Forecasting Reference Evapotranspiration Using LSTM and Transformer

Reference evapotranspiration (ET) helps manage irrigation water supply, minimize freshwater waste, maximize crop yields, and promote sustainable agriculture. However, machine learning (ML)-based ET forecasting models are only effective and accurate if localized weather/climatic data are available, which is not always the case in most cases, especially in many parts of Africa. In this study, we obtained a historical climate record of 23 years from the gridded AgERA5 dataset and calculated ET using the Penman-Monteith (PM) method. Thereafter, two different ET prediction algorithms (transformer and long short-term memory network (LSTM)) were trained using the AgERA5 to overcome the challenges caused by limited localized data. According to the results, LSTM outperformed the transformer in terms of forecasting ET, with lower mean absolute error (MAE) values, root mean square errors (RMSE), and a higher correlation coefficient (R2). Therefore, the proposed approach demonstrates how gridded climate observations can be integrated with soft computing techniques for precise ET forecasting within areas that lack access to climate records from local meteorological stations.