← All work

Beyond the heat map

The hottest neighborhood is only part of the story.

ROLE

Student researcher

CONTEXT

Pioneer Academics · June–September 2026

TOOLS & METHODS

Landsat 8/9 · NLCD land cover · Census ACS · k-means clustering · Spatial statistics

A hand-built island model for Beyond the heat map
THE ISLAND, EXPLAINED

A desert city, seen from orbit and from the street.

An illustration built for this site in Blender. It explains the idea; it is not data from the project.

THE QUESTION

A surface-temperature map tells us where a city is hot. It does not, by itself, tell us which neighborhoods are most vulnerable. My research brought satellite, land-cover and census information together for 369 Phoenix census tracts to examine that difference, and to ask how much a cooling program’s priorities change when the definition of “priority” does.

369Phoenix census tracts analysed
17 / 37top-decile tracts shared by both rankings
0.749rank correlation between the two rankings
3vulnerability profiles
01

A different way to ask the question

I studied 369 Phoenix census tracts. Rather than treating temperature as the whole problem, I compared two ways of choosing where to act first: a ranking by surface temperature alone, and a ranking by a score that also counts how built-up a tract is, how many people live there, and how sensitive they are to heat.

02

Building one picture from eleven

Surface temperature comes from Landsat 8 and 9 Collection 2 Level-2 scenes. Eleven scenes from June to August 2024 passed quality screening; cloud, shadow and saturated pixels were masked, and the per-pixel median of what remained gave a summer composite of temperature and vegetation. The 2024 National Land Cover Database supplied paved surface and tree canopy, and the American Community Survey supplied age, poverty and housing figures with their margins of error.

03

A score that keeps its parts visible

The priority score is the plain average of four percentile ranks: heat, built burden (paved surface and lack of vegetation), exposure (population density) and sensitivity (people over 65, children under five, and poverty). Equal weights are a value choice, not a fitted model, and the paper says so. Keeping the four parts separate means anyone can see why a tract ranks where it does.

04

When the rankings diverge

Across all tracts the two rankings agree fairly well, with a rank correlation of 0.749. At the top, where decisions get made, they do not: only 17 of the 37 hottest tracts are also among the 37 highest-priority tracts. Moving from one ranking to the other shifted the median tract by 42 places, and the largest moves were about 250 places in each direction.

05

Three kinds of neighborhood

Clustering the tracts on seven standardised features gave three readable profiles: 206 tracts with compounded heat and socioeconomic burden; 45 low-density tracts that are far from cooling centres; and 118 greener, slightly cooler tracts with the oldest residents. The grouping stayed stable when the census inputs were perturbed within their margins of error 50 times, which is the check that lets the profiles be discussed at all.

INSIDE THE ENGINEERING

One city. Three ways to read it.

  1. 01

    From radiance to a tract

    Landsat thermal values are converted to kelvin with the product’s published scale and offset, masked with its pixel-quality bands, and combined as a per-pixel median over the summer. Each raster is then summarised inside tract boundaries in a common map projection. Of 373 tracts assembled, 369 passed the coverage and quality checks.

  2. 02

    A change in the decision rule

    A temperature-only ranking and the four-part priority ranking are compared at the top decile and top quintile. Overlap is reported as counts and as a Jaccard index (0.298 at the top decile), so the result is about which places a program would visit first, not about an average difference.

  3. 03

    Profiles that survive perturbation

    k-means with three clusters was chosen after comparing three to six clusters against Ward linkage. Stability was measured by re-clustering after perturbing census variables within their margins of error: the median adjusted Rand index was 0.797 for k-means against 0.580 for Ward. Both scores are spatially clustered (Moran’s I of about 0.5), so neighboring tracts are not independent samples.

Surface temperature, demographic vulnerability and intervention priority answer related but different questions. The study compares those perspectives; it does not measure the effects of a completed cooling program.

WHAT THIS DOES NOT SHOW

The honest edges.

  • Satellite surface temperature is not air temperature, night-time heat or indoor heat.
  • The equal weights in the priority score are a stated value choice. Different weights would move tracts.
  • The study does not link any ranking to health outcomes, and distance to a cooling centre is straight-line, not travel time.
NEARBY ISLANDS

Where this connects.

CONTEXT & FURTHER READING

Paper: “Remote Sensing and Interpretable Machine Learning for Neighborhood-Level Urban Heat Vulnerability in Phoenix, Arizona.” Submitted and accepted by my supervising professor; unpublished. “Interpretable” refers to the readable cluster profiles and the transparent score.