Conventional ground-based ET measurements are inadequate for large-scale areas, while remote sensing models exhibit inconsistent accuracy over heterogeneous surfaces due to variations in physical mechanisms and parameterization schemes. This study, based on remote sensing data from 2012 to 2016, utilized the Surface Energy Balance Algorithm for Land (SEBAL), Surface Energy Balance System (SEBS), Penman-Monteith (PM) and Priestley-Taylor (PT) models to estimate evapotranspiration (ET) in the Heihe River Basin. After the multi-model results were fused and optimized using Triple Collocation (TC) and Maximum Likelihood Estimation (MLE) methods, the model accuracy was systematically evaluated across various spatial (site and basin) and temporal scales, as well as across diverse land cover types, with a further sensitivity analysis conducted under extreme heat conditions. The results showed that, the SEBAL achieved the highest accuracy (R = 0.76, RMSE = 1.18 mm/day), followed by PM (R = 0.74, RMSE = 1.28 mm/day) and SEBS (R = 0.72, RMSE = 1.34 mm/day), and the PT model had the lowest accuracy (R = 0.65, RMSE = 1.47 mm/day); Among the single models, SEBAL showed slightly higher accuracy, but each exhibited biases under specific land cover types or climatic conditions. To overcome this limitation, fusion models (TC_PET and MLE_PET) were developed, leveraging the strengths of individual models to achieve superior overall accuracy, the fusion models demonstrated significantly higher accuracy (R >= 0.79, RMSE < 0.9 mm/day) than single models. Spatially, ET exhibited a decreasing trend from upstream to downstream regions, with the highest values in the upper reaches and the lowest in the lower reaches. Under extreme high-temperature conditions, single-model accuracy significantly declined. Fusion models can effectively reduce such errors and uncertainties, and they achieve the most significant improvement in accuracy over single models in complex surface scenarios such as croplands and grasslands. The proposed fusion framework substantially enhances the reliability and robustness of ET estimation in complex environments, providing theoretical and technical support for water resource management and ecological conservation in arid/semi-arid regions.