Modelling ambient PM2.5 exposure at an ultra-high resolution and associated health burden in megacity Delhi: exposure reduction target for 2030

Deriving hyperlocal information about fine particulate matter (PM _2.5 ) is critical for quantifying exposure disparities and managing air quality at neighborhood scales in cities. Delhi is one of the most polluted megacities in the world, where ground-based monitoring was limited before 2017. Here we estimate ambient PM _2.5 exposure at 100 m × 100 m spatial resolution for the period 2002–2019 using the random forest model. The model-predicted daily and annual PM _2.5 show a ten-fold cross-validation R ^2 of 0.91 and 0.95 and root mean square error of 19.3 and 9.7 μ g m ^−3 , respectively, against coincident ground measurements from the Central Pollution Control Board ground network. Annual mean PM _2.5 exposure varied in the range of 90–160 μ g m ^−3 in Delhi, with shifts in local hotspots and a reduction in spatial heterogeneity over the years. Mortality burden attributable to ambient PM _2.5 in Delhi increased by 49.7% from 9188 (95% uncertainty interval, UI: 6241–12 161) in 2002 to 13 752 (10 065–19 899) in 2019, out of which only 16% contribution was due to the rise in PM _2.5 exposure. The mortality burden in 2002 and 2019 are found to be higher by 10% and 3.1%, respectively, for exposure assessment at 100 m scale relative to the estimates with 1 km scale. The proportion of diseases in excess mortality attributable to ambient PM _2.5 exposure remained similar over the years. Delhi can meet the United Nations Sustainable Development Goal 3.4 target of reducing the non-communicable disease burden attributable to PM _2.5 by one-third in 2030 relative to 2015 by reducing ambient PM _2.5 exposure below the World Health Organization’s first interim target of 35 μ g m ^−3 . Our results demonstrate that machine learning can be a useful tool in exposure modelling and air quality management at a hyperlocal scale in cities.