Improving Quantitative Accuracy of Ammonia from Open-Path Fourier Transform Infrared Spectroscopy by Incorporating Actual Spectra into Synthetic Calibration Set of Partial Least Squares Regression

Abstract

Partial least squares (PLS) regression is often employed for quantification of ammonia (NH3) from open-path Fourier transform infrared (OP/FT-IR) spectra. In this work, PLS models built with both synthetic and actual calibration sets are evaluated, and it is found that the root mean square error of prediction (RMSEP) of the actual model is 86% lower than that of the synthetic model. Although the quantitative accuracy is mediocre, synthetic calibration set is readily constructed, and still required in the beginning when there are insufficient actual OP/FT-IR spectra. However, the RMSEP of the synthetic model is significantly reduced by incorporating some actual spectra into the synthetic calibration set, i.e. by constructing a mixed calibration set. Results show that a mixed calibration set composed of three quarters of synthetic and one quarter of actual spectra reduces the RMSEP by 80%, which performs similarly to the actual calibration set. Synthetic, mixed, and actual calibration sets all have pros and cons, and should be utilized according to respective advantages, which results in a novel strategy for NH3 quantification. The strategy starts with a synthetic model, later builds mixed models by incorporating actual spectra into the synthetic calibration set, and eventually builds an actual model after a large number of OP/FT-IR spectra are collected. This strategy is more accurate than a synthetic calibration set, and more efficient than an actual calibration set, the preparation of which is time-consuming.