DNN-MET: A deep neural networks method to integrate satellite-derived evapotranspiration products, eddy covariance observations and ancillary information

Accurate estimates of the spatiotemporal distribution of evapotranspiration (ET) are essential for understanding terrestrial energy, carbon and water cycles. Station-based observations are limited for their spatial coverage whereas satellite-derived ET products exhibit large discrepancies and uncertainties. Here we presented a Deep Neural Networks based Merging ET (DNN-MET) framework that combines information from satellite-derived ET products, eddy covariance (EC) observations and ancillary surface properties to improve the representation of the spatiotemporal distribution of ET, especially in data-sparse regions. DNN-MET was implemented over the Heihe River Basin (HRB) from 2008 to 2015, and the performance of DNN-MET and eight input state-of-the-art satellite-derived ET products (i.e., MOD16, ET-SEMI, ET-JPL, ET-MS, ET-HF, GLEAM, ETMonitor and EB-ET) was evaluated against observations from 19 EC flux tower sites. The results showed that DNN-MET improved ET estimates over HRB, and decreased the RMSE by 0.13 to 1.02 mm/day (14%-56%) when compared with eight products. DNN-MET also yielded superior performance compared to the products derived by other merging methods (i.e., Random Forest, Bayesian model averaging and a simple averaging method). When DNN-MET was validated for data-scarce regions, its performance remained better even when the training samples were decreased to 20% of the available EC sites. An innovation of our approach is by building a multivariate merging model with ancillary surface properties, DNN-MET incorporated geographical proximity effects and spatial autocorrelations into merging procedure, which can be used as a “spatial knowledge engine” to improve ET predictions. The approach can be readily and effectively applied elsewhere to improve the spatiotemporal representation of various hydrometeorological variables.