The Shuttleworth-Wallace two-source (S-W) model has been widely applied to estimate evapotranspiration (ET) and its components in a variety of vegetation-covered surface conditions. However, significant uncertainties occur in calculation of vegetation canopy resistance (r_sc) and soil surface resistance (r_ss). In this study, an enhanced version of the S-W model is proposed to simulate and partition ET. The deep learning (DL) approach which combined soil, vegetation, and meteorological observation data, is employed to simulate r_sc. A two-objective Monte Carlo-based Bayesian parameter optimization (TOMCBP) is developed to determine the empirical parameters for the r_ss calculation. The enhanced S-W model was verified based on the three years of eddy correlation (EC) and stable water isotope observations in an urban forest land located in Tianjin, China. Results suggest the machine learning-based r_sc parametric scheme effectively improved the performance of the S-W model for the ET simulation. The TOMCBP-based r_ss parameterization scheme can improve the performance of the S-W model for ET partition. Furthermore, this study used the Bayesian model evidence (BME) to evaluate the performance of different models. BME is shown to balance the model complexity and fitting accuracy when compared with traditional statistical parameters, thus showing superiority in model evaluation and selection. This study improves the physical mechanism and performance of the S-W model and proposes a new method for r_sc and ET components simulation based on machine learning. The enhanced S-W model is more precise and provides guidance for irrigation measures in urban woodland areas.