Increasing the accuracy of atmospheric ammonia concentrations calculated from Open-Path Fourier transform infrared spectra using partial least squares model by scanning and removing interference spectral data

Monitoring atmospheric pollutants has raised worldwide concerns in last decades, due to deterioration of air quality. Multivariate regression methods have gained extensive applications in calculating and monitoring concentrations of air pollutants, for example ammonia (NH3), using Open-Path Fourier Transform Infrared (OP/FT-IR) spectra data. However, the prediction accuracy of multivariate regression models is interfered heavily by the dominant and omnipresent absorption bands of atmospheric H2O vapor and CO2 in OP/FT-IR spectra. Hence, a new method of variable selection, referred to as window of scanning and removing interference information (WSRII), is developed to remove interference data of OP/FT-IR spectra. The key of WSRII is to confirm informative variables according to the change of root mean square error of calibration in the partial least squares regression (PLSR) model after removing a spectral window. If the change is greater than 0, the spectral window is reserved as an informative information window. Then, the new matrixes of spectral data are reconstructed by using the informative information windows, which are selected by changing position of the spectral window in full wave number range. Based on this, a PLSR model is rebuilt to predict accurately NH3 concentrations. The results showed that the proposed method was able to eliminate uninformative information and improved prediction accuracy of PLSR models. Moreover, this process of variable selection is significantly potential to improve prediction accuracy of PLSR model for monitoring atmospheric NH3 concentrations in real time.