Daily reference evapotranspiration estimation using sequential and ensemble learning approaches: A case study in a high-evaporation region

Study Region This study was conducted using observations of meteorological variables for daily reference evapotranspiration (RET) estimation across the state of Florida, United States ( 24 . 39 6 ∘ N − 31 . 00 0 ∘ N , 80 . 03 1 ∘ W − 87 . 63 4 ∘ W ) , covering the period from January 1 to December 31, 2018. Study Focus RET estimation is fundamental for many community planning activities, such as water-use permitting regulation and estimating agricultural irrigation demands. The Penman–Monteith 56 (FAO-56PM) method is considered the standard method for computing RET. This research proposes two artificial intelligence models, i.e., CatBoost and Long Short-Term Memory (LSTM), for daily RET estimation based on different input combinations: temperature-based, humidity-based, and hybrid-based. New Hydrological Insights for the Region The results demonstrate that the proposed models provide daily RET estimation across different high evaporative climatic regions of the study area. CatBoost achieved R 2 values ranging from 0.9569 to 0.9817 and RMSE values from 0.1023 to 0.0653 mm day −1 , while LSTM achieved R 2 values ranging from 0.9653 to 0.9897 and RMSE values from 0.0896 to 0.0470 mm day −1 . The RET values ranged from 0.76 to 6.72 mm day −1 , with an average of 3.74 mm day −1 across all regions of Florida, highlighting both models’ adaptability to heterogeneous land surface conditions. These findings provide valuable insights into water balance dynamics in high-evaporative regions, thereby supporting scientific evaluations of ecosystem resilience and hydrological modeling of surface and groundwater systems.