Bridging Data Gaps for Air Quality Monitoring: Daily PM2.5 Estimates for 10 km Grid Cells in India

India has experienced elevated levels of Particulate Matter (PM) _2.5 concentrations. Despite increased efforts by the Indian government,  the current monitoring network remains limited, impeding a comprehensive understanding of PM_2.5 variations throughout the country. Limited PM_2.5 data has led previous health studies to rely on publicly-available monthly PM_2.5 estimates. However, these estimates have large uncertainties over the under-monitored regions, including India because PM_2.5 observations have been calibrated into their model. The coarse temporal resolution of existing datasets makes it challenging to assess short-term effects of exposure to PM_2.5. To bridge these gaps, it is imperative to develop daily PM_2.5 datasets with robust spatial and temporal certainty.

This study develops open-source daily PM_2.5 datasets at a 10 km resolution for India spanning almost two decades (2005 - 2023). Leveraging two-stage machine learning model with 10-fold spatial cross-validation (CV), we generate PM_2.5 estimates for regions without ground measurements. In contrast to random k-fold CV, widely used in previous studies, spatial CV is implemented in this study to control for spatial auto-correction, which could lead to overfitting to the training data and underestimation of spatial prediction errors. The first stage fills missing observations for daily MODIS AOD, Sentinel-5P mission's TROPOPOMI NO₂, and TROPOMI CO. The second stage predicts daily ground-measured PM_2.5 concentrations. Two models are constructed for the second stage: the AOD model and the Full model, the latter incorporating TROPOMI features in addition to AOD.

The Full model exhibits a spatial out-of-sample performance with an R² of 0.68, effectively predicting local and temporal PM_2.5 variations rather than just average differences between locations, months, or years (within R² = 0.49). The AOD model performs similarly, with an R² of 0.64 and within R² of 0.45. At the monthly level, our model outperforms the existing monthly PM_2.5 dataset, with an R² of 0.74 and within R² of 0.52. 

Utilizing our PM_2.5 predictions, we identified that 31% of 10 km grid cells across the country demonstrated a more than 5% reduction in PM_2.5 concentrations in 2018-2022 compared to 2005–2010, and any decrease in PM_2.5 was observed in 75% of the locations. Additionally, population-weighted annual average PM_2.5 concentrations indicate a decline since 2018, except for a notable increase in 2021. Despite an overall declining trend since 2018, approximately 60% of the population remains exposed to PM_2.5 concentrations above the national annual guideline (40 µg/m³), with 10% facing extreme levels of 80 µg/m³ annually.

Our method is useful for resource-constrained countries to understand nationwide air quality trends and identify areas with elevated pollution. To address this, we established the optimal number of air quality monitors using multiple machine learning models with randomly-sampled incremental training data. Our findings show a polynomial increase in within R² for test data, ranging from 0.24 at 25 monitors to 0.54 at 300 monitors in the training data.

Our predictions offer valuable insights into air quality trends in India from 2005 to 2023. Importantly, our estimates contribute to understanding the number of ground monitors needed to explain variations in PM_2.5 concentrations across the country, offering insights for other countries.