Cities are often described as “heat islands,” with media reports warning that some neighborhoods can run 20°F hotter than others. But those figures are usually based on satellite data rather than the conditions people actually experience on the ground — a gap that hampers efforts to understand public health risks during heat waves, plan for energy demand, and build climate resilience.

A new study aims to close that gap. Led by University of Illinois Urbana-Champaign civil and environmental engineering professor Lei Zhao, graduate student Yiwen Zhang, and Pierre Gentine at Columbia University, the research introduces the first high-resolution urban air temperature dataset that reflects what heat actually feels like to people and infrastructure — not just what satellites see from orbit.

Why Weather Stations Miss the Point

World Meteorological Organization guidelines require standard weather stations to sit in open, unobstructed areas far from buildings — meaning many urban stations end up at airports or on the outskirts of town, nowhere near the dense neighborhoods where most people actually live and work.

“Yes, cities are hot, and some neighborhoods are hotter than others. But not always as extreme as surface-temperature maps suggest.”

Lei Zhao, University of Illinois Urbana-Champaign

To fill the gap, the team built a physics-informed transfer learning model — an AI framework that blends physical understanding of the atmosphere with data-driven methods — capable of estimating near-surface air temperature at very high spatial resolution across more than 380 cities in the contiguous United States. The resulting Urban High-Resolution Air Temperature dataset, or U-HAT, maps how heat varies block by block, and can support public health studies, urban climate research, energy planning, and further machine-learning applications.

“This data allows for pixel-by-pixel comparison between satellite land surface temperature and true urban air temperature. And it shows that satellite-based data often overestimates heat stress and exaggerates disparities between neighborhoods, helping explain why some past maps and media stories may have unintentionally misled the public about how extreme urban heat differences really are.”

Lei Zhao

Beyond Wealthy, Data-Rich Cities

The researchers emphasize that the framework isn’t only useful for well-monitored American cities — it can be applied anywhere weather and climate observations are sparse, which carries particular significance for global equity.

“Many regions, particularly in the Global South, are highly vulnerable to climate change and extreme heat and lack dense networks of weather stations. These are often the very places that most need reliable climate and weather information for planning, health and development decisions.”

Lei Zhao

By using physics-informed AI to fill in missing data, the framework could deliver more accurate, decision-relevant climate information to underserved communities without requiring costly deployments of physical sensors.

The research was supported by the U.S. National Science Foundation, NASA, and the U.S. Department of Energy.

---

The study, “Transfer learning reveals large discrepancies between air and land surface temperatures in cities,” was published May 27, 2026, in Nature Communications (DOI: 10.1038/s41467-026-73716-7).