My Journey: Cooling Urban Heat Islands with AI

Hello all,

This hackathon didn’t end with a top 20 finish for me, but it turned out to be one of the most rewarding learning experiences I’ve had in a while. Here’s how I approached the challenge and what I learned along the way.

When I first saw the problem predicting Urban Heat Island (UHI) intensity I didn’t treat it as just a machine learning task. I saw it as a geospatial problem. UHI isn’t isolated; it’s influenced by surroundings like buildings, vegetation, elevation, and materials.

The Data I Used

• DEM (~5m) and DSM (~20cm)
• Landcover and NDVI (Sentinel-derived)
• Building shadow data
• Weather time-series data

For each location, I cropped a ~100×100 grid from all raster layers. Overlap between points was intentional to preserve spatial continuity.

Experiment 1: CNN + LSTM Hybrid

I approached this as a multi-modal learning problem. Instead of merging everything, I grouped inputs based on resolution:

• High-resolution: Landcover, DSM
• Lower-resolution: DEM, NDVI

Each group passed through its own CNN (ResNet50). Weather time-series data was handled using an LSTM. The outputs were concatenated and passed through fully connected layers.

Result: ~84% accuracy — decent, but clearly improvable.

Experiment 2: Building Spatial Awareness with GNNs

Since UHI is inherently spatial, I moved to a graph-based approach. The idea was to generate embeddings for each point and connect them in a graph.

This time, I added building shadow data and redesigned the encoder:

• Branch 1: DSM + DEM + Shadow (3-channel)
• Branch 2: Landcover + NDVI (2-channel)
• Used attention-based fusion instead of simple concatenation

The CNN generated embeddings, and I removed the final layer to use it purely as a feature extractor (128-dimensional vectors).

I then built a graph using K-Nearest Neighbors (k=4) based on spatial location. Prediction points were masked during training, and I tested both:

• Vanilla GNN
• Graph Attention Network (GAT)

Result: ~94% accuracy.

Why This Approach Matters

UHI isn’t just about a single point—it’s about relationships between locations.

GNNs naturally capture this by learning:

• What exists at a location
• How it is influenced by nearby areas

This makes them particularly powerful for urban climate problems.

Also, since the embedding stage is modular, this approach can scale to other cities with similar datasets.

What I Wish I Had Time For

• Feature importance analysis before GNN input
• Faster training (GNNs are computationally expensive)

If I could redo this, I’d definitely collaborate. Doing everything solo—from data prep to modeling was challenging. A team would have made this smoother (and more fun).

Final Thoughts

This journey reinforced one key idea: geospatial problems need spatial thinking. Combining deep learning with graph structures opens powerful directions for urban climate modeling.

If you're working on geospatial AI, climate resilience, or GNNs, I’d love to connect.

Project Code

The code for this project is available on GitHub: EY-UHI-Hackthon-Code