The 2025 EY Open Science AI & Data Challenge

This project was developed as part of the EY Open Science AI & Data Challenge 2025, where global teams worked to design machine learning models that predict the Urban Heat Island (UHI) effect across New York City.

---

Cooling Urban Heat Islands - Team Vietnamese

Mason Phung (Leader), Macquarie University, Master of Data Science

Claire Dinh, Macquarie University, Master of Business Analytics

Wendy Nguyen, Macquarie University, Master of Business Analytics

Challenge Achivements

📍 Ranked 67th globally out of 10,000+ participants

🇦🇺 Top 7 in Australia

---

Project Overview

Urban Heat Islands (UHIs) represent a critical climate issue in urban centers. Our solution aims to predict UHI intensity across NYC using multi-source data including:

• Ground-traverse temperature data

• Sentinel-2 and Landsat-8 satellite imagery

• Building footprint data

• Local weather station measurements

Through careful data preprocessing, feature engineering, and model tuning, we developed an regression model of predicting UHI index at meter-scale precision.

---

Workflow Summary

I. Data Acquisition & Processing

• Extract and merge Sentinel-2, Landsat LST, KML-based building data, and weather data
• Compute grid-based building coverage and building count

II. Feature Engineering

• Create vegetation indices (e.g., NDVI), normalize timestamps, derive satellite-derived features
• Integrate spatial features from KML overlays and building polygons
• Apply focal buffer technique to reflect real-world surrounding effects (created based on Shandas et al, 2019)

III. Model Development

• Tune and train XGBoost regressors using SHAP-driven feature selection
• Evaluate using R², MAE, and RMSE on held-out data

IV. Final Prediction

• Predict UHI index on test set with top 50 most important features
• Generate final CSV for leaderboard submission

---

Directory

Open notebooks/proj.ipynb for full project workflow

├── README.md          <- Project README
├── data
│   ├── interim        <- Intermediate data after initial transformations.
│   ├── processed      <- Final datasets prepared for modelling.
│   ├── raw            <- Original, unprocessed data.
│   └── test           <- Final submission results or test outputs.
│
├── docs               <- Project documents, references and challenge questions.
│
├── models             <- Trained models, model predictions, or model summaries
│
├── notebooks          <- Jupyter notebooks contain project process and KML processing guide.
│
├── pyproject.toml     <- Project configuration file with package metadata for 
│                         src and configuration for tools like black
│
├── requirements.txt   <- The requirements file for reproducing the analysis environment, e.g.
│                         generated with `pip freeze > requirements.txt`
│
├── setup.cfg          <- Configuration file for flake8
│
└── src   <- Source code for use in this project.
    │
    ├── __init__.py             <- Makes src a Python module
    │
    ├── data_manipulation.py     <- Important class and functions for extracting satellite band values and calculating vegetation indices. 
    │
    ├── data_processing.py      <- Script to collect, process and transform Sentinel-2 and Landsat-8 data.
    │
    ├── kml_processor.py        <- Code to process KML files and extract relevant features for the model.
    │
    ├── modeling.py             <- Code to train, tune and evaluate the model.
    │
    ├── utils.py                <- Utility functions to support the project.

---

Results

Dataset	R² Score	MAE	RMSE
Local Train Set	0.9990	0.0004	0.0000
Local Test Set	0.9610	0.0023	0.0000
Competition Test Set	0.9680	—	—

Key Takeaways

“The success of this project came not from exotic modeling techniques, but from rigorous data preprocessing and feature engineering. Proper handling of spatial features and grid-cell overlays using building data proved crucial to improving model performance.” — Team Vietnamese

Discussion

Feel free to reach out to Mason through LinkedIn or GitHub.