The measurement sites represented eight crop types, namely, alfalfa, almond, citrus, corn, pasture, rice, tomato, and beardless winter wheat .The eddy covariance system uses a sonic anemometer and infrared gas analyzer to measure three-dimensional wind velocities and high-frequency fluctuations of water vapor concentrations . It measures evapotranspiration by monitoring the vertical flux of water vapor. High-frequency eddy covariance measurements in two alfalfa, two corn, and one rice AmeriFlux sites were collected and preprocessed into half-hourly evapotranspiration data as outlined in Eichelmann et al. and Hemes et al. . Of the five AmeriFlux sites, net radiation for alfalfa and corn was measured with four-channel net radiometers. Most sites were located in the Sacramento-San Joaquin Delta region was also employed in the Delta Consumptive Use Comparative Study supported by the California State Water Resources Control Board Office of the Delta Watermaster and other agencies. Less expensive surface renewal systems were deployed over 14 sites for corn, alfalfa, and pasture. They use thermocouples to measure sensible heat flux, an NRLITE2 Net Radiometer for net radiation, and either measure ground heat flux with a combination of ground heat flux plates and soil thermocouples or assume it is zero for daily estimates. Evapotranspiration is then estimated as the residual of the energy balance. For each crop type, an eddy covariance tower was deployed to calibrate the sensible heat flux relationship between eddy covariance and surface renewal measurements . Evapotranspiration measurements were compiled from two specialty crop research projects in Tulare and Kern county of the southern Central Valley, including surface renewal measurements in citrus orchards from 2001 to 2004 and eddy covariance measurements in an almond orchard from 2009 to 2012 . We used only data collected after February 2003 in this study, considering the data availability of California Irrigation Management Information System Spatial product data. The most recently available eddy covariance tower measurements by NASA JPL were also added. The JPL sites were located at the Russell Ranch research field, near Davis,plant raspberry in container including one over tomato from February to October 2017, and the other over winter wheat from December 2016 to October 2017.

These towers have advanced thermal infrared radiometers to measure land surface temperature, and two sets of four channels net radiometers to reduce measurement uncertainty. High-frequency evapotranspiration data were automatically processed using Campbell Scientific Inc.’s standard Eddy-Covariance Datalogger Program software and various quality control procedures. All half-hourly measurements were preprocessed and aggregated into daily evapotranspiration if <20% of the half-hourly measurements were missing within a day.We obtained the daily gridded meteorological data, including minimum and maximum-air temperature at 1.5 m, and daily dew point, from Spatial-CIMIS at a 2-km resolution . The DWR manages a network of over 145 automated weather stations over well-maintained and well-watered grass sites across California providing reference evapotranspiration for pasture. The station data were spatially interpolated to produce the 2-km gridded data set since 2003. We also used the Spatial-CIMIS cloud cover and incoming solar radiation for both clear-sky and all-sky conditions, derived from Geostationary Operational Environmental Satellite visible channel imager data, for our radiation component calculation.All available surface reflectance and surface temperature products, and the corresponding quality assessment layers at 30 m were downloaded from USGS Landsat Analysis Ready Data set . A total of eight tiles covered the whole study area. The land surface temperature retrieval from the Landsat thermal data is based on a radiative transfer model with an improved surface emissivity estimate . Each active Landsat satellite takes snapshots between 9:53 and 10:55 a.m. Pacific Standard Time every 16 days. Invalid or high uncertainty pixel values were filtered based on the quality assessment rasters, including SLC gaps , snow, cloud/cloud shadow, for example, a high value for cloud or cirrus confidence, or with a surface temperature uncertainty greater or equal to 6 K. For model calibration and validation purposes, a single pixel near each measurement site was extracted.During cloud-free days with Landsat overpasses, Landsat-derived LAI and NDMI were fed into Equation to estimate the actual Priestley-Taylor coefficient for each pixel, which was then combined with available energy to estimate daily evapotranspiration . For days between Landsat overpasses without valid or high-quality values such as cloudy days or over scan-line corrector data gaps , a temporal interpolation approach was adopted . First, daily evapotranspiration estimates, during the adjacent clear-sky Landsat days and within ±2 months search window, were divided by the concurrent Spatial-CIMIS daily reference evapotranspiration to derive the fraction of reference evapotranspiration .

A shape-preserving piece wise cubic interpolation was applied to this discrete time series of EToF to obtain a continuous time series of daily EToF. We set a requirement of a minimum of 2 valid observations within the search window for a robust interpolation. This temporal interpolation was needed mostly during rainy season in winter and early spring in California, an off-season for the majority of the crops. Finally, daily evapotranspiration for missing days was estimated as a product of the interpolated EToF and Spatial-CIMIS reference evapotranspiration.The Priestley-Taylor method optimized here was applied over the whole California Central Valley to estimate crop evapotranspiration during the 2014 and 2016 water years. The crop-specific actual Priestley-Taylor coefficient parameterization results were used for daily averaged evapotranspiration estimation over alfalfa, almond, corn, citrus, pasture, and rice areas during Landsat over passing days. For remaining crop types, including but not limited to grapes, walnut, pistachio, tomatoes, wheat, and cotton, where no field evapotranspiration data were available for crop-specific optimization, the generalized actual Priestley-Taylor coefficient parameterizations was applied. Temporal interpolation was applied to derive a complete time series of daily evapotranspiration for each Landsat pixel. For each month, an EToF pixel is interpolated only if there are at least two estimates on clear-sky over passing days with a ±2 months moving time window; the uninterpolated pixels were gap-filled by multiplying daily reference evapotranspiration by EToF averaged by corresponding month and crop within each Landsat Analysis Ready Data tile. Daily evapotranspiration estimates were further averaged to annual time scales to analyze the regional patterns. Evapotranspiration was summarized for each crop type and compared the differences among crops by evaluating the annual evapotranspiration, reference evapotranspiration, and EToF. Specifically, the per-area water consumptive use average was computed by dividing the sum of annual evapotranspiration by crop area over nongap-filled pixels, while total consumptive use was computed over all crop area pixels. We further summarized annual evapotranspiration by GSA boundaries to provide agricultural water use information for water planning. This was achieved by quantifying annual water use and variability for each planning area and compared across areas.

We also analyzed the association of water use with corresponding land use, Rn, actual Priestley-Taylor coefficient, EToF,plastic seedling pots and reference evapotranspiration, to understand what contributed to water use differences among GSAs. While GSAs manage local groundwater resources, DWR oversees water resources regionally by water planning area. We summarized our annual crop evap-otranspiration estimates by water planning areas in the Central Valley and compared them with DWR’s estimates for the water year 2014.The seasonal dynamics of the actual Priestley-Taylor coefficient typically followed the plant growth curve, as shown by the values derived from both the field measurements and satellite observations . For example, the actual Priestley-Taylor coefficients of alfalfa frequently fluctuated from 0.5 to 1.5, likely due to the multiple cuttings throughout the growing season, as shown by the similar variations in LAI . Field measurements showed a substantial seasonal variation in the actual Priestley-Taylor coefficient for the corn and rice sites, e.g., with towering peaks in summer growing season, a relatively small peak in spring, and much lower values in between fall and winter . In general, the remote sensing-derived actual Priestley-Taylor coefficients, from the crop-specific optimization, could explain 56% of the variance observed across sites and time periods, with an RMSE and RMAD of 0.23% and 17.7%, when compared with the field-based estimates over the testing data set . For the generalized optimization, the uncertainties of actual Priestley-Taylor coefficient estimates increased slightly . Among crop types, both crop-specific and generalized actual Priestley-Taylor coefficient estimation performs best for almond . The performance of the crop-specific actual Priestley-Taylor coefficient is significantly better than the generalized actual Priestley-Taylor coefficients for corn and citrus. The actual Priestley-Taylor coefficient estimates showed significant improvement when compared to those derived from PT-0, which only captured small seasonal variation and had a higher bias of 0.24 and a larger RMAD of 34.7% over the irrigated cropland in the valley . In contrast, PT-JPL estimates showed a reasonable seasonal pattern for alfalfa and corn , although it was not calibrated for any land cover type . Across all sites, the crop-specific PT-UCD showed an overall improvement over PT-JPL, as shown by the empirical cumulative distribution function of the absolute errors when compared to both testing and independent testing data . For example, 88% of testing samples had an absolute error were below 0.30 from crop-specific PT-UCD estimates, compared to 62% and 59% from PT-JPL and PT-0 estimates, respectively. The generalized PT-UCD performed only slightly better than PT-JPL .Two types of cross-validation testing further showed the optimization of the parameters in Equation 3 for estimating the actual Priestley-Taylor coefficient was reasonably robust. The distribution of the estimated parameters showed a very small variance, for the majority of the crops and the generalized optimization . One exception was parameter D, which represented the moisture regulation over the coefficient, for citrus and pasture . The estimated actual Priestley-Taylor coefficients were shown to be stable among the repeat and leave-two-out cross-validations , with an Inter Quantile Range of RMAD of <5% .We found a good agreement between field measurements of evapotranspiration and satellite-based estimates during the clear-sky days with Landsat acquisitions. When evaluated with the testing data set, both the crop-specific and generalized evapotranspiration models captured the seasonal variability well . Across all sites, the crop-specific evapotranspiration had an R2 of 0.79, RMSE of 0.90 mm day−1, and RMAD of 20.5% . Only a small bias of 0.14 mm day−1 was found. When using the generalized actual Priestley-Taylor coefficients, slightly higher uncertainties were found, with an R2 of 0.76, RMSE of 0.98 mm day−1, and RMAD of 23.1% . The performance of evapotranspiration estimates varied by crop types. When using the crop-specific Priestley-Taylor optimization, the RMSE and RMAD ranged from 0.68 to 1.34 mm day−1 and 13.3% to 28.4%, based on the comparison with the testing data set . The best performance was found for alfalfa, citrus, and pasture sites, while the weakest performance in rice. The generalized approach also performed the best for alfalfa and citrus and performed the poorest for rice and corn . The leave-two-out cross-validation showed relatively small differences in RMSEs of daily ET estimates from site to site , e.g., 0.7 mm day−1 in alfalfa site #6 vs. 0.9 mm day−1 in site 5 based on the results from alfalfa-specific optimization, and 0.7–1.2 mm day−1 among the corn sites. Crop-specific PT-UCD showed an improvement over PT-0, PT-JPL, and generalized PT-UCD. About 80% of crop-specific evapotranspiration estimates in the testing and independent data set had an error of <1 mm day−1, as shown by the empirical cumulative distribution functions of the absolute errors between the daily crop-specific evapotranspiration estimates and field measurements . In contrast, both generalized PT-UCD and PT-JPL appeared to perform similarly, that is, about 70%–76% of samples had an evapotranspiration error <1 mm day−1, and about 85%–90% <1.5 mm day−1. However, for the PT-0 evapotranspiration estimates, only 55% and 70% of samples had an error <1 and 1.5 mm day−1, respectively.The interpolation of EToF from adjacent overpassing days introduced a small overall uncertainty in daily evapotranspiration estimates, for example, RMSE increased by 0.10–0.17 mm day−1 and decreased R2 by 0–0.08 when estimating evapotranspiration for alfalfa, citrus, corn, and pasture . When further aggregated to weekly and monthly time scales, the satellite-derived evapotranspiration estimates agreed better with those from the field measurements . For example, across all sites, R2 was increased to 0.83 and 0.88, and RMSE reduced to 0.79 and 0.65 mm day−1, respectively, for weekly and monthly evapotranspiration values based on the crop-specific Priestley-Taylor optimization.