Innovative application of the composite Bezier GSXG hybrid machine learning model for daily evapotranspiration Estimation implementing satellite image data

Abstract

Estimating reference evapotranspiration (ET₀), a vital hydrological parameter, is particularly challenging in regions with scarce meteorological data, such as developing countries. Remote sensing data is a valuable resource for obtaining climatic and vegetation parameters. By using MODIS data (LST and NDVI), we aim to improve ET0 estimation accuracy. Four interpolation methods (spline, cubic spline, Bezier, and composite Bezier) are used to enhance the temporal resolution of MODIS data for improved daily ET₀ estimation. Conducted at the Yazd station in Iran, using data from 2003 to 2024, this study implements the traditional XGBoost (eXtreme Gradient Boosting) model and its optimized variant, GSXG (GridSearch- XGBoost), which incorporates GridSearch for superior parameter tuning. The results demonstrate the GSXG model’s significant performance enhancements over the base XGBoost, with the Bezier function achieving an RMSE of 0.855 mm/day and R² of 0.531 using only remote sensing data, and the cubic spline method reaching an RMSE of 0.208 mm/day and R² of 0.972 when combining meteorological and remote sensing inputs. These findings underscore the potential of GSXG to minimize errors and improve predictive reliability. This study demonstrates the value of integrating remote sensing data with optimized machine learning for improved ET₀ estimation, providing a valuable approach for hydrological assessments in data-scarce regions.