1.1.

Hardware

We have developed over the last few years a device based on differential reflectometry (DR) which is capable of detecting explosive materials on surfaces such as luggage, parcels, shoes, garments, etc. from a distance.¹ The DR (Ref. 1) measures the normalized difference between the reflectivities of two adjacent parts of the same specimen or two slightly different samples. Ultraviolet (UV) and blue light are shone on a surface, for example, on a piece of luggage situated on a moving conveyer belt. Upon reflection, the light is collected then diffracted by a spectrometer and its intensity recorded with a charge coupled device (CCD) camera. A computer processes the resulting data and produces in turn a differential reflection spectrum. This differential technique is highly sensitive and provides spectroscopic data of materials, specifically of explosives. As an example, 2,4,6-trinitrotoluene (TNT) displays strong and distinct features in differential reflectograms near 420 nm, that is, at the edge of the UV and visible region of the light spectrum.² Similarly, but distinctly different features are observed for other explosives, whereas nonexplosives display essentially no structure in the pertinent wavelength region (280 to 450 nm).³ We have developed an improved hardware specifically designed for the fast detection of explosives on surfaces such as luggage for airport security or parcels for air shipping.

DR (also known as differential reflection spectroscopy) is a surface analytical technique that uses visible and UV light as a probing medium to reveal details about the electronic structure. In other words, the instrument detects the energies that electrons absorb from photons as they are raised into higher, allowed quantum (energy) states. Since each material has a specific electronic structure, the measurement of the characteristic energies for “electron transitions” serves as a fingerprint for identifying these substances. The differential reflectivity, which is proportional to the complex dielectric constant¹ as a function of wavelength, is unique to each material, making DR a viable investigative tool in materials science and particularly to explosives.

A number of different designs for a DR device are conceivable. For example, the object to be investigated is moved (on a conveyor belt for airport baggage scanning) and two reflectivities $R_1$ and $R_2$ are measured milliseconds apart (see Fig. 1). A broadband light source (Energetiq EQ99), 200 to 500 nm, is focused on the exterior of the luggage to be scanned due to a quartz collecting lens (50.8 mm diameter and 60 mm focal length) combined with a flat aluminum mirror (set at 90 deg) and a quartz focusing cylindrical lens (30 mm length, 20 mm width, and 25 mm focal length). The projected line, which is perpendicular to the direction of the scan as shown in Fig. 1, is 600 mm long by 3 mm wide on the conveyor belt. The focusing optics is located at 450 mm above the conveyor belt. Since the beam is diverging from the lens downward, at midheight (where the top of the luggage is located) the beam is about $300\times 3{\mathrm{mm}}^2$.

Fig. 1

Schematic representation of the differential reflectometry scanner (a). Picture of an actual device scanning a bag (b). (Adapted with changes from Ref. 4.)

A custom-designed spectrograph (f-number of 3) which includes a custom design collection optics (a three-mirror fold system matching the f-number of the spectrograph with a primary mirror of about $10\times 40\mathrm{mm}$) made by Horiba Jobin-Yvon coupled with an UV CCD camera (Sarnoff CAM512UV, $512\times 512\text{pixels}$, $10\times 10\mathrm{mm}$ sensor size, 10 to $400\text{frames}/\mathrm{s}$, quantum efficiency $>50\%$ over the range of interest, i.e., 250 to 450 nm) collects the reflected photons. Since there are 512 pixels along the scanning line, at midheight (i.e., on the top of the luggage) each scanning pixel is $0.58\times 3\mathrm{mm}$. Our custom analysis software analyzes and classifies the signal which outputs an acoustic and visual indicator if traces of explosive are detected.

1.2.

Spectral Data

Figure 2(a) depicts characteristic spectral features for C4 (91% RDX [1,3,5-trinitroperhydro-1,3,5-triazine] and 9% plasticizer, binder, and oil), TNT, penta-erythritol tetranitrate (PETN), and ANFO (94% ammonium nitrate and 6% diesel fuel). It can be observed that each type of explosive has a different critical energy range for optical absorption, that is, PETN, ANFO, C4, and TNT have absorption edges at 290, 335, 340, and 420 nm, respectively. These absorption edges have different center wavelengths and different full widths from one another, therefore, the differential reflectograms allow us to distinguish each explosive (see below differentiation algorithm). The origin of the absorption edges stem from electronic transitions which are observed due to hydrogen bonds between explosive molecules in condensed state.⁵ Without hydrogen bonds, the energy levels are different and the absorption edges listed above are not observed. Figure 2(b) displays DR spectra of white powder materials (flour, Splenda, and salt) which are similar in physical appearance to explosives but have substantial different spectral features which allows one to distinguish them from one another. Additionally, Fig. 2(b) displays the DR spectra of a red jacket and denim jeans as examples of background materials on which explosives residue can be found. The explosives spectral fingerprints are also clearly different than these background clothes.

Fig. 2

(a) Differential reflectograms of some common explosive materials (TNT, C4, ANFO, and PETN). (b) Differential reflectograms of various nonexplosive materials (white powders: flour, Splenda, salt, and typical background materials, such as a red jacket and denim jeans).

The quality of the spectral fingerprints (shown in Fig. 2) is hardware dependent. That is, the signal-to noise-ratio depends on hardware parameters. Most of these parameters are set by the manufacturer. For example, the read-out noise from the detection camera or the light source power, fixed at 20 W, imposes a fixed number of photons interacting with the sampled surface and a given noise floor for any sample. The most important adjustable parameters are the acquisition time and the scanning speed. Figure 3 displays DR spectra of TNT measured with acquisition times ranging from 5 to 100 ms (corresponding to 200 and $10\text{frames}/\mathrm{s}$, respectively). It is observed that the relative noise level increases as the acquisition time decreases. Furthermore, when plotting the signal-to-noise ratio as a function of the acquisition time a linear relationship is observed with a correlation coefficient of 0.82, see Fig. 4. Note that the signal is taken to be equal to the absorption edge amplitude.

Fig. 3

Unfiltered differential reflectograms of TNT at various acquisition times (equal to one over frame rates). The curves have been shifted upward for clarity.

Fig. 4.

Signal-to-noise ratio for a given TNT sample as a function of the acquisition time. A linear trend is observed with a correlation coefficient of 0.82. No filter was applied.

The other important adjustable parameter is the conveyor scanning speed. Fig. 5(a) shows DR spectra of TNT for various scanning speeds ranging from 10 to $30\mathrm{cm}/\mathrm{s}$. This range is limited by the electro-mechanical design of the conveyor. Figure 5(b) displays the signal intensity of the TNT absorption edge as a function of the scanning speed. A linear trend is observed with a correlation coefficient of 0.81. This result is counter intuitive. We would expect that for a faster scanning speed the signal strength would decrease since the amount of time an explosive threat is exposed to the light beam drives the device’s response. The shorter the amount of exposure time, the smaller the signal should be. But the opposite is observed. This is explained by the fact that the scanning range which we use is narrow. In this range, the signal strength is correlated with the camera frame rate. If we could expand the scanning range we would expect a result closer to our intuition. For high speeds, explosive residues on the suitcase being scanned would have a short interaction time with the light beam and therefore would generate a small signal intensity. On the other hand, for a low speed, the explosive residues would be measured in two consecutive frames and therefore when we calculate the differential reflectivity, a smaller signal would be observed. In other words, the signal strength is dependent on the difference in reflectivities between the two parts being measured. For similar materials, the difference is small whereas for dissimilar materials, the difference is large.

Fig. 5

Differential reflectograms of TNT for various conveyor scanning speeds (a). Signal-to-noise ratio of a given TNT sample as a function of the conveyor speed (b). The acquisition time was 10 ms for all measurements and no filter was applied.

In summary, the acquisition time and scanning speed have both a strong influence on the signal-to-noise ratio. The signal-to-noise ratio has an impact on the limit of detection. False positive rates and true positive rates of detection will be discussed in a later section.

4.1.

Limit of Detection

The absolute limit of detection of the present DR system was determined to be about 100 ng for TNT.¹² Briefly, the limit of detection is defined as the mass of explosive material necessary to obtain a signal-to-noise ratio of three. In a past experiment,¹² we acquired the DR spectra of TNT samples of different sizes ranging from filling a full square pixel, $0.58\times 3\mathrm{mm}$ ($1.74{\mathrm{mm}}^2$) by 4-μm thick corresponding to about 11 μg, down to filling about 1% of a pixel, a triangular shape $0.58\times 0.05\mathrm{mm}$ ($0.0145{\mathrm{mm}}^2$) by 4-μm thick⁴ which is about 100 ng. In this experiment, the samples were prepared by packing a fine powder on a flat substrate. The thickness of 4 μm was determined to be an upper limit of the minimal interaction thickness between the TNT and the probing beam. Below this thickness, it becomes very difficult to control the flatness and packing of the fine TNT powder. Therefore, it is difficult to determine the minimal interaction thickness of TNT in the UV region of the spectrum. After acquiring the spectra, we extracted the signal-to-noise ratio for each sample size. We found a linear relationship between the sample size and the signal-to-noise ratio.⁴ From this linear relationship, we evaluated the limit of detection as per its definition. It should be said that this definition may not always lead to a usable limit of detection. We have observed that, for the algorithm presented in the previous section, the practical limit of detection is about two times larger at about 200 ng for TNT (i.e., signal-to-noise ratio of about 5) than the absolute limit. Similar limit of detection values has been found for other explosives. It should be added that the sample shape has a very strong influence on the detection limit.¹⁴ First, the geometry of the sample matters. Only the top 4 μm of explosive material will be sensed by the scanning beam. Therefore, detecting thick samples will increase the detection limit. Second, the roughness was found to have a strong influence on the reflectivity of the sample¹⁴ and as a consequence has an effect on the limit of detection. The DR is most sensitive for strongly diffusive and weakly specular samples, that is, for rough rather than smooth sample surfaces. Real threats, such as explosive residues in the form of particles or finger prints, are thin with diffusive surfaces and therefore the method presented above is of practical significance.

Future development is being considered to improve the limit of detection of DR in order to reach trace level detection. This means being able to sense a few tens of micrometer size diameter particles¹⁵ weighing between one and a few tens of nanograms. In order to attain such sensitivity, one could increase the intensity from the light source, for example, using more efficient optics (i.e., custom designed) or a more powerful source, or increase the size of the collection optics, thus increasing the photon flux going to the camera. Our current system is only using about 10% of the dynamic range of the camera. Using 100% of the dynamic range would allow our detection system to reach a 10-ng detection limit.

Other standoff optical techniques have been applied to explosive detection by hyperspectral imaging. In particular, Raman [deep UV¹⁶and coherent anti-Stokes Raman spectroscopy¹⁷] and short-wave infrared¹⁸ have been developed to detect nanogram range threats, however, these techniques suffer from slow scanning speed and they cannot at the present time handle large surface screening at these detection levels. On the other hand, DR is about two orders of magnitude faster with the ability to screen the top surface of carry-on size luggage in about 3 s.

In addition to the limit of detection, the performance of a detection system is evaluated by its capacity to differentiate threats and nonthreats. Such performance is evaluated using receiver operating characteristic (ROC) curves which display the true positive rate as a function of the false positive rate for any detection system.

4.2.

ROC Curves

The ROC curves for each explosive are displayed in Fig. 8. These results were generated for each classifier (i.e., each explosive individually). The classifiers were trained for each explosive as described in Sec. 2. For each explosive, two different data sets were used to train the classifiers and therefore generated two ROC curves. In the first case, the data set contains explosives spectra of sample sizes ranging from about 0.03 to 100 μg. That is from below the limit of detection to above the maximum dynamic range of our system. In the second case, the data set contains only explosives spectra with SNR above 5 (sample sizes ranging from 0.2 to 100 μg), which is above twice the limit of detection. Removing samples near the limit of detection have a small effect on the system performance. For example, in the case of C4, at 5% false positive rate, the true positive rate increases only by 0.05 (from 0.92 to 0.97). This demonstrates that the robustness of the algorithm developed is in the 0.2 to 100 μg range.

Fig. 8

Receiver operating characteristic curves of three materials: PETN, C4, and AN. The curves labeled “All samples” represent the classifier trained with every explosives sample in the set, sample sizes range from 0.03 and 100 μg (0.3% and above fill factor). The curves labeled “Large samples only” represent the classifier trained without the samples close to or below the limit of detection. Large sample sizes range between 0.2 and 100 μg (2% and above fill factor).

Figure 8 displays a good performance of the detection system as long as the sample size is above about 0.2 μg ($\mathrm{SNR}=5$). Below this value, the classification algorithm loses its ability to differentiate explosives from nonexplosives materials due to noise confusion. That is, the algorithm detects noise where the spectral features of explosives are normally located which causes a dramatic increase in the false positive rate.

The data set of each explosive varied in regards to the number of explosive and nonexplosive samples used in training. PETN was trained with 30,000 non-PETN samples and 825 PETN samples. C4 was trained with 14,000 non-C4 samples and 520 C4 samples. AN was trained using 20,000 non-AN samples and 1200 AN samples. Approximately 10% of each explosives sample sets was comprised of samples near the limit of detection and was removed to demonstrate the difference shown in Fig. 8. The nonexplosives samples were generally comprised of clothing materials of different compositions and colors.⁹