Grantee Research Project Results

Final Report: Orthogonal Background Suppression Technique for EPA's Field Infrared Data Processing

EPA Grant Number: R825366
Title: Orthogonal Background Suppression Technique for EPA's Field Infrared Data Processing
Investigators: Blatherwick, R. D.
Institution: University of Denver
EPA Project Officer: Hahn, Intaek
Project Period: October 1, 1996 through September 30, 1999 (Extended to September 30, 2000)
Project Amount: $248,743
RFA: Analytical and Monitoring Methods (1996) RFA Text |  Recipients Lists
Research Category: Ecological Indicators/Assessment/Restoration , Environmental Statistics , Water , Air , Land and Waste Management

Objective:

The research project, which the University of Denver entitled "Spectral Background Suppression" (SBS), was intended to develop tools which can be used to extract information on pollutant gases from Fourier Transform Infrared (FTIR) spectra obtained at field sites having highly variable concentrations of water vapor.

Summary/Accomplishments (Outputs/Outcomes):

This work dealt primarily with investigation of orthogonal background suppression (OBS) and model matching techniques as a means to remove background effects in infrared spectral data. The background in infrared transmission measurements is a combination of source, instrument and atmospheric effects. Since water vapor is the dominant atmospheric absorber in the mid infrared, a significant part of this work was devoted to finding ways to remove the contribution of water vapor from spectra obtained over horizontal paths ranging from a few meters to a few hundred meters under a variety of temperature, pressure and humidity conditions.

One of the primary mathematical techniques employed in this study is that of singular value decomposition (SVD). This is a standard technique, and routines to perform SVD are readily available in many mathematical packages. The one used in this study was MATLAB_. In this technique, a matrix of m vectors, each of length n, is decomposed into m vectors (also of length n) which are all orthogonal to each other. In addition, the singular vectors obtained from SVD are ordered according to the amount which they contribute to the original set of vectors. The first singular vector represents the average of all of the original vectors, and has the largest singular value (or, eigenvalue). Subsequent singular vectors represent deviations from the average. Much of this study deals with sets of infrared spectra recorded over a fixed path at the same location on different days. These spectra are the "vectors" upon which SVD is performed. Since the spectra differ in relatively small details, only a few singular vectors can be used to represent the entire set of spectra (i.e., the singular values associated with these vectors diminish rapidly).

Since the primary goal of this project was to investigate methods to facilitate the analysis of actual field spectra for the presence of spectral signatures arising from pollutants, a considerable effort was undertaken working with closed path spectra provided to us by the EPA's Office of Research and Development Research Triangle Park (RTP) facility. These spectra fell into two classifications: a group of about 400 spectra recorded during the early autumn of 1994 at the EPA RTP site, and approximately 270 spectra recorded during the early summer of the same year. By performing SVD analysis on subsets of these spectra, it was determined that if a SVD was performed on a set of about 200 spectra, then any member of the larger group could be adequately reconstructed using only 30-40 singular vectors.

A test was performed by degrading the spectral resolution of one of the spectra by 25 percent and attempting to regenerate it from the singular vectors. It was found that the spectrum could not be adequately reconstructed, leading to the conclusion that the spectra used in the SVD analysis must all have very similar resolution.

The spectral signature of a possible pollutant gas was introduced into one of the field spectra by multiplying it by a computed spectrum of ammonia, and after using OBS to remove the background, the ammonia signature was clearly visible. Thus, if a data base of unpolluted spectra exists, this technique can be very useful in quickly isolating spectral signatures of potential target gases.

We also investigated a related technique, that of orthogonal filtering. In this analysis, a computed or measured absorption spectrum is used to generate a "filter" which is orthogonal to all of the 30-40 singular vectors needed to reconstruct a set of spectra, and therefore very nearly orthogonal to any spectrum which can be represented by those singular vectors. This filter can then simply be multiplied (in a vector dot-produce sense) by a test spectrum, and both the presence and amount of target gas present in the test spectrum deduced from the result. We generated orthogonal filters for ammonia and sulfur dioxide and applied them to field spectra.

A different approach to the problem of identifying a target gas in an infrared spectrum investigated in this study is the method of model matching. In this method, one attempts to simulate the observed spectrum by generating a synthetic spectrum for the conditions under which the actual data were obtained using a line-by-line (or other) computer code. Inputs to these codes consist of a set of atmospheric parameters (temperature, pressure, mixing ratios of atmospheric gases, etc.) and a set of spectroscopic parameters (positions of the absorption lines in frequency, intensity of the lines, etc.) for the gases. In this study, we investigated a method of using line-by-line simulations to remove the combined affects of the instrument and the source to convert raw field data into transmission spectra by reducing the spectral resolution of both the observed and calculated spectra by a factor of about 25. We demonstrated that the same line-by-line calculation at full instrument resolution can be then be used to isolate spectral signatures of suspected target gases and to quantify those gases. These calculations are fairly computer intensive, however, and several iterations are generally required during which the amount of the interfering atmospheric water vapor, spectral resolution, and instrument line shape function must all be varied to maximize the agreement between the computed and observed spectrum. (Other atmospheric gases such as carbon dioxide, methane, carbon monoxide, etc., may need to be included in the simulations, but the abundances of these are usually known.)

Since water vapor is the predominate interferent over much of the mid infrared spectral region, we performed a series of laboratory measurements on water vapor and used the results to verify that state-of-the-art line-by-line computer codes, in conjunction with the most recent spectroscopic line parameters, can reliably be used to simulate closed path field spectra over the range 700-4500 wavenumbers (~2.2 to 7 microns). This study was conducted using a White cell about 85 cm long configured for 40 reflections, giving a total path length of about 34 meters. The cell was temperature controlled, and was filled with water vapor by passing dry nitrogen gas over the boil off of a flask of distilled water. The measurements were performed at a spectral resolution of 0.1 wavenumbers, and at temperatures of 22C, 30C and 40C. In order to determine spectral transmittance from such laboratory measurements, one must divide the spectra recorded with water vapor in the cell by spectra recorded with no water vapor in the cell. Since part of the optical path (~2 meters) passed through the air in the laboratory, regions of very strong water absorption (primarily ~1400-1900 wavenumbers, and 3500-4000 wavenumbers) could not be treated in this work. This is not a serious limitation, however, since in any field spectra over a distance of ~50 meters or longer would essentially be completely absorbed in these regions anyway. A comparison of the measured water vapor transmittance with the results of line-by-line model simulations showed excellent agreement over the spectral region studied. We concluded from these comparisons that computer generated simulations of atmospheric transmission can reasonably be applied for background suppression using model matching if needed.

However, OBS using SVD remains the easiest and fastest method of background suppression if a sufficient data base of pollutant free background spectra are available. SVD is especially powerful if it is used to generate orthogonal filters for target gases of interest, since then one need only perform the simple mathematical operation of calculating the dot product between the filter and a test spectrum to identify those target gases in the test spectrum.

Journal Articles:

No journal articles submitted with this report: View all 1 publications for this project

Supplemental Keywords:

measurement methods, modeling, principal component analysis., RFA, Scientific Discipline, Ecosystem Protection/Environmental Exposure & Risk, Air, Toxics, Engineering, Chemistry, HAPS, Environmental Chemistry, tropospheric ozone, Monitoring/Modeling, water vapor, field portable monitoring, Fourier transform infrared, air quality data, atmospheric chemistry, environmental monitoring, spectroscopic, measurement methods , ambient particle properties, real-time monitoring, atmospheric monitoring, remote sensing, aerosol analyzers, analytical chemistry, ambient air, FTIR, orthogonal background suppression, spectroscopic studies

Progress and Final Reports:

Original Abstract

1997 Progress Report

1998 Progress Report

1999 Progress Report