Reconstruction of the pan evaporation based on meteorological factors with machine learning method over China

Measurements of evaporation from pans have traditionally been used to represent the evaporative demand of the atmosphere when estimating the crop water requirements. In China, Pan evaporation (Epan) has been observed routinely at meteorological stations since the 1950 s with D20 pans, but since 2002, the pans have been replaced by E-601B. To explore the effective reconstruction of missing daily D20 Epan over China from 1951 to 2020, this study employed three types of Epan models: the widely used physical model PenPan, two popular machine learning (ML) models (multivariate adaptive regression splines (MARS) and random forest (RF)), and multiple linear regression (MLR). Daily Epan data were predicted based on the daily wind speed (U), atmospheric pressure (AP), relative humidity (Rh), air temperature (Ta), and sunshine hours (n) of 2410 meteorological stations. The results showed that the MARS and RF predictions were superior to those of PenPan, and the results of MLR were the worst. The average determination coefficient for RF, MARS, PenPan, and MLR values were 0.95, 0.91, 0.88, and 0.86, respectively, and the average root-mean-square difference were 0.62, 0.91, 1.17, and 1.15 mm day−1, respectively. Thus, the missing daily Epan were predicted using RF and the reconstructed Epan had the same probability density function as the observed Epan. The annual Epan first showed a downward trend (at a rate of 6.17 mm yr−1) from 1961 to 1993 and then a reverse upward trend (at a rate of 1.84 mm yr−1) from 1994 to 2020. Epan predictions by PenPan are limited by regional characteristics, making it difficult to transfer between regions. However, ML methods are less affected by regional characteristics and can be used across regions. Furthermore, ML methods can effectively reconstruct missing Epan providing support for verification of PenPan, which is beneficial for the study of driving factors of Epan.