A silent or low reading is inconsistent with its neighbours, not proof of intent.
PM2.5, shaded between monitors. Rings are silent monitors.
Around 1 am most silent monitors are estimated at the 1,000 limit, so the estimate sits at the top of its interval. Replays on earlier Diwalis show estimates run low, not high.
Oct to Dec 2025, in the hours when the city median was at or above 300 µg/m³. Pick a monitor on the map.
Share of high-pollution hours reported.
A flag means PM10 fell relative to nearby monitors, more than PM2.5 did, for 1 to 6 daytime hours. It says nothing about cause: weather, local activity, instrument drift and water sprinkling can all do this.
Every number comes from held-out tests. The targets were set in advance and did not change after seeing results.
Tested on Oct to Dec 2025 with models trained on earlier winters and other stations. About 10% of real readings were hidden in 2 to 8 hour blocks. Errors are in µg/m³.
Diwali 2025's exact pattern of silent stations was hidden on Diwali 2023 and 2024, where the readings are known.
PM10 reduced 15%, PM2.5 2 to 7%, for 1 to 6 daytime hours. Thresholds were fixed for 1 false alarm per station-month on separate clean data.
Uses the neighbour-offset fill on 2025 hourly data, the benchmark's best method without learning. The trained model and its intervals cover Delhi only.