1. Introduction

Hyperspectral imaging (HSI) technique, as an important non-contact detection method, has been widely and successfully applied in various areas such as environmental monitoring [1], geological exploration [2], biology [3], food science [4] for images acquisition and spectra measurement of a target scene or a sample. In ultraviolet (UV) waveband, it has been mainly investigated for space science in far UV (120~200 nm) [5] and extreme UV (10~120 nm) [6]. However, near UV (300~400 nm) and middle UV (200~300 nm) are important spectral wavebands. Many organic substances and inorganic substances have prominent absorptions in these UV wavebands, for example, harmful gas SO₂ has strong absorption in 300~325 nm [7], and essential part of organisms protein in 280 nm [8]. For the applications of forensic science, the UV imaging and spectra measurement have been developed for latent images detection, traces of blood, explosives, fingerprints, and drugs detection [9–12]. For these UV applications, when the spectra measurement is combined with images acquisition through a imaging spectrometer, the different optical characteristics provide a more comprehensive assessment for the measured object.

However, in UV spectra range, developing an imaging spectrometer is challenging due to a low level of incident power of photon flux, low UV throughput of optical system, relatively low quantum efficiency of imaging sensor in UV waveband, large chromatic aberration, and fewer suitable optical filters for the desired broad UV wavebands [13]. Previously, many researchers have investigated both grating-based dispersion configuration and interferometric configuration for the developments of UV imaging spectrometers. Johansson et al. designed a UV-near-infrared (250~975 nm) imaging spectrometer for microspectroscopy, which was featured with three (UV, visible and near-infrared) parallel optical recording channels. Their instrument consisted of an entrance slit and a complicated UV illumination system including a xenon lamp and several filters [13]. Similarly, Dubroca et al. developed a differential HSI system in UV waveband of 250~450 nm for the detection of explosives using a UV-blue (250~500 nm) light source [14]. They also used an entrance slit in their optical system. Since the radiation intensity of UV waveband from light sources, including the deuterium lamp, the xenon lamp, and the laser-driven light source, is much lower than the radiation intensity of visible and near-infrared waveband, such typical configuration equipped with an entrance slit aperture can't provide sufficient UV spectral intensity for detectors. Research of Fourier transform imaging spectrometers (FTIS) without entrance slit is required for UV HSI. Zucco et al. presented a hyperspectral imager based on a Fabry–Pérot interferometer working in the UVA range (315~400 nm) [15]. Lyu et al. designed a close-range UV imaging spectrometer based on a spatial-modulated Michelson interferometer for latent image analysis applications [11]. Both systems have good performances without entrance slits, but their interferometric stabilities were relatively low. Moreover, their UV energy utilizations and the wavenumbers performances still need to be improved. Kohen et al. reported a solid ultra-stable Sagnac interferometer for the fluorescence spectroscopy in 1999 [16]. After that, the Air Force Research Laboratory launched a small-satellite (MightySat II.1) for terrain classification in September 2000. Its primary payload was a spatially modulated Fourier transform hyperspectral imager, which was the first imaging spectrometer on satellite to use a solid block Sagnac interferometer as the heart of the instrument [17]. This solid Sagnac interferometer has improved stability of the interferometry in various environments. As of now, it has not been used for imaging spectrometer in UV waveband.

In this paper, we present a large-aperture UV (250~400 nm) imaging spectrometer, which focuses on close-range (on the order of centimeters) hyperspectral sensing with high spatial resolution and decent spectral resolution. To do so, we designed and implemented a modified solid interferometer working in waveband of 250~400 nm to optimize the optical system and improve the interferometric stability, and used air-spaced achromatic doublets for collimating and imaging lenses to address chromatic aberration. The optical system has a high level of incident power of photon flux, thus causes little UV energy loss. Combined with reflective-transmissive filtering system, the radiometric energy efficiency of the optical system is significantly improved in the UVA and middle UVB waveband. In section 2, the new design of the optical system is elaborated. In section 3, the data acquisition and spectral calibration are presented. The results of UV hyperspectral imaging are presented in Section 4. The analysis of experiments results demonstrate the usability of the designed UV imaging spectrometer. The conclusion is in Section 5.

2. Instrument design and implementation

2.1 Overview

The schematic diagram of the designed UV (250~400 nm) imaging spectrometer is shown in Fig. 1. The instrument consists of a collimating lens, a imaging lens, a modified solid Sagnac interferometer, a reflective-transmissive filtering system, a moving motorized stage, a scientific complementary metal-oxide semiconductor sCMOS imager (Indego S-UV-M-R-U, China), and a computer with the accompanying digital hardware and software. The quantum efficiency of the sCMOS imager is enhanced and is higher than 48% in 220~400 nm and greater than 80% at 280 nm. This sCOMS imager provides sufficient spectral response at the actual working waveband of the spectrometer. The incident object beam is firstly reflected and filtered by a custom-built ultra-broadband dielectric mirror with >97% reflectivity in 245~420 nm, and 80%~90% transmittance factor in visible and near infrared range. Then the object beam transmits through a custom-built transmissive filter (transmittance >80% in 250~400 nm and <0.8% in 400~730 nm) and is filtered again. Because the output spectra of the UV light sources include disturbing sharp spectral peaks, such as the xenon lamp in 800~1050 nm, and the deuterium lamp in 650~660 nm. The transmissive filter and the reflective filter guarantee the useless visible and near infrared wavebands can be effectively filtered out, while the intensity of the working UV waveband is reduced slightly. Next, the filtered object beam passes through the collimating lens and becomes parallel beam. In the UV waveband, the traditional transmissive lenses cause relatively large chromatic aberration, which severely reduces the spatial resolution. To address this issue, Lyu et al. adopted a all-reflective optics method, which was featured with a Cassegrain structure and a Offner relay lens [11]. But such structure has relatively low UV energy utilization in close-range application. Another choice of achromatic reflective optics is the off-axis parabolic mirrors, which can completely eliminate the chromatic aberration. But the off-axis parabolic mirrors require high-precision spatial registration and are very complicated for assembling and adjustment. Therefore, we designed two sets of air-spaced achromatic doublets as the collimating lens and the imaging lens. The simulation by ZEMAX shows that the MTF (modulation transfer function)of the imaging lens is larger than 0.35 at Nyquist frequency (45.5 lp/mm) of the detector. The RMS (root meant square) spots radii in the all fields are lower than 7 μm in the waveband of 250~400 nm. It proves that all imaging spots are completely enclosed in a square pixel of side of 11 μm of the sCMOS imager. The maximum field curvature is shorter than 0.45 mm and the maximum distortion is next to zero. The focal shift of the air-spaced achromatic doublets is less than 0.083 mm between 325 nm and 250 nm and less than 0.125 mm between 325 nm and 400 nm. It has good optical performance and is easy for assembling. Subsequently, the parallel object beam is spatially modified into two virtual sources with a lateral displacement by the Sagnac interferometer. The sCMOS imager is placed at the focal plane of the imaging lens, and the imaging lens focuses the two virtual sources onto the plane of the detector. The acquired image is superimposed with interference fringes. Different points of the objective scene are imaged onto different pixels with different optical path differences.

 

Fig. 1 The design and configuration of the designed UV (250~400 nm) imaging spectrometer. The subfigure shows the structure of the modified solid Sagnac interferometer.

Download Full Size | PDF

2.2 Design of interferometer

The performance of a Fourier transform spectrometer is mainly determined by the interferometer, which is a core component of the spectrometer. The zero-order position of the interferogram and the interference fringes are easily affected by the stability of the interferometer [18]. The interferometers reported before consisted of several detached optical components [11,15], which limit the stability for the interferometry. For improvement, we designed a modified solid Sagnac interferometer working in the waveband of 250~400 nm. The solid Sagnac interferometer without moving parts has good long term stability and is robust in harsh mechanical environment, e.g. stable against vibration. The design of the solid Sagnac structure is depicted in the insert in the bottom left of Fig. 1. The interferometer is assembled by cementing two same half-penta prisms with UV fused silica. The longest surface of one half-penta prism is divided into three areas B1, AR, and B2 for different coatings. A beam-splitting film (50:50 in UV 250~400 nm) is coated onto the area B1. The same film is coated onto area B2 to guarantee the parallelism between two cemented planes, when the two half-penta prisms are cemented together. The area AR is coated with anti-reflection film in UV 250~400 nm. This coating method can increase the throughput of optical system rather than the method of coating a beam-splitting film onto the whole surface. Another two surfaces M1 and M2 of the two half-penta prisms are coated with the UV enhanced aluminum coating to reflect the beams inward to the prisms. These two surfaces are also covered with a protective outer coating. For the spatially modulated interferometry, the interferogram is generated by the lateral displacement $d$ of the two virtual sources. For the designed Sagnac structure, when the two half-penta prisms are cemented together along the longest surfaces, a prism is slightly translated in one direction along the longest side. Thus, the two half-penta prisms have a small translation $t$ and generate a lateral displacement $d$ for the two virtual object sources.

The spectrometer is designed to acquire spectral samples over the range of 250~400 nm. The lateral displacement $d$ is determined from the maximum optical path difference (OPD) $OP{D}_{\max }$, half of the interferogram dimension $N$, the detector pixel spacing $s_{\text{px}}$, and the focal length of the imaging lens $f$. Base on the geometric relationship between lateral displacement $d$ and translation $t$ [19], the translation $t$ can be calculated by the equation:

There are few optical cement that can provide sufficient spectral transmittance covered both near and middle UV waveband. The two half-penta prisms in our system are cemented by a custom-build optical cement, which has a spectral transmittance greater than 91% in the 250~400 nm wavelength. The transmittance of the optical cement is lower than 20% in the waveband below 250 nm. Hence, the actual working waveband of our imaging spectrometer is from 250 to 400 nm. For interferometer assembly, we used a low pressure mercury lamp (a prominent 253.7 nm characteristic peak), the collimating lens, the imaging lens, and the sCMOS imager to build up an alignment system to generate interference fringes during the translation adjustment and the cementation. Monitoring the interferogram during the adjustment is also necessary to keep the interference fringes vertical.

2.3 Spatial resolution and field of view

The UV imaging spectrometer is designed for the detection of close-range targets, thus high spatial resolution is required to distinguish small details of an object. The sCMOS imager has an array of 1024 × 1024 pixels with 11μm pixel spacing. The distance between the collimating lens and the target is 400 mm and the focal length of imaging lens is 200 mm. The size of the target corresponding to the imaging are in one dimension is measured to be 24 mm. Thus, the field of view (FOV) of the imaging system is 59.98 mrad and the instantaneous field of view (IFOV) is 0.059 mrad based on calculation. The spatial resolution that the size of the object projected on one pixel is 23.44 μm. This spatial resolution is good enough for various applications to distinguish micron-level details.

3. Data acquisition and analysis

3.1 Push-broom scanning and hyperspectral data cube reconstruction

Hyperspectral imaging can be viewed as a measurement of a three dimensional (3D) data cube, which contains two spatial dimensions and one spectral dimension. In this study, the data cube acquisition is conducted through a push-broom scanning, which is performed in an orthogonal Fourier representation of the wavenumbers dimension. Furthermore, to acquire the whole interferogram of the target, each column of the object image needs to occupy each possible position over the OPDs that the translation t allows. In our designed imaging spectrometer, a moving motorized stage (MMS) is used to drive the spectrometer for the push-broom scanning. As shown in Fig. 2, the MMS carries the imaging instrument and shifts the object image on the sCMOS imager plane laterally for one pixel per snapshot. This process is repeated until the whole object area is covered totally. The moving motorized stage is controlled by a computer equipped with control and image acquisition interfaces, which coordinates the motor motion of the moving stage and the exposure of the sCMOS imager. Because the sCMOS imager has 1024 columns of pixel, 1023 extra exposures are needed for any column of the object image to achieve the maximum OPD for the maximum spectral resolution. For example, an object with size of 20 mm × 20 mm (approximately 854 columns × 854 rows in pixel on the detector plane) need to be examined. A total of 1877 times of exposure are needed for the raw data cube acquisition.

 

Fig. 2 The procedure of push-broom scanning and hyperspectral data cube reconstruction.

Download Full Size | PDF

The procedure of the data cube reconstruction is depicted in Fig. 2. The acquired raw image data cube contains both the interferogram information and the redundant portion. The interferogram of each object column is reorganized by extracting one pixel column from each image. Through array processing, a reorganized data cube that contains the interferogram information of all the columns of the object can be obtained. The spectrum of each pixel on the detector plane is calculated by spectral inversion algorithm based on a fast Fourier transform (FFT). Before this calculation, the interferogram of each pixel is processed by a self-adaptive differential filtering procedure to inhibit the useless zero-frequency component [21] and optimized by the apodization algorithm [22]. Then the interferogram of each pixel is calculated by the FFT to get the spectra-image data cube. The instrument control, the data preprocessing and the FFT calculation are efficiently operated by a developed multithreading program written by C# language.

3.2 Spectral calibration and the SNR

The structure of the interferometer causes a linear distribution of the phase difference in resulting interferogram, so do the wavenumbers of the spectral samples. To calibrate the spectral resolution and the actual wavenumber (or wavelength) of each spectral sampling point, we measured the spectrum of a low pressure mercury lamp for reference. The low pressure mercury lamp has prominent characteristic peaks at 253.7 nm and 365.0 nm [23]. The peaks were measured and verified by a UV fiber optics spectrometer (Ocean HR4000). In experiment, we used a reference sample (WS-1 Reflectance Standards, reflectivity >98% from 250 nm to 1500 nm, Ocean optics, USA) as the target and illuminated it by the low pressure mercury lamp. The beams of the lamp were expanded to the surface of the reference sample, and showed diffuse reflection, then imaged by the proposed spectrometer. Because the surface of reference sample is homogeneous, all the columns are equal in spectral reflectance and all OPDs can be recorded in one snapshot. The captured interferogram of the reference sample is shown in Fig. 3(a). The spectra were reconstructed through inversion algorithm based on the FFT. In the reconstructed spectrum, the 365.0 nm characteristic peak located near the 105th spectral sampling point and the 253.7 nm characteristic peak located near the 151th spectral sampling point. To achieve high spectral calibration accuracy, a method base on Gaussian distribution was applied to locate the subpixel positions of the two prominent peaks [24]. The position of 365.0 nm was relocated at the 105.82th spectral sampling point and the position of 365.0 nm was relocated at 151.36th spectral sampling point. The normalized spectrum from the same row of the interferogram is shown in Fig. 3(c). Two prominent characteristic peaks can be observed at 365.0 nm and 253.7 nm in the curve trace. Besides a relatively low peak can be discerned at 313.2 nm. This peak is illustrated in the manual of the low pressure mercury lamp. The reconstructed spectrum acquired by the our spectrometer perfectly matches the actual spectrum of the low pressure mercury lamp. Therefore, the wavenumber resolution of the spectrometer can be determined as 255.73 cm⁻¹. The wavelength resolution is 1.59 nm at 250 nm and 4.09 nm at 400 nm.

 

Fig. 3 Results of spectral calibration. (a) Interferogram of the reference sample with the illumination of a low pressure mercury lamp. (b) Spectrum of the calibration lamp, the curve was normalized to arbitrary units.

Download Full Size | PDF

The SNR is an important factor that reflects the detecting performance of the spectrometer. We obtained the mean signal and standard deviation of a homogeneous, illuminated surface, to estimate the SNR for each sampling wavelength [25,26]. The reference sample was used as the homogeneous surface and illuminated by a collimated UV light source equipped with a deuterium lamp. There are total of 110 spectral data cubes (x × y spectra for each data cube, x and y are column and row numbers of the sCMOS imager pixels) acquired by 110 times of push-broom scanning. Exposure time of each snapshot is 100 ms. Then the SNR of each sampling wavelengths (λ) at each pixel can be calculated from SNR_x,y,λ = m_x,y,λ/STD_x,y,λ, where the m_x,y,λ and STD_x,y,λ are the mean values and the standard deviation of the 110 sets of spectra at each pixel respectively. By averaging the SNR results of all pixels, the averaged SNR of each sampling wavelength can be obtained, which was higher than 40:1 in the waveband of 250~400 nm except at three narrow wavebands (322~330.3 nm, 365~368.5 nm, 375.8~379.5 nm). The SNR in these three wavebands is between 34.95:1 and 38.15:1, which is influenced by the fluctuations in the sCMOS imager quantum efficiency.

In summary, our UV imaging spectrometer has excellent spectra response in UV waveband 250~400 nm and can provide 59 precise wavelength samples with acceptable SNR. The imaging spectrometer has a high spatial resolution of 23.44 μm on the target plane and a wavelength resolution of 1.59 nm at 250 nm, and little chromatic aberration can be observed.

4. Experiments and results discussion

4.1 1951 USAF resolution test chart

In order to validate our imaging spectrometer, examples of UV hyperspectral imaging were performed with a representative sample of 1951 USAF resolution test chart, and a cotton sample with vitamin C (VC) and vitamin B6 (VB6) traces. The pattern of the 1951 USAF resolution test chart was printed on a blank paper. Its actual measured size was 13 mm × 20 mm. The sample was illuminated by a deuterium lamp light source, then recorded by push-broom scanning in 315 seconds. The reconstruction procedures produce spectrum for each point of the two dimensional target, which leads to the spatial dependence of the reflective spectra of the sample. As a result, monochromatic images of each sampling wavelength can be presented and the hyperspectral data cube can be reconstructed. The results are shown in Fig. 4. In Fig. 4(a), the monochromatic images in spectral dimension (λ axis) show the UV reflective intensity distribution of the target. The hyperspectral data cube clearly presents the two dimensional spatial information (X and Y axis) and one dimensional reflective spectral intensity(λ axis). Figure 4(b) shows examples of the reflective spectra distribution of five selected pixels from the picture. Each color spectral curve is corresponding to the same color point in Fig. 4(a). The spectra curves were normalized to arbitrary units. Figure 4(c) shows examples of monochromatic images at 250.4 nm, 271.8 nm, 285.8 nm, 299.3 nm, 330.3 nm, and 379.5 nm, respectively.

 

Fig. 4 UV hyperspectral imaging of a 1951 USAF resolution test chart. (a) The hyperspectral data cube. (b) The reflective spectra extracted from selected pixels of the picture. The spectra curves were normalized to arbitrary units. The color of curves refers to the colors of the points in (a). (c) Monochromatic images of the sample from the hyperspectral data cube.

Download Full Size | PDF

The red and magenta spectra curves are extracted from two pixels of the picture, which are corresponding to two tiny blank points in the printing paper. The similar distributions of red and magenta curves show that the reflective intensities of the two tiny points of the target are almost the same. The small difference of intensity might be due to the nonuniform of illumination. The spectral curves show that the spectrometer has an excellent spectral response, especially in the waveband of 250~320 nm. As far as we know, the reported imaging spectrometers didn't have such performance in this waveband [11,13]. The green curve is the spectral distribution of another blank point of the printed paper, and there is an invisible olive oil stain in the region of this point. It can be seen that the reflective spectra intensity in the waveband 250~290 nm is much lower than that in the waveband 290~400 nm, indicating that the oil stain has a unique absorptive spectrum in UV spectral range. In the two dimensional images, the oil stain is visible in the monochromatic images at 250.4 nm, 271.8 nm, and 285.9 nm. In contrast, it is indistinguishable in the monochromatic images at 299.3 nm, 330.3 nm, and 379.5 nm. So the oil stain can be inspected based on its spectral characteristics at various wavebands. The blue and cyan curves represent the reflective spectra of two black points printed with toner by a laser printer. The spectral distribution shows that black toner has low spectral reflectivity in UV waveband. The reflective intensities of these two toner points are quite different from the reflective intensities of blank points. Based on the characteristics of spectral reflectivity of different materials in the UV waveband range, our imaging spectrometer can identify or inspect the material information of the target object.

4.2 Trace detection and classification

In this experiment, example of UV hyperspectral imaging was performed with a white cotton sample with VC powder and VB6 powder traces. Before the UV hyperspectral analysis, the sample images was captured by a color CCD camera with a waveband of 400~700 nm and a sCMOS imager with 250~400 nm broad bandpass filter. Figure 5(a) shows the image captured by the color CCD camera with a D65 illumination and the Fig. 5(b) shows the image captured by the sCMOS imager with a deuterium lamp illumination. In the color image, the traces are slightly visible. In the UV image, the traces are obvious. But from both images, we can't distinguish the difference between VC powder and VB6 powder.

 

Fig. 5 Polychromatic measurement results of the vitamin traces sample. (a) Image captured by a color CCD camera with D65 illumination. (b) Image captured by a sCMOS imager with 250~400 nm broad bandpass filter and a deuterium lamp illumination.

Download Full Size | PDF

Then we acquired the UV hyperspectral data cube of the sample cotton surface. The actual measured area of the sample was 10 mm × 20 mm. The sample was illuminated by a deuterium lamp light source, and then the interferograms were recorded by push-broom scanning in 185 seconds. Figure 6(a) shows a view of the interferogram data cube. The reconstructed monochromatic images at 250.4 nm, 279.7 nm, 304.1 nm, 345.2 nm, 365.0 nm and 399.2 nm are presented in Figs. 6(b)-6(g), respectively. The spectral distribution of the VC, VB6 and the background are shown as the blue, red and green curves in Fig. 6(i). It shows that the reflective intensity of the VC and VB6 are different from that of the white cotton. In the UV waveband of 250~295 nm, the reflective intensity of the VC is lower than that of the background. In the UV waveband of 302~325 nm, the reflective intensity of VC and the background are almost the same. But in the UV waveband of 330~390 nm, the reflective intensity of VC is obviously higher than that of the background. The reflective intensity of the VB6 is lower than that of the background in the waveband of 250~325 nm and is significantly higher than that of the background in the waveband of 360~390 nm. In addition, the characteristics of the reflective spectra of VC and VB6 are quite different from each other. In the UV wavebands of 250~270 nm and 370~390 nm, the reflective intensity of VB6 is a little higher than that of VC. But in UV waveband 280~350 nm, the reflective intensity of VB6 is obviously lower than that of VC. In the spatial dimensions, monochromatic images emphasize reflective intensity variations of substances in the UV waveband of 250~400 nm. The spectral characteristics among the vitamin traces and the white cotton are quite different. Because spectra-images data cube contains the spectral distribution of each pixel, we used the k-Means pattern recognition algorithm to analyze the data. The classification algorithm generates three cluster centers, which are corresponding to the VC, VB6, and the white cotton. The clustering results of different substances are shown in Fig. 6 (h). The blue pixels represent the VC trace, and the red pixels represent the VB6 trace. In contrast to the RGB color image in Fig. 5(a), the two vitamin traces can be identified with the information of the UV hyperspectral images. It proves that the UV imaging spectrometer provides much more information for materials identification, which some time only have subtle difference in the spectral domain.

 

Fig. 6 UV hyperspectral imaging of the white cotton sample with vitamin traces. (a) A view of the interferogram data cube. (b) Monochromatic image at 250.4 nm. (c) Monochromatic image at 279.7 nm. (d) Monochromatic image at 304.1 nm. (e) Monochromatic image at 345.2 nm (f) Monochromatic image at 365.0 nm. (g) Monochromatic image at 399.2 nm. (h) Clustering result of the vitamin traces based on the hyperspectral data cube. (i) Normalized reflective spectra of the vitamin B6, the VC, and the white cotton.

Download Full Size | PDF

These two experiments show that the proposed imaging spectrometer is able to acquire hyperspectral images in UV waveband 250~400 nm. The experimental results verify the spatial and spectral accuracy of the proposed UV imaging spectrometer.

5. Conclusion

To conclude, we have introduced a new type of UV Fourier transform imaging spectrometer based on a modified solid interferometer, a reflective-transmissive filtering system, and air-spaced achromatic doublets. The advanced structure of solid Sagnac interferometer is first implemented in the design of UV (250~400) imaging spectrometer. The imaging spectrometer acquires 59 wavelength samples with excellent spectral response in UV waveband from 250 to 400 nm. Moreover, the maximum spatial resolution of the imaging spectrometer is 23.44 μm on the target plane, the wavelengths resolution is 1.59 nm at 250 nm, and little chromatic aberration can be observed. Two experiments of UV hyperspectral imaging are successfully carried out. The samples are a 1951 USAF resolution test chart, and a cotton sample with VC and VB6 traces. Through the analysis of spectra, monochromatic images and k-Means clustering results, the spectral and spatial accuracy of the spectrometer are satisfactory. The experimental results demonstrate the effectiveness and the reliability of our spectrometer. This imaging spectrometer might be a potential tool for the identification of materials with UV spectral characteristics. Further work will focus on promoting the spectral resolution and reducing the push-broom scanning time. In addition, by replacing the collimating lens in the imaging spectrometer system, it is expected to be used in long-range UV hyperspectral imaging as well.