Improving PM2.5 Exposure Modeling by Hyperlocal Monitoring Using Low-Cost Sensors in the Kolkata Metropolitan Area

Personal exposure to PM_2.5 poses a significant health risk, necessitating assessments at very high spatial and temporal resolutions. However, existing monitoring techniques, like Continuous Ambient Air Quality Monitoring Systems (CAAQMS), provide highly accurate data but are expensive and limited by their sparse distribution. Low-cost sensors (LCS) offer dense spatial data but often encounter reliability challenges and need extensive calibration. These limitations prevent the precise tracking of PM_2.5 exposure at the personal level. To overcome these challenges, we have developed a hybrid framework integrating calibrated LCS data with CAAQMS observations. This approach aims to generate a unified, high-resolution spatiotemporal PM_2.5 database, bridging existing gaps and significantly improving exposure assessments at the personal scale.

This study developed a high spatial (1 km × 1 km) and temporal (1 h) scale PM_2.5 estimates by integrating calibrated static low-cost sensor (LCS) data with hourly ground-based PM_2.5 measurements from CAAQMS across Kolkata, India, for the winter season (1^st December 2023–31^st January 2024). To harmonize these datasets and create a spatiotemporal database of PM_2.5 estimates, we utilized hourly PM_2.5 data from seven CAAQMS stations and calibrated data from 22 static LCS stations. The LCS calibration incorporated meteorological data, precisely temperature (T) and relative humidity (RH), sourced from the nearest CAAQMS stations. 

We employed a Random Forest machine learning model, an ensemble algorithm that effectively captures complex non-linear relationships in the data and improves accuracy by combining multiple decision trees. Our model achieved an approximately 24% reduction in RMSE and an R² of 0.90, validated using an 80:20 train-test split, ensuring robust evaluation of its accuracy. This reduction demonstrates the efficacy of the integrated approach for high-resolution air quality mapping. On 4^th December 2023, PM_2.5 exposure estimates for a common grid point (88.34°E, 22.54°N) were derived using two approaches: one with only CAAQMS data and another with a hybrid of CAAQMS and LCS data. Without LCS, the exposure range at this grid point was 51.34 µg/m³, with an average exposure of 89.40 µg/m³. By integrating LCS data with CAAQMS, the exposure range was reduced to 30.02 µg/m³, and the average exposure increased to 103.39 µg/m³. This increase suggested that LCS might have captured more localized variations, contributing to the higher average exposure value. The reduction in the exposure range indicated a more consistent exposure pattern, highlighting the importance of integrating sparse, accurate CAAQMS data with spatially dense LCS data. This integration enhanced the spatial variability of PM_2.5 and provided a more accurate estimate for personal exposure assessments.