Seasonally optimized calibrations improve low-cost sensor performance: long-term field evaluation of PurpleAir sensors in urban and rural India

Abstract

Lower-cost air pollution sensors can fill critical air quality data gaps in India, which experiences very high fine particulate matter (PM2.5) air pollution but has sparse regulatory air monitoring. Challenges for low-cost PM2.5 sensors in India include high-aerosol mass concentrations and pronounced regional and seasonal gradients in aerosol composition. Here, we report on a detailed long-time performance evaluation of a popular sensor, the Purple Air PA-II, at multiple sites in India. We established three distinct sites in India across land use categories and population density extremes (in urban Delhi and rural Hamirpur in north India and urban Bengaluru in south India), where we collocated the PA-II model with reference beta attenuation monitors. We evaluated the performance of uncalibrated sensor data, and then developed, optimized, and evaluated calibration models using a comprehensive feature selection process with a view to reproducibility in the Indian context. We assessed the seasonal and spatial transferability of sensor calibration schemes, which is especially important in India because of the paucity of reference instrumentation. Without calibration, the PA-II was moderately correlated with the reference signal (R2 = 0.55–0.74) but was inaccurate (NRMSE ≥ 40 %). Relative to uncalibrated data, parsimonious annual calibration models improved the PurpleAir (PA) model performance at all sites (cross-validated NRMSE 20 %–30 %; R2 = 0.82–0.95), and greatly reduced seasonal and diurnal biases. Because aerosol properties and meteorology vary regionally, the form of these long-term models differed among our sites, suggesting that local calibrations are desirable when possible. Using a moving-window calibration, we found that using seasonally specific information improves performance relative to a static annual calibration model, while a short-term calibration model generally does not transfer reliably to other seasons. Overall, we find that the PA-II model can provide reliable PM2.5 data with better than ±25 % precision and accuracy when paired with a rigorous calibration scheme that accounts for seasonality and local aerosol composition.