How AQI Forecasts Work: Models, Satellites, and Why They Sometimes Fail
Every AQI forecast is the output of a multi-step pipeline involving emissions inventories, chemical transport models, weather prediction, and satellite data. Here's what's actually happening behind the number.
Key Takeaways
- • 24-hour AQI forecasts are ~78% accurate in categorical terms (±1 AQI band)
- • 5-day forecasts should be used for trends only, not specific numbers
- • Sudden pollution spikes — fires, inversions, malfunctions — are nearly unforecastable until they happen
- • Satellite data (TROPOMI, MODIS, GOES) now enables global forecasts where ground monitors are absent
- • The Copernicus CAMS ensemble (11 models) is the global standard for 5-day forecasts
The Four-Step AQI Forecasting Pipeline
When an AQI app shows you a forecast for tomorrow, that number has passed through four distinct modeling steps — each with its own data sources, error sources, and resolution limits.
Emissions Inventory
A gridded database of emission rates for every pollutant from every source type in every location. Road traffic, coal plants, industrial facilities, residential heating, agricultural burning, natural dust and sea spray — all quantified in tons/year per grid cell. Major global inventories: CAMS-GLOB-ANT (Copernicus), EDGAR (EU Joint Research Centre), EPA NEI (USA). Updated annually for anthropogenic sources; fires require near-real-time updates from satellite fire detections (FIRMS/VIIRS).
Numerical Weather Prediction (NWP)
The atmosphere carries pollutants around. Without accurate wind fields, boundary layer heights, precipitation rates, and temperature profiles, chemical transport models cannot simulate where pollution goes. AQI forecasters use the same NWP models as weather agencies — ECMWF's IFS, NOAA's GFS/HRRR, the UK Met Office's Unified Model — as the physics 'engine' that drives pollutant transport. The quality of the AQI forecast is bounded by the quality of the weather forecast.
Chemical Transport Model (CTM)
The CTM takes emissions (from Step 1) and meteorology (from Step 2) and simulates atmospheric chemistry: how NOx + VOCs form ozone in sunlight, how SO₂ oxidizes to sulfate PM2.5, how sea salt and dust particles settle. Common CTMs: CMAQ (EPA), CAMx (ENVIRON), GEOS-Chem (Harvard), WRF-Chem. Each CTM has 100+ chemical reactions and species. Running a CTM at 3km resolution over the continental USA for 48 hours requires supercomputer time — this is why commercial forecasts use lower-resolution global models and only high-priority regions get high-resolution runs.
Data Assimilation (Observations Correction)
Model output is corrected using real observations — satellite aerosol optical depth (AOD) from MODIS/VIIRS, NO₂ columns from Sentinel-5P TROPOMI, and ground-level PM2.5 from surface monitors. This 'nudging' process (technically: optimal interpolation or ensemble Kalman filter) significantly improves AQI accuracy, especially in the first 12–24 hours. Data assimilation explains why a forecast issued at 6am is much more accurate for 6pm that day than a forecast issued the previous evening.
Major AQI Forecast Models Compared
Different platforms use different underlying models. Understanding which model powers a forecast app tells you a lot about its strengths and weaknesses.
| Model | Region | Resolution | Horizon |
|---|---|---|---|
NOAA HRRR High-Resolution Rapid Refresh | USA | 3km | 48 hours |
Copernicus CAMS Copernicus Atmosphere Monitoring Service | Global | 40km | 5 days |
ECMWF GEMS Global and regional Earth-system Monitoring using Satellite and in-situ data | Global | 20km | 5 days |
WRF-Chem Weather Research & Forecasting + Chemistry | Regional (configurable) | 1–12km | 72–120 hours |
BerkeleyEarth ML Berkeley Earth Machine Learning AQI | Global | ~10km | 72 hours |
Forecast Accuracy by Time Horizon
Accuracy degrades non-linearly with forecast horizon. The first 24 hours are significantly more reliable than days 3–5 because weather uncertainty compounds over time — and weather drives pollutant transport.
Near-real-time sensor fusion — highly reliable
Good for typical patterns; misses sudden emission events
Useful guidance; ±1 AQI category expected error
Reliable for large pollution events; local spikes uncertain
Trend direction useful; specific values unreliable
Climatological averages more useful than specific model output
Accuracy defined as: correct categorical AQI level (Good/Moderate/USG/Unhealthy/Very Unhealthy/Hazardous) ±1 category. Based on CAMS validation studies and EPA ForecastReady assessment 2023.
The Satellite Revolution in Air Quality Monitoring
Daily global NO₂ maps that track industrial and traffic emission hotspots with unprecedented resolution. First instrument to image individual power plants' SO₂ plumes from space.
20+ years of global AOD data — the foundation of satellite-based PM2.5 estimation. Fire detection (FIRMS) powers real-time wildfire smoke forecasts used by IQAir and AirNow.
Geostationary — 15-minute repeat coverage over the Americas. Enables continuous wildfire smoke tracking. GOES smoke detection now drives NOAA AQI nowcasts updated hourly.
The satellite gap problem: Polar-orbiting satellites (MODIS, TROPOMI) pass each location once or twice per day. This creates a fundamental gap: a fire that ignites at 10am may not be seen by a satellite until 2pm or the following day, during which it could have grown significantly.
The solution: Geostationary satellites (GOES, Himawari) compensate with near-continuous coverage but lower resolution. The combination of both types — geostationary for real-time detection, polar for accurate measurement — now gives global AQI systems near-complete fire detection within 1–4 hours of ignition in monitored regions.
Machine Learning vs Physics-Based Models
Since 2020, machine learning models have entered AQI forecasting — trained on years of sensor + satellite + weather data to directly predict AQI without simulating atmospheric chemistry explicitly.
Physics-Based CTMs (CMAQ, WRF-Chem)
Machine Learning (LSTM, GNN, Transformer)
Frequently Asked Questions
How does an AQI forecast actually get made?
An AQI forecast combines four types of data and models: (1) Emissions inventories — databases of how much SO₂, NOx, PM2.5, VOCs, and CO are emitted by each source type (power plants, vehicles, industry, fires) in each grid cell. These are updated annually or sometimes daily for fires. (2) A chemical transport model (CTM) — software that simulates how pollutants move through the atmosphere, react chemically (forming secondary PM2.5 and ozone), and deposit on surfaces. Examples: CAMx, CMAQ, GEOS-Chem. (3) A numerical weather prediction (NWP) model — this provides wind fields, temperature profiles, precipitation, and boundary layer height to drive the CTM. Without accurate weather, chemical transport predictions fail. (4) Observations — satellite data (MODIS, Sentinel-5P), surface monitors, and aircraft measurements that correct ('nudge') model outputs in a process called data assimilation. The result is ingested by AQI apps and displayed as forecasts.
Why do AQI forecasts miss sudden pollution spikes?
Sudden AQI spikes — from a factory malfunction, a new wildfire ignition, an unusual temperature inversion, or wind-driven dust storm — are almost impossible to forecast because they require predicting rare events. Models work from emission inventories that assume 'average' emission rates. When a new fire starts, models have no data until satellite instruments detect it — typically 2–6 hours after ignition. Even then, satellite coverage is not continuous; MODIS orbits twice daily. For inversions, NWP models can fail to predict the exact height and persistence of a temperature inversion that traps pollution close to ground. The result: forecast AQI of 80, actual measured AQI of 200. This is why real-time sensor data always supersedes forecast data when available.
How does satellite data improve AQI forecasts?
Three satellite instruments are transformative for air quality forecasting: (1) MODIS (Terra + Aqua) — measures aerosol optical depth (AOD), a proxy for total column PM. MODIS data has been used since 2000 and is the backbone of global AQI estimates in regions without ground monitors. (2) Sentinel-5P TROPOMI — launched 2017, measures SO₂, NO₂, CO, methane, ozone, and aerosol at 3.5km resolution globally, once daily. Its NO₂ maps correlate strongly with vehicle and industrial emissions. (3) GOES-16/18 (USA) and Himawari (Japan/Asia) — geostationary satellites that provide near-continuous aerosol data at 15-minute intervals. GOES smoke detection is now the backbone of USA wildfire AQI nowcasts. Satellites are most valuable in data-sparse regions (Africa, Central Asia) where ground monitors are rare — allowing global air quality mapping with uncertainty estimates.
What is ensemble forecasting and does it work for AQI?
Ensemble forecasting runs the same model many times with slightly perturbed initial conditions and compares the spread of results. A 'tight' ensemble (all members agree) means the forecast is confident. A 'wide' ensemble (members spread apart) signals uncertainty. For AQI, ensemble methods work well for: (a) synoptic-scale pollution transport events (dust storms, transboundary smoke), where the physics are well-constrained; (b) winter inversion forecasting in cities, where the boundary layer depth is the key uncertain variable. Ensemble methods are less effective for: (a) fire behavior prediction — whether a fire will grow or die determines everything but is highly nonlinear; (b) localized urban pollution peaks driven by micro-meteorology below model resolution. The Copernicus CAMS ensemble currently uses 11 models, and its ensemble mean outperforms any individual model by 10–15% RMSE.
How should I actually use a 5-day AQI forecast?
The correct interpretation changes by time horizon: Days 1–2: treat the forecast like a weather app — the categorical level (Good/Moderate/Unhealthy) is usually right, and the specific AQI number is meaningful. Plan outdoor activities accordingly. Days 3–4: use for planning major decisions (should I schedule the outdoor event on Thursday or Friday?) but expect the number to shift by ±1 category. If two days look similar, it probably doesn't matter which you choose. Day 5+: look for trends only — 'does the pollution look like it's building toward a high-AQI episode?' is answerable. 'Will Tuesday be AQI 145 or 158?' is not answerable with useful precision. For health management: if a 5-day forecast shows persistent elevated AQI (all days above 150), that's a real signal to adjust plans — even if individual day values will shift. If all days look below 100, you're likely safe to plan outdoor activities without daily checking.