A novel surface reflectance calibration approach to improve the estimation of maize actual evapotranspiration in a semi-arid USA region

This study aimed to develop a novel surface reflectance (SR) data calibration approach to improve maize actual crop evapotranspiration (ETa, mm h−1) predictions in agricultural settings. The remote sensing (RS) platforms used in the study included Landsat-8 (30 m spatial resolution), Sentinel-2 (10 m), Planet CubeSat (3 m), a handheld radiometer (1 m), and a small uncrewed aerial system or sUAS (0.03 m). A one-source surface energy balance (OSEB) was applied to estimate ETa and assess any improvements when using original and calibrated SR data. Five machine-learning (ML) algorithms were evaluated as part of the calibration process. The ML approaches were linear regression, regression tree, random forest, support vector machine, and Gaussian process regression. The SR calibration approach incorporated a novel dual-source pixel decomposition model that accounts for the contributions of soil and vegetation light reflectance captured by a given SR band. Two maize research sites, one in Greeley and another in Fort Collins, Colorado, U.S.A. provided the grounds for data collection over a four-year period. Results indicated that the best ML model, for a given RS sensor, depended on the SR contributions from plants and bare soil. Among the evaluated algorithms, the Regression Tree and Gaussian Process Regression provided the most accurate pixel decomposition and reflectance adjustments, consistently demonstrating the highest statistical agreement (R2 = 0.70–0.94) and lowest residual errors (RMSE = 0.004–0.022). Improvements in maize hourly ETa had a 34% error reduction in maize ETa estimation for data from the handheld radiometer and sUAS, compared to a 12% error reduction for the spaceborne sensors, on average.