Determining actual evapotranspiration based on machine learning and sinusoidal approaches applied to thermal high-resolution remote sensing imagery in a semi-arid ecosystem

Abstract

Evapotranspiration (ET) is key to assess crop water balance and optimize water-use efficiency. To attain sustainability in cropping systems, especially in semi-arid ecosystems, it is necessary to improve methodologies of ET estimation. A method to predict ET is by using land surface temperature (LST) from remote sensing data and applying the Operational Simplified Surface Energy Balance Model (SSEBop). However, to date, LST information from Landsat-8 Thermal Infrared Sensor (TIRS) has a coarser resolution (100 m) and longer revisit time than Sentinel-2, which does not have a thermal infrared sensor, which compromises its use in ET models as SSEBop. Therefore, in the present study we set out to use Sentinel-2 data at a higher spatial-temporal resolution (10 m) to predict ET. Three models were trained using TIRS' images as training data (100 m) and later used to predict LST at 10 m in the western section of the Copiapo Valley (Chile). The models were built on cubist (Cub) and random forest (RF) algorithms, and a sinusoidal model (Sin). The predicted LSTs were compared with three meteorological stations located in olives, vineyards, and pomegranate orchards. RMSE values for the prediction of LST at 10 m were 7.09 K, 3.91 K, and 3.4 K in Cub, RF, and Sin, respectively. ET estimation from LST in spatial-temporal relation showed that RF was the best overall performance (R-2 = 0.710) when contrasted with Landsat, followed by the Sin model (R-2 = 0.707). Nonetheless, the Sin model had the lowest RMSE (0.45 mm d(-1)) and showed the best performance at predicting orchards' ET. In our discussion, we argue that a simplistic sinusoidal model built on NDVI presents advantages over RF and Cub, which are constrained to the spatial relation of predictors at different study areas. Our study shows how it is possible to downscale Landsat-8 TIRS' images from 100 m to 10 m to predict ET.