Chesapeake Bay · NO₂ and ozone · 2026
We combine ground and satellite measurements into an estimate of nitrogen dioxide and ozone for every 10-km square of the Bay, every hour, then test it against EPA's own monitors on months it never saw.
Research prototype: not for health or regulatory decisions.
Study grid
348
of about 10 km, each estimated every hour
Ground instruments
5 + 6 + 104
spectrometers, sun photometers and in the region
Jan–Mar 2026 · NO₂ / O₃
+0.79 / +0.93
next-hour at 7 and 17 EPA sites · +0.81 / +0.91
Apr–Oct 2026 · new data
+0.73 / +0.93
six months after training · persistence +0.68 / +0.90
Why the Bay
from traffic, power plants and ships settles out of the air as nitrate. In the Bay, that extra nitrogen feeds algal blooms, which use up the water's oxygen and leave “dead zones” behind. The same pollution helps form , which can irritate the lungs, in a watershed that is home to more than 18 million people.
Official monitors are few and far between, mostly in cities. This project asks how well open ground and satellite measurements can fill in the gaps, hour by hour, and shows the answer honestly, including where it falls short.
How we got here
Five steps, each one changing what the model learned from or how it was scored.
01 / 05
+0.72 / +0.80
NO₂ / O₃ against its own training target, on hours it had never seen
Learn from instruments
+0.72 / +0.80
NO₂ / O₃ against its own training target, on hours it had never seen
The first model learned from instruments, which measure NO₂ and ozone in the whole column of air above them. Shown hours it had never seen (January–March 2026), it reproduced those measurements well.
7 sites
EPA ground monitors it never trained on, where it did worse than guessing each site's average
The real test is the official at ground level. There, the first model did worse than simply guessing each site's usual value. Matching a column of air overhead is not the same as matching the air at street level.
5 of 348
grid cells had a live reading of their own
Away from those few instruments, the model's most useful input was replaced by a fixed placeholder, so it fell back on a generic daily rhythm. Rescaling its output site by site ()only brought it to about each site's average: the missing signal had to come from data, not a correction.
+0.79 / +0.93
NO₂ / O₃ next-hour R² at EPA monitors, Jan–Mar 2026 ( +0.81 / +0.91)
Adding those monitors' 2018–2024 readings to training (the test months stay hidden) lifted the NO₂ score to +0.56. A signal borrowed from the nearest measured cell took it to +0.65, and a better ozone target to +0.79 (O₃ +0.93): about as good as persistence, the “next hour = this hour” guess.
+0.73 / +0.93
NO₂ / O₃ next-hour R², April–October 2026 (persistence +0.68 / +0.90)
On April–October 2026, months after training ended, the model beat persistence at 4 of 5 NO₂ sites and 16 of 17 ozone sites.
An earlier version of the EPA scoring compared each estimate with the wrong hour, which overstated these numbers. Every value on this site uses the corrected method; the validation page documents the change.
Read this first
New to air quality?
Dotted-underlined words anywhere on the site open the same short definitions.
A gas that comes mostly from vehicle exhaust, power plants and ships. It irritates the lungs, helps make ozone, and some of it ends up in the Bay as nitrate.
Not released directly: it forms when sunlight acts on NO₂ and other pollutants, so it peaks on hot, sunny afternoons. (This is the ozone we breathe, not the protective layer high up.)
Micrograms per cubic metre: millionths of a gram of pollutant in each cubic metre of air. Both pollutants are shown in this unit.
How well the estimates follow the real ups and downs. 1 is perfect; 0 means no better than always guessing that site's average; below 0 is worse than that guess.
The simplest possible forecast: assume the next hour equals this hour. One hour ahead it is hard to beat, so it is the bar every model here has to clear.
Each estimate looks at the previous 24 hours of data and predicts the following hour, never further ahead. It never sees the reading it is predicting.
One square of the map, 0.1° (about 10 km) on a side. The Bay is covered by 348 of them, each with its own hourly estimate.
Official ground-level air stations, run by state agencies under EPA rules, that sample the air people breathe. Their quality-checked readings are archived in EPA's Air Quality System (AQS).
A ground instrument that looks up at the sun and measures how much NO₂ and ozone are in the whole column of air overhead. Five of them feed this model.
Long short-term memory network: a type of neural network built to learn patterns in sequences, here the last 24 hours of readings at each grid cell.
