Comparison of deep neural networks for reference evapotranspiration prediction using minimal meteorological data

Abstract

Evapotranspiration, the process of water loss through surface evaporation and plant transpiration, is hugely influenced by numerous meteorological variables in different climatic zones. A precise and accurate prediction of reference-evapotranspiration (ET0), a crucial factor in the crop evapotranspiration estimation, facilitates efficient management of agricultural water supplies. This paper proposes an ET0 prediction method, employing minimal meteorological data, as well as exploring the potential of deep learning to learn the time series data pattern. In this study, four variants of a deep neural network model were developed using different feature combinations of two datasets of California Irrigation Management Information System (CIMIS) weather stations in California, USA for ET0 modeling and evaluated their predictive performance. The results showed that among the four deep neural network model variants (DnnV1, DnnV2, DnnV3, and DnnV6), the two input deep neural network, DnnV2 (RMSE =0.36 Millimeter/day and 0.52 Millimeter/day, R2=0.94 and 0.94) showed a comparable performance to the six input neural network, DnnV6 (RMSE =0.3 Millimeter/day and 0.43 Millimeter/day, R2=0.96 and 0.96). However, by considering the minimalism factor in the selection of input variables, we recommend DnnV2 for ET0 modeling.