UrbanHeat.be
Predicts and explains urban heat anomalies across five Belgian cities, from citizen-science weather stations and Sentinel-2 satellite imagery to an interactive dashboard.
- Engineered geospatial features (NDVI, NDBI, SWIR, humidity, altitude, wind, solar radiation) with rasterio.
- Compared Random Forest, XGBoost, a stacked ensemble with a Ridge meta-learner, and a PyTorch CNN on 16×16 satellite patches trained on a GPU in Google Colab.
- Validated with 5-fold and Leave-One-Station-Out cross-validation (LOSOCV RMSE 0.575 ± 0.460), which revealed a spatial-leakage gap of about 18%.
- Explained predictions with SHAP and plain-language summaries from Llama 3.3 70B via the Groq API.
- 155
- weather stations
- 5,575
- station-day observations
- 0.599
- R² of the stacked RF + XGBoost model
- ~18%
- spatial-leakage gap found



