AI connects hourly satellite pollution signals to air near the ground

A preprint posted on 23 October estimates hourly nitrogen dioxide near the ground across the contiguous United States. It combines TEMPO satellite observations with weather, land-use and time information, learning from 474 monitoring stations between August 2023 and June 2025.[1]

The satellite measurement requires translation. NASA’s retrieval document explains how TEMPO reads nitrogen dioxide’s absorption of sunlight scattered back from Earth. The result describes gas along a path through the atmosphere. Calculations then account for the light’s route and separate the lower atmosphere’s contribution from the stratosphere above.[2]

A column amount is therefore different from the concentration in air at street level. The retrieval already uses a modelled vertical gas profile, alongside information about clouds, surface reflectivity and viewing geometry. Cloud cover changes how strongly the instrument can sense different altitudes. These are properties of the measurement before the new machine-learning model attempts its surface estimate.[2]

The study trained boosted decision trees, including a separate model for hours without usable satellite data. Its clearest satellite benefit appeared when whole groups of stations were withheld in 100-kilometre spatial blocks. Adding TEMPO increased R² in all ten folds, by 0.07 on average, to 0.48. A value of 1 indicates a perfect fit.[1]

NASA’s algorithm depends on reflected sunlight, restricting this view to daylight. Hourly sampling can expose changes that a single daily pass misses, but it cannot supply a nighttime observation.[2]

The new estimates often understate high concentrations and evening peaks. The satellite helps fill spatial gaps, but the map still depends on ground measurements for learning and checking. This is an estimation study under review, without a demonstrated health outcome.[1]

Report an error or suggest a correction