Nation-scale reference evapotranspiration estimation by using deep learning and classical machine learning models in China

Accurately estimating the reference evapotranspiration (ET0) is a basic requirement for precision irrigation and the correct planning of regional water resources. This study aimed to investigate the spatiotemporal variations in ET0 in China and to improve the accuracy of ET0 calculations on different spatiotemporal scales. Meteorological data collected at 100 stations in China during 1961 to 2019 were used to calculate ET0 with the Penman–Monteith model, and the temporal and spatial patterns in ET0-PM were analyzed with the Mann–Kendall nonparametric trend test method. Three machine learning models comprising convolutional neural network (CNN), extreme learning machine (ELM), and multiple adaptive regression splines (MARS), and seven empirical models calibrated with mind evolutionary algorithm (MEA) were compared to assess their suitability for calculating ET0 on different spatiotemporal scales in China. The results showed that the annual mean ET0-PM value (413.29–2772.35 mm) in China gradually increased from north to south and from west to east. ET0 exhibited an upward trend in the temperate continental zone (TCZ) and mountain plateau zone (MPZ) but a downward trend in the temperate monsoon zone (TMZ) and subtropical monsoon region (SMZ). By comparing the global performance indicators (GPI), the machine learning models generally performed better than the empirical models at different spatiotemporal scales. And CNN was the best model for calculating ET0 in terms of the model accuracy and stability. On the daily scale, MARS performed well in MPZ, whereas ELM performed well in TMZ and TCZ. On the monthly scale, MARS performed well in TMZ, whereas ELM performed well in SMZ and MPZ. At the annual scale, the accuracy of ELM was higher than that of MARS.