Comparison of Biases in Filter-Based Aerosol Light Absorption Measurements over a Rural and Urban Area

JOSHIN KUMAR, Nishit Shetty, Patrick Sheridan, Allison Aiken, Manvendra Dubey, August Li, Ganesh Chelluboyina, Benjamin Sumlin, Joseph V. Puthussery, Rajan K. Chakrabarty, Washington University in St. Louis

     Abstract Number: 508
     Working Group: Remote and Regional Atmospheric Aerosol

Abstract
Measuring aerosol light absorption is crucial to assess direct radiative forcing that impacts local and global climate. At the ARM Southern Great Plains (SGP) site, light absorption by aerosols has been continuously measured using a three-wavelength (467, 530, and 660 nm) Particle Soot Absorption Photometer (PSAP). While filter-based instruments like the PSAP are low-cost and simple to operate, they suffer from unquantifiable artifacts due to the filter medium and complex interactions between filter fibers and deposited aerosols.

Various correction algorithms (Bond et al., 1999; Virkkula et al., 2005; Li et al., 2020) have been introduced to correct for the filter-based absorption coefficient measurements toward predicting the particle-phase absorption coefficient (Babs). These algorithms have a fixed analytical form, which limit their universal applicability across diverse environments (e.g. rural, urban, marine).

In this study, we analyze and compare aerosol light absorption measurements over the rural SGP site and the urban La Porte area of Texas (site of TRACER campaign), respectively, using ARM’s PSAP and PI-deployed guest instruments, namely, multiwavelength photoacoustic spectrometers (PASS). Photoacoustic spectroscopy is a first-principles measurement technique of particle-phase Babs. We find that the PSAP overestimates the absorption measurements by a factor of four and three, respectively, over the SGP and La Porte sites in comparison to PASS measurements.

To correct for this overestimation in PSAP measurements, we implemented the widely used Virkkula/Bond analytical correction algorithms, as well as a Random Forest Regression (RFR) machine-learning algorithm. Our results show that the RFR algorithm significantly outperforms the extent of corrections achieved using the Virkkula/Bond algorithms.

By applying the RFR algorithm, the wavelength averaged Root Mean Square Error (RMSE) for predicting particle-phase Babs is reduced by approximately 88% and 60% of the RMSE of the Virkkula algorithm at the SGP and La Porte sites, respectively. The RFR algorithm predicted Babs values within approximately 25% and 40% of the reference Babs measured by the multiwavelength PASS at SGP and La Porte sites, respectively.

Machine learning offers a promising path to correct for biases in long-term filter-based absorption datasets and accurately quantify their variability and trends needed for robust radiative forcing determination.