Multi-spatial scale assessment and multi-dataset fusion of global terrestrial evapotranspiration datasets

Abstract. Evapotranspiration (ET) is an important component of the terrestrial water cycle, carbon cycle, and energy balance. Currently, there are four main types of ET datasets: remote sensing–based, machine learning–based, reanalysis–based, and land–surface–model–based. However, most existing ET fusion datasets rely on a single type of ET dataset, limiting their ability to effectively capture regional ET variations. This limitation hinders accurate quantification of the terrestrial water balance and understanding of climate change impacts. In this study, the accuracy and uncertainty of thirty ET datasets (across all four types) are evaluated at multiple spatial scales, and a fusion dataset BMA(Bayesian model averaging)-ET, is obtained using BMA method and dynamic weighting scheme for different vegetation types and non-common cover years among ET datasets. ET from FLUXNET as reference, the study recommends remote sensing– and machine learning–based ET datasets, especially Model Tree Ensemble Evapotranspiration (MTE) and Penman-Monteith-Leuning (PML), but the optimal selection depends on season and vegetation type. At the basin scale, land–surface–model–based ET datasets have less relative uncertainty compared to other types of ET. At the global scale, the uncertainty is lower in regions with larger ET, such as the Amazon, Central and Southern Africa, and Southeast Asia. The BMA-ET dataset accurately captures trends and seasonal variability in ET, showing a global terrestrial increasing trend of 0.21 mm·yr⁻¹ over the study period. BMA-ET has higher correlation coefficients and lower root-mean-square errors than most individual ET datasets. Validation using ET from FLUXNET as reference shows that correlation coefficients of more than 70 % of the flux sites exceed 0.8. Overall, BMA-ET provides a comprehensive, long-term resource for understanding global ET patterns and trends, addressing the limitation of prior ET fusion efforts. Free access to the dataset can be found at https://doi.org/10.6084/m9.figshare.28034666.v1 (Wu and Miao, 2024).