Global evapotranspiration simulation research using a coupled deep learning algorithm with physical mechanisms

Abstract

Evapotranspiration (ET) and actual evapotranspiration (AET) serve as critical parameters in the water vapour exchange between terrestrial surfaces and the atmosphere. ET denotes the theoretical maximum evapotranspiration achievable under ideal conditions, whereas AET represents the actual evapotranspiration observed, factoring in the constraints imposed by available water resources. Precise estimation of AET is imperative for the optimization of water resource management and the advancement of sustainable development initiatives. In recent years, deep learning techniques have been extensively utilized in AET estimation. However, traditional deep learning models often lack the incorporation of essential physical constraints. We proceeded to enhance the loss function of the temporal convolutional network (TCN) by taking into account the physical relationships that exist among soil water content (SWC), potential evapotranspiration (PET) and AET, thereby introducing a novel physically coupled deep learning model (AET, SWC after kernel principal component analysis, PET, TCN and AKP-TCN), and checked the rationality of the model with the FLUXNET 2015 dataset. These findings underscore that the AKP-TCN model exhibits heightened sensitivity to peak fluctuations in AET under the imposition of physical constraints. This approach notably enhances the precision of AET simulations in areas marked by complex and variable climatic conditions, such as the Mediterranean climate zone and Oceania, achieving determination coefficient (R²) values surpassing the threshold of 0.900. Compared to traditional models, which include long short-term memory (LSTM), convolutional neural networks (CNN) and TCN, the AKP-TCN delivers substantial R² improvements of 16%, 16% and 9%, respectively. This advancement offers a novel perspective for coupling deep learning with physical mechanisms.