The city is now being asked to report its own heat
the city is now being asked to report its own heat: a physics informed transfer model maps air temperature block by block across 380 u.s. cities, a deep net nowcasts land surface temperature at one kilometer every fifteen minutes over europe, and a light cnn beats random forest by 12 points on built up area from ecostress.
the street now files its own heat ledger.
Context
The Illinois study (Nature Communications, July 2026, Lei Zhao, Yiwen Zhang and Pierre Gentine) builds the Urban High-Resolution Air Temperature dataset, U-HAT, covering more than 380 cities in the contiguous United States. A physics-informed AI framework fills the gap left by sparse city weather stations. The authors say urban heat islands may be less extreme than satellite surface temperature maps suggest.
The TU Delft preprint (arXiv, May 2026) combines a geostationary satellite (SEVIRI on Meteosat, 3 km, every 15 minutes) with a polar orbiter (MODIS, 1 km) using a U-Net, to get land surface temperature at 1 km every 15 minutes for large European cities. Downscaling RMSE is 1.92 C with near-zero bias. A ConvLSTM nowcast for 15 to 75 minute lead times has RMSEs of 0.57 to 1.15 C and beats persistence and rolling median baselines.
The ECOSTRESS paper (Sustainability, July 2026) uses a lightweight CNN to classify local climate zones in Xuzhou, China, and analyses ECOSTRESS surface temperatures. It reports overall accuracy of 84.46% against 72.07% for random forest, and for built-up areas 69.41% against 46.91%.
The note joins three works with different targets: air temperature in the US, surface temperature in Europe and land cover classification in one Chinese city. The note says the CNN beats random forest 'by 12 points on built up area'. The abstract gives a 12.4 point gap in overall accuracy (84.46 against 72.07) and a 22.5 point gap for built-up areas (69.41 against 46.91), so the figure fits the overall number, not built-up. The Delft result is a nowcast out to 75 minutes, and the note's 'every fifteen minutes' describes the output step. The Illinois accuracy numbers were not in the press article read, so no error figure is given for U-HAT here. 'Over Europe' applies to cities above 1 million people in the Delft training set.
Related work
- Block-by-block AI maps uncover real urban air temperatures across 380 U.S. cities (Phys.org, University of Illinois) ↗Source for U-HAT, the 380 cities and the finding about air versus surface temperature.
- Yiwen-Zhang97/U-TL (GitHub) ↗Code for the transfer learning framework behind U-HAT.
- Spatiotemporal downscaling and nowcasting of urban land surface temperatures with deep neural networks (arXiv) ↗Source for the 1 km, 15 minute land surface temperature and the nowcast errors.
- LCZ-Informed Analysis of Surface Urban Heat Island Intensity and Daily Thermal Dynamics Using CNN-Based Mapping and ECOSTRESS Data (Sustainability, listing) ↗Source for the CNN versus random forest accuracies in Xuzhou.
- Transfer learning reveals large discrepancies between air and land surface temperatures in cities (listing) ↗The U-HAT paper itself, as listed.
Watch next
- Read the U-HAT paper for its validation error against station data. Check whether the Xuzhou result holds in other cities.
Sources
- Block-by-block AI maps uncover real urban air temperatures across 380 U.S. cities (Phys.org, University of Illinois)phys.org
- Spatiotemporal downscaling and nowcasting of urban land surface temperatures with deep neural networks (arXiv)arxiv.org
- LCZ-Informed Analysis of Surface Urban Heat Island Intensity and Daily Thermal Dynamics Using CNN-Based Mapping and ECOSTRESS Data (Sustainability, listing)exa.ai
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 00:17 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →