Improving actual evapotranspiration estimates through an integrated remote sensing and cutting-edge machine learning approach

Recent technological advances have allowed the production of many studies on evapotranspiration, resulting in improvements in reference evapotranspiration estimates and crop coefficients with remote sensing data. However, these two areas of research often work independently, producing valuable studies, but without an effective integration to predict actual evapotranspiration directly, without the need for weather stations. Thus, this study aimed to model actual evapotranspiration in sugarcane crop using machine learning techniques, independently of weather stations and thermal sensor data. To achieve this goal, data from the OLI (Operational Land Imager) and TIRS (Thermal Infrared Sensor) sensors aboard the Landsat-8 and 9 satellites were used to produce the variable observed from the METRIC model, and data from the Sentinel-2A and 2B satellites, NASA POWER, WorldClim and astronomical variables, latitude, elevation, day of the year and month were used to generate the explanatory variables and feed 13 machine learning models for three different biomes: Atlantic Forest, Caatinga and Cerrado. The results indicated that the brnn (Bayesian regularized neural networks) model with R2 and RMSE of 0.73 and 1.10, respectively, and the XgbLinear (extreme gradient boosting - linear method) model, which obtained values of 0.74 and 1.25 for these metrics, in that order, showed the best overall performance. Specific analyses indicated that brnn was superior for cultivated areas in the Atlantic Forest and Caatinga biomes, while XgbLinear was superior in the Cerrado biome. These results show that machine learning algorithms are able to predict actual evapotranspiration without the need for using weather stations and thermal data.