Estimation of PM10 and PM2.5 Using Backscatter Coefficient of Ceilometer and Machine Learning

ABSTRACT

Air quality issues, including health and environmental challenges, have recently become more relevant in urban areas with large populations and active industries. Therefore, particulate matter (PM) estimation with high accuracy using various methods is required. In this study, PM₁₀ and PM_2.5 in Cheongju city, South Korea, were estimated using the attenuated backscatter coefficient of the ceilometer and meteorological observation data from an automatic weather station with supervised machine learning (ML). The backscatter coefficient data were obtained from the vertical layer with the highest correlation with PM₁₀ and PM_2.5. The estimation methods utilized were tree-, vector-, neural-, and regularization-based supervised ML. The extreme gradient boosting method yielded the highest PM estimation accuracy. The estimation of PM₁₀ and PM_2.5 for the test data set was more accurate than that in previous studies that used satellite and ground-based meteorological data (bias = 0.10 µg m^–3, root mean square error (RMSE) = 14.44 µg m^–3, and R = 0.92 for PM₁₀; and bias = 0.12 µg m^–3, RMSE = 7.16 µg m^–3, and R = 0.91 for PM_2.5). Particularly, the correlation coefficient was the highest for the estimation results for strong haze cases (1 km < visibility ≤ 5 km) (R = 0.95 for PM₁₀; R = 0.89 for PM_2.5). Therefore, PM estimation using meteorological observation data can help obtain meteorological and PM information simultaneously, making it useful for air quality monitoring.