eds_project2

An academic earth and environmental data science project investigating post-2050 changes in monthly mean of maximum near-surface urban temperature (TREFMXAV_U) across climate scenarios. The project combines climate-data preprocessing, supervised learning, and post-2050 scenario, seasonal, spatial, and Manchester-focused analysis.

Research question

How do projected post-2050 increases in TREFMXAV_U vary across scenario and season, and are these patterns in Manchester consistent with those across the wider study region? As a supporting analysis, which variables appear most important for explaining projected variation in TREFMXAV_U?

Project aims

• inspect and preprocess NetCDF climate scenario data
• aggregate daily data to monthly scale
• build a supervised learning table
• evaluate XGBoost and linear baselines using rolling time-series cross-validation on pre-2050 data
• fit a final model on pre-2050 data and evaluate on post-2050 rows where TREFMXAV_U is available
• analyse post-2050 scenario, seasonal, spatial, and Manchester-focused patterns in predictions

Repository structure

• data/ raw, interim, and processed data
• notebooks/ exploratory and reporting notebooks
• src/ reusable source code
• scripts/ runnable pipeline scripts
• outputs/ figures, tables, and saved models
• presentation/ contains the final project slide deck

Data

Raw NetCDF files are not stored in the repository. Place them in data/raw/.

Reproducibility

Create the environment from environment.yml, place the raw NetCDF files in data/raw/, and then run the notebooks/scripts in workflow order.

Suggested workflow:

1. Run preprocessing to create the monthly dataset.
2. Run 03_modelling_xgboost.ipynb to:
   • validate models on pre-2050 folds
   • fit the final model on pre-2050 data
   • score the post-2050 holdout
   • export row-level and grouped post-2050 evaluation tables
3. Run 04_post2050_predictions.ipynb to:
   • compare post-2050 truth vs predictions
   • analyse scenario, seasonal, spatial, and Manchester-specific patterns

Key results

• XGBoost outperformed linear and ridge baselines for predicting TREFMXAV_U during pre-2050 validation.
• Model performance remained strong when tested on true post-2050 heldout data.
• Removing TREFHT reduced model performance substantially, showing it is a major predictor, while reduced-model skill remained strong enough to indicate useful contributions from other variables.
• For post-2050 predictions:
  • Post-2050 warming is positive across all scenarios, with stronger increases by 2070–2080 than in 2050–2059.
  • Projected warming is seasonally uneven, with larger increases in late spring, summer, and early autumn than in late autumn and winter.
  • Manchester follows the broader regional warming pattern but remains consistently warmer in absolute terms and shows some seasonal differences in warming magnitude.

Authors

Jacob Woodland, Yuhui Duan, Ruiqi Huang