Potential evapotranspiration (PET) is a key component of the terrestrial water cycle, and its estimation accuracy directly affects the reliability of drought monitoring. The FAO56 Penman-Monteith method is widely regarded as a standard PET model in drought monitoring. However, it employs a static, non-water-limited canopy resistance parameter (r(cNWL)), which fails to fully capture the regulatory effects of vegetation on PET and drought. In this study, we investigated the canopy resistance (r(c)) models from empirical, process-based, machine learning, and physically constrained machine learning approaches based on worldwide eddy flux data and remote sensing data to simulate dynamic r(cNWL), aiming to build up a PET model that accounts for vegetation dynamics and hence represents drought realistically. The results indicated that the hybrid machine learning and optimal stomatal behavior (r(c_USO_ML)) is the best model to simulate dynamics of canopy conductance behavior (highest Global Performance Indicator, GPI), as rc_USO_ML model significantly outperforming traditional empirical (r(c_Jarvis)) and process-based (r(c_USO_Lin)) models. Following the PET assumption, we formulated the PETrcNWL_ USO_ ML model by removing the soil water stress on r(c_USO_ML). Compared with conventional PET models, the PETrcNWL_ USO_ ML model more accurately reproduced observed PET (R-2 > 0.85, RMSE < 0.66 mm/day), and showed stronger increasing trends in PET and drought characteristics. Notably, in the arid regions of northwest China, neglecting vegetation dynamics led to an underestimation of increasing trends in drought duration, severity, and intensity by 74%, 51%, and 23%, respectively. Overall, this study highlighted the necessity of fully accounting for the integrated effects of vegetation physiology in drought monitoring and future risk assessments.