Daily Evapotranspiration Estimation Using HJ-1 Data for Irrigation Water Management in a Desert-Oasis Irrigation District, Northwest China

Irrigation of agricultural land is the main water consumer in the arid and semiarid regions. The accurate time series of daily evapotranspiration (ET) at the field scale is crucial for irrigation water management. Here, we presented an integrated approach to field-scale ET mapping by combing METRIC (Mapping Evapotranspiration at High Resolution with Internalized Calibration) with trapezoidal framework of vegetation fraction and land surface temperature (fv-Trad). This approach fed with Chinese satellite HJ-1 (environmental and disaster monitoring and forecasting with a small satellite constellation) images was used to map daily ET over the desert-oasis irrigation fields in the middle of the Heihe River Basin, Northwest China. The results showed that time series of daily ET derived from HJ-1 were well consistent with the in situ measurements (coefficient of determination (R2) of 0.95, root mean square error (RMSE) of 0.31 mmd–1, mean absolute percentage error (MAPE) of 12.0% at Daman station; R2 of 0.89, RMSE of 0.16 mmd–1, and MAPE of 13.2% at Huazhaizi station). The rapid variation of ET could be precisely captured by HJ-1, especially when rainfall or irrigation events occurred between successive acquisitions of satellite data. Meanwhile, the spatiotemporal variation of monthly ET is closely related to crop growth. Daily ET at field scale will support a variety of local interests in water use and irrigation management for both planning and regulatory purposes.


Introduction
Agriculture heavily relies on irrigation of water drawn from rives or aquifer in arid and semiarid areas. In the past 50 year, the river runoff showed the declining trend with the change in intensive climate. But water consumption increases due to booming socio-economy and fast-growing population, which will further deteriorate the water resource status in China [1]. Evapotranspiration (ET), the total amount of land surface evaporation and plant transpiration, is the most active process in the terrestrial hydrological cycle and the major component of energy and water balance in agricultural system. In the north of China, agricultural water consumption accounts for about 70% of total water consumption, and more than 90% of agricultural water is consumed via ET [2]. Therefore, it is necessary to determine the accurate spatiotemporal ET in irrigated fields for recognizing the crop consumption patterns, evaluating water use efficiency and optimizing irrigation scheduling, especially in the semi-arid and arid areas.
Typically, ground observation systems including lysimeters, Bowen ration systems, eddy covariance systems (EC), and large aperture scintillometers are considered to be accurate for measure ET at individual sites or for small footprint [3][4][5]. However, these observation systems are not applicable to a large region due to their high cost. New approaches to take advantage of remote sensing (RS) to map ET have been developed, especially in recent decades. Major improvements in ET estimates have been obtained through RS methods based on thermal infrared (TIR) data. A suite of algorithms that were based on surface residual of the energy balance, derived from optical and TIR observations, has been extensively developed and applied. One-source models include surface energy balance algorithm for land (SEBAL) [6], surface energy balance system (SEBS) [7], mapping evapotranspiration at high resolution with internalized calibration (METRIC) [8], and simplified surface energy balance (SSEB) [9]. Two-source models, such as the two-source energy balance model (TSEB) [10] and the disaggregated atmosphere-land exchange inverse (DisALEXI) [11], integrate the biophysical processes and RS data to estimate the plant transpiration and evaporation from non-vegetated surfaces separately. As reported in several reviews (e.g., Kalma et al., 2008;Liang et al., 2010), these approaches appear to accurately reproduce ET over a wide range of conditions at both the satellite overpass time and daily time scales [12,13]. Here we selected METRIC algorithm to estimate the field evapotranspiration. METRIC algorithm evolved from SEBAL but uses ground-based meteorological measurements of reference ET acquired within scene to establish the latent heat flux at the cold pixel, rather than assuming a priori that sensible heat flux equals zero for the condition. This innovation makes the best use of existing technology in agricultural areas and establishes a "reality check" for the satellite-based ET estimates, which makes it more robust [14]. It performed well in estimating high-resolution ET with the relative errors of 10%-30% at different ecosystems around the world [15][16][17].
Local scale water management usually requires temporally continuous high spatial resolution ET in order to understand the crops ET variation, which is related to spatial heterogeneity in crop type, phenological stage, meteorological conditions, as well as soil moisture conditions [18]. The 100-m resolution characteristic of TIR imagery such as Landsat series, which we define as "moderate resolution" to distinguish from the high-resolution (meter scale) shortwave imagery collected by the Landsat series of satellites has proven critical to many operational needs for ET data because it will typically enable discrimination of individual agricultural fields [19]. But moderate resolution cannot realize the temporally continuous daily ET data because of the long revisit period and cloud contamination [20]. The Chinese Environment and Disaster Monitoring and Forecasting with a Small Satellite Constellation (HJ-1) can provide global coverage every two days with a spatial resolution of 30 m, which makes it a good balance of temporal and spatial resolution. The characteristics of short repeat interval and high spatial resolution are very useful for monitoring the regional land surface variables, such as vegetation indexes (VI), drought indexes, land surface temperature (LST), and ET [21][22][23].
Here METRCI fed with HJ-1 data was applied in the middle reach of Heihe River Basin to estimate daily ET at field scale for irrigation water management in 2015. A trapezoidal framework of vegetation and land surface temperature (f c -T rad ) was integrated into METRIC to reduce ambiguity and uncertainty in the latent heat flux estimation due to the differing extremes possibly selected by the operator. The main objectives of this paper were to: 1) obtain and validate continuous daily ET at field scale through METRIC model driven by HJ-1 data; 2) characterize the spatial and temporal variations of ET values for different land cover types.

Study Area
Heihe River Basin (HRB) is the second largest inland river basin in the arid region of northwest China. The experimental area (100.03•-100.68• E, 38.33•-39.2• N) including Daman, Yingke, Ganjun, Shangshan, Wujiang and Xigan irrigation districts is located in the middle reaches of HRB, as shown in Figure 1. The study area has a continental, dry, temperate climate with a mean annual temperature, precipitation, and open water evaporation of 7.6 ℃, 116.8 mm, and 2390.0 mm, respectively [24]. During the growth period, the crop needs much water and irrigation is required in these months to meet crop growth requirements. Take 2012 as example, irrigation for oasis agriculture in the middle of HRB consumed 10.6×10 8 m 3 of surface water diverted from Heihe River and 5.36×10 8 m 3 from pumped groundwater [25]. Ganzhou irrigation district is an important maize seed production base in Northwest China. Maize is predominant crop in the irrigation districts. Other plants include spring wheat, alfalfa, cotton, woods, vegetables and fruits.

Fields Measurement
There are five ground observation stations in the study area: Daman, Huazhaizi, Shenshawo, Zhangye and Gobi. An automatic weather station (AWS) was installed at each site and equipped with sensors for collecting data including temperature, wind speed and direction, air pressure, humidity, precipitation, soil moist profile, solar radiation, four-component radiation, soil heat flux, and infrared temperature every 10 min over the whole year of 2015. Daman Station and Huazhaizi Station were also equipped with an eddy covariance (EC) system. The EC sensors were installed at the height 4.5 m above the ground. All EC data were sampled at a frequency of 10 Hz and turbulent fluxes (sensible heat flux and latent heat flux) were recorded every 30 min. The Gaussian fitting method was used to interpolate the missing data (Xu et al., 2017) [26]. In the case of the energy balance of EC system, the closure error of EC-based measurements at the daily scale was calculated based on the energy balance ratio (EBR) to correct the daily flux data (Liu et al., 2011) [27]. devices (CCDs) and a hyper-spectrum image (HIS) loaded on HJ-1A, and there are two CCDs and an infrared scanner (IRS) on HJ-1B. HJ-1 can provide visible/near-infrared/thermal/infrared (VIS/NIR/TIR) images with 30 m, 150 m, and 300 m spatial resolution respectively, with every 2-4 days' revisit time covering China and the surrounding areas. Based on the consideration of the satellite image quality of the HJ-1 and Landsat 8, the 36 available HJ-1 images and 9 available Landsat 8 images obtained in the growth season of 2015 from April to October were selected. All the images were subjected to geometric correction and atmospheric correction.

Satellite Data
A DEM (30m) covering Heihe River Basin was provided by the "Heihe Plan Science Data Centre, National Natural Science Foundation of China" (http://www.heihedata.org). DEM data were used to demonstrate the terrain of the study area and applied in METIRC. Cropping pattern data were derived from HJ-1 and Landsat 8 data with support vector machine (SVM), object-based method, and time-series analysis to improve the classification accuracy and the overall accuracy for the crop classification reached 84% [28].

Brief Outline of METRIC Algorithm
METRIC is a one-source model, which considers soil and vegetation as a sole source in the estimation of ET. It is widely used to estimate daily ET at the field scale. The latent heat flux (λET) is computed as a residual of the energy balance equation on a pixel by pixel basis: (1) where R n is net surface radiation, G is soil heat flux and H is sensible heat flux. Their units are all Wm −2 . Each energy component in Equation 1 can be calculated as follows: (1 ) (1 ) Where R s ↓is the down-welling shortwave radiation (Wm −2 ); R L ↓is the down-welling long wave radiation (Wm −2 ); R L ↓is the up-welling longwave radiation (Wm −2 ); α is the surface albedo for shortwave radiation; ɛ is broad band surface emissivity; T rad is the surface radiometric temperature (K); LAI is leaf area index; ρ air is air density (kgm −3 ); C p is air specific heat at constant pressure (1004 Jkg −1 K −1 ); r ah is the aerodynamic resistance to heat transport (sm −1 ); and dT represents the near surface to air temperature difference (K). dT is assumed to have a linear relationship with land surface temperature (Equation 6) as developed by Bastiaanssen et al. (1998) [6].
In METRIC, an iterative method was applied to calculate both r ah and H to reduce the effects of buoyancy of heated, light air at the surface. The algorithm started with finding a set of "hot" and "cold" pixels. It assumed that λET equals to 0.05 and 1.05 times reference ET (ET 0 ) for "hot" and "cold" pixels at the beginning of the iteration, and then calculated dT hot and dT cold , followed by calculating a and b, and then updating dT, and H. The Monin-Obuknov length (L) was used to define the stability conditions of the atmosphere during iteration in METRIC. The iteration stops when the value of r ah remains stable. It is critically important to note that METRIC or all studies pertaining to SEBAL, recognize the key role of extremes in the resulting estimates [29]. A large uncertainty in the model outputs could be introduced by differing extremes manually identified by the operator. Here we introduced a trapezoid framework of f c -T rad to replace the intractable step of the selection of two extremes.
Instantaneous ET (ET inst ) of each pixel at satellite overpass time is calculated from Equation 1-6. It is extended to daily values (ET daily ) by fraction of reference ET (ET 0 F) in METRIC, which is calculated by dividing ET inst by hourly reference ET (ET 0 _ inst ). It is assumed that the value of instantaneous ET 0 F is the same as the average value within 24 h of the day satellite overpass. Finally, the daily ET at each pixel is calculated as a product of ET 0 F and daily ET 0 : More details about METRIC algorithm can be found in Allen et al. (2007) [8]. Irrigation management requires continuous daily ET especially during growth seasons so that ET losses can be compensated by applying the proper quantity of water at the right time to meet the water demands of plants. Continuous daily ET can be derived from a series of ET 0 F images by interpolating ET 0 F between processed images and multiplying, on a daily basis, by the ET 0 for each day. Here we applied a cubic spline interpolation method to generate a smoothed ET 0 F curve based on the nonlinearity of the temporal dynamics of ET 0 F [30]. Then continuous daily ET was derived from the smoothed curve and multiplied by daily ET 0 (ET 0_daily ).

Brief Outline of METRIC Algorithm
METRIC is set by some training and manual image interpretation of a hot and cold pixel to determine the temperature difference (dT) between the surface and the lower atmosphere, which is assumed to be linearly correlated with land surface temperature throughout a scene. Coefficients a and b are determined by the characteristic variables for two anchor points selected from the satellite image. The hot anchor pixel is selected where bare land is assumed to be dry enough and ET could be assumed to zero. The cold anchor pixel is selected among the open water or well-watered, full- , rad cold b aT = − (10) The variables in Equation 9 and 10 are the corresponding variables in Equation 5 and 6 for hot or cold extremes. In this study, to reduce the arbitrariness caused by the subjectivity in the selection of hot and cold pixels, a method based on the trapezoidal framework of the temperature-vegetation index space (f c -T rad ) proposed by Long and Singh (2012) was introduced [31]. It is assumed that the coefficients a and b vary with fc but remain fairly invariant under the same fc or within the same f c class:  (10) and (11) are the corresponding variables for hot or cold extremes. Pixel with the same fc value or within class f c,i , use a i and b i to infer their temperature gradient dT i . The detailed calculation of f c -T rad space could be found in Long and Singh (2012) [31].

Evaluating the Performance of Continuous Daily ET Estimates from HJ-1
The continuous daily ET during the growth periods of 2015 were generated by the methodologies illustrated above. Totally 36 cloud-free days were selected to generate daily ET and ET 0 F. Then the simulated daily ET 0 F were employed to extrapolate ET for all cloudy days. Continuous daily ET estimates were compared with ground measurements at the Daman and Huazhaizi stations. Also, the outputs of METRIC including sensible heat flux (H inst ) and instantaneous latent heat flux (λET inst ) were compared with the ground measurements. Note that some observed fluxes could not be measured due to bad weather conditions or to instrument failure.
The scatter plots in Figure 2 show the comparison between measured and modelled instantaneous heat flux including H inst and λET inst . The values of the coefficient of determination (R2), root-mean-square error (RMSE), and mean absolute percentage error (MAPE) between the HJ1B-retrived and EC-observed H inst were 0.93, 15.63 Wm −2 , and 11.7% respectively. It was suggested a good agreement between simulated and measured H inst . However, METRIC underestimated H inst values when these were compared with the eddy covariance measured values and the slope for the linear regressions is 0.87 which is less than 1. R 2 , RMSE and MAPE between measured and modelled λET inst were 0.86, 24.17 Wm −2 and 13.8% respectively, indicating a good performance in estimating λET by HJ-1 data through METRIC. Corresponding to the underestimation of H inst , METRIC also overestimated λET inst values when compared with ground measurements. The underestimation of H and overestimation of λET compared with the ground measurements were confirmed and the slopes for line regression were 0.87 and 1.13. Li et al. (2015) and Huang et al. (2015) found the RS-based surface energy balance models may overestimate ET due to the underestimation of sensible heat flux in arid and semiarid regions where water limiting situation is prevailing [32,33]. Integrating the information of soil moisture into METRIC may be considered in the future. For the Daman station (cropland) in Figure 3(a), R2 is 0.95, indicating a strong correlation between estimated and measured daily ET. And MAPE and RMSE are 0.31 mmd −1 , and 12.04%, respectively, again suggesting good performance. The larger absolute difference (AD) occurred in the images of 6 June (1.85 mmd −1 of AD) and 8 August (2.14 mmd −1 of AD). The overestimation of latent heat flux (λET) and the missing 30-min data in EC system might explain the big errors. In addition, the simulated ET was overestimated compared with EC measurements during mid-September and late September when the field was covered by maize residue, Odhiambo and Irmak (2012) also reported similar overestimation results of ET because of soybean residues. Figure 3(a) indicates that the continuous daily ET method with HJ-1 data captured the similar dynamics of daily ET as field measurements. Daily ET showed an increasing trend after early May and a decreasing trend since September and varied from 0.27 mmd −1 to 7.51 mmd −1 . There was always an impermanent increase after rainfall and irrigation events in the maize growth period, such as days after rain . Rain events influenced ET because of the decreased water vapor pressure deficit (VPD), which caused temporal activation of the stomata, thus leading to the expedition of transpiration [34]. Rain and irrigation also increased the soil moisture, which would reduce soil resistance and soil water stress so that ET increased [16]. For the Huazhaizi station (sandy desert), R 2 , RMSE and MAPE were respectively 0.89, 0.16 mmd −1 and 13.2%, indicating a strong correlation between measured and modelled ET. Most of the time daily ET stayed fairly low and was less than or around 1 mm. The maximum of ET occurred after intensive rainfall from June 29 (DOY 181) to July 8 (DOY 190), during which there was intense rainfall.  ET0 derived from meteorological data was also shown here in black line.

Assessing the Performance of HJ-derived ET0F
Continuous daily ET 0 F were obtained through the cubic interpolation method with available HJ-1 images. The temporal patters of the variables with HJ-derived ET0F, NDVI, LAI and soil moisture (SM) of 10 cm were plotted in Figure 4. Compared with the Landsat-derived ET 0 F, HJ-derived ET 0 F could fit the characteristics of ET 0 F better. At the Daman Station, ET 0 F derived from HJ-1 showed an increasing trend since the beginning of the growing season (late April) and a decreasing trend since the starting of the non-growth season, especially the harvesting season in mid-September. The ET 0 F was around 0.    The mean and standard deviation of daily ET over different land use type from April was illustrated in Figure 6. The mean represented the average strength of ET, whereas the standard deviation illustrated the degree of its spatial variability in the experimental region. A smaller value of the standard deviation (short error bar) signified a more uniform spatial ET distribution. On the contrary, a larger value of the standard deviation (long error bar) means a more heterogeneous spatial ET distribution. Also, the spatiotemporal variability of ET for the whole experimental area can be seen more easily and clearly exhibited from the sum of daily ET for every month during crop growth period than day-by-day ET from April to September (Figure 7). The high daily ET values represented well-irrigated wheat parcels in April and May, when winter wheat was in the tillering or jointing stage and maize was in the seeding stage. As can be seen Figure 7, the pixels of wheat stayed an obviously higher value than maize. However, the mean values of daily ET for maize gradually increased from late April, the trend of which began to be different from that of bare land. In early-June, the spatial mean values of daily ET on wheat pixels reached up to 7.34mm, when winter wheat was in the heading stage. The daily ET for maize increased quickly and the spatial distribution of daily ET was discrete during vigorous growing jointing stages of maize, with the standard deviation can reach up to 1.01 mm (16.8% of the mean value). The pattern of daily ET in July is strongly linked to the crop planting structure. The high ET values occurred at well-irrigated maize parcels in mid-season growth stages. The spatial distribution of ET in July was more heterogeneous than that in June, with a standard deviation of 1.67mm (25.3% of the mean value). The high spatial variability in the distribution of ET reflects different stages of crop growth and soil moisture conditions over the studied region. The daily ET at pixels of wheat planting parcels drastically decreased in July. Even the pattern of daily ET for wheat appeared to be similar to that for bare land from late July according to Figure 8. This was because winter wheat was already harvested and maize was in the heading stage, with NDVI for wheat less than 0.2 and that for maize around 0.65. The spatial mean values of daily ET for maize stayed steadily at high value in August when maize was in the mature stage. Maize harvesting in September caused a rapid decline in mean of daily ET, to the mean of daily ET less than 2mm. Moreover, the trend of daily ET for maize and wheat was similar to that for bare land since September, staying at a low level (around 1mm). This indicated that daily ET for crops was closely related to soil condition during non-growing season. The spatial mean values of daily ET for water always stayed at absolute high level than other land use types. It showed an increasing trend since late April. Then a high evaporation intensity happened during July and August, with the peak up to 13.05 mm. Since September, the mean of daily ET for water also showed a decreasing trend. Conversely, bare land always stayed at a low evapotranspiration under low soil moisture condition and low vegetation coverage. But, ET for bare land would fluctuate with the increasing soil moisture after rain. Monthly spatiotemporal variation of ET over agriculture land during crop growth period can be seen in Figure 8. Similar seasonal dynamic trends were apparent for all plants in agriculture land, with ET increasing first to peak, and then decreasing. From April to May, all of the monthly ET values increased significantly as vegetation coverage. Wheat had the

Conclusions
A continuous daily ET at a high spatiotemporal resolution estimation is essential for growers in their efforts to apply the right amount of water at the right time. A method combing METRIC with f c -T rad space fed with HJ-1 data was used to convert satellites-retrieved ET on clear days into continuous time series daily ET. This study mapped continuous daily ET distribution in a representative area in the oasis of the HRB using the METRIC model and 36 HJ-1 images acquired during the growing season of 2012. The METRIC estimated surface energy flux and daily ET were compared with in situ measurements.
High accuracy was found by comparing the modelled results including instantaneous sensible heat flux (H), instantaneous latent heat flux (λET) and daily ET with the ground measurements. Our studies suggested that it was practicable and feasible to estimate continuous daily ET based on METRIC with f c -T rad space driven HJ-1 data. METRIC-based outputs agreed well with flux tower measured ET both instantaneously and at daily time step. High accuracy of daily ET was found during the growing season. Meanwhile, the fraction of potential ET (ET 0 F) was derived from the available clear sky HJ-1 overpass dates using a cubic interpolation technique. HJ-1 with visit cycle at 2-4 days is more capable of capturing the dynamic variation of ET 0 F than Landsat 8. The results demonstrated that HJ image data was very useful in estimating continuous daily ET for crops at the field scale. The present work supports the use of the remote sensing for simple, rapid, and relatively accurate estimation of daily ET at field scale and the water consumption for crops in every growth stage. Such accumulated values are critical, for example, for water use planning, irrigation scheduling and water management, especially in arid and semiarid areas.
However, error and uncertainty of HJ-derived estimates still exist. On the one hand, METRIC model itself has the uncertainty with variables/parameters in these equations varying with the domain scale, which would largely determine the mechanisms of error propagation and the magnitudes of errors in the resulting surface flux estimates. Second, optical remote sensing would always suffer from the clouds. When the HJavailable interval is more than 8 days, the fitted ET from HJ cannot catch the intensive change of ET caused by the rainfall and irrigation events, which often represent the change of weather and soil moisture conditions ( Figure 4). In future, while work is done to enhance model, a combination of HJ and other high temporal resolution satellite image data such as MODIS should be considered. Moreover, the main factors controlling spatial and temporal variation of ET such as rain, irrigation, temperature, relative humidity, wind speed and soil moisture should be further examined.