Key Details
The World Bank’s Policy Research Working Paper 11448, Tracking Air Pollution from Space: High-Resolution Daily Estimates for PM₂.₅, O₃, NO₂ and CO, develops a machine-learning method for estimating pollution where ground monitoring is limited.
Coverage: More than 1.5 million terrestrial grid cells at 0.1-degree resolution—approximately 11 kilometres at the equator.
Pollutants: Fine particulate matter (PM₂.₅), ozone (O₃), nitrogen dioxide (NO₂) and carbon monoxide (CO).
Urban reach: Population-weighted estimates for more than 8,000 functional urban areas, meaning cities together with their surrounding commuting zones.
Pilot period: Daily estimates from 10 April to 17 November 2025. The underlying satellite data could support a longer series from May 2018.
Validation: Across cities with sufficient monitoring stations, median correlations between predicted and monitored daily pollution exceeded 0.69 for all four pollutants; upper-quartile correlations exceeded 0.82.
India coverage: Ahmedabad, Asansol, Delhi, Kolkata and Mumbai feature in the city-level analysis as validation/illustration cases.
Five Data Sources Produce One Daily Pollution Grid
The method combines five sources to estimate pollution between scattered ground monitors:
Ground observations + satellite data + emissions + weather + geography → daily pollution grid → population exposure
Ground observations from IQAir train the model, while ESA’s TROPOMI provides satellite measurements and the Emissions Database for Global Atmospheric Research supplies emissions estimates. Weather and geographic variables capture how pollutants disperse under different atmospheric and physical conditions.
A random-forest model with 500 decision trees combines these inputs separately for each pollutant.
Different Pollutants Depend on Different Signals
The importance of each input varies by pollutant:
PM₂.₅, NO₂ and CO: Satellite observations were the strongest predictor group, followed by emissions, weather and geography.
Ozone: Weather was most influential because ozone forms through atmospheric reactions rather than being emitted directly.
All four pollutants: Location and terrain remained important, showing that similar emissions can produce different ambient concentrations.
PM₂.₅ recorded the strongest statistical fit in the test data. Performance for ozone, NO₂ and CO was weaker, though the models still captured meaningful daily variation.
Indian Cities Show Both Potential and Uneven Accuracy
Across Ahmedabad, Asansol, Delhi, Kolkata and Mumbai, estimated PM₂.₅ showed a broadly similar seasonal pattern: declining from May to August, stabilising through September and rising towards November.
Performance varied by city and pollutant. Ozone estimates tracked observations particularly well in Kolkata, with strong results in Mumbai; Ahmedabad showed a close fit for CO, while Delhi’s performance was more moderate.
The results show that the method can reproduce pollution patterns and turning points, but accuracy in one city or for one pollutant cannot be assumed elsewhere.
The Dataset Could Extend Pollution Evidence Beyond Monitoring Stations
Longer time series could help identify pollution hotspots and population exposure beyond existing monitoring networks, combine exposure with health and poverty data, assess changes following pollution-control measures and support short-term forecasting using weather predictions.
The accompanying PM₂.₅ platform has already extended estimates to more than 30,000 cities for October 2018–August 2026. The paper itself remains a seven-month pilot covering four pollutants.
Where Caution Is Needed
Predictions depend partly on the coverage and quality of ground monitors used for training. IQAir combines regulatory stations with sensors operated by institutions, researchers, civil-society organisations and individuals, and coverage varies considerably across regions.
The 0.1-degree grid is useful for city- and regional-scale analysis but cannot reliably represent exposure beside individual roads, industrial facilities or neighbourhood sources.
Most importantly, the model estimates pollutant concentrations, not their sources. Determining whether pollution comes from traffic, industry, dust, waste burning, household fuels or transported pollution still requires source-apportionment, inspections and local monitoring.
Sulphur dioxide is excluded because satellite coverage was considered insufficient.
Policy Relevance
India’s ambient air-quality network had expanded to 1,646 manual and continuous stations across 603 cities by March 2026, but substantial spatial gaps remain —particularly across smaller cities, rural areas and cross-border airsheds.
Best use: Satellite-derived estimates can help CPCB and state pollution-control boards identify where additional monitors are needed, compare exposure across districts and assess whether pollution-control measures coincide with sustained changes.
Necessary guardrail: Modelled concentrations should supplement, rather than substitute for, calibrated ground stations used for compliance and enforcement. Accuracy must be tested by pollutant, season and region.
Policy test: Integration with the National Clean Air Programme would be valuable only if the data improve decisions—such as targeting funds, revising city action plans or identifying underserved populations—not simply increase the number of available pollution maps.
Follow the Full Paper Here: Tracking Air Pollution from Space: High-Resolution Daily Estimates for PM₂.₅, O₃, NO₂ and CO