Temporal Reconstruction and Machine‐Learning Emulation of a METRIC‐Derived Daily Actual Evapotranspiration Benchmark Using MODIS and Climatic Data in a Semi‐Arid Irrigation District

ABSTRACT Accurate daily actual evapotranspiration (ETa) is important for irrigation‐district water accounting, drought assessment, and seasonal water allocation in semi‐arid agricultural regions. This study developed a district‐scale machine‐learning framework to emulate a temporally reconstructed METRIC‐derived ETa benchmark for the Akarsu Irrigation District, Türkiye, using MODIS NDVI, land surface temperature and station‐based climatic variables. The benchmark was anchored by 79 Landsat/METRIC anchor observations. Between satellite acquisitions, the short‐grass reference evapotranspiration fraction (EToF) was linearly interpolated and combined with daily FAO‐56 reference evapotranspiration (ETo), yielding 1476 reconstructed daily values for 2021–2024. We evaluated five machine‐learning models and five predictor scenarios using random holdout, cross‐validation and leave‐one‐water‐year‐out (LOWYO) validation. A leave‐one‐satellite‐date‐out sensitivity analysis additionally quantified reconstruction uncertainty. The full climatic scenario with XGBoost achieved a random‐holdout R 2 of 0.880 and RMSE of 0.447 mm day −1 , whereas the LOWYO validation yielded mean R 2 = 0.686 and mean NSE = 0.503. The results demonstrate strong internal predictive performance but weaker interannual transferability, supporting the framework for continuous district‐scale ETa reconstruction, water accounting and operational assessment under comparable hydro‐climatic and cropping conditions, while highlighting the importance of temporally independent validation.