Simulation of forest evapotranspiration based on Attention-LSTM model

Abstract

Forest evapotranspiration is one of the important components of terrestrial evapotranspiration, and it is also the key link of hydrological cycle and energy cycle in forest ecosystem. Accurate simulation of forest evapotranspiration is of great significance for the management and effective utilization of forest water resources and the sustainable development of forestry. The 3-hour evapotranspiration of Luya Mountain Nature Reserve of Shanxi Province was calculated in this study, which was based on the six environmental parameters of air temperature, relative humidity, water vapor pressure, wind speed, net radiation and soil heat flux, along with the latent heat flux of eddy covariance system of the study site. Then, the accuracy of support vector machine (SVM), random forest (RF) and long short-term memory (LSTM) were compared. In addition, the LSTM model was optimized by using Attention Mechanism to explore the optimal model for simulating the forest evapotranspiration. The results showed that: compared with SVM model and RF model, LSTM model had the best simulation performance on forest evapotranspiration, with Root Mean Square Error (RMSE) of 0.137, Mean Absolute Deviation (MAE) of 0.073 and coefficient of determination (R2) of 0.735. The RMSE, MAE and R2 of the Attention-LSTM model were 0.110, 0.055 and 0.832, respectively. Compared with the LSTM model, the RMSE, MAE and R2 of the Attention-LSTM model were reduced by 19.71%, 24.66% and 13.20%, respectively. Attention-LSTM can be used as an accurate model to simulate forest evapotranspiration and provide a reference for water cycle prediction in forestry management of Shanxi Province. Attention-LSTM can be used as an accurate model to simulate forest evapotranspiration and provide a reference for water cycle prediction in forestry management of Shanxi Province.