How We Built It

So what did we do about it?

We put on our thinking caps and built a machine learning model that accurately showcases "Sol" in California based on vegetation and urban island heat influencing factors.

TLDR: Check for trees and building density around a specific location. Mix it in with temperatures and some special ML magic and you get a heat vulnerability score.

Normalized Difference Vegetation Index (NDVI)

Land Surface Temperature (LST)

Building Footprint Density

Google Earth Engine

Vulnerability ML-Model

Gemini-2.5-Flash Tuner

Tuned Weights

{
  w1: 0.6,  // temperature
  w2: 0.2,  // vegetation
  w3: 0.2   // buildings
}

1NDVI (Vegetation Index)

2LST (Land Surface Temp)

3Building Footprint Density

4Google Earth Engine

5Random Forest ML Model

6Gemini-2.5-Flash Tuner

7Tuned Weights → Heat Score

Data at each stage

NDVI Input

Vegetation Index

LST Input

Land Surface Temp

Building Density

Urban Footprint

Vulnerability Output

Heat Score Map

Remote Sensing Data

Both NDVI and LST rasters were exported directly from Google Earth Engine using Landsat 8/9 imagery for California.

Google Earth Rasterization of NDVI

NDVI raster shows vegetation health and density across California, with brighter areas indicating more vegetation.

Google Earth Rasterization of LST

LST raster shows land surface temperature across California, with brighter areas indicating higher temperatures (hotter urban heat islands).

*Notice how areas near the coast are darker, corresponding to cooler temperatures.

Machine Learning Model Dataframes

Data Classifications for ML Model

As part of feature extraction, we rasterized both building density and tree density into uniform 100 m grid cells across California. These become two of the three columns in our feature matrix X (the third being normalized temperature). We then fed X into our Random Forest, which learned how those spatial patterns combine to predict heat‐vulnerability scores.

Tree Density Raster

Each cell = 100m × 100m patch. Scattered green clusters correspond to urban tree canopy. The darkest cell peaks at 0.006 (density units).

In ML terms: Vegetation input (proxy for NDVI), normalized tree/vegetation cover per cell.

Building Density Raster

A concentrated urban cluster (center-left) reaches 300 m² of roof per 10,000 m² cell. Surrounding cells show moderate suburban density.

In ML terms: Urbanization input — m² of roof area per 10,000 m² cell.

So what did we do about it?

Remote Sensing Data

Data Classifications for ML Model

Enough Technical Talk. Check out the 3D City Visualization.