Improved remote sensing reference evapotranspiration estimation using simple satellite data and machine learning

Abstract

Reference evapotranspiration (ET0) estimation is crucial for efficient irrigation planning, optimized water management and ecosystem modeling, yet it presents significant challenges, particularly when meteorological data availability is limited. This study utilized remote sensing data of land surface temperature (LST), day of year, and latitude, and employed a machine learning approach (i.e., random forest) to develop an improved remote sensing ET0 model. The model performed excellently in 567 meteorological stations in China with an R2 of 0.97, RMSE of 0.40, MBE of 0.00, and MAPE of 0.11 compared to the FAO-PM ET0; it also performed well globally, yielding an average R2 of 0.97 and RMSE of 0.43 across 120 sites in mid-latitude (20°-50°) regions. This model demonstrates simplicity, accuracy, robust and generalization, holding great potential for widespread application, especially in the large-scale, high-resolution estimation of ET0. This study will contribute to advancements in water resources management, agricultural planning, and climate change studies