Deep learning for daily potential evapotranspiration using a HS-LSTM approach

Abstract

Accurate estimation of potential evapotranspiration (ET0) is important for the sound design of irrigation schedules, management of water resources, assessment of hydrological drought, and research on atmospheric variations. The present study proposed a novel deep learning (DL) approach for daily ET0 estimations with limited daily climate data: HS- LSTM. This approach was constructed based on a classic ET0 model and a long short-term memory neural network (LSTM). Specifically, the Hargreaves-Samani (HS) model was employed as the classic model, and the predictors were restricted to the daily maximum and minimum air temperature data. Ground truth data for ET0 were employed to train, validate, and test the models. Traditional machine learning (ML) algorithms comprising adaptive neuro-fuzzy inference system (ANFIS), genetic programming (GP), multi-gene genetic programming (MGGP), and one-dimensional CNN (1D-CNN), as well as the HS-ML models (HS-ANFIS, HS-GP, HS-MGGP, HS-1D-CNN), were also established and assessed for daily ET0 estimations. Compared to the other tested approaches, the errors of the HS-LSTM technique significantly decreased, demonstrating that the novel HS-LSTM approach significantly outperformed the other techniques beyond the study area (in Songliao Basin, Northeast China, which is a semi-humid zone with temperate continental climate). The developed models can then be used to estimate future ET0 with only air temperature forecasts, which can be readily obtained from public weather forecasts. The present study provides a new and promising strategy that can provide more accurate estimations of daily ET0 with limited meteorological data, along with significant implications for enhancing atmospheric research.