Determination of carbonaceous aerosols in IMPROVE network samples based on FTIR spectroscopy - Assessing the impact of spectral processing methods

摘要

Ambient particulate matter samples collected in the IMPROVE network are routinely processed by FTIR spectroscopy coupled with multivariate calibration to assess the mass of organic and elementary carbon deposited on PTFE filters. Since PTFE absorption and scattering contributions are the main sources of variability in the raw spectra, additional factors are required in the calibration model to account for PTFE interferences. Although the predictive capability of the model is satisfactory for raw spectra, interpretation is not always possible. Both model interpretability and prediction capability may be enhanced by considering processing techniques which either estimate and correct the slope of the spectra or simply remove any baseline contributions. Compared with raw spectra, a significant reduction in model dimensions is obtained by combining second derivative spectra with an extra step associated with variable selection. Although not always suitable for interpretation, models based on second derivative spectra readily minimize the number of PLS factors but tends to remove broad features associated with particulate matter. For a better band assignment, improvement can be claimed using either smoothing splines or the null space projection techniques. Despite featuring a larger number of factors, baseline corrected spectra provided by both approached are more prone to interpretation due to their absorption spectra-like shape. Selecting the best processing techniques is thus defined as a tradeoff between predictive capability and model interpretability. For practical applications implementing spectral processing techniques instead of using the raw spectra can, for example, enhance the prediction accuracy of heavily loaded samples which scatter light and contribute to the baseline signal to a greater extent compared with calibration samples.