蒸散发是地表水热碳平衡的重要组成部分，是陆-气相互作用的重要纽带，也是区域和全球尺度气候变化的重要驱动因素。无人机遥感技术获得的甚高分辨率数据使得表征亚田块尺度蒸散发的大小与变化成为了可能，开展长期、大范围无人机遥感飞行实验对农田灌溉决策和植被生长状况监测等有着重要的意义。本研究基于2019-2020年黑河中游三个实验区植被生长季飞行实验获取的无人机遥感数据，结合站点数据实现了地表阻抗参数（空气动力学粗糙度z_0m和热传输附加阻尼kB^-1）的数据驱动估算，结合CFD模型实现了气象驱动数据的降尺度；从地表阻抗参数化方案和气象驱动数据降尺度角度着手优化了SEBS模型，精准估算了研究区域甚高分辨率ET，评估了不同组合方案的优化效果，探索了模型精度的影响机制，并开展了无人机遥感估算ET空间分布特征的精细分析。主要结论如下：

（1）根据站点数据推算了地表阻抗参数（空气动力学粗糙度z_0m和热传输附加阻尼kB^-1），对不同下垫面z_0m的季节变化特征以及kB^-1的日、季节变化特征进行了分析，并优选了XGBoost机器学习方法来估算z_0m和kB^-1。在站点尺度上，XGBoost方法估算的z_0m与观测值的RMSE为0.03 m，R²为0. 929。数据驱动方案1的XGBoost方法估算的kB^-1结果与观测值的RMSE为2.372，R²为0.619；数据驱动方案2的XGBoost方法估算的kB^-1结果与观测值的RMSE为1.978，R²为0.75。结合无人机遥感数据和优选的机器学习模型，获得区域尺度甚高分辨率z_0m和kB^-1。在区域尺度上，z_0m验证结果表明：数据驱动的z_0m估算值与站点推算值的RMSE 为0.034m，MRE为2.24%，R²为0.878。其中花寨子站精度最佳，大满站略微高估，张掖湿地站稍有低估。kB^-1验证结果表明：数据驱动方案1的kB^-1估算值的RMSE为2.406，MRE为70.351%，R²为0.464；数据驱动方案2的kB^-1估算值的RMSE为3.234，MRE为92.134%，R²为0.496；SEBS模型原参数化方案的kB^-1估算值的RMSE为4.576，MRE为77.654%，R²为 0.098。数据驱动的kB^-1比SEBS模型原参数化方案的估算效果改善较为明显。总之，数据驱动的z_0m和kB^-1估算值较为准确地捕捉了地表阻抗参数的大小与空间分布特征。

（2）本研究提出了6种SEBS模型估算方案（4个优化组合、2个对照组合），使用这6种估算方案分别生产区域尺度甚高分辨率地表蒸散发产品并开展验证。基于 EC 观测数据验证结果表明：“ML1+Site”方案估算的潜热通量与EC观测数据的RMSE为39.02W/m²，MRE为1.353%，R²为0.96；“ML2+Site”方案估算的潜热通量与EC观测数据的RMSE为54.16W/m²，MRE为-4.684%，R²为0.928；“PHY+Site”方案估算的潜热通量与EC观测数据的RMSE为75.77 W/m²，MRE为-6.541%，R²为0.863；“ML1+CFD”方案估算的潜热通量与EC观测数据的RMSE为83.99W/m²，MRE为12.782%，R²为0.79，R²为0.6；“ML2+CFD”方案估算的潜热通量与EC观测数据的RMSE为82.11W/m²，MRE为21.001%，R²为0.784；“PHY+CFD”方案估算的潜热通量与EC观测数据的RMSE为97.77W/m²，MRE为20.889%，R²为0.704。从验证结果可知：“ML1+Site”方案估算的潜热通量和感热通量精度均为最优。数据驱动的地表阻抗参数化方案可以有效改善SEBS模型估算效果，而使用CFD降尺度气象数据的3个方案由于WRF大气驱动数据的不确定性，估算精度比使用站点气象数据的3种方案偏低。与同样使用站点气象数据的方案相比，“Wei”方案可以在一定程度上提高SEBS模型的估算精度，但由于未对z_0m参数化方案进行优化，其精度仍比数据驱动方案差。OMS验证结果表明：在更大空间范围内，使用“ML2+Site”方案可以有效改善感热通量的估算精度。

验证结果也表明：改进z_0m和z_0h（kB^-1）参数化方案均可有效地提高模型精度，且改进z_0m参数化方案的优化效果更佳。此外，区域尺度甚高分辨率ET空间分布特征对比分析结果表明：使用站点气象数据估算的甚高分辨率ET空间分布主要受土地利用类型影响；使用CFD降尺度的高分辨率气象数据估算的甚高分辨率ET能够更真实地刻画下垫面水热状况的空间分布，尤其在中高度异质性下垫面上空间分布格局改善效果更为明显。因此，在均质下垫面使用站点观测气象数据足以表征气象条件的真实情况，而在中高度异质性下垫面则需使用CFD降尺度的高分辨率气象数据才能真实地反映气象要素对ET空间分布的影响。

Evapotranspiration is an important component of the surface water-heat-carbon balance, an important link in land-atmosphere interactions, and an important driving factor for regional and global climate change. The very high-resolution data obtained by drone remote sensing technology makes it possible to characterize the size and variation of evapotranspiration at the sub-field scale. Long-term and large-scale drone remote sensing flight experiments are of great significance for farmland irrigation decision-making and vegetation growth monitoring. This study is based on drone remote sensing data obtained from three experimental areas throughout the vegetative period, within the central segment of the Heihe River from 2019 to 2020, combined with site data to achieve data-driven estimation of surface impedance parameters (aerodynamic roughness z_0m and thermal transmission additional damping kB^-1), and combined with CFD models to achieve downscaling of meteorological forcing data. From the perspective of surface impedance parameterization schemes and meteorological forcing data downscaling, the SEBS model was optimized to accurately estimate the very high-resolution ET in the study area, evaluate the optimization effects of different combinations, explore the impact mechanism of model accuracy, and conduct a detailed analysis of the spatial distribution features of ET estimated by Unmanned Aerial Vehicle(UAV) remote sensing. The main conclusions are as follows:

(1) Based on the site data, the surface impedance parameters (aerodynamic roughness z_0m and thermal transmission additional damping kB^-1) were estimated, and the seasonal variation characteristics of z_0m and the daily and seasonal variation characteristics of kB^-1 were analyzed for different underlying surfaces. The XGBoost machine learning method was optimized to estimate z_0m and kB^-1. At site scale, the RMSE and R² of the XGBoost method for estimating z_0m were 0.03 m and 0.929, respectively. The RMSE and R² of the XGBoost method for estimating kB^-1 were 2.372 and 0.619, respectively. The RMSE and R² of the XGBoost method for estimating kB^-1 were 1.978 and 0.75, respectively, for data-driven scheme 2. Combining the UAV remote sensing data and the optimized machine learning model, the regional very high-resolution z_0m and kB^-1 were obtained. At rigional scale, the z_0m verification results showed that the RMSE of the data-driven z_0m estimation value was 0.034 m, with a MRE of 2.24% and an R² of 0.878. Among them, Huazhaizi station had the best accuracy, with a slightly overestimated value at Daman station and a slightly underestimated value at Zhangye Wetland station. The kB^-1 verification results showed that the RMSE of the data-driven kB^-1 estimation value was 2.406, with a MRE of 70.351% and an R² of 0.464 for data-driven scheme 1, and the RMSE of the data-driven kB^-1 estimation value was 3.234, with a MRE of 92.134% and an R² of 0.496 for data-driven scheme 2. The RMSE of the original parameterization scheme of the SEBS model for kB^-1 was 4.576, with a MRE of 77.654% and an R² of 0.098. The estimation result of data-driven kB^-1 improved significantly compared to the original parameterization scheme of the SEBS model. Overall, the data-driven z_0m and kB^-1 estimation values accurately captured the size and spatial distribution characteristics of the surface impedance parameters.

(2) This study proposes six SEBS model estimation schemes (four optimized combinations and two control combinations). Using these six estimation schemes, we produced regional scale very high-resolution evapotranspiration products and conducted verification using EC observation data. The results indicate that the RMSE of the latent heat flux estimated by the "ML1+Site" scheme is 39.02W/m², with a MRE of 1.353% and an R² of 0.96. The RMSE of the latent heat flux estimated by the "ML2+Site" scheme is 54.16W/m², with a MRE of -4.684% and an R² of 0.928. The RMSE of the latent heat flux estimated by the "PHY+Site" scheme is 75.77 W/m², with a MRE of -6.541% and an R² of 0.863. The RMSE of the latent heat flux estimated by the "ML1+CFD" scheme is 83.99W/m², with a MRE of 12.782% and an R² of 0.79. The RMSE of the latent heat flux estimated by the "ML2+CFD" scheme is 82.11W/m², with a MRE of 21.001% and an R² of 0.784. The RMSE of the latent heat flux estimated by the "PHY+CFD" scheme is 97.77W/m², with a MRE of 20.889% and an R² of 0.704. The verification results show that the accuracy of the estimated latent and sensible heat fluxes for the "ML1+Site" scheme is optimal. The data-driven surface impedance parameterization scheme can effectively improve the estimation accuracy of SEBS models, while the three schemes using CFD downscaled data as meteorological input have lower estimation accuracy than the three schemes using site data as meteorological input due to the uncertainty of WRF atmospheric forcing data. Compared with the same scheme using site meteorological data, the "Wei" scheme can improve the estimation accuracy of SEBS models to a certain extent, but due to the lack of optimization of the z_0m parameterization scheme, its accuracy is still lower than that of the data-driven scheme. The OMS verification results indicate that using the "ML2+Site" scheme can better improve the estimation accuracy of sensible heat flux in a larger spatial range.

The verification results also show that both the improved z_0m and z_0h (kB^-1) parameterization schemes can effectively improve the model accuracy, and the improved z_0m parameterization scheme has better optimization effect. In addition, the comparative analysis results of the spatial distribution features of very-high resolution ET at the regional scale indicate that the spatial distribution of very high-resolution ET estimated using station meteorological data is mainly affected by land use types; the spatial distribution of very high-resolution ET estimated using CFD downscaling high-resolution meteorological data can more accurately depict the spatial distribution of underlying surface hydrothermal conditions, especially on medium/high heterogeneity underlying surfaces, where the improvement effect is more pronounced. Therefore, using station-based meteorological data on homogeneous underlying surfaces is sufficient to characterize the true meteorological conditions, while on medium-to-highly heterogeneous underlying surfaces, it is necessary to use high-resolution meteorological data downscaled by CFD to truly reflect the impact of meteorological factors on the spatial distribution of ET.