Sensitive Algorithm for Multiple-Excitation-Wavelength Resonance Raman Spectroscopy

Transcription

1 Sensitive Algorithm for Multiple-Excitation-Wavelength Resonance Raman Spectroscopy Balakishore Yellampalle *, Hai-Shan Wu, William McCormick, Mikhail Sluch, Robert Martin, Robert Ice and Brian E. Lemoff WVHTC Foundation, 1000 Technology Drive, Suite 1000, Fairmont, WV, USA ABSTRACT Raman spectroscopy is a widely used spectroscopic technique with a number of applications. During the past few years, we explored the use of simultaneous multiple-excitation-wavelengths (MEW) in resonance Raman spectroscopy. This approach takes advantage of Raman band intensity variations across the Resonance Raman spectra obtained from two or more excitation wavelengths. Amplitude variations occur between corresponding Raman bands in Resonance Raman spectra due to complex interplay of resonant enhancement, self-absorption and laser penetration depth. We have developed a very sensitive algorithm to estimate concentration of an analyte from spectra obtained using the MEW technique. The algorithm uses correlations and least-square minimization approach to calculate an estimate for the concentration. For two or more excitation wavelengths, measured spectra were stacked in a two dimensional matrix. In a simple realization of the algorithm, we approximated peaks in the ideal library spectra as triangles. In this work, we present the performance of the algorithm with measurements obtained from a dual-excitation-wavelength Resonance Raman sensor. The novel sensor, developed at WVHTCF, detects explosives from a standoff distance. The algorithm was able to detect explosives with very high sensitivity even at signal-to-noise ratios as low as ~1.6. Receiver operating characteristics calculated using the algorithm showed a clear benefit in using the dual-excitation-wavelength technique over single-excitation-wavelength techniques. Variants of the algorithm that add more weight to amplitude variation information showed improved specificity to closely resembling spectra. Keywords: Raman Albedo, Deep Ultraviolet, Resonance Raman Spectroscopy, Multiple Excitation Wavelengths, Explosive Detection. 1. INTRODUCTION Resonance Raman spectroscopy 1 refers to Raman spectroscopy performed with an excitation laser source whose frequency is close-to or overlapping with the absorption resonance of molecules being studied. Under resonance excitation conditions, Raman cross-sections of molecules can be enhanced by orders of magnitude 2,3. In multipleexcitation-wavelength (MEW) resonance Raman spectroscopy, two or more lasers, whose frequencies are sufficiently far apart so as not to interfere with Raman spectra of interest, are used as excitation sources. Since Raman intensities are dependent on the excitation wavelength, intensity variation between corresponding Raman bands at different excitation wavelengths can be used as an orthogonal signature 4 in addition to the traditional use of the Raman spectrum itself. For the past few years we have been studying the use of MEW techniques for standoff detection of explosives 5. MEW techniques improve specificity 4, which is especially important in detecting small quantities of explosives among unknown background materials. WVHTC Foundation has built a resonance Raman explosive sensor based on two excitation wavelengths 6. In this article we present a very sensitive algorithm that is used to process the MEW resonance Raman signatures obtained from this dual-excitation-wavelength (DEW) sensor. 1.1 Multiple excitation wavelength resonance Raman spectroscopy An ideal implementation of the (MEW) resonance Raman detection technique consists of a multiple-wavelength DUV source that simultaneously generates collinear excitation beams directed onto a sample, and a multiple-band detection * byellampalle@wvhtf.org; phone ;

2 spectrometer that measures the multiple Raman spectra simultaneously. The simultaneous detection of spectra from a single interrogation area ensures that any spatial inhomogeneities of explosives in the sample, or any time-dependent phenomena, such as motion or photo-degradation present in a real detection scenario, contribute equally to the detected signal level at each excitation wavelength. Figure 1 shows the schematic of an ideal implementation of the MEW concept. Figure 1: Concept of multiple-excitation-wavelength deep UV resonance Raman detection scheme. The Raman scattering intensity from an explosive depends on its Raman cross-section, laser penetration depth and the resonance Raman self-absorption. Resonance Raman cross-sections and the absorption of both the excitation and scattered light varies strongly with excitation wavelength. This results in Raman band intensities that depend on excitation wavelength in a complex way, forming unique signatures for different compounds. In our earlier work, we measured these signatures for a variety of explosives and demonstrated the utility of multi-excitation-wavelength resonance Raman measurements for explosive detection. 2. ALGORITHM Figure 2: (a) Flowchart of the algorithm. (b) Example of arranging two wavelength measurements into a 2D matrix. (c) Calculation of ratio factors. and refer to ratios for the sample and library spectra.

3 A number of chemometric techniques exist for Raman spectral analysis Examples include principal component regression, principal component least squares, hybrid least squares, support vector machines 11, genetic algorithms 12 etc. Regression techniques work well when the SNR of the measurements is high. This is because Raman spectra of most explosives result in sparse data, which cause regression techniques to be ill-posed to matrix inversion, requiring the use of regularization techniques 7. If a technique explicitly tries to locate the peak position, like for example our previous algorithm for MEW Raman 4 or the algorithms in References 13 and 14, high SNR measurements are even more important. In a typical detection scenario, spectra of background surfaces on which an explosive might be present is not known. There may be additional unknown interferent materials with broad spectral features or partially overlapping peaks. Therefore, a Raman detection algorithm needs to be able to detect known materials among unknown substances. The algorithm developed as part of this work has the following features: a) low sensitivity to slowly varying backgrounds (either from interferents or auto-florescence) b) good performance at very low SNR, and c) interferent or background materials need not be characterized. The Raman spectrum of a typical explosive contains narrow peaks corresponding to vibrational modes of molecular species. Often, broad features in the Raman spectra of explosives are non-specific and may arise from unwanted interferents or degraded products. In our analysis target library, signatures are treated as consisting of a limited number of sparsely located peaks with good signal-to-noise ratio. Smoothly varying background was not considered as signal and removed during preprocessing. Several approaches could be used to create ideal library spectra from the characterized spectra. One approach includes using triangles as approximating the narrow Raman peaks. The ideal spectra lower sensitivity to noise in the measurements. Figure 2 shows a flow chart of a typical implementation of the algorithm. Initially the measured DEW spectrum is preprocessed to remove background. The background is assumed to be smoothly varying. Following background removal, any cosmic rays or fixed pixel noise with variations smaller than the design optical resolution of the spectrometer are removed. A combination of median filtering and intensity threshold are used for cosmic ray removal. The resulting spectrum is cast as a 2D matrix with each row corresponding to the Raman spectrum of one wavelength excitation. If the SNR of the measurement is high, the second spectrum is interpolated into the same Raman frequency grid as the first Raman spectrum. If SNR of the measurement is low, interpolation is not performed but a same vector for the second row of the matrix is used. An identically prepared library spectrum is used to compare and calculate the scores. A key step in the algorithm is calculation of scores by correlational least squares. In this approach scores are calculated from two sequences; 1) cross-correlation of measured spectrum with pre-recorded target library spectrum and 2) auto-correlation of target library spectrum. Scores are calculated as sum of sequence-one multiplied by the corresponding element of sequence-two divided by sum of the second sequence. Further these scores are multiplied by a ratio factor, which includes the orthogonal information of intensity variation, to improve specificity. The method used in calculation of ratio factors is shown in Figure 2(c). Higher weightage is given to ratios, with the help of an exponential factor, when the SNR is high. For low SNR signals the ratio factor converges to one as SNR approaches one. The exponential factor was empirically determined using hundreds of measurements with different SNR levels. 3. RESULTS We show a typical measurement from the dual-wavelength excitation sensor in Figure 3. Raman spectra from nm wavelength excitation and nm wavelength excitation are marked in Figure 3(a). These spectra are arranged in a two dimensional matrix and processed by the algorithm. 3.1 Measurements at low signal-to-noise ratio Figure 3 also illustrates the sensitivity of the algorithm to a low signal-to-noise situation. Spectra (a) and (b) in the figure correspond to spectra obtained from bulk Ammonium Perchlorate packed on an aluminum foil. The noise spectra in (c) and (d) were obtained using a black aluminum foil target that produced no Raman signatures.

4 (a) (c) score = 1.98 SNR = nm Raman spectrum nm Raman spectrum score = 0.2 (b) (d) score = 1.6 SNR = 1.6 score = 0.8 Figure 3: Illustration of algorithm sensitivity. (a), (b) Ammonium Perchlorate measured at exposure levels of 1 sec and 0.1 sec respectively. The algorithm produced a high score of 1.98 and 1.6 respectively. (c), (d) Background noise from a black Aluminium substrate obtained with the same exposure times (1s and 0.1s). The algorithm produced a low score of 0.2 and 0.8 respectively. The plots also indicate scores obtained using the 2D algorithm for all four situations. Even when the SNR was small, approaching one, as in (b), the algorithm produced a high score of 1.6, which is well separated from the pure noise score of Receiver operating characteristics at low SNR In order to understand the detection sensitivity of the algorithm we constructed receiver-operating characteristics of the sensor under a low signal-to-noise ratio situation, similar to the one shown in Figure 3(b) and (d). We performed 4000 repeated measurements creating two sets of spectral populations, one consisting of Ammonium Perchlorate and second consisting of only black aluminum foil. Detection algorithm analyzed each spectrum in the populations and produced a score for each measurement. For the analysis, scores were calculated without the threshold step of the algorithm (last step of Figure 2(a)). By plotting histograms of these scores (see Figure 4) for the two sets of measurements, we analyzed

5 how well the two populations separated. We calculated true-positive-rate verses false-positive-rate using the histograms by varying the threshold score for detection. A plot of true-positive-rate verses false-positive-rate is shown in Figure 5. (a) (b) (c) Exposure time: 0.1 s Distance between sample and collection mirror: 1 meter Average laser power at nm: 51µw Average laser power at nm: 20µw Samples: Ammonium Perchlorate or black Al foil Figure 4: 2000 spectra similar to Figure 3(b) and 2000 spectra similar to Figure 3(d) were measured and analyzed. Histograms of scores from algorithm using (a) only nm information, (b) only nm information, and (c) both nm and nm information. The results in Figure 4 and Figure 5 clearly show the sensitivity of the algorithm as well as the advantage of the MEW technique over the single excitation wavelength techniques. The histogram from the nm score (Figure 4(a)) has two broad peaks with significant overlap between the positive (green, indicating presence of explosive) and negative (red, indicating absence of explosive) populations. Similarly, the overlap between the two populations is large when the nm score (Figure 4(b)) alone was used. However when spectra from both the excitation wavelengths are combined, the 2D algorithm produced scores where the two populations have less overlap indicating lower falsepositive-rates for the same true-positive-rates. The ROC curve plotted in Figure 5, using the information in the histograms (Figure 4) clearly shows that the algorithm is capable of separating the low SNR (1.6) Ammonium perchlorate signatures from noise with a low false-positive-rate of 0.1% at a true-positive-rate of 78%. The curves also show a significant improvement with the dual-excitationwavelength technique (red line) over the single excitation wavelength technique (green and blue lines). For example at a false-positive-rate of , true-positive-rate from the techniques are three times better compared to nm and >8 times compared to nm alone.

6 Figure 5 ROC curves derived from the three score histograms. The experimental conditions for the plot are the same as in Figure Other explosive detection We have also used the DEW algorithm developed here to detect several explosives [15]. A library was created with fourteen different explosives or precursor materials. The algorithm was tested with materials in the library and with typical background surfaces or representative interferent materials that were not part of the library. All the explosives and precursor materials were correctly identified. Nearly all tested background and interferent materials produced scores below the threshold. 4. SUMMARY We have developed a new algorithm for detection or concentration estimation of materials using multiple-excitationwavelength resonance Raman spectroscopy. The algorithm calculates scores for each measurement based on correlational least squares approach. We applied the algorithm to stand-off explosive detection using a sensor in our laboratory. The sensor was based on dual-wavelength-excitation Raman spectroscopy. Measurements indicate that the algorithm is capable of detection even under very low (approaching one) SNR situations. Receiver operating characteristics show the dual-excitation-wavelength technique to be superior to any one single excitation wavelength technique. 5. ACKNOWLEDGEMENT This work was funded by the Office of Naval Research under contract number N C-0069.

7 REFERENCES [1] Asher, S. A., UV RR, in Handbook of Vibrational Spectroscopy, J.M. Chalmers and P.R. Griffiths (Eds), John Wiley & Sons, Ltd., Vol (2001). [2] Asher, S. UV Raman Spectroscopy for Analytical, Physical, and Biophysical Chemistry, Part 1, Analytical Chemistry, 65, (1993). [3] Asher, S. In Ultraviolet raman spectrometry, Chalmers, J., and Griffiths, P., Eds., John Wiley & Sons, Vol. 1, pp (2002) [4] Yellampalle, B., Sluch, M., Asher, S., Lemoff B., Multiple-excitation-wavelength resonance-raman explosives detection, Proc. of SPIE Vol. 8018, (2011). [5] Yellampalle, B., and Lemoff, B. E., "Raman albedo and deep-uv resonance Raman signatures of explosives, " Proc. SPIE 8734, 87340G (2013). [6] Yellampalle, B., Sluch, M., Wu, H. S., Martin, R., McCormick, W., Ice, R., and Lemoff, B., " Dual-excitation wavelength resonance Raman explosives detector," Proc. SPIE 8710, 87100Z (2013). [7] Frank, LLdiko E., and Jerome H. Friedman. "A statistical view of some chemometrics regression tools," Technometrics 35.2, , (1993). [8] Sompel, D. V., Garai, E., Zavaleta, C., Gambhir, S. S., A Hybrid Least Squares and Principal Component Analysis Algorithm for Raman Spectroscopy, PLoS ONE. 7, e38850, (2012). [9] Ferraro, J. R., Nakamoto, K., and Brown, C., Introduction to Raman Spectroscopy, Second edition, Academic Press. [10] Zavaleta C, Smith B, Walton I, Doering W, Davis G, et al. Multiplexed imaging of surface enhanced Raman scattering nanotags in living mice using noninvasive Raman spectroscopy, Proceedings of the National Academy of Sciences 106, (2009) [11] Widjaja, Effendi, Wei Zheng, and Zhiwei Huang. "Classification of colonic tissues using near-infrared Raman spectroscopy and support vector machines," International journal of oncology 32(3), (2008). [12] Lavine, Barry K., et al. "Raman spectroscopy and genetic algorithms for the classification of wood types," Applied Spectroscopy 55(8), (2001). [13] Kwiatkowski, A.., Gnyba, M., Smulko, J., and Wierzba, P., Algorithms of chemicals detection using Raman spectra, Metrol. Meas. Syst. Vol. XVII, (2010). [14] Kode, Kranthi, et al. "Raman labeled nanoparticles: characterization of variability and improved method for unmixing," Journal of Raman Spectroscopy 43(7), (2012). [15] Yellampalle, B., McCormick, W., Wu, H. S., Sluch, M., Martin, R., Ice, R., and Lemoff, B., High Sensitivity Explosives Detection using Dual-Excitation-Wavelength Resonance-Raman Detector," Proc. SPIE 9073, (2014).