National, satellite-based land-use regression models for estimating long-term annual NO2 exposure across India

In India, scarcity of ground-based measurements of nitrogen dioxide (NO 2 ) is a major challenge for estimating long-term exposure and associated health impacts. This study aimed to develop and validate a national-scale annual NO 2 exposure model for India for 2019 and determine if model cross-validation predictive ability was improved by including non-continuous (manual) measurements along with reference-grade, continuous measurements. We used a supervised forward-addition linear regression method to fit land use regression (LUR) models developed with up to 804 Central Pollution Control Board ground monitoring stations (n = 157 continuous, n = 647 manual) and 209 spatial predictor variables, including satellite-based tropospheric NO 2 columns. Two models were developed: one using continuous sites only and one using continuous and manual sites, with standard diagnostics and cross-validation (CV) methods. We also assessed if the kriging of final model residuals reduced spatial autocorrelation and improved model CV results. LUR coefficients for the best-performing model were applied to predictors for 2015–2021 and gridded at 100 m to estimate population-weighted exposure. The continuous sites-only model and combined continuous and manual sites models had CV-R 2 values of 0.59 (root-mean-square error [RMSE]: 9.4 μg/m 3 ) and 0.54 (RMSE: 8.3 μg/m 3 ), respectively, and both included the satellite NO 2 predictor. Kriging residuals increased the CV-R 2 of the combined model to 0.70 (RMSE: 7.2 μg/m 3 ) but offered no improvement for the continuous site model. National population-weighted average NO 2 was 22.1 μg/m 3 in 2019. We estimated over 92% of the Indian population was exposed to annual NO 2 exceeding the WHO air quality guideline (10 μg/m 3 ). In Delhi, Mumbai, and Kolkata, an estimated 45%, 100%, and 100% of the population, respectively, experienced annual NO 2 levels that surpassed Indian standards (40 μg/m 3 ). To our knowledge, this is the first such long-term NO 2 LUR model specific to India, and predictions are available to interested researchers. • LUR model developed using predictors and CPCB continuous and manual sites for annual NO 2 predictions gridded at 100 m. • Merging two types of monitors leaveraged a better understanding of NO 2 spatial variations across larger portion of population. • Kriging improved the model performance and reduced RMSE. • In 2021, Mumbai, Kolkata, and Delhi had peak population-weighted annual NO 2 of 51.5, 50.3, and 42.6 μg/m 3 , respectively.