Daily field-scale estimation of actual evapotranspiration (ETₐ) in dryland agricultural systems remains challenging due to limited in situ observations, strong spatiotemporal variability in hydroclimatic forcing, and data constraints. This study develops a footprint-aware, data-driven framework for estimating daily ETₐ using readily available remote-sensing and meteorological variables in dryland agricultural fields in Colorado, USA. Harmonized Landsat Sentinel surface reflectance imagery was integrated with two-dimensional eddy covariance flux footprint climatology to extract vegetation indices consistent with the effective source area contributing to measured turbulent fluxes. Model performance was evaluated using two input configurations, a six-variable configuration and a parsimonious three-variable configuration, to assess tradeoffs between predictive accuracy and input dimensionality. Four machine learning models, random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), and long short-term memory (LSTM), were trained and evaluated using spatial and temporal blocking across years and instrumented fields. Feature attribution analysis identified a water variable representing short-term hydrologic memory (10-day antecedent precipitation) as the dominant control on ETₐ variability, followed by a vegetation variable (enhanced vegetation index) and an energy variable (incoming solar radiation). Using the six-variable configuration, the LSTM and RF models achieved coefficients of determination up to 0.78 with root mean square error as low as 0.38 mm day⁻¹. When reduced to three variables, the LSTM model retained comparable accuracy and had more-stable performance across years and sites than other models. Model feature selection accurately reflected the coupled roles of water availability, vegetation dynamics, and energy supply in regulating ETₐ while emphasizing the dominance of short-term precipitation in water-limited environments. Although model development relied on two eddy-covariance towers, the physically based input structure and footprint-aware feature extraction support broader applicability across dryland agricultural systems.
Lamichhane et al. (Mon,) studied this question.