Balancing performance and uncertainty in multi-source evapotranspiration fusion across the Qinling–Daba Mountains, China

Study region The Qinling–Daba Mountains, China. Study focus In heterogeneous mountainous regions, how weighting strategies, bias-correction references, and error dependence affect the accuracy, propagated uncertainty, and robustness of multi-source evapotranspiration (ET) fusion remains unclear. We integrated seven ET products and compared seven fusion schemes, including arithmetic averaging, linear and exponential-score weighting, and bias-corrected variants. Product performance and reliability were assessed using water-balance-derived ET (WB_ET), Kling–Gupta efficiency, and Triple Collocation-derived uncertainty. Parameters calibrated during the training period were evaluated across seven basins for monthly and seasonal accuracy, propagated uncertainty, seasonal dependence, and leave-one-out robustness. New hydrological insights for the region The ET products showed strong temporal but moderate spatial consistency. Product-mean relative uncertainty ranged from 8% to 21%, compared with 13%–16% among basins. Fused products achieved R values of 0.90–0.93, RMSE of 14–20 mm·month⁻¹, MAE of 9–15 mm·month⁻¹, and KGE mainly between 0.60 and 0.85. EWF provided the most balanced accuracy, whereas GBC schemes more effectively reduced BIAS, RMSE, and MAE. ABC-EWF yielded the lowest covariance-based relative propagated uncertainty (6.18%), although this did not consistently translate into higher validation accuracy. Accuracy gains were more stable in the dry season, while covariance-aware propagation produced systematically higher uncertainty than the independent-error assumption. LOOCV KGE deviations remained within ±0.04. Overall, fusion outcomes depended on weighting formulation, bias-correction reference, seasonal conditions, and inter-product error dependence.