Air · Method
How the air numbers were checked
Every number on the air page comes from public data and code anyone can rerun. This page lists the tests that could have broken the results, what the data looks like, and what the analysis can't tell you.
Tests that could have broken the result
If the crop-fire share were an artefact of one choice, changing that choice would change the answer. If it were just "November is smoggy", the two fake tests would show a large share too. They don't.
How smog rises with fire heat on the air's path
The data, and how it was checked
Method, step by step
- Measured air. Hourly readings from CPCB and UPPCB stations, via OpenAQ. City values average the stations reporting that day. The model uses the four stations present in every winter (Talkatora, Lalbagh, Central School, Gomti Nagar).
- AQI. India's National AQI (CPCB, 2014). Sub-indices come from 24-hour averages of PM2.5, PM10, NO₂ and SO₂, and from the highest 8-hour average of CO and ozone. A day needs at least 16 hours of data and 3 pollutants, one of them PM2.5 or PM10. The highest sub-index is the AQI. The figures shown here are calculated by this site and may differ slightly from CPCB's own.
- Air paths. Eight times a day, an air parcel is traced 120 hours back from Lucknow using 100 m winds (ERA5 reanalysis and Open-Meteo forecasts).
- Fires. NASA VIIRS satellite fire detections within 50 km of each path on the same day, weighted by their heat output (fire radiative power).
- Model. Gradient-boosted trees predict daily PM2.5 from mixing height, wind, temperature, humidity, rain, 3-day stagnation, where the air came from, and each winter's own baseline. Fire heat on the air's path is added with a rule that more fire can only mean more smoke.
- Crop-fire share. For each day: the prediction with that day's fires, minus the prediction for a quiet-fire day (the 25th percentile of fire heat). Days with almost no fire on the path are mostly rainy days, so comparing against "zero fires" would credit rain's cleaning to the fires. That version gives about 30%, the same in December as in November, and is shown as rejected in the tests table.
- Diwali. A weather-only model, trained without the Diwali days, predicts what Diwali and the days after should have looked like. The extra is measured minus predicted.
- Testing. Every reported share comes from days the model never saw in training: two-week blocks are held out one at a time. Ranges come from 40 refits on resampled weeks.
Limits
- The gap from November 2022 to February 2025. OpenAQ has no Lucknow station records for this period, so the 2023 and 2024 winters are left out rather than filled with modelled data.
- Fires and cold weather arrive together. November brings the most fires and the stillest air at the same time. The model and the fake tests separate them as far as the data allows, but the spread in the tests table is the real uncertainty.
- Satellites miss fires. Small fires, fires under smoke or cloud, and fires between satellite passes go uncounted. Smoke from them would show up as "local and regional".
- "Crop fires" includes Uttar Pradesh's own fields, not only Punjab and Haryana.
- Station changes. Winters since 2025 read much lower than 2019–2021. Part of that may be real improvement and part may be weather or changes at the stations. Each winter gets its own baseline in the model, so this doesn't affect the shares, but it does mean the years aren't directly comparable yet.
- The current season. Satellite fire records for a season build up day by day from the start of the daily runs, so early in a season the crop-fire share covers only the days with a full fire record. Outside October–December the 3-day forecast uses the CAMS air-quality model instead, adjusted to Lucknow's stations.