H-ConvLSTM to Estimate Reference Evapotranspiration From Air Temperature and Relative Humidity

The estimation of evapotranspiration reference is significant importance in the agricultural sector, as it allows for the precise determination of irrigation water distribution times. A few deep learning architecture models are employed by researchers in the estimation of evapotranspiration reference (ETo), including Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). These three algorithms have been extensively evaluated and validated due to their ability to predict and estimate ETo data using temperature (T), relative humidity (RH), and solar radiation (Rs) as variables. This paper presents a comparison of the results of the three algorithms using only two variables, namely temperature and relative humidity, without the inclusion of solar radiation variables. This paper presents a novel approach to compare the ConvLSTM algorithm with a newly developed algorithm, called Hybrid Convolutional Neural Network Long Short-Term Memory (H-ConvLSTM). Furthermore, this study employs a minimal set of variables, comprising temperature and relative humidity. This study aims to facilitate the computational process on devices with limited processing capabilities. The results of the comparison demonstrate that the determinant coefficient values for RNN, LSTM, GRU, and ConvLSTM are 0.8354 (84%), 0.8506 (85%), 0.8496 (85%), and 0.8536 (85%), respectively. In contrast, the new architecture of the proposed algorithm exhibits a value of 0.8942 (89%), which is superior to that of other algorithm architectures. The objective of this paper is to serve as a reference for determining the reference value of evapotranspiration with limited device computation. Additionally, this paper presents the results of measuring the error comparison of each algorithm utilized.