term trend in ET o was quantified for all individual counties in the U.S. When averaged across all counties, T had the largest relative contribution at more than one-third (36%), followed by R s (29%), U 2 (19%) and RH (16%).At the CONUS level, there was no singular predominantly important driver and each driver considerably contributed to ET o change, as shown by frequency distribution presented in Figure S8.To be able to evaluate driver importance while accounting for spatial distribution of harvested cropland within the CONUS, we weighted county-level relative contribution of each driver by harvested cropland acreage in each county.This resulted in a slight change in contributions, but the order of importance largely remained the same: T (35%), R s (26%), U 2 (21%), and RH (18%).To further interpret findings in the context of food security, we weighted county-level relative contribution of each driver by total kilocalories produced in each county.County-specific kilocalories produced included all possible end-uses for crops, i.e., direct human consumption, animal feed, and biofuel feedstocks, of which are affected by increased evaporative demand.The findings remained largely consistent with what was found when all data were averaged across CONUS: T (36%), R s (29%), RH (18%), and U 2 (17%).To better understand the exact degree to which each of the drivers contributes to ET o change, ranks (1st, 2nd, 3rd, and 4th) were assigned to each driver in order of their relative contribution in each county (Figure 4a-d).Across much of the CONUS, T is ranked either 1st or 2nd among all the drivers.Within western and central U.S., RH is ranked 2nd after dominant T, while R s is ranked last.In the Southeast and portions of Northern Plains, Upper Midwest, and Ohio Valley, R s is ranked 1st followed by T on 2nd rank.In the Northeast, R s is ranked at 2nd place after T. U 2 is ranked 2nd in much of the Northern and Central Plains and the South.There were some regional hotspots where one driver's  relative contribution was dominant (>60%) and others where quite the opposite was true where one driver showed extremely low (<10%) relative contribution, especially so for R s , RH, and U 2 .Nevertheless, in most of the counties, at least three drivers were substantially important for ET o change, based on the sum of their relative contributions.Although, counties may or may not show substantial difference between relative contributions from rank 1 and rank 2 drivers, a dominant driver (ranked 1st) in each county was identified to create a driver dominance map for CONUS (Figure 4e).T and R s together account for the dominant drivers for >90% of the counties, with 52% and 39% of the counties, respectively.RH and U 2 were found to dominate in only 6% and 4% of the counties.

| How does importance of ET o drivers change during the year?
Monthly-level relative contribution from each of the drivers can reveal within-season heterogeneity in driver importance that cannot be inferred from annual observations.Figure 5 shows CONUS-aggregated relative contribution of each driver as a weighted by total kilocalories produced across all CONUS counties (a) and counties with significant ET o change (b).T remained the dominant driver throughout the year except summer months (June-July-August) and November (Table S1).T was the most important indicator during winter (58% relative contribution) and gradually decreases to a minimum relative contribution (15%) during August, before returning to larger values in the fall.Importance of R s shows quite the opposite seasonal behavior to that of T, where the relative contribution of R s has a characteristic bell curve: lowest in winter and peaking in summer.Interestingly, R s relative contribution troughed at a negligibly low value (2%-3%) during winter, which demonstrates the minimal importance of winter R s to drive ET o change.This is contrasting to T, which maintained reasonable relative contribution (15%-30%) even during its annual minima in summer.R s was the dominant driver during the summer months (relative contribution of 40%-50%), with the other three drivers roughly being similarly important.RH did not have a clear seasonal trend as with T and R s and remained in the range of 15%-28% relative contribution during the year.November RH is an exception with the highest relative contribution (46%), while May, June, and July showed lower contribution (15%-16%).Importance of U 2 was equally low during winter and summer months (14%-20%) and peaked (25%-30%) in spring (February-April) and fall (October).
As seen with annual assessments, while CONUS-level inferences are useful, these can differ by different degrees from regional-level observations.The spatial patterns of relative contribution across CONUS can be inferred from driver-specific and month-specific maps given in Figures S9-S12.These relative contributions were also translated into and presented as county-specific ranks (Figures S13-S16) for each driver and dominant drivers (Figure 6).These maps can be employed to tease out local/regional level drivers and their relative ranks to drive long-term ET o change.For the most part, local inferences are consistent with CONUS-level findings, especially for regions with substantial share of calorie production/crop acreage.For example, Figure 7 reveals that the CONUS-level dominance of R s during summer is largely because of its importance in the eastern half of CONUS.

| DISCUSS ION
A key takeaway from this research is the unequivocal observations of increased evaporative demand across agriculturally important regions in the CONUS.We found that annual ET o increased by more than 2 SD on average from 1981 to 2000 baseline across U.S. croplands that account for roughly one-third of total agricultural acreage and total calorie production.Figure 7 presents the distribution of county-level annual ET o change within each major cropping region (belt) in the CONUS, represented as both relative change (Figure 7a) and absolute change (Figure 7b).While each crop region (Figure S17) is shown to be undergoing a considerably wide range of ET o change (due to wide spatial distribution), all trends are positive and often greater than 1.5 SD.These observations hold critical importance for agricultural sector's consumptive use as evaporative demand has likely translated into actual water use or ET, given sufficient moisture availability.A common approach for estimating crop water requirements is linear scaling of evaporative demand using a crop coefficient, assuming no biological feedback due to soil moisture limitations, CO 2 increase, or higher VPD.In the absence of these feedbacks, crop water requirements can be assumed to proportionally increase in the U.S. Several independent evidence supports this hypothesis, including reduced streamflow (Lehner et al., 2017) precipitation patterns (Zangvil et al., 2004), and greening trends (Zeng et al., 2016).For instance, ET o during the summer months of June, July, and August has increased by 1.7 SD on average in the U.S., although county-specific values range between 0.6 and 4.2 SD.In terms of absolute depth units, this translates to an additional 57 mm (range 15-70 mm) of crop water requirements on average across CONUS.In absence of water supply constraints, this 13% additional summer ET o may mean greater crop water stress in rainfed and limited-irrigation scenarios, and typically two additional irrigation events (considering 1 inch application depth for center-pivot irrigated field) requiring greater water abstraction and associated energy costs.These additional water requirements may be even higher for crop regions such as oranges, sugarcane and rice, which show much larger rates of ET o increase (Figure 7).Moreover, increase in evaporative demand during the dormant season (winter and spring months) likely have resulted in lower soil water availability during pre-planting and early growth periods.Such water deficit in early crop vegetative growth periods may imply an earlier start to the irrigation season than before, resulting in a longer water abstraction period during the growing season.With novel high-resolution datasets on crop water use such as OpenET (Melton et al., 2022) becoming available, it will be useful to investigate interannual variability in ET and comparative meteorologic versus biological drivers in the future.
In addition to direct impacts from higher ET o , ET o drivers can also show independent control on plant functioning in agroecosystems.
For instance, increase in ET o has also accompanied increase in VPD (Ficklin & Novick, 2017;Kukal et al., 2023), especially in areas with increasing temperatures and decreasing humidity (Figure 1).VPD has an independent control on stomatal conductance, affecting both photosynthesis and transpiration, and thus, water use efficiency (Novick et al., 2016;Zhang et al., 2019).Thus, periods of high VPD suppress water uptake because of lower stomatal conductance under both sufficient and limited soil water.Under agricultural drought conditions (low soil water availability) and high VPD periods, crops limit transpiration to conserve water, and thus, ET o increases might not linearly translate into actual water use and cause yield penalties (Singh et al., 2021).Similarly, solar brightening trends have shown to improve radiation absorption and an increase in crop yields have been attributed to this phenomenon in the U.S. (Tollenaar et al., 2017).Lastly, increasing wind speeds have been shown to reduce transpiration and enhance CO 2 uptake, especially under high radiative energy scenarios due to more efficient convective cooling (Schymanski & Or, 2015).This contrasts with the widespread understanding of a positive relationship between wind speeds and transpiration, as also mathematically represented in Penman-type models, including the ASCE-PM ET o method used in this study.However, this hypothesis while evaluated at leaf scales by Schymanski and Or (2015), must be tested at canopy to ecosystem scales.Overall, while the widespread increase in ET o is widespread and holds significant multidimensional implications, however, its manifestation as actual crop water use is also dependent on crop biological response and independent impacts and interactions of ET o drivers on plant processes.
Temperature and insolation (R s ) represented the two dominant controls on ET o across vast regions of the CONUS, underscoring dominance of radiative controls.Nevertheless, relative contribution of aerodynamically relevant inputs, i.e., U 2 and RH cannot be ignored.In fact, they represent dominant controls on ET o in a small fraction of counties.This holds strong significance for appropriate selection of potential/reference ET methods to decompose long-term trends.The observations of T and R s dominance may be interpreted as supporting the use of temperature-based evaporative demand formulations such as e.g., Thornthwaite (1948), Hargreaves (Hargreaves & Samani, 1985), Hamon (1961) Blaney and Criddle, 1950).However, the relative role of U 2 and RH cannot be considered negligible given their combined relative contribution of more than one-third on average across US croplands.Thus, a temperature-based formulation of evaporative demand will result in inaccurate long-term trends and eventually misleading assessments of agricultural water use.It has also been shown that the covariance between T and R s is not always positive (which underlies T-based formulations) depending on region and timeframe (Hobbins, 2016).have undergone significant ET o increase (Figure 1), which has primarily resulted from interplay between T, RH, and U 2 change (Figure 3).
Having ignored RH and U 2 dynamics, it is suspected that the decision support systems may not realistically represent the evolution of crop water requirements in this region, which can hinder management efforts against the water stress experienced in these basins (Averyt et al., 2013;National Research Council, 2007;Norton et al., 2021).Overall, assigning explicit relative contributions for each county (Figure 3 and Figures S5-S8) aid in understanding region and timescale-specific importance of T, R s , RH, and U 2 for driving ET o increase in the context of water planning efforts.
The framework and approach used to quantify relative contributions from ET o drivers in this investigation differs in some respects from other relevant literature.Distinctions among these methodological aspects must be clearly articulated for effective interpretation of corresponding findings.et al. (2022) in that how relative contribution for each driver is quantified.In their approach, three out of four inputs are constrained so that ET o trends can be interpreted as a sole function of the unconstrained input.This approach may be vulnerable to presence of any interactive effects between two or more input drivers that were forced to their climatologies.In fact, such interactions have been evidenced by findings of substantial covariance between input meteorological variables by Hobbins (2016).In contrast, our approach is based on allowing three out of four inputs to represent actual observations and only one constrained input to represent climatological values.This approach was also adopted by Vicente-Serrano et al. ( 2014) to quantify sensitivity of ET o in Spain and allows for accounting for any interactive effects of two or more input drivers.It seeks to infer the importance of a given driver by relative loss of information or deviation in ET o trends (described analytically by Equation 3) that results by constraining each input to its climatology.Thus, a greater deviation of driver-constrained ET o from observed ET o trends implies greater relative contribution of that driver toward ET o change.It is very likely that these conceptual and analytical differences in methodologies translate to differences in findings.Our findings concur with those of Albano et al. (2022) who also found that T was the most important driver of ET o increase in the U.S., although the specific relative contributions varied largely in intensity.Our findings indicate that R s closely followed T as the second most important driver, whereas Albano et al. (2022) ranked it the least important.These differences should be interpreted while also acknowledging the differences of scales and aggregation methods used.Since the fundamental scale of analyses vary vastly between our study (county-level) and that of Albano et al. (2022) (water resources regions), one-to-one comparison is quite challenging.Thus, we calculated county-level relative driver contributions using the alternative methodology of constraining three variables at a time for comparison.Upon comparison, the root mean squared differences between relative contributions (fractional) derived from the two approaches were 0.12 (T), 0.09 (RH), 0.07 (R s ), and 0.10 (U 2 ).These difference in relative contribution outcomes were subject to significant spatial variation, as presented in Figure S18.Specific reasons for why these differences would appear in some regions like the northern plains is unclear but might be a consequence of considering/ignoring presence of driver interactions in these regions.
There are certain caveats associated with this investigation that deserve mention for cautious interpretation.First, it has been shown that choice of gridded datasets for obtaining meteorological data for ET o derivation is a major dictator of robustness of trends and driver attribution (Albano et al., 2022;Blankenau et al., 2020).While an ensemble of datasets is best to encounter uncertainties as accomplished by Albano et al. (2022), it is well documented that gridMET has shown reasonable performance in estimating ET o when compared against ground-based station data (Blankenau et al., 2020), especially for T. Second, we did not account for changes in CO 2 concentrations during 1981-2021, that has been shown to affect stomatal resistance (Yang et al., 2019).However, it is worth noting that ASCE-PM ET o represents evapotranspiration from a hypothetical surface with a fixed stomatal resistance (70 sm −1 ), whereas CO 2 -induced stomatal dynamics are considered biological feedback and are of significance to actual ET rather than ET o .Nevertheless, inclusion of CO 2 effects on ET o has been shown to not significantly affect ET o during the study period (Albano et al., 2022) and thus this research has ignored these minimal effects.Third, VPD (e s − e a ) used in calculating ET o in this research was derived from T and RH data as also suggested by procedures outlined in the ASCE manual of practice 70 (Jensen & Allen, 2016) and used extensively in operational practice.However, it has been shown that using T and RH data to derive VPD can be suboptimal in some environments, where air and dewpoint temperatures perform the best (Howell & Dusek, 1995;Kukal & Irmak, 2022;Yoder et al., 2005).However, lack of robust dewpoint temperatures gridded data on a sufficient spatial resolution limited us to use these superior methods.Lastly, it is possible that approaches based on forcing input variables to climatologies may distort covariance structures among input variables.While our approaches tend to minimize this distortion by allowing three out of four input variables to represent observations, we acknowledge that this issue may persist to some degree.

| CON CLUS ION
A CONUS-wide county-level investigation was conducted aiming at (1) assessing comparative trends in annual and monthly total ET o and its drivers during 1981-2021 and (2) quantifying relative contribution of each of the four drivers to ET o trends.We found that annual total ET o has undergone a mean increase of 2.1 SD (111 mm) relative to the 1981-2000 baseline across 53% of CONUS counties.As for the drivers of ET o , R s showed the greatest change (increased by 2.7 SD), closely followed by U 2 (increased by 2.5 SD) and RH (decreased by 2.3 SD), while T showed the least change (increased by 1.1 SD) over regions with significant ET o change, on average.The ET o increase was driven primarily by T as evidenced by the largest relative contribution (36%), followed by R s (29%), U 2 (18%) and RH (17%) when weighted for total calorie production in the CONUS.Substantial seasonal variation was encountered in relative contribution of the four drivers to ET o increase.During summer when ET o is a strong predictor of agricultural productivity and water use efficiency, R s was the dominant driver with relative contribution ranging between 40%-50%, with the other three drivers roughly being similarly important.CONUS-aggregated findings serve to mask strong spatial heterogeneity in trends as well as importance of drivers, which was communicate via maps for driver-specific relative contribution, driver ranks, and dominant drivers.Overall, this research comprehensively addresses not only the demand (atmospheric)-side aspect of agricultural water use and its long-term changes, but also allows for understanding what changes its drivers have undergone and how important they have been to bring about change in evaporative demand.Since the resulting data and findings are presented in agriculturally relevant terms, i.e., weighing of quantitative indicators by cropland acreage and total calorie production, these are oriented to be digested and utilized for agricultural and natural resources policymaking.The implications of the findings herein can assist in discussions surrounding food and water security importance of greater evaporative demand and its translation into actual agricultural water use, independent and interactive effects of observed changes in T, R s , RH, and U 2 with ET o change, and negative consequences of inappropriate evaporative demand modeling approaches given importance of ET o drivers.A retrospective addressal of evaporative demand and its causes allow agricultural and water managers to anticipate and understand the emerging importance and impacts of future ET o projections for rainfed crop sustainability, irrigation abstraction, limited irrigation transitions, crop physiological impacts, sectoral water impacts, and land-atmospheric feedbacks.
practitioners and producers by focusing on different agricultural commodities and their calorific footprints.The specific objectives of this research were to (a) investigate change in monthly and annual total ET o during 1981-2021 at county scale across conterminous U.S. (CONUS); (b) quantify relative contribution of T, RH, R s , and U 2 toward the assessed change in evaporative demand for each county; and (c) identify dominant drivers of change in evaporative demand across CONUS.
b. Daily observed ET o (mm/day) during 1981-2021 was calculated for all 3108 counties using the following equation(Allen et al., 2005):where, ET o = standardized grass-reference ET (mm/day), Δ = slope of saturation vapor pressure versus air temperature curve (kPa/°C), R n = net radiation (MJ/m 2 /day), G = heat flux density at the soil surface (assumed to be zero for daily time step), T = mean daily air temperature (°C), U 2 = mean daily wind speed at 2-m (m/s), e s = saturation vapor pressure (kPa), e a = actual vapor pressure (kPa), e s − e a = VPD, γ = psychrometric constant (kPa/°C).900 is the value for C n constant that represents the time step and aerodynamic resistance of the grass reference surface, 0.34 is the value for C d constant that represents the time step, bulk surface resistance, and aerodynamic resistance of the grass reference surface, and 0.408 is constant in m 2 mm/MJ (1/λ, where λ is the latent heat of vaporization (2.45 MJ/m 2 /mm).c.Using the same procedure as outlined in b, daily ET o was recalculated four more times for all counties, each time constraining one of the four ET o drivers (T, RH, R s , and U 2 ) to its climatology from step a, while actual observations were used for the other three drivers.This approach resulted in four constrained series of daily ET o for each county, one each for T, RH, R s , and U 2 constrained to climatology.d.The daily observed ET o and four constrained ET o time series for all 3108 counties were aggregated into monthly and annual mean ET o .

F
I G U R E 1 Total change in annual (a) total grass-reference (short crop) evapotranspiration (ET o ); (b) mean air temperature (T); (c) mean relative humidity (RH); (d) total shortwave incoming solar radiation (R s ); and (e) mean wind speed at 2-m height (U 2 ) during 1981-2021 expressed as standard deviations from the 1981 to 2000 baseline.Striped lines indicate that the change detected for a given county was statistically significant at 95% confidence interval.

F
Total change in monthly total ET o during 1981-2021 expressed as standard deviations from the 1981 to 2000 baseline.Striped lines indicate that the change detected for a given county was statistically significant at 95% confidence interval.3.3.2 | Spatial variation in driver-specific relative contribution to ET o change County-level relative contributions for T, RH, R s and U 2 were mapped and presented in Figure 3.Each driver has a distinct signature of spatial gradient in relative contribution to ET o change.T showed high relative contribution in western half of CONUS and well as the Northeast, with exceptionally high contribution in the Pacific Northwest region.Interestingly, much of this region in the west coincides with extremely low (<10%) relative contribution of R s .On the other hand, R s shows strong relative contribution in the Southeast and portions of Northern Plains, Upper Midwest, and Ohio Valley.In most of the regions where R s had strong relative contributions, RH appears to be less important.In contrast, the South, Southwest, Northern Rockies and Plains show moderately high relative contribution of RH.U 2 has shown moderate relative contribution in the High Plains region, but much greater importance in the Rocky Mountains.

F
Relative contribution (fraction) of annual (a) mean T; (b) mean RH; (c) total shortwave incoming R s ; and (d) mean U 2 toward change in total ET o during 1981-2021.The relative contributions are expressed as fractions that add to 1.

F
Ranks assigned to each county based on decreasing order of relative contribution of (a) mean T; (b) mean RH; (c) total shortwave incoming R s ; and (d) mean U 2 toward change in total ET o during 1981-2021.Rank 1 means highest relative contribution and Rank 4 means least relative contribution.(e) shows the dominant driver for each county, i.e., the driver with the highest relative contribution among all four drivers.
, higher aridity (Milly & Dunne, 2020), increased irrigation needs and groundwater abstraction (Aeschbach-Hertig & Gleeson, 2012; Kustu et al., 2010), F I G U R E 5 Seasonal patterns of relative contribution of mean T, mean RH, total shortwave incoming R s and mean U 2 toward change in total ETo during 1981-2021.The monthly-level relative contribution is presented as (a) mean across all CONUS counties and (b) mean across counties with statistically significant (α = 0.05) monthly ET o change.

F
Dominant driver of change in monthly total ET o during 1981-2021 for each county, i.e., the driver with the highest relative contribution among all four drivers.F I G U R E 7 Crop region-specific distribution of annual total change in total ET o during 1981-2021 expressed as (a) standard deviations from the 1981 to 2000 baseline, and (b) total absolute change in mm.Only those counties are included where statistically significant (α = 0.05) ET o change was recorded.
Such assumptions employed by T-based formulations will not hold true under these scenarios and result in flawed estimates.It is not uncommon to encounter water management agencies relying on T-based reference ET estimates to compute crop water requirements.For instance, the consumptive use model used by Colorado Decision Support Systems developed by the Colorado Water Conservation Board and the Colorado Division of Water Resources relies on Blaney-Criddle (USDA, 1950) formulation.Our results show Colorado's water basins (Hammond Wagner et al., 2019)oped using global extent and productivity of 41 major agricultural crops (accounting for >90% of total calorie production around the world) during 1997-2003 (accessed at http://www.earth stat.org/crop-allocation-food-feed-nonfood/);(b)county level reported acreage for 21 key crops (FigureS1b) from the United States Department of Agriculture's 2012 Agricultural Census(Hammond Wagner et al., 2019)accessed at https://bmcre snotes.biomedcent ral.com/artic les/10.1186/s1310

Long-term change in ET o and its drivers during the year
Annual ET o masks within-season distribution of ET o , which can vary substantially, and thus can overlook the intra-annual variability that may exist in ET o trends.Addressing change in ET o and its drivers at month-specific scale also aids in decomposing sector-specific implications such as agricultural growing season versus dormant season evaporative demand, or maize early vegetative (April-May), late vegetative (June), early reproductive (July), and late reproductive phases (August-September).To allow for such inferences, monthly ET o trends during 1981-2021 are Mean statistically significant (α = 0.05) ET o change (in standard deviations, SD) for annual and monthly time-steps and corresponding significant change in T, RH, R s , and U 2 within areas of significant ET o change.
TA B L E 1 Albano et al. (2022),Hobbins (2016), andIrmak et al. (2006)have addressed sensitivity of ASCE ET o within CONUS using varying approaches.First,Irmak et al. (2006)andHobbins (2016)focus on how meteorological drivers account for "variability" aspect of daily evaporative demand at different timeframes, as opposed to their role in driving "long-term trend" in evaporative demand.While Irmak   et al. (2006)quantified sensitivity as change in ET o to unit change in drivers,Hobbins (2016)relied on a method of moments technique that also accounted for drivers' observed variabilities and their covariances.Albano et al. (2022)on the other hand, focused on meteorological drivers' contribution toward long-term ET o change, which is consistent with our research.However, our approach differs from that of Albano