Marine data science · Supervised learning

Predicting significant wave height from environmental conditions

Built a reproducible preprocessing and model-selection pipeline to predict significant wave height from environmental conditions.

Completed course project · Imperial College London2024

Random-forest predictions against observed significant wave height on the held-out test subset.

68,751observations

3model families

5-foldcross-validation

0.967test R²

01 / Question

Why this problem mattered

Significant wave height is a practical summary of sea state used in marine safety and operational planning. The project treated it as a supervised-regression problem and tested whether environmental conditions—including wind, direction, temperature and season—could support accurate predictions in a reusable modelling pipeline.

02 / Approach

From physical fields to usable evidence

Removed duplicate observations, created a reproducible 70/30 train–test split and used exploratory analysis to identify constant, highly incomplete and strongly correlated variables.
Built a ColumnTransformer pipeline with median imputation and scaling for numerical features, one-hot encoding for categorical features and a spline transformation for wave direction.
Compared linear regression with tuned random-forest and XGBoost regressors using five-fold cross-validation, R² and mean-squared error.

03 / Result

The tuned random forest provided the strongest held-out result.

The final random-forest model achieved test R² = 0.96678 and MSE = 0.003821, equivalent to an RMSE of approximately 0.062 m. XGBoost was close at R² = 0.96524, while the linear baseline reached R² = 0.88183.

Correlation heatmap for environmental wave-height predictors — Training-set feature correlations used to guide collinearity checks and feature selection.

Random-forest predictions compared with observed wave heights — Predicted versus observed significant wave height for the selected random-forest model.

Implications

The improvement over linear regression indicates that nonlinear interactions between wind, direction, temperature and seasonal conditions were important within this dataset. The study demonstrates disciplined preprocessing, cross-validated model comparison and clear selection of a final operationally interpretable metric set.

Limitations

The notebook uses a random train–test split. If the observations are temporally autocorrelated, that can overstate performance on genuinely future conditions. A stronger follow-up would use blocked temporal validation, encode direction explicitly as sine and cosine, report MAE and residual diagnostics, and evaluate rare high-wave cases separately.

Project material

Download project notebook ↗View source on GitHub ↗