Data and sources

Every number, traced.

VayuDrishti uses only real, free, public data. No synthetic values: gaps are filled by models with uncertainty bands and labels, never invented.

Provenance

Sources

Source	What we use	Auth
OpenAQ S3 archive	CPCB hourly history: PM2.5, PM10, NO2, SO2, O3, CO, from Feb 2025	None
data.gov.in CPCB	Live station snapshot	Free key
Open-Meteo, ERA5, CAMS	Wind, temperature, humidity, precipitation, boundary-layer height. CAMS is a covariate only, never a validation target	None
NASA FIRMS	VIIRS 375m active fires and fire radiative power	None
NASA GIBS WMTS	Satellite raster overlays: Sentinel-5P NO2, MODIS AOD, VIIRS	None
Google Earth Engine	Numeric satellite features (Sentinel-5P, MODIS MAIAC AOD) and WorldPop population	Service account
Datameet and OSM	Ward boundaries, roads, built-up area, vulnerable sites	None
US Diplomatic Post (via OpenAQ)	Independent validation network, never trained on	Free key

Honesty

Coverage and gaps

Ward vintage is best-available and differs by city: Mumbai uses 24 BMC wards, Bengaluru 198 BBMP wards, Delhi 290 wards. Granularity is noted per city.
Training data is a single winter (Feb 2025 onward). No multi-season generalization is claimed; this is stated on the receipts page.
Live AQI is dull in the monsoon. The demo pairs it with a Nov 2025 replay window built from out-of-fold predictions.
Stations and grid cells outside any ward polygon are kept as an unassigned bucket, never dropped.

Satellites

Numeric satellite path

checking

Five real satellites: Sentinel-5P (TROPOMI), Terra and Aqua (MODIS MAIAC AOD), Suomi-NPP and NOAA-20 (VIIRS). Visuals come from zero-auth GIBS and are always on. Numeric features run through Google Earth Engine when the service account is live; parquet columns are always present, NaN when unavailable, so absence never blocks.

CPCB methodology