Estimating and forecasting daily reference crop evapotranspiration in China with temperature-driven deep learning models

Accurately estimating and forecasting short-term daily reference crop evapotranspiration (ETo) is crucial for real-time irrigation decision-making and regional agricultural water management. Although the Penman-Monteith formula shows high accuracy, the requirement for excessive meteorological factors by this formula restricts its practical application. Previous studies have developed many ETo estimation models using deep learning (DL) algorithm, which only require temperature data as input. Subsequently, temperature forecast data is used to drive these models for ETo forecasting. However, these models are often limited to the specific locations of their training sets due to significant climatic variations across regions. Besides, weather forecasts at different lead days typically exhibit different biases. It remains unclear whether train ETo forecasting models for different lead times will enhance the overall forecasting accuracy. Hence, in this study, we innovatively utilized an extensive array of weather forecast data to develop customized ETo forecasting models for each day of the next 15 days, while incorporating both location and seasonal features into the model training procedure. Five deep learning (DL) models were employed in this study, namely Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Networks Bi-LSTM (CNN-BiLSTM), and CNN-BiLSTM-Attention. The results revealed that the differences in the performance of estimating ETo among the DL models were less pronounced compared to the variations that existed between diverse training strategies. By integrating location and seasonality information into the training set, we found a notable improvement in the accuracy of ETo estimating, with the average Root Mean Square Error (RMSE) of the five DL models decreasing from 0.55 mm d−1 to 0.48 mm d−1. Furthermore, when we directly employed a larger volume of weather forecast data to train the models, the forecasting accuracy of ETo was significantly improved, and among the five DL models, GRU performs the best. Specifically, the RMSE values for the ETo forecasts made by GRU model for the 1st, 4th, 7th, and 15th days in the future have decreased from 0.70, 0.87, 1.00 and 1.33 mm d−1 to 0.51, 0.56, 0.61 and 0.67 mm d−1, respectively. Additionally, compared to previous studies, we have successfully extended the lead time of ETo forecasts from 7 days to 15 days. These results indicate that the ETo estim