Planetary boundary layer height (PBLH) is an important parameter for understanding land–atmosphere interactions and for weather, climate, and air quality modeling. Radiosonde-based PBLH estimates can differ substantially across diagnostic variables and methods, particularly when stable boundary layers (SBLs) coexist with overlying residual layers (RLs). We develop an ensemble method that obtains candidate PBLHs from nine radiosonde-based algorithms and evaluates their physical support using six consistency scores. These scores assess thermodynamic structure, shear–buoyancy conditions, surface-layer similarity, temperature and humidity distributions, gradient contrasts, and wind structure. The resulting scores are combined with algorithm-specific quality and agreement among candidate heights to determine each candidate's contribution to the final PBLH estimate, while accounting for redundant information shared among related methods. Convective and neutral boundary layer heights are obtained by weighted averaging, while SBL heights are determined by first selecting a layer supported by multiple physical criteria and then averaging estimates within that layer. RL tops are detected separately. Applied to 12,266 radiosonde profiles at Beijing station during 2009–2023, the method achieves a retrieval rate of 79.41 % at approximately 5 m vertical spacing. Increasing the vertical grid spacing to 25 m raises the SBL retrieval rate from 66.55 % to 74.26 %, whereas further coarsening to 100 m reduces it to 14.48 %. At 100 m, retrieval rates for convective and neutral boundary layers remain above 99 %, but estimated heights shift upward overall. The method provides an interpretable framework for PBLH estimation under complex atmospheric conditions and highlights the importance of high-resolution observations for identifying shallow stable layers and distinguishing vertically separated structures. Such an approach could also strengthen the observational basis for understanding boundary layer processes and evaluating their representation in numerical models.