Satellite-Based Innovative Methodological Procedure for Estimation of Daily Actual Crop Evapotranspiration in Vulnerable Agroecosystems: A Case Study in Castilla-La Mancha, Spain

1. Introduction

Evaporation from land surfaces and water bodies and transpiration by the plants constitute one of the most important factors of the water cycle. Evapotranspiration (ET) as a unique element presents the main losses from a hydrological system. Several types of ET can be considered; however, the most used today is the actual evapotranspiration (ETa) [1,2] defined as the ET that occurs under actual environmental and management conditions, which is inclusive of both standard and nonstandard conditions [3]. ET estimations are necessary in many hydrological applications, as well as agricultural water management, or drought identification. Based on the above, ET has a very important influence on water management, especially in semi-arid climate conditions as the study area which is described in session 3. As a result, the estimation of ET_a is generally very important, and it is based mainly on measurement procedures like Eddy Covariance system (EC), Lysimeters or Bowen ratio energy balance system [2]. Nevertheless, it is very common that all those procedures are not available in many parts of the world, and alternatives such as indirect methodologies are utilized [4]. Due to the difficulties of collecting good field observations, ET is generally calculated from weather data. A vast number of empirical or semi-empirical equations have been devised to estimate crop or reference crop evapotranspiration using meteorological data. Some approaches are limited to specific meteorological and agronomic conditions and cannot be applied under other conditions

On the other hand, with regards to cultivation, the crop evapotranspiration (ET_c) parameter is imported, using either clipped grass (ET_o) or alfalfa (ET_r) as a reference evapotranspiration value. The methodology relates then the crop coefficient (K_c) depending on the crop type as well as the different vegetation phases [3],([4]). Crop water productivity as stated in ([5]) is a very important step to connect water management with water and food security, as well as economic growth.

Remote sensing (RS) techniques being already a powerful tool for land surface monitoring can provide a continuous coverage of land surfaces and have been used for many decades for the estimation of ET_a worldwide [6,7] ([6])[5,7,8,9,10,11,12,13,14,15,16,17,18,19,20]. The most frequently used RS platforms are available mostly free of charge for every part of the world, where all the available spatial and temporal patterns of ET derived from remote sensing platforms. Most of these platforms suffer from the fact that they are not identical for use at the field level because of their spatial and/or temporal resolution [18,21]. The best practice is to use a more precise methodology, like that of the family of Energy balance models ([22]). There are several variations between energy balance models mainly due to the related spatiotemporal capabilities analyzing energy balance fluxes (EB) for different crop types. According to [23,24] land surface temperature-based energy balance models are the prevailed methodologies nowadays, and between them: (a) the Surface Energy Balance Algorithm for Land (SEBAL) [25,26]; (b) Surface Energy Balance System (SEBS) [27] and (c) Mapping EvapoTranspiration at high Resolution with Internalized Calibration (METRIC) [3], are the most common ones [28]. METRIC model is a modification of the SEBAL model. METRIC has been used worldwide in different crops and ET_a is computed as a residual of the surface energy balance equation using meteorological data and data derived from satellite images [28]. One of the main advantages of METRIC model is that it is suitable for both plain and high elevated terrain ([29]). Results showed a good correlation between the METRIC model results and the Penman–Monteith model [29]. Likewise, [30] showed that results from METRIC model were also in accordance with the Eddy-Covariance methodology. The work of [31] demonstrates a sensitivity analysis according to the terrain of the study areas, while an alternative approach to model surface energy fluxes using Land Surface Temperature (LST) and radiometric heat fluxes radiometric surface temperature is described in [33,34]. The study of [35]well describes the history of TSEB and its extensions ALEXI and disALEXI which nowadays are considered among the most powerful methodologies for ET computations. In addition, a very recent study [36]evaluates the two remote sensing Energy Balance models, METRIC and TSEB in two ecosystems, estimating ET_a and energy fluxes using the eddy covariance (EC) as a reference. The comparison showed a good agreement at both test sites located in central Europe, empowering the use of the specific methodologies for the estimation of ET_a values for addressing various practical hypotheses and challenges associated with water balance.

Another common practice is using RS from vegetation indices (VI). Generally, the estimation of ET is made using the K_c ET_o approach where the value for K_c is estimated from the VI, commonly is the NDVI, which is the NDVI_ET method [36,37]. Reyes-Gonzalez et al. [38]showed the appearance of a high correlation between NDVI and K_c reported by FAO-56.

Artificial Intelligence-based models are also considered an interesting alternative to the most common based surface energy balance models. Artificial Neural Networks (ANNs) have been successfully utilized in modelling reference evapotranspiration, i.e., ET_o [37,38,39,41] A very recent study shows that [23] there are also significant results using METRIC model as the validation procedure.

After considering all the available methodologies, and authors’ capabilities, the Two Source Energy Balance (TSEB) based on the EB methodologies has been selected for further processing in this study for the final estimation of ET. The idea is to use the combination of Sentinel-2 and Sentinel-3 imagery for the computation of ET based on the studies of [35,42] and modify the well described methodology using meteorological data from Weather Research and Forecast (WRF) model instead of the initially proposed ERA-5 [42]. The objective of this study is to estimate ET based on high-resolution (10 by 10 m) Sentinel data and to evaluate the ET values using observations from a large weighing lysimeter available at the study region.

The paper is organized as follows: a brief description of the study is presented in the “Introduction”, the procedure for the computation of ET has been described in “Materials and Methods” section, the study area follows, and finally, the results are analyzed and validated at the “Results and Discussion” section.

2. Materials and Methods

2.1. SEN-ET (Copernicus)

The basic idea used in this study is the utilization of Sentinel-2 and Sentinel-3 imagery for the computation of ET based on the studies of [35,43]. Specifically, TSEB is incorporated within the Sen-ET plug-in (https://www.esa-sen4et.org/) according to European Union’s Copernicus Earth Observation program [45]. Among relevant Copernicus platforms Sentinel-2 and Sentinel -3 satellites were utilized. Sentinel-2 mission is a pair of satellites [4][46]demonstrating multispectral shortwave observations with a spatial resolution of 10–60 m and a temporal resolution of 5 days. Sentinel-3 on the other hand contributes to the Thermal Infrared (TIR) wavebands, with a 1-km resolution and a 1-day temporal resolution [47]. Finally, the Data Mining Sharpener (DMS) tool is incorporated to improve the spatial resolution to a 20mx20m [48]. DMS methodology enhances spatial information from thermal bands using information retrieved from surface reflectance bands, which usually have a finer spatial resolution. The well described Copernicus’ methodology is modified at this study using meteorological data from Weather Research and Forecast (WRF) model instead of the initially proposed ERA-5 [42].

The reliability of the Sen-ET approach was initially validated with satisfactory results in Denmark, [43] leading to an open-source plugin for COPERNICUS’ SNAP Software under the name “Sen-ET approach” (https://www.esa-sen4et.org) [35,36,37,38,39,40,41,42]. Other studies using Sen-ET approach have been conducted in India [47] the Italian Alps [49], Spain and Tunisia [50,51]. The Sen-ET approach used in this study consists of thirteen steps presented in Table 1.

Figure 1. ET computation flow chart (modified from [35]).

Figure 1. ET computation flow chart (modified from [35]).

2.2. Use of WRF

In this study, WRF data are used instead of ERA5 (ECMWF Reanalysis 5) to improve local uncertainties of the model. Assuming that ERA5 has a fixed configuration, WRF can allow the user to select between various physics related themes (planetary boundary layer, radiation etc.) which can be adapted to the specific region of interest. The performance of the WRF model based on a large ensemble consisting of physics parameterizations and initial and boundary datasets for various events is already evaluated at a region like that of the study area [52].

Generally speaking, the choice between WRF and ERA5 can be based on the scale of the study. WRF improves the assessment of selected meteorological parameters at a local level compared to ERA5, especially in areas where a complex terrain is present [53,54,55]

2.3. Validation with the Use of Large Weighing Lysimeter ET Measurements

It is essential that even if remote sensing ET products can offer satisfactory spatial coverage, they need to be validated against more traditional computational methodologies. One of the most prevailing ground measurement methodologies is the use of weighing lysimeters, which have already been used for reference measurements in many studies around the world for measuring actual ET considering precipitation, irrigation and drainage [54,55]

Both data sets of ET_a (observations and estimations) were compared by using the lineal regression analysis (the interception, slope, coefficient of determination (R²)). In addition, the Root Mean Square Error (RMSE) and the Mean Bias Error (MBE) are introduced as the appropriate metrics for the validation process. The validation process can be characterized as successful when a low RMSE is found and an MBE is close to zero, while R² should be close to 1 [14,56]

3. Study Area

Albacete, Spain, is the study domain occupied 14,858 km². The region is in the southern part of the autonomous community of Castile–La Mancha (Figure 2), located at the South-east part of the Iberian Peninsula. It is characterized by relatively flat surfaces averaging 510 to 700 meters in altitude. It has semi-arid climate with annual precipitation below 350 mm and evapotranspiration above 1200 mm. In the region there is a strong agriculture sector with annual crops, such as Wheat, Barley, Alfalfa, Onion, Garlic, and Legumes and trees like Almont, Olive, Pistachio and Vineyard. According to the Köppen-Geiger classification [56]the climate in the study area is cold semi-arid (steppe; type “BSk”).

The lysimeter used at his study is installed in the center of a 10,000 m² plot of grass maintained in optimum growth conditions (Figure 3). There is an automated 15 × 12.5 m sprinkler irrigation system of total underground coverage used for overnight watering to avoid the high evaporative hours [57]. The entire plot is regularly irrigated and kept as near as possible to the reference standard conditions. In this way, the soil can be maintained close to field capacity between 0.10 and 0.15 m height. All the above conditions are according to FAO guidelines, and from this point all the related ET measurements can be assumed as ET₀ [57]. The lysimeter has been also used for similar studies with successful results [58,59].

4. Results and Discussion

Table 1 summarizes the results of ET_o using the proposed methodology for 2021 growing season where Sentinel images were available for the study region. ET values from lysimeters and FAO Penman-Monteith methodology are also presented at the table. From the simple linear relationship Equation 1 is derived:

ET0,model =0.8722* ET0,lysimeter

(1)

Equation (1) was based on previous work [60] proposing a linear approach (Figure 4).

The next step of the study is the validation of the model. For that reason, available Sentinel scenes from 2022 and 2023 were considered, selected from all the available clear days of that period. Table 2 illustrates the results for the validation period consistently with Table 1, while Figure 5 illustrates the above relationship.

After examining the statistics from the model (Table 3), MBE and RMSE provide reasonable but not satisfactory results. Initial modeling using Sentinel data for the year 2021 showed a high correlation between ET_o derived from satellite data and lysimeter measurements respectively, however, a direct 1:1 comparison showed a systematic underestimation of ET_o values, as evidenced by the negative mean bias error (MBE) of -1.61 mm/d and the root mean square error (RMSE) of 2.64 mm/d. For that reason, an interception is incorporated into the model, to avoid the bias, which is usually generated in the 1:1 approach. The new result producing equation (2), describes the approach differently (Figure 6):

ET₀,_model = 0.6173*ET₀,_lysimeter +2.0178

(2)

The new statistics are shown inat Table 4:

The results after the modification seem to be more reliable than the previous approach. Incorporating the interception value significantly improved Mean Bias Error. The Mean Bias Error (MBE) was reduced from -1.61 mm/d to -0,52 mm/d representing a 67,7% improvement of the model’s overall accuracy, reducing also the related underestimation. Furthermore, the Root Mean Square Error (RMSE) decreased from 2.64 mm/d to 2.48 mm/d. This result suggests that, while the model effectively detects the temporal behavior of evapotranspiration, the result requires calibration to match the actual measurements provided by the well-established large weighing lysimeter.

As shown in Table 1 and Table 2, the ET_o values derived from the FAO Penman-Monteith (FAO PM) method (following the guidelines in FAO56) correlate very well with the lysimeter data (R²=0.99). The proposed model, which uses high spatial resolution Sentinel images, cannot achieve the same level of accuracy as the FAO PM methodology. This is a very common situation in remote sensing models, where local environmental conditions (e.g., soil moisture conditions or temperature anomalies) may not be fully captured by the model parameters [50] The introduction of and interception in Equation 2 proved critical for the model’s validation, because the model was now able to account for the bias inherent in the firstly 1:1 proposed relationship. The statistical improvements were substantial: MBE was reduced to -0.52 mm/d, reducing systemic underestimation, while RMSE slightly decreased too. The modified model can be considered more physically represented by the relative data, considering atmospheric or sensor noise producing errors. The improved model increases the proposed methodology’s accuracy, bringing the estimating values closer to ground-based lysimeter data and/or the FAO Penmann-Monteith standard computations, but significant work has yet to be conducted for even better accuracy.

It was already stated [33] that one of the main disadvantages of using Sentinel-3 imagery is that LST sharpening from 1 km to 20 m causes larger deviations when extreme temperatures are evident. This is due to systematic overestimation over colder pixels and the relevant underestimation over the hottest pixels. This can be justified because LST downscaling approaches cannot describe the whole dynamic image captured under the finer resolution [61]. To minimize this issue, Landsat LST imagery has been utilized during the sharpening process [50]. With the advancement of technology, new Sentinel sensors and new WRF schemes will improve the existing anomalies, and the errors are expected to minimize improving the accuracy of the methodology.

5. Conclusions

This paper presents a new evapotranspiration (ET) estimation methodology based on high spatial resolution Sentinel-2/Sentinel-3 images for Albacete, central Spain area in 2021, validated with lysimeter measurements and the standard FAO Penman-Monteith ET_o equation. Initial results showed a high correlation (R²=0.75) between the proposed model and lysimeter measurements, but with systematic underestimation (MBE=--1.61 mm/d, RMSE=2.64 mm/d), which was corrected by introducing an interception in the proposed equation, critically improving the measurements (MBE=-0,52 mm/d, while RMSE=2,48 mm/d). Initially, a multiplicative adjustment factor of 0.90 was observed, suggesting a 10% systematic underestimation by the model. This approach is consistent with the recent methodology of [60], who utilized empirical adjustments factors ranging from 0.59 to 0.63 to cotton, maize and hay. The goal was to calibrate satellite-based ET estimates against ground observation to Greece and France which are Mediterranean countries with a similar climate to Spain. [60] suggested the 1:1 line, and this is the reason for the first part of this study. However, the subsequent incorporation of an interception allowed a more refined calibration, reducing MBE to -0.52 mm/d and improving the model compared to a simple linear scaling. As a result, a model optimization has been achieved, integrating a non-zero interception which significantly improved the linear regression. Similarly, bias reduction also occurred eliminating systematic underestimation, as well as error minimization and predictive reliability with a strong coefficient of determination (R²=0.75). The model, although not as accurate as FAO Penman-Monteith (R²=0.99), offers a significant advantage on a spatial scale for agricultural and hydrological applications.