Does it work?

Method

1. 1A season replay around the 11 Nov 2025 smoke episode (window shown above). Definitions were fixed before looking at the data.
2. 2Observed smoke signal: PM2.5 minus a rolling 7-day 10th-percentile baseline at the same station.
3. 3Calibration: least squares PM2.5 excess ≈ α × model smoke, fitted on the first part of the window; skill is scored on the held-out remainder only.
4. 4An episode is an excess of at least 60 µg/m³ for 3 or more hours; a hit is a forecast crossing within ±6 h.
5. 5A uses VIIRS/MODIS detections only, as official counts do. B adds hourly afternoon geostationary detections. The A-to-B difference is the measured cost of the afternoon gap.

Limitations

• Smoke from farm fires is one part of PM2.5. Traffic, industry, dust and winter inversions are not modelled; the excess baseline only partly removes them.
• If reanalysis weather was used instead of archived forecasts, skill here is better than a live forecast would achieve.
• Fires hidden by cloud or too small for the satellites are invisible to the model.
• Monitors are sparse in rural Punjab, where the smoke lands first, so village-level forecasts are not directly verified.
• Some stations fail. They are shown here, not dropped.