Particulate matter concentrations derived from airborne high spectral resolution lidar measurements using machine learning regression

We use measurements of near-surface aerosol backscatter, extinction, and depolarization acquired by four NASA Langley Research Center airborne High Spectral Resolution Lidars (HSRLs) in machine learning (ML) regression algorithms to derive concentrations of particulate matter (PM) with aerodynamic diameters less than 2.5 µm (PM2.5 ), 10 µm (PM10 ), and the PM2.5 / PM10 ratio. The ML regression models are trained using airborne HSRL measurements acquired over major metropolitan regions in the United States and Asia that are coincident with hourly surface PM2.5 and PM10 measurements from the EPA air quality system and similar networks in other countries. We examine several regression methods and find that exponential Gaussian Process regression (GPR) algorithms consistently give the best performance in terms of the lowest root-meansquare (RMS) errors and the highest correlations. When evaluated using surface measurements withheld from the training sets, ML models that use the HSRL near-surface measurements of aerosol backscatter and aerosol intensive properties such as depolarization, backscatter color ratio, and lidar ratio typically give the best performance with RMS differences in PM2.5 retrievals around 5 µg m−3 and correlation coefficients above 0.8, respectively. Corresponding RMS differences and correlation coefficients for PM10 retrievals are 11 µg m−3 and 0.7 and corresponding RMS differences and correlation coefficients for PM2.5 / PM10 are 0.17 and 0.75. This retrieval performance is achieved using airborne HSRL measurements alone and so does not depend on external knowledge of or