Urban Heat Island Mapping with Sensor Networks

ISEF Category: Earth and Environmental Sciences

Subcategory: Atmospheric Science  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

A city can be 10 degrees hotter than a nearby park, and the hottest spots can hide on one block, not one neighborhood. If you only use one weather station, you miss that detail. This project lets you map heat where people actually walk, bike, and wait for buses. You also get to mix field data with satellite features, which is a strong research move.

What Is It?

Urban heat islands happen when cities trap and store more heat than nearby rural or green areas. Dark roofs, pavement, and tall buildings can raise local temperature, while trees and open soil can cool it. In this project, you place small sensor nodes around a city and measure tiny temperature differences across space.

Think of the city like a patchwork quilt. Each sensor gives you one stitch, but the full pattern only appears when you connect the points. A graph neural network is a machine learning model that learns from connected locations, so it can estimate temperature in places where you do not have a sensor. You add features such as building footprint density and NDVI, a satellite-based greenness index, to help the model understand why some places run hotter than others.

Why This Is a Good Topic

This makes a strong science fair topic because you can measure a real environmental pattern, compare places with different land use, and test whether extra spatial features improve predictions. You can collect data with affordable hardware, then analyze it with methods that many students do not try. The topic connects to public health, city planning, and climate adaptation, so your results have clear real-world value. You can also grow the project step by step, which helps you build a cleaner study over time.

Research Questions

• How does land-cover type affect near-surface temperature anomalies across a city block?
• What is the effect of building-footprint density on the accuracy of temperature interpolation between sensors?
• Does adding NDVI from Sentinel-2 improve graph neural network predictions of urban heat compared with distance-only models?
• To what extent do bicycle-mounted measurements differ from fixed lamppost sensors in mapping microclimate variation?
• Which combination of neighborhood features best predicts hotspots that sit between sparse sensors?
• How does time of day change the spatial pattern of temperature anomalies in a mixed residential and commercial area?

Basic Materials

• ESP32 development boards with Wi-Fi capability.
• BME280 temperature, humidity, and pressure sensors.
• Breadboards or soldered prototyping boards.
• Jumper wires and headers.
• USB cables for data upload and power.
• Bicycle mount or weatherproof pouch for mobile sampling.
• Zip ties, tape, and small brackets for fixed mounting.
• Portable power banks.
• Laptop for coding and data cleaning.
• Digital thermometer for spot checks.
• Notebook or field logging app for location and time notes.
• GPS-enabled phone for geotagging sensor runs.

Advanced Materials

• Calibrated reference thermometer or meteorological probe.
• Weatherproof enclosure rated for outdoor deployment.
• SD card module or local data logger.
• More than one ESP32 and BME280 node for network sampling.
• Soldering tools and multimeter.
• Optional anemometer for wind context.
• GIS software for spatial joins and map layers.
• Sentinel-2 land-cover and NDVI datasets.
• Street-network or building-footprint data from city open-data portals or OpenStreetMap.
• Compute access for graph neural network training.
• Python environment with geospatial and machine learning packages.

Software & Tools

• Python: Cleans sensor logs, merges map layers, and trains the prediction model.
• Jupyter Notebook: Keeps code, charts, and notes in one place for analysis.
• QGIS: Plots sensor points, building footprints, and satellite layers on maps.
• Google Earth Engine: Helps access and process Sentinel-2 imagery for NDVI work.
• ImageJ: Can help inspect exported maps or heat plots if you need quick image-based checks.

Experiment Steps

1. Define the spatial question you want to answer, such as block-level heat variation, sensor interpolation, or hotspot prediction.
2. Choose the sensor layout and sampling route so you cover different land uses, shading patterns, and street widths.
3. Plan your calibration strategy so every node reports temperatures on the same scale.
4. Decide which map features you will pair with each reading, such as building density, tree cover, distance to roads, and NDVI.
5. Design a baseline model first, then compare it with the graph neural network so you can prove the model adds value.
6. Set up a validation plan that tests predictions on locations and days the model did not see before.

Common Pitfalls

• Letting different sensor nodes run uncalibrated, which makes network differences look like real heat patterns.
• Mixing moving bicycle measurements with fixed-site readings without tagging them separately, which blurs the interpretation.
• Sampling at different times on different days, which confounds weather change with location effects.
• Using too few sensors for the study area, which leaves the graph model with weak spatial coverage.
• Matching satellite and street-level data poorly, which can assign the wrong NDVI or building-footprint value to each sensor point.

What Makes This Competitive

A stronger version of this project does more than draw a heat map. It tests whether graph-based prediction beats simpler interpolation methods, and it explains why. If you compare multiple neighborhoods, seasons, or sampling modes, your results get much richer. Strong sensor calibration, careful validation, and a clear link between city form and temperature will push the work beyond a basic mapping exercise.

Project Variations

• Use only fixed rooftop, lamppost, or balcony sensors to compare how sensor height changes the heat map.
• Swap NDVI for tree-canopy cover or impervious surface fraction to test which land-cover feature predicts heat better.
• Replace the graph neural network with kriging or random forest interpolation to compare model performance on the same dataset.

Learn More

• NASA Earthdata: Search for satellite land-surface and vegetation products, including Sentinel-2 and related guides.
• NOAA Climate.gov: Read background articles on urban heat islands, heat risk, and climate data interpretation.
• USGS EarthExplorer: Find satellite and land-cover datasets for map-based environmental projects.
• PubMed: Search review articles on urban heat islands, microclimate mapping, and temperature exposure.
• NOAA National Weather Service: Use local weather observations to compare your field measurements with official station data.
• MIT OpenCourseWare: Look for free material on machine learning, graph methods, and environmental data analysis.