Neighborhood Exposure Driver Models

ISEF Category: Earth and Environmental Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Some neighborhoods breathe dirtier air, even when they sit only a few blocks apart. That gap can come from traffic, land use, heat, and housing patterns all mixing together. Your project asks which factors matter most, and where. That turns a big social problem into a data puzzle you can test.

What Is It?

This project studies environmental justice, which means how pollution exposure is not shared equally across communities. You combine satellite and ground data to compare air quality, surface heat, and neighborhood demographics. Then you train an interpretable model, which is a model you can explain, not just one that gives an answer.

Think of each neighborhood like a recipe. NO₂ from TROPOMI can point to traffic and combustion sources. PurpleAir PM2.5 can capture fine particles people actually breathe. MODIS land surface temperature can flag hotter places that often line up with fewer trees and more pavement. Census data adds the social layer, such as income, race, age, and housing patterns. Your goal is to find which ingredients best explain exposure differences.

The strongest version of this project does not stop at mapping hotspots. It asks why the hotspots form. That makes the work more useful for local planning, public health, and environmental policy.

Why This Is a Good Topic

This is a strong science fair topic because you can test a real question with public data and clear outputs. You do not need a chemistry bench to start, but you do need careful data cleaning, mapping, and model interpretation. The project connects directly to air pollution, heat risk, and environmental fairness, so your results have real-world meaning. You can also learn spatial analysis, feature importance, and basic machine learning in one project.

Research Questions

• How does neighborhood traffic density relate to TROPOMI NO₂ exposure gradients?
• What is the effect of land surface temperature on predicted PM2.5 across neighborhoods?
• Does adding Census demographic variables improve exposure predictions beyond satellite and sensor data alone?
• To what extent do tree cover and impervious surface explain neighborhood differences in heat and air pollution?
• Which variables, NO₂, PM2.5, surface temperature, or demographics, most strongly drive model predictions in each area?
• What is the effect of nighttime versus daytime land surface temperature on exposure pattern estimates?
• Which neighborhoods show the largest mismatch between pollution burden and demographic vulnerability?

Basic Materials

• Laptop with internet access.
• Spreadsheet software or Google Sheets.
• GIS software such as QGIS.
• Python installed with Jupyter Notebook.
• Public Census data access through the Census API or data.census.gov.
• TROPOMI NO₂ data from NASA or Copernicus sources.
• PurpleAir PM2.5 data access through the PurpleAir map or API.
• MODIS land surface temperature data from NASA Earthdata.
• NOAA or EPA air quality reference data for comparison.
• Notebook for logging data cleaning decisions and map settings.

Advanced Materials

• Access to a school or university workstation for large geospatial files.
• Python with geopandas, pandas, scikit-learn, matplotlib, and shapely.
• QGIS or ArcGIS Pro for spatial joins and map making.
• Google Earth Engine for large-scale remote sensing analysis.
• Census TIGER or ACS boundary files.
• Satellite raster data from TROPOMI, MODIS, or other NASA products.
• Air monitor reference data for validation.
• External storage for large raster and tabular datasets.
• Access to a geostatistics or spatial statistics package.
• Optional one-class or explainable ML tools for sensitivity testing.

Software & Tools

• QGIS: Maps neighborhood boundaries, overlays exposure layers, and checks spatial patterns.
• Python: Cleans data, joins datasets, and runs the interpretable model.
• Jupyter Notebook: Keeps code, notes, and charts in one place.
• Google Earth Engine: Helps you handle large satellite datasets without downloading everything.
• ImageJ: Can help inspect image-based outputs or exported maps when needed.

Experiment Steps

1. Define a neighborhood unit, such as census tracts or block groups, so every dataset uses the same spatial scale.
2. Choose a small set of exposure and vulnerability variables, then decide which one you will treat as the outcome and which ones you will treat as predictors.
3. Build a data-cleaning plan that handles missing satellite pixels, sensor gaps, and mismatched dates before you analyze any patterns.
4. Create a spatial join workflow so each neighborhood gets one combined record from the pollution, heat, and Census layers.
5. Train an interpretable model first, then compare its predictions with simpler baselines so you can tell whether the extra variables really help.
6. Plan how you will test variable importance, uncertainty, and geographic bias so your conclusions hold up across different neighborhoods.

Common Pitfalls

• Mixing datasets from different dates, which can make a heat map and a pollution map describe different conditions.
• Using raw PurpleAir readings without checking whether nearby sensors cluster in only a few wealthy or dense neighborhoods.
• Comparing neighborhoods of different sizes without standardizing by area or population, which can blur exposure gradients.
• Treating one satellite pixel as if it exactly matches one neighborhood boundary, which creates edge and averaging errors.
• Letting the model fit demographic proxies so strongly that it predicts social patterning without truly explaining exposure drivers.

What Makes This Competitive

A competitive version of this project goes beyond a simple map. You would test several models, compare them with a plain baseline, and report which variables matter under different conditions. Strong entries also separate correlation from explanation by checking spatial autocorrelation, missing-data bias, and sensitivity to neighborhood scale. If you can tie your model to a local planning question, your work gets even stronger.

Project Variations

• Use only one city and compare exposure drivers across high-income and low-income tracts.
• Swap MODIS land surface temperature for tree cover or impervious surface to test whether built environment features explain more variation.
• Focus on seasonal changes, then ask whether exposure drivers shift between summer heat and winter combustion patterns.

Learn More

• NASA Earthdata: Search for MODIS and other satellite data, plus tutorials on downloading and reading remote-sensing products.
• TROPOMI data on Copernicus or NASA resources: Look for NO₂ product guides and user documentation.
• CDC Environmental Justice Index: Use this to compare pollution burden and social vulnerability patterns by place.
• U.S. Census Bureau ACS and TIGER/Line data: Find neighborhood demographics and boundary files at data.census.gov and the Census geography pages.
• PubMed: Search for review articles on environmental justice, air pollution exposure, and spatial modeling.
• QGIS Training Manual: A free guide for mapping, joining layers, and making spatial analysis maps, available through the QGIS documentation site.