README for Supplementary Table S1

Manuscript title
----------------
A dropsonde-based machine-learning framework for estimating marine atmospheric boundary layer height around the Korean Peninsula

Supplement contents
-------------------
This supplement contains the following file:

TableS1_ML_input_642profiles.csv

This file provides the per-profile machine-learning (ML) dataset used in the study. Each row corresponds to one ML-complete dropsonde profile. The table includes the 11 ML input variables, the expert-interpreted marine atmospheric boundary layer (MABL) height, the T850–LTS-based thermodynamic regime, ERA5 planetary boundary layer height (PBLH), cross-validation grouping information, and out-of-fold predictions from each ML model.

File format
-----------
File name: TableS1_ML_input_642profiles.csv
Format: comma-separated values (CSV)
Encoding: UTF-8
Number of rows: 642
Number of columns: 25
Missing values: none

Column descriptions
-------------------
profile_id
    Unique identifier for each dropsonde profile, corresponding to the available profile or file identifier used in the analysis.

date_utc
    Dropsonde release date and time in Coordinated Universal Time (UTC).

obs_date_kst
    Observation date in Korea Standard Time (KST). This date was used as the grouping variable for the date-grouped five-fold GroupKFold cross-validation.

fold_id
    Cross-validation fold number assigned by the date-grouped five-fold GroupKFold validation. All profiles observed on the same obs_date_kst were assigned to the same fold.

lat
    Dropsonde release latitude in degrees north.

lon
    Dropsonde release longitude in degrees east.

sst
    ERA5 sea surface temperature used as an ML input variable (°C).

ta
    Near-surface air temperature from the lowest valid dropsonde observation within 2–30 m above sea level (°C).

rh
    Near-surface relative humidity from the lowest valid dropsonde observation within 2–30 m above sea level (%).

u
    Near-surface scalar wind speed from the lowest valid dropsonde observation within 2–30 m above sea level (m s⁻¹). Despite the column name, u denotes wind-speed magnitude rather than the zonal wind component.

t925
    Temperature at 925 hPa derived from the dropsonde profile using interpolation in log-pressure coordinates when needed (°C).

rh925
    Relative humidity at 925 hPa derived from the dropsonde profile using interpolation in log-pressure coordinates when needed (%).

t850
    Temperature at 850 hPa derived from the dropsonde profile using interpolation in log-pressure coordinates when needed (°C).

lts
    Lower-tropospheric stability (K), defined as θ700 − θsurface, where θsurface was calculated from the lowest valid dropsonde observation within 2–30 m above sea level.

wind_speed_difference
    Wind-speed difference (ΔWS), defined as the wind speed at 850 hPa minus the near-surface wind speed from the lowest valid dropsonde observation within 2–30 m above sea level (m s⁻¹). Positive values indicate stronger winds at 850 hPa than near the surface, whereas negative values indicate stronger near-surface winds.

shf
    Sensible heat flux calculated from the near-surface dropsonde observations and ERA5 SST using the COARE 3.0-based iterative bulk-flux formulation (W m⁻²). Positive values indicate upward sensible heat transfer from the ocean surface to the atmosphere.

lhf
    Latent heat flux calculated from the near-surface dropsonde observations and ERA5 SST using the COARE 3.0-based iterative bulk-flux formulation (W m⁻²). Positive values indicate upward latent heat transfer from the ocean surface to the atmosphere.

mabl_height
    Expert-interpreted MABL height used as the supervised-learning target variable (m).

regime
    T850–LTS-based thermodynamic regime assigned by K-means clustering for post hoc physical interpretation. The regime label was not used as an ML input feature. Possible values are:
      - Warm-stable
      - Transitional
      - Cold/weakly stable

era5_pblh
    ERA5 PBLH collocated with the dropsonde profile and used as the reanalysis baseline (m).

pred_svr
    Out-of-fold prediction from the support vector regression (SVR) model (m).

pred_rf
    Out-of-fold prediction from the random forest (RF) model (m).

pred_xgb
    Out-of-fold prediction from the Extreme Gradient Boosting (XGBoost) model (m).

pred_gb
    Out-of-fold prediction from the gradient boosting model (m).

pred_dnn
    Out-of-fold prediction from the deep neural network (DNN) model (m).

Reproducibility notes
---------------------
The pred_svr, pred_rf, pred_xgb, pred_gb, and pred_dnn columns contain out-of-fold predictions from the date-grouped five-fold GroupKFold validation. They are not predictions from models retrained using the complete 642-profile dataset.

The performance statistics for the ML models and ERA5 PBLH can be reproduced by comparing mabl_height with each of the following columns:
  - pred_svr
  - pred_rf
  - pred_xgb
  - pred_gb
  - pred_dnn
  - era5_pblh

Mean bias should be calculated as:

    prediction_or_ERA5_PBLH - mabl_height

The following values were obtained from TableS1_ML_input_642profiles.csv:

Model                 N     R     RMSE (m)   MAE (m)   Mean bias (m)
SVR                  642   0.766    241.7     162.7       -34.4
Random Forest        642   0.792    220.6     155.5       -12.3
XGBoost              642   0.791    221.8     155.9       -18.6
Gradient Boosting    642   0.788    223.0     158.5       -18.0
DNN                  642   0.815    209.5     155.6        -6.8
ERA5 PBLH            642   0.796    341.2     263.5       243.3

Important limitations
---------------------
The table contains processed per-profile scalar quantities only and does not include the full raw dropsonde vertical profiles. The raw dropsonde observations were collected by the Korea Meteorological Administration/National Institute of Meteorological Sciences (KMA/NIMS) for research purposes and are subject to the data policy of the operating agency.

The expert-interpreted MABL height should be interpreted as a physically based reference assigned to profiles for which a single MABL top could be defined relatively clearly, rather than as an absolute truth.

Contact
-------
For questions about the supplementary table or access to the underlying raw dropsonde observations, please contact:

Min-Seong Kim
Observation Research Department
National Institute of Meteorological Sciences
33, Seohobuk-ro, Seogwipo 63568, Jeju-do, Republic of Korea
Email: givemeasong@korea.kr
