Estimating Regional Evapotranspiration by Combining Tower-based observations and Remotely Sensed Data

Five upscaling models based on five different machine learning methods (regression tree, random forest, artificial neural networks, support vector machine, depth belief network) are constructed to estimate regional evapotranspiration (ET) by combining tower-based observations and remotely sensed data. The upscaling models are trained and tested based on the site observation data (ET observations from eddy covariance instruments and meteorology observations from automatic weather stations) accumulated in the observational experiments carried out in the Heihe River Basin in recent years. The results indicate that the random forest based model is the best upscaling model with lowest RMSE and MAPE values compared with in-situ observations. With the best upscaling model (with random forest method), the ET products (ETMap) over Heihe river basin are generated based on regional reanalysis data and remotely sensed observations. The ETMap products are evaluated with ET observations from large aperture scintillator (LAS), and indicate the average RMSE of ETMap is 0.86 mm/d, MAPE is 24.8% with no systematic deviation, and the temporal trend is reasonable.