1. Differential Optical Absorption Spectrometer for Monitoring Urban Atmospheric Pollutants and Their Distribution Christopher P. Beekman and Heather C. Allen, Environmental Science Graduate Program Chemistry Department The Ohio State University

2. Motivation Differential Optical Absorption Spectroscopy –DOAS Development of smaller size instrument Applicability of small size spectrometers Improve mathematical techniques Confidence in retrieval techniques Trace Gas Light SourceDetector kilometers

3. Differential Optical Absorption Spectroscopy DOAS Step 1: Remove Broadband Features

4. Differential Optical Absorption Spectroscopy DOAS Step 2: Fit Laboratory References Also fit SO2, CH 2 O, Ring Absorber

5. Miniature DOAS (Test Platform) BG-7 Red Absorbing Filter Quartz Window Plano-Convex Lens Adjustable X-Y-Z Fiber Mount Ocean Optics USB 2000 310-520 nm Advantages: Low Cost Stable Optics USB Connection Portability Disadvantages: Stray Light Noise Thermal Instability Limited Channels B. Galle, C. Oppenheimer,A. Geyer,A.J.S. McGonigle, M. Edmonds, L. Horrocks Journal of Volcanology and Geothermal Research, 119(1-4), 2003. V. Tsanev, Personal Communication

6. Spectral Processing Pre-Processing –Spectral Alignment –Noise Removal Fit Analysis and Optimization –Factor Selection –Wavelength Region Selection Post-Processing –Bootstrap Evaluation –Monte-Carlo Evaluation

7. DOAS Fitting

8. Alignment Procedure Random Shift References Calculate Fit Statistics OK? Y/N Accept Fit Y Eliminate Combination From Search Set Start J. Stutz, U. Platt, Applied Optics, 35(30), 1996. N

9. DOAS Fit After Alignment

10. Noise Removal Difference between Raw and Optimal Smoothed Spectra Calculated Difference is analyzed for buried signals T. Iwata, J. Koshubu, Applied Spectroscopy, 50(6), 1996. Y. Termonia, J. Deltour, J. Phys. D: Appl. Phys, 7, 1974.

11. Noise Removal A B Accepted Rejected As Noise Singular Value Analysis of Raw Data – Smoothed Data = noise Can detect buried signal otherwise lost J.L. Castellanos, S. Gomez, V. Guerra, Applied Numerical Mathematics, 43, 2002. C

12. Noise Removal 5 – 8% Decrease in Residuals and Error Over Savitzky-Golay 1 – 2% Decrease in Residuals 2 – 4% Decrease in Fit Error Minimal Impact on Retrieval with Gas Cell Data

13. Wavelength Region Selection Through Partial Least Squares Wavelength Selection Based On Knowledge of Absorbing Features A Moving, Unfixed Size PLS Window, Larger Region Initially Selected

14. Ordinary Least Squares vs. Partial Least Squares with Wavelength Selection

15. Conclusions Modified smoothing improves retrieval error with minimal data loss Partial least squares allows advanced data selection and analysis Miniature spectrometers show promise, still limited by resolution Further mathematical techniques needed

16. Acknowledgements Heather C. Allen and the Allen Group NSF-ATM Tao Shi of the Statistics Department, The Ohio State University