Improving terrestrial evapotranspiration estimation across China during 2000–2018 with machine learning methods

• Five process-based ET models were integrated based on Gaussian process regression. regression was better than the other five machine learning methods. Integrated estimates more accurate currently available eight products. China terrestrial with a spatial and temporal resolution of 1 km 10 days is generated. Estimating evapotranspiration (ET) accurately at various scales crucial for understanding hydrological cycle water resource management. The have some uncertainties need to be further improved. In this study, six methods including random forests, support vector machine, (GPR), ensemble trees, general neural network, Bayesian Model Averaging, are applied evaluated improve estimation by integrating algorithms SEMI-PM, RS-PM, RRS-PM, MOD16, PMLv2. Then evaluations conducted eddy covariance flux observations 14 tower sites distributing in forest, shrub, wetland, grassland, cropland, as well balance-based basin scale. According multiple training, validation, testing, GPR method superior all Compared individual algorithms, can reduce root mean square error (RMSE) 0.45 mm day −1 (for SEMI-PM) ~0.81 PMLv2), coefficient determination (R 2 ) 0.061 PMLv2) ~ 0.33 MOD16), decrease absolute relative percent (RPE) 8.32% RS-PM) ~42.47% test data. At scale, results demonstrate that annual GPR-merged reliable = 0.88, RMSE 57.18 year , RPE −0.26%) has higher accuracy high-resolution products from single models. average across 2000–2018 estimated approximately 397.65 . More ground-based covering land types should collected update estimates. resultant product (ChinaET1km10days) produced https://doi.org/10.6084/m9.figshare.12278684.v5