This study presents an observation-based machine-learning framework for estimating marine atmospheric boundary layer (MABL) height around the Korean Peninsula using aircraft dropsondes collected during 2020–2024. Of 1193 profiles, 1179 were collocated with ERA5 and 666 had well-defined MABL tops assigned through expert interpretation of thermodynamic and moisture profiles; 642 complete cases were used for machine learning. Dropsonde observations and ERA5 sea surface temperature were combined to estimate air–sea heat fluxes using a COARE 3.0-based formulation. Selected profile-based diagnostics and ERA5 planetary boundary layer height (PBLH) were evaluated against the expert-interpreted reference. Five machine-learning algorithms were trained using all thermodynamic regimes together and evaluated by five-fold GroupKFold cross-validation grouped by observation date. For the same 642 profiles, the machine-learning models reduced mean absolute error by 38.3–41.0 % relative to ERA5 PBLH. Random Forest, XGBoost, and a deep neural network yielded mean absolute errors of 155.5–155.9 m; the neural network had the lowest root-mean-square error (209.5 m) and the highest correlation (0.815). Random Forest permutation importance identified near-surface relative humidity and relative humidity at 925 hPa as the leading predictors. Moisture-related variables remained important across three thermodynamic regimes defined using 850 hPa temperature and lower-tropospheric stability. The proposed framework thus reproduced the expert-interpreted MABL height reference more closely than ERA5 PBLH for well-defined cases. The expert heights are physically informed references, not independent direct measurements of turbulent boundary layer depth.