Testing the performance of a simplified distributed model to assess actual evapotranspiration in a Mediterranean orchard using ground and remotely sensed data

Accurate estimations of actual crop evapotranspiration are essential to evaluate crop water requirements, to improve water use efficiency in agriculture, and to optimize the use of available freshwater resources. To this aim, several models were developed to allow quantifying crop water requirements based on the knowledge of actual crop evapotranspiration rates, ET_a.

The objective of this research was to estimate ET_a using a simplified distributed model combining ground and remotely sensed data.

The experiment was carried out in a Mediterranean commercial citrus orchard (C. reticulata cv. Tardivo di Ciaculli) located in the Northwest of Sicily, Italy, during the whole 2019. The experimental layout consisted of: i) a WatchDog 2000 standard weather station (measuring the main climate variables and the precipitation depths, P); ii) a database of irrigation volumes, I, scheduled by the farmer; iii) an Eddy Covariance tower equipped with an open patch gas-analyzer, a three-dimension sonic anemometer, a four-component net radiometer, and a soil heat flux plate iv) a dataset of 75 Sentinel-2 multispectral images, acquired in clear sky condition.

In particular, the daily crop reference evapotranspiration, ETo, was calculated according to the FAO-56 Penman-Montheith equation using the climate variables; the crop coefficient, K_c, the Fractional Vegetation Cover, FVC, and, thus, the potential evapotranspiration, ET_p, were computed via the processing of reflectance values in the RED, NIR and SWIR spectral bands. The Available Water, AW, the short-term water stress factor, Cws, and the ET_a, were computed by analyzing cumulated ET_p and water-supplying values using moving temporal windows characterized by different sizes (from 5 to 400 days).

The validation of the model outputs was carried out by taking into account the ET_a of the pixels within the flux tower footprints estimated at each satellite acquisition day (i.e. by selecting the pixels on the basis of the footprint shape and extension). The performance of the model was evaluated for each temporal window size using the following metrics: the Root Mean Square Error, RMSE, the Mean Absolute Error, MAE, the angular coefficient of the regression line forced to the origin, b, and the determination coefficient, R².

Results suggest that the best temporal window size for this crop is around 85 days allowing to achieve an RMSE of 0.51 mm d^-1, a MAE of 0.38 mm d^-1, a b value of 0.94 and an R² of 0.96. The comparison with the model outputs over the whole field (all the pixels within the crop field) revealed that a strong decrease in all the metrics occurs if the validation of the remote sensing products is not properly carried out.