Estimation and mechanism analysis of global evapotranspiration based on a physics-informed deep-learning model

초록

As a key component of the water cycle, evapotranspiration (ET) plays a critical role in agricultural management and climate prediction. Several machine learning approaches have been developed for ET estimations based on in-situ observations. However, most proposed models focus on the accuracy competition in the same domain and neglect the extrapolation of cross-domain estimation, which training and testing dataset are from the same region. This paper proposed a physics-informed deep-learning model, namely Self-attention Influence (SAI) method which coupled environment cognition and parameter calibration learning. Compared to pure machine learning methods, SAI model had the advantages of spatial extrapolation, environmental adaptability and estimation robustness at both site- and basin-scale. At site-scale, the Root Mean Square Error (RMSE) was reduced by 25.15 % and the Coefficient of Determination (R2) was increased by 38.35 % in the best case over data-poor regions (e. g. South America and Africa). At basin-scale, the SAI model showed the lowest estimation uncertainty (2.79 mm/ month) in data-poor basins. Based on the water balance method, the results indicate that the SAI model had the highest estimation accuracy (20.97 mm/month of centered RMSE). Hence, a collaborative breakthrough of accuracy-adaptability-explainability is realized to overcome data imbalances problem over region and the globe. The explainable method was further used to analyze ET mechanism, indicating the physical consistency of physics-informed embedded in SAI model for model's spatial extrapolation. Finally, the effects of El NinoSouthern Oscillation and temperature on ET were revealed on a global scale from 2000 to 2021.

키워드