Hyperspectral Imaging Infrared Sensor Used for Enviromental Monitoring M

Vol. 124 (2013) ACTA PHYSICA POLONICA A No. 3

Optical and Acoustical Methods in Science and Technology 2013 Hyperspectral Imaging Infrared Sensor Used for Enviromental Monitoring M. Kasteka,∗, T. Piatkowskia, M. Zyczkowskia, M. Chamberlandb, P. Lagueuxb and V. Farleyb aInstitute of Optoelectronics, Military University of Technology, S. Kaliskiego 2, 00-908 Warsaw, Poland bTelops Inc., Quebec, Canada This paper presents detection, identication and quantication of gases using an infrared imaging Fourier- -transform spectrometer. The company Telops has developed an imaging Fourier-transform spectrometer instrument, Hyper-Cam sensor, which is oered as short or long wave infrared sensor. The principle of operation of the spectrometer and the methodology for gases detection, identication and quantication has been shown in the paper, as well as theoretical evaluation of gases detection possibility. The variation of a signal reaching the imaging Fourier-transform spectrometer caused by the presence of a gas has been calculated and compared with the reference signal obtained without the presence of a gas in the imaging Fourier-transform spectrometer eld of view. Some result of the detection of various types of gases has been also included in the paper. DOI: 10.12693/APhysPolA.124.463 PACS: 07.57.Ty, 78.20.Ci, 78.40.−q, 78.47.jg

1. Introduction

A problem of remote detection of chemical substances appears in many, sometimes extremely dierent elds of human activities. Applications of detection devices include monitoring of technological processes, diagnostics of industrial installations, monitoring of natural environ- ment, and military purposes. Such diversity has caused development of many detection methods employing vari- Fig. 1. The infrared IFTS Hyper-Cam. ous physical phenomena due to which detection and identication of chemical compounds are possible. Among these methods, a signicant group constitutes the solutions using a phenomenon of selective absorption of elec- tral characteristics and resolution matched to the absorp- tromagnetic radiation by chemical compounds. Here, we tion bands of compounds to be detected. Two types can distinguish the solutions based on absorption of laser of such devices can be distinguished. One, it is a sys- radiation emitted by an illuminator, being an element of tem similar to typical thermal camera but additionally a measuring system, and thermovision methods record- equipped with the lter system ensuring the required ing thermal radiation. The methods of the rst group are spectral resolution and the signal analysis system. The the active methods employing absorption of laser radia- other type includes the device based on the principles tion by chemical compounds. Thermovision methods are of the Fourier spectroscopy (Fourier transform infrared based on using the absorption bands of chemical com- spectroscopy FTIR) that is expensive and so not read- pounds in IR. ily available. Thermovision methods, employing arrays detectors, Fourier-transform spectrometers (FTS) are renowned provide an image (in real time) of the observed scenery instruments, particularly well-suited to remotely pro- with the marked regions of searched chemical substances. vide excellent estimates of quantitative data. Many au- Usefulness of this method results from the fact that a lot thors have presented how they are using conventional of substances and chemical compounds have their absorp- (non-imaging) FTS to perform quantication of distant tion bands in IR. Figure 1 shows absorption bands of the gas emissions. Amongst others, we should mention the selected chemical compounds. important contributions made by Harig et al. [1, 2]. The devices operating in IR, used for detection of His group performed many measurement campaigns for chemical substances (gases) in the atmosphere have spec- which they obtained excellent results by ensuring proper modeling of the scene and by paying attention to under- stand (and to take into account) the instrument signature. Other groups developed similar approaches how- ∗corresponding author; e-mail: [email protected] ever limiting their study to optically thin plumes [3, 4] to dedicated instruments performing optical subtraction

(463) 464 M. Kastek et al.

[5, 6], or introducing Bayesian algorithms to maximize image. The spectral distribution for an entire image is the use of a priori information [7]. On the other hand, obtained by performing the analysis according to (2) for detection and quantication activities with an imaging every pixel of an FPA array. The results can be treated as Fourier-transform spectrometer (IFTS) have been rst 3-dimensional data (two image coordinates and a wave- presented by Spisz et al. [8]. length) and can be analysed using the special software The present paper deals with the remote gas identica- for detection and identication of gases. tion and quantication from turbulent stack plumes with Imaging Fourier-transform spectroradiometer Hy- an IFTS. It presents rst the modeling that is required perCam uses the layout of the Michelson interfero- in order to get an appropriate understanding of both the meter. Its schematic diagram, showing the elements re- scene and the instrument. Next it covers the method- sponsible for the change of the length of optical path, is ology of the developed quantication approach. Finally, shown in Fig. 2. results are presented to demonstrate the capabilities and the performances of the remote gas quantication by using hyperspectral data obtained from the Telops Hyper- Cam. The latter is described in much detail in prior Refs. [912].

2. Imaging Fourier transform spectroradiometer

The infrared IFTS Hyper-Cam LWIR used in experiment was built by Telops Inc. This IFTS use 320 × 256 pixels mercury cadmium telluride (MCT) focal plane arrays (FPA) with a 6◦ × 5◦ FOV. The FPA has Stirling cooler to provide good noise gures in a eld ready package. Spectral information is obtained using Fig. 2. Block diagram of imaging Fourier-transform a technique called Fourier transform infrared radiometry spectroradiometer. (FTIR). FTIR is a classical interference based technique applied to gas spectroscopy that uses a Michelson inter- In the presented system the retro-reectors are used ferometer to mix an incoming signal with itself at several instead of at mirrors. As a result it is easier to move dierent discrete time delays. The resulting time do- M2 mirror keeping the reecting plane in constant po- main waveform, called an interferogram, is related to the sition with respect to the optical axis. It should be power spectrum of the scene through the Fourier trans- mentioned here that the interferometer works in the far form. An interferogram for each pixel in an image is cre- infrared range, thus the necessary dierence in optical ated by imaging the output of the interferometer onto a paths requires large mirror travel. For such long moves focal plane array and collecting data at each discrete time in a classical setup with at mirrors the skew of mirror delay. Advantages of using an FTIR sensor over lter or plane is dicult to avoid which may result in consid- grating based system include higher resolution for equal erable measurement errors. The solution presented in cost and the absence of misalignment of dierent color Fig. 2 minimizes such errors. images due to platform motion. However, FTIR does The main dierence between standard and imaging have the disadvantage of producing a slower frame rate interferometer is the application of FPA detector type. than lter based systems because twice as many points The imaging device can be described as classical inter- are taken for the same number of spectral points. For ferometer multiplied by the number of pixels in the ar- atmospheric tracking of gases however, the LWIR sensor ray. As a result the spectral analysis of interferograms has sucient frame rate [13]. at single pixels gives the spectral information of the en- Data for this collection can be taken from 0.25 cm−1 to tire image. The analysis of incident radiation at pixel 150 cm−1 spectral resolution between 830 cm−1 (12 µm) level is identical for every array element. The change in and 1290 cm−1 (7.75 µm) at a frame rate of 0.2 Hz. In amplitude as a result of optical path dierence x, for λi addition to the infrared data, visible imagery was taken wavelength is given by using a rewire camera that is boresighted to the IR sensor. Figure 1 shows a picture of a Hyper-Cam LWIR. The x A(x) = Ai cos 2π . (1) sensor is controlled by eld computer and data is stored λi on a RAID drive to guarantee integrity of data [13]. Detector response is proportional to the radiant intensity The main dierence between the standard and imaging of incident radiation interferometer is the application of FPA detector type. 2 x I(t) = Ii cos 2π . (2) The imaging device can be described as classical inter- λi ferometer multiplied by the number of pixels in the ar- Considering the trigonometric relations for the cosine of ray. As a result the spectral analysis of interferograms at double angle and by limiting the analysis to variable com- single pixels gives the spectral information of the entire ponent only, the radiant intensity as the function of op- Hyperspectral Imaging Infrared Sensor Used . . . 465 tical path dierence can be written as ment can be achieved by applying longer optical path x dierences and by increasing the number of measurement I(x) = Ii cos 4π . (3) points. Unfortunately both methods lead to longer mea- λi For the sources with continuous radiation characteristics surement times, which are still shortest in case of Fourier- in the range from to , whose spectral characteristics -transform spectrometers than in any other type. The λ1 λ2 advantage of imaging spectrometer over scanning ones is is described by a function S(λ), the radiant intensity can be described as that for a given optical path dierence the entire image is analyzed simultaneously. Another advantage of the de- Z λ2 x I(x) = S(λ) cos 4π dλ. (4) vices using the Michelson interferometer layout is quite λ1 λ large eld of view. The telescope placed in front of the Typical interferogram is shown in Fig. 3. The curve illus- interferometer (Fig. 2) is optional and it is used only for trates radiant intensity as the function of the translation the analysis of object having small angular dimensions. of M2 mirror. Maximum of the output signal is achieved Summing things up, there are three ways to obtain for identical lengths of optical paths (x = 0). imaging spectroradiometer: thermal camera and a set of optical lters, thermal camera with a tunable lter or Fourier-transform device described above. Simple camera plus lters setup has the worst parameters of these three. It has xed imaging windows dened by the number of lters (usually not more than eight). Its operation is slow because lters are mechanically inserted into optical path, e.g. by a rotation of a lter wheel. However it is a simple and cost-eective solution which can be used when the application does not require high spectral resolution, as in the detection of previously-selected gases in the chosen absorption band. The described devices are referred to as spectroradiometers, but sometimes the more general spectrometer Fig. 3. Sample interferogram. term is used. In case of prism or grating-type spectrometers the functional fragment called monochromator Spectral distribution for a given pixel detector can be can be distinguished. Such devices can be used as recreated using inverse Fourier transform, by calculating sources of quasi-monochromatic radiation. This cannot the integral be achieved with Fourier-transform spectroradiometers. Z +∞ x S(λ) = I(x) cos 4π dx. (5) 3. Method of gas detection −∞ λ In practical applications the optical path dierence is The target detection and identication algorithm is realized in the nite, symmetrical range ±d, so the rela- generally based on three key factors: the composition tion (5) assumes the following form: of the analyzed pixel, the type of model used to estimate the variability of the target and background spaces, and Z +d x S(λ) = I(x) cos 4π dx. (6) the model used to describe the pure and mixed pixels. −d λ The mathematical representation of a mixed pixel de- The spectral distribution for an entire image is ob- pends on whether the background (or target) space is tained by performing the analysis according to (6) for estimated statistically or geometrically. The considered every pixel of an FPA array. The results can be treated as sub-pixel target detection algorithms are of a stochastic 3-dimensional data (two image coordinates and a wave- nature. When the background is entirely represented by length). its statistics, the detection problem consists in extracting The device described above is an imaging Fourier- the targeted spectral signatures from a background noise -transform infrared spectrometer. Such devices have sev- term: ε . Equations (7) and (8) express this symbolically eral advantages. One of the most important param- b (7) eter of any spectrometer is the product of its resolu- H0 : x = εb, tion and transmission. This coecient is smallest for H1 : x = Ta + εb. (8) prism-type spectrometers. Grating-type spectrometers Here, a is a weighting vector to be estimated and T is have this coecient larger by several percent, whereas a matrix of k targets: [t1, t2, t3, . . . , tk]. It is assumed Fourier-transform devices boast up to two orders of mag- that the noise component has a mean value of zero and nitude advantage. It is particularly important in the in- covariance C. frared range, where the radiant intensity levels are usu- The basic idea behind the clutter-matched lter ally much lower compared with visual range. The res- (CMF) is to minimize the response to the unknown back- olution of Fourier-transform spectrometers is compara- ground signatures while accentuating the response to the ble with grating-type counterparts. Further improve- target spectrum. To do so, the following mathematical 466 M. Kastek et al. operator was developed: detect the presence of targeted species. As its name in- −1 dicates, SAM is an algorithm that computes the angle C ti qi = , (9) formed by the projection of a target on an image pixel p 0 −1 tiC ti xit = kxik ktik cos(θ) (12) where qi is the CMF related to the i-th target ti taken from the matrix of targets T and the N × N estimated or, in a more common form covariance matrix, , is given by Eq. (10): x t C cos(θ) = i . (13) M kxik ktk 1 X C = (x − µ)(x − µ)0. (10) The results are limited between 0 and 1. A value of 0 M − 1 i i i=1 signies that the image pixel and the target are not at all correlated whereas a score of 1 means that the two Here, xi and µ are, respectively, the spectrum of the i-th pixel and the mean pixel of the scene and M is the vectors are perfectly correlated. SAM is obviously not a number of pixels from the image used to estimate the co- linear lter and it cannot indicate the strength of the sig- variance. This model falls under the stochastic methods nal. The detection and identication of chemical agent category. is performed through post-processing of the recorded hy- The signal to clutter ratio is simply calculated with perspectral. The algorithm integrates a double-stage de- Eq. (11): cisional scheme used to analyze the hyperspectral data Signal produced by Hyper-Cam sensor. The algorithm is pre- = q(x − µ). (11) sented in Fig. 4. Measurement of the background is per- Clutter i If the CMF is not calculated for each scene, which is formed to obtain the covariance matrix of the scene. This sensible if real-time processing is required, then the mean covariance is used to calculate the clutter-matched lters pixel found in Eq. (10) must be the one introduced in for several gas signatures from a library based on Pa- Eq. (11) when the covariance matrix was introduced. cic Northwest National Laboratory (PNNL) database. The CMF is normalized so that, when the signal is ab- On each scene measurement, the background is removed sent, the probability density function of the ltered image and the calculation of the image of clutter-matched lter has a standard deviation of 1. The score of each pixel de- scores and correlation (SAM) scores is performed. To de- notes the number of standard deviations that this pixel tect and identify a given chemical agent, the score from parts from zero. As can be ascertained with Eq. (9), the both calculations (CMF and SAM) must exceed a given response of the CMF is linearly dependent on the target threshold [1416]. signature. The results obtained with this model give an approximate concentration of the targeted species occu- 4. Field tests results pying the pixels. Eectively, the scores obtained when the CMF is applied are a good indicator of the strength During the eld tests the following gases were mea- of the signal, which, in turn, is proportional to the tem- sured: CO2, propanebutane mixture, and freon 134 perature contrast and to the column density of the target. (CH2FCF3 tetrauoroethane). In each case gases concentrations were tested in order to verify the eciency of gas measurements in the open space by HyperCam LWIR.

Fig. 5. The theoretical absorption CO2 for dierent Fig. 4. Double-stage detection and identication of concentration 1% and 5%. scheme used to process hyperspectral data. Theoretical absorption spectra for two concentrations

When the target mean is known, it is possible to use of CO2 (1% and 5%) were calculated by HITRAN soft- the spectral angle mapper (SAM, or, as it will sometimes ware and are presented in Fig. 5. As expected, higher be called in this article, the correlation coecient) to gas concentration signicantly increases the absorption Hyperspectral Imaging Infrared Sensor Used . . . 467 and broadens its spectral range. During the experiment dierent conditions (dierent concentration) and during measured absorption characteristics of CO2 for two con- dierent environments conditions. For the over side the centrations were measured, too. Freon 134 was another results of tests conrmed possibility to use this kind of gas used during the eld measurements. This gas exhibits spectroradiometers to automatical detection of gases, af- absorption bands in long wave infrared range, thus it can ter the development method and software for it. be detected using Hyper-Cam LWIR. The registrations The measurements conducted by means of IFTS spec- were made at two distances: 60 m and 10 m, for two gas trometers have lower accuracy due to lower sensitivity, concentrations of 3% and 6%. During the eld measure- the inuence of weather conditions (mainly wind speed) ments the weather conditions (humidity, pressure, wind and unknown thickness of a visualized gas cloud. Cur- speed, and direction) were monitored by an automatic rently the new quantication algorithm is under develop- weather station Vantage Pro. It was observed that wind ment, in which the cloud thickness will be automatically speed has the most signicant inuence on the measure- estimated on the basis of spatial distribution of areas with ment results. The results are presented in Fig. 6a and emissivity dierent than that of a background. This new Fig. 7a. The Matlab procedure applied for the analysis method combining mathematical analysis, data from ref- of data from Hyper-Cam utilized the reference signature erence signature database (PNNL database), CMF mod- database and CMF module [7]. ule and additionally SAM module should improve the eectiveness of gas detection by IFTS-based solutions.

References

[1] R. Harig, G. Matz, Field Anal. Chem. Technol. 5, 75 (2001). [2] R. Harig, G. Matz, P. Rusch, Proc. SPIE 4574, 83 (2002). [3] M.K. Grin, J.P. Kerekes, K.E. Farrar Burke, K.H.- Fig. 6. The results of measure of freon 134 (concen- H., Proc. SPIE 4381, 360 (2001). tration 3%) at distance 60 m (left), freon 134 results of [4] T. Burr, N. Hengartner, Sensors 6, 1721 (2006). analysis in Matlab (right). [5] R.L. Lachance, J.-M. Thériault, C. Lafond, A.J. Ville- maire, Proc. SPIE 3383, 124 (1998). [6] J.-M. Thériault, Technical Report Defence Re- search Establishment Valcartier (DREV) TR-2000- 156 (2001). [7] P. Heasler, C. Posse, J. Hylden, K. Anderson, Sensors 7, 905 (2007). [8] T.S. Spisz, P.K. Murphy, C.C. Carter, A.K. Carr, A. Vallières, M. Chamberland, Proc. SPIE 6554, 655408 (2007). [9] V. Farley, M. Chamberland, P. Lagueux, A. Vallières, Fig. 7. The results of measure of freon 134 (concen- A. Villemaire, J. Giroux, Proc. SPIE 6661, 66610L tration 6%) at distance 10 m (left), freon 134 results of (2007). analysis in Matlab (right). [10] A. Vallières, A. Villemaire, M. Chamberland, L. Bel- humeur, V. Farley, J. Giroux, J.-F. Legault, Proc. 5. Conclusion SPIE 5995, 59950G (2005). [11] M. Chamberland, C. Belzile, V. Farley, J.-F. Legault, IFTS technique is a very useful method for gas detec- K. Schwantes, Proc. SPIE 5416, 63 (2004). tion in the open space. Its sensitivity is of order of magni- [12] V. Farley, C. Belzile, M. Chamberland, J.-F. Legault, tude lower than typical laboratory methods, mainly due K. Schwantes, Proc. SPIE 5546, 29 (2004). to the fact that lock-in detection is not possible. However [13] M. Kastek, T. Pi¡tkowski, P. Trzaskawka, Metrol. the possibility to present the results in a form of an im- Measur. Syst. 18, 679 (2011). age is a clear advantage, because the gas can be not only [14] P. Tremblay, S. Savary, M. Rolland, A. Villemaire, detected, but also located in space. Even its lower ac- M. Chamberland, V. Farley, Proc. SPIE 7673, curacy is still sucient to detect and identify particular 76730H (2010). gases in the mixed gas cloud. On the basis of conducted [15] M. Kastek, T. Piatkowski, R. Dulski, M. Chamber- experiments it can be stated that the gas concentrations land, P. Lagueux, V. Farley, Phot. Lett. Poland 4, exceeding 2% can be detected in the open space using 146 (2012). the Hyper-Cam sensor. [16] M. Kastek, T. Piatkowski, R. Dulski, M. Chamber- After the analysis of result tests eld eciency of the land, P. Lagueux, V. Farley, in: Symp. on Photonics IFTS technology was conrmed. During the experiment and Optoelectronics SOPO 2012, IEEE Photonics So- measurement of characteristics of gases were made with ciety, Shanghai 2012, art. no. 6270545.