Evapotranspiration (ET) is an important component of connecting hydrological cycle and surface energy balance. Accurate estimation of soil evaporation (LE_s) and vegetation transpiration (LE_v) is of great significance to urban water resources allocation and management. This study, based on the MOD16 model, proposed an improved MOD16 dual-source model, which is applied to estimate the ET of urban woodland areas. The improved MOD16 dual-source model more accurately described the energy distribution process of the complex underlying surface in the urban woodland areas. The model was applied to estimate LE_s and LE_v respectively in 20 cloudless days in Guiwan Garden area of Shenzhen City, by using Sentinel-2 remote sensing satellite images which had high spatial and temporal resolution. The Shuttleworth-Wallace (S-W) and dual crop coefficient (FAO dual-K_c) model were used to evaluate model performance. This study also analyzed the sensitivity of the input variables. The results show that the root mean square error (RMSE) between the improved MOD16 model and the S-W model is 21.39 W/m², the average absolute error (MAE) is 18.81 W/m², R²=0.801; the root mean square error (RMSE) between the improved MOD16 model and the FAO dual-K_c model is 20.41 W/m², the average absolute error (MAE) is 19.05 W/m², R²=0.634. The improved MOD16 model was successfully applied to the estimation of LE_s and LE_v of urban forest. The monthly average value of total evapotranspiration in Guiwan Garden is between 85 and 165 W/m². The LE_s and LE_v present substantially seasonal variation, with the increased evapotranspiration in spring and summer, and the decreased evapotranspiration in autumn and winter. The range of LE_s is 0-50 W/m² and the range of LE_v is 0-120 W/m². The high values of LE_v are scattered, mostly located along the coast of Guiwan Garden. The high values of LE_s are mainly located in the northwest, northeast and southeast of the study area. Sensitivity analysis showed that the improved MOD16 model was more sensitive to vegetation coverage, solar radiation and humidity. The LE_v simulation results were most affected by solar radiation and vegetation cover, meanwhile, the LE_s simulation results were most affected by humidity and vegetation cover. Therefore, it is necessary to prioritize the accurate inputs of these parameters when applying the improved MOD16 model. The improved MOD16 model greatly improved the accuracy of LE_s and LE_v simulation in high-resolution, small-scale regional urban forestland. This model provides a powerful tool for accurately obtaining vegetation water consumption information to scientifically guide urban forestland irrigation and solve urban water resource allocation and management.